Errata for The Linux Programming Interface

Toward the end of this page, a typographical glitch means that the "Registered Trademark" symbol after the word "UNIX" was incorrectly rendered. The (tiny) characters after "UNIX" need to be replaced by a circled "R".

Explanation:

This typo first appeared in print run 5.

Reported by Alex Pinkney.

2018-06-21

9

Under the heading Web site and source code of example programs, change:

http://www.man7.org/tlpi/.

to:

http://man7.org/tlpi/.

2011-01-08

2

Under the heading Feedback, following the sentence:

Book bugs and general suggestions about how the explanations in the book can be improved are also welcome.

add the sentence:

(A current list of errata can be found at http://man7.org/tlpi/errata/.)

2011-01-08

2

Above the heading Feedback, insert the following text:

Training courses

I teach courses on system programming, based on the content of this book, as well as a number of other related topics, including network programming. For details, see http://man7.org/training/.

2014-02-17

5

This is a note to the text, rather than an erratum.

As a consequence of the (July 2011) renumbering of the Linux kernel series from 2.6.x to 3.x and then 4.x, 5.x, and (most recently) 6.x, the discussion of kernel numbering on this page is now out-of-date. However, all that has changed is that the numbering scheme has been simplified (replacing an invariant 2.6 with 3 and, subsequently, 4, 5, and 6); the kernel development model remains unchanged. As Linus Torvalds noted of the 3.0 release, there is otherwise nothing special about the release (i.e., no changes more significant than those in Linux 2.6.39 and each of the preceding 2.6.x releases).

In the bulleted list at the top of the page, each of the instances of 2.6.z could simply be replaced by 6.z and the description on this page would still be accurate for the current kernel development model.

Explanation:

(This note was last updated in December 2022.)

2011-07-29

WONTFIX

This is a note to the text, rather than an erratum.

The small-font note at the bottom of the page states "At the time of writing, the 2.4 stable Linux kernel is still supported…" This has now ceased to be true: in April 2012, Willy Tarreau, the Linux 2.4 maintainer, noted that the 2.4 kernel series has now reached end-of-life.

2012-04-10

WONTFIX

In the second paragraph, change:

which was subsequently adopted in 1990 as an International Standards Organization (ISO) standard

to:

which was subsequently adopted in 1990 as an International Organization for Standardization (ISO) standard

Explanation:

Although often referred to as the International Standards Organization, the correct title is International Organization for Standardization.

2011-08-04

3

In the paragraph just above the heading Legacy features, change:

When using interfaces that are unspecified or weakly specified, we have few guarantees when porting applications to other UNIX implementations.

to:

When using interfaces that are unspecified or weakly specified, we have few guarantees when porting applications to other UNIX implementations, and portable applications should avoid relying on the behavior of a specific implementation.

Explanation:

The added text simply emphasizes a point that was already implied.

2011-02-28

2

This is a note to the text, rather than an erratum.

As a follow-on to the discussion in Section 1.3.5, note that the Austin group published Technical Corrigendum Number 1 for POSIX.1-2008/SUSv4, in 2013, Technical Corrigendum Number 2 in 2016. The standard can be found online at http://www.unix.org/version4/.

Explanation:

(This note was updated on 14 October 2017.)

2014-01-17

WONTFIX

On the fourth line from the bottom of the page, remove the superfluous right parenthesis following "mid-1990s".

Reported by Justin Pryzby.

2010-12-28

2

In the fourth paragraph, change:

the operation performed execve()

to:

the operation performed by execve()

Reported by Sun Jian.

2011-02-25

2

In the last sentence of the first paragraph in Section 2.12, change:

However, each thread has it own stack…

to:

However, each thread has its own stack…

Reported by Jaewook Yu.

2011-04-09

3

In the first bullet point of Section 2.16, change:

Universal Coordinated Time

to:

Coordinated Universal Time

Reported by Yaakov Eisenberg.

2022-08-17

14

This is a note to the text, rather than an erratum.

In the list item "5" (toward the bottom of the page), change:

In response to the trap to location 0x80, the kernel invokes its system_call() routine (located in the assembler file arch/x86/kernel/entry.S) to handle the trap.

to:

In response to the trap to location 0x80, the kernel invokes its system_call() routine to handle the trap.

Explanation:

Over the years, the details about which source files contain various kernel functions have changed. Those details are in any case not critical to the discussion on this page, so, rather that trying to keep the text constantly up to date with the source code changes, it seems better to remove such details.

An earlier version of this erratum (reported by Leif Dyvik on 2012-04-25 and incorporated in the fourth print run) is shown below.

In the list item "5" (toward the bottom of the page), change:

arch/i386/entry.S

to:

arch/x86/kernel/entry.S

Reported by Alex Colomar.

2021-04-05

WONTFIX

This is a note to the text, rather than an erratum.

Because of ongoing changes in the kernel and in the GNU C library, all of the pathnames mentioned in Figure 3-1 are no longer accurate. Since such changes are likely to be ongoing in the future, I won't make further updates in this figure, and will just note that the pathnames in the upper right, lower left, and lower right boxes are all nowadays out of date.

Explanation:

An earlier erratum for this figure (dated 2012-05-25, for a change that was made in the fourth print run) is shown in the following paragraphs. By now, that erratum has itself become out of date.

In the bottom left hand of Figure 3-1, under the heading System call service routine, change:

arch/x86/kernel/process_32.c

to:

arch/x86/kernel/process.c

Erratum originally written on 2012-04-25; a small correction was made on 2012-05-25 after a note from Robert P. J. Day.

Reported by Alex Colomar.

2021-04-05

WONTFIX

In the fourth line, change:

(in #ifdef statements)

to:

(in #if statements)

Reported by Richard Moore.

2011-01-26

2

About a quarter of the way through Listing 3-2 (lib/error_functions.h), change:

/* This macro stops 'gcc -Wall' complaining that "control reaches

to:

    /* This macro stops 'gcc -Wall' complaining that "control reaches

2012-09-04

4

Two related changes are made on this page.

About two thirds of the way down the page, change:

However, this approach is inefficient because errno is defined in threaded programs as a macro that expands into a function call that returns a modifiable lvalue.

to:

However, this approach is inefficient because errno is defined in threaded programs as a macro that produces a modifiable lvalue via a function call that returns a pointer to a thread-specific storage area.

In the small-font note at the bottom of the page, change:

Under the POSIX threads API, errno is redefined to be a function that returns a pointer to a thread-specific storage area (see Section 31.3).

to:

Under the POSIX threads API, errno is defined as a macro that produces a modifiable lvalue by dereferencing the result of calling a function that returns a pointer to a thread-specific storage area (see Section 31.3).

Explanation:

Yaakov Eisenberg pointed out that neither of the original sentences was quite accurate. To resolve this, I reduced the level of detail in the main text, and provided the full details in the small-font note.

Reported by Yaakov Eisenberg.

2022-12-04

15

In Listing 3-3 (lib/error_functions.c), change:

#ifdef __GNUC__
__attribute__ ((__noreturn__))
#endif
static void
terminate(Boolean useExit3)

to:

NORETURN 
static void
terminate(Boolean useExit3)

Explanation:

The NORETURN macro that is already defined in lib/error_functions.h is exactly equivalent to the code that is replaced here.

Reported by Alejandro Colomar.

2021-05-16

13

Several small changes are required to fix some formatting issues that were introduced in the thirteenth print run.

In the first paragraph, change "out-put" to "output".

In the second paragraph, change "pro-gram" to "program".

In the first line of Listing 3-3, the string "error_functions.c" is misplaced. It should appear at the end of the end of the previous line (so that the end of that line reads "lib/error_functions.c").

In Listing 3-3, the following lines:

NORETURN
static void
terminate(Boolean
useExit3) {

should read:

NORETURN
static void
terminate(Boolean useExit3)
{

Reported by Yaakov Eisenberg.

2022-12-04

15

In the first line of text below the example program, change:

The file enames.c.inc

to:

The file ename.c.inc

Reported by Kiju Kim.

2011-02-13

2

In the third line of the second normal-size font paragraph on the page (i.e., a bit more than half way down the page), change:

names separate by a slash.

to:

names separated by a slash.

Reported by Subu Rama.

2010-12-03

2

At the top of the page (in the final part of Listing 3-6, lib/get_num.c), change:

    return (int) res;

to:

    return res;

Explanation:

The cast serves no purpose, especially given the range check in the immediately preceding lines of this function.

Reported by Alejandro Colomar.

2021-05-18

13

In the second sentence of the first paragraph of Section 3.6.1, change:

Some of these standards are defined by standards bodies such The Open Group

to:

Some of these standards are defined by standards bodies such as The Open Group

Reported by Murray McAllister.

2012-04-30

4

Not quite half way down the page, under the description of the _BSD_SOURCE macro, change:

If defined (with any value), expose BSD definitions. Defining this macro also defines _POSIX_C_SOURCE with the value 199506.

to:

If defined (with any value), expose BSD definitions.

2014-01-06

5

This is a note to the text, rather than an erratum.

The description of the _BSD_SOURCE feature test macro about one third of the way down this page says that it causes BSD definitions to be favored in a few cases where standards conflict. Starting with glibc 2.19, _BSD_SOURCE no longer has this effect.

2014-01-06

WONTFIX

This is a note to the text, rather than an erratum.

Starting with version 2.19, glibc adds a new feature test macro, _DEFAULT_SOURCE. Defining this macro provides an effect similar to the feature test macros that are defined by default—that is, _BSD_SOURCE, _SVID_SOURCE, and _POSIX_C_SOURCE=200809. This macro can be defined to ensure that the "default" definitions are provided even when the defaults would otherwise be disabled, as happens when individual macros are explicitly defined, or the compiler is invoked in one of its "standard" modes (e.g., cc -std=c99).

2014-01-06

WONTFIX

This is a note to the text, rather than an erratum.

In glibc 2.20 and later, _BSD_SOURCE and _SVID_SOURCE are deprecated. They are redefined to have the same meaning as _DEFAULT_SOURCE, and thus continue to expose all of the definitions that they exposed earlier. However, the use of these macros now produces a warning, unless _DEFAULT_SOURCE is also defined.

This means that all function prototype boxes in TLPI that show _BSD_SOURCE or _SVID_SOURCE as being required should instead show _DEFAULT_SOURCE. However, that change will be left to a future edition of TLPI. In the meantime, a workaround fix to the makefiles has been added to the code in the source distribution so that now all code is compiled with _DEFAULT_SOURCE defined. Note also that starting with release 4.05 (March 2016), the Linux manual pages provide up-to-date documentation of the feature test macro requirements since glibc 2.20.

2014-03-12

WONTFIX

Near the start of Table 3-1, in the second column of the row that starts dev_t, change:

an arithmetic type

to:

integer

Explanation:

POSIX.1-2001 specified that dev_t shall be an arithmetic type. In practice, implementations have used an integer type. POSIX.1-2008 tightened the specification to say that dev_t shall be an integer type.

2020-10-18

13

Toward the end of Table 3-1, in the second column of the row that starts time_t, change:

integer or real-floating

to:

integer

Explanation:

POSIX.1-2001 specified that time_t shall be an integer or real-floating type. In practice, (all?) implementations have used an integer type. POSIX.1-2008 tightened the specification to say that time_t shall be an integer type.

2020-10-18

13

In the first paragraph of the small-font note in the middle of this page, change:

Thus, we could write %zd instead of using %ld plus a cast for these types.

to:

Thus, we could write %zd instead of using %ld plus a cast for ssize_t, and analogously %zu for size_t.

2015-03-17

6

A bit more than half way down the page, immediately following the heading Using macros that may not be present on all implementations change:

In some cases, a macro may be not be defined

to:

In some cases, a macro may not be defined

Reported by Markus Boje.

2010-11-22

2

Toward the end of Listing 4-1 (fileio/copy.c), change:

    while ((numRead = read(inputFd, buf, BUF_SIZE)) > 0)
        if (write(outputFd, buf, numRead) != numRead)
            fatal("couldn't write whole buffer");

to:

    while ((numRead = read(inputFd, buf, BUF_SIZE)) > 0)
        if (write(outputFd, buf, numRead) != numRead)
            fatal("write() returned error or partial write occurred");

Explanation:

This program is intended as an early (simple) example that demonstrates the use of the key I/O system calls. Nevertheless, the fact that two different cases—write() returning an error and a partial write—are combined in the same error diagnostic has been the source of questions from a few readers. The aim of this change to the text is to more clearly alert the reader that there are two separate cases here. If there had been sufficient space on the page, ideally I might also have included an explicit forward reference to Section 61.1, which discusses partial writes on sockets and shows a function (writen()) that restarts write() calls after a partial write.

Partial writes can also occur when writing to regular files if, for example, the disk is nearly full, or the process resource limit on file size (RLIMIT_FSIZE) has been reached. However, in such cases, restarting the write() is not useful, since it will either result in an ENOSPC error if the disk is no full, or a SIGXFSZ signal if the RLIMIT_FSIZE resource limit has been exceeded.

Reported by Hrvoje Nikšić.

2020-01-14

12

This is a note to the text, rather than an erratum.

Some further explanation may be useful for readers that try running a command similar to this command shown in the shell session near the top of the page:

$ ./copy /dev/pts/16 /dev/tty

In this case, /dev/pts/16 is a pseudoterminal device that is being used as the input file for the copy program. (You can discover the name of the pseudoterminal being employed in a terminal window by running the tty command.)

If trying this example, you may see results that are puzzling until various topics are covered later in the book. In particular, if, as is likely the case, a shell is running in the foreground of the source device, then the following points mean that some of the text that you type on the source device may not be copied to /dev/tty:

• On the source device, the shell and the copy program are competing for input.
• Assuming that a line-editing shell such as bash(1) is running on the source device, the shell will have placed the terminal in noncanonical mode (see TLPI section 62.6), so that input is available a character at a time. This means the shell will get some of the characters typed on the source device and copy will get some.

One way to avoid this confusing behavior is to run a command such as sleep 1000 on the source device. This will cause the shell to set the terminal to canonical mode (input available a line at a time) and place itself in the background. This will stop the shell competing for input, and will make all input typed on the source device available (a line at a time) to the copy program.

Reported by Henri van de Water.

2015-03-12

WONTFIX

Near the end of Listing 4-2, change:

    fd = open("w.log", O_WRONLY | O_CREAT | O_TRUNC | O_APPEND,
                       S_IRUSR | S_IWUSR);

to:

    fd = open("w.log", O_WRONLY | O_CREAT | O_APPEND,
                       S_IRUSR | S_IWUSR);

Explanation:

The code as originally given was technically correct, but the presence of O_TRUNC was superfluous for the purpose of the example, and its semantics were not explained in the accompanying comment.

Reported by Jessica T McKellar.

2011-05-26

3

In Table 4-3, move two rows to a different part of the table:

Flag	Purpose	SUS?
O_RDONLY	Open for reading only	v3
O_WRONLY	Open for writing only	v3
O_RDWR	Open for reading and writing	v3
O_CLOEXEC	Set the close-on-exec flag (since Linux 2.6.23)	v4
O_CREAT	Create file if it doesn't already exist	v3
O_DIRECT	File I/O bypasses buffer cache
O_DIRECTORY	Fail if pathname is not a directory	v4
O_EXCL	With O_CREAT: create file exclusively	v3
O_LARGEFILE	Used on 32-bit systems to open large files
O_NOATIME	Don't update file last access time on read() (since Linux 2.6.8)
O_NOCTTY	Don't let pathname become the controlling terminal	v3
O_NOFOLLOW	Don't dereference symbolic links	v4
O_TRUNC	Truncate existing file to zero length	v3
O_APPEND	Writes are always appended to end of file	v3
O_ASYNC	Generate a signal when I/O is possible
O_DSYNC	Provide synchronized I/O data integrity (since Linux 2.6.33)	v3
O_NONBLOCK	Open in nonblocking mode	v3
O_SYNC	Make file writes synchronous	v3

to give the following:

Flag	Purpose	SUS?
O_RDONLY	Open for reading only	v3
O_WRONLY	Open for writing only	v3
O_RDWR	Open for reading and writing	v3
O_CLOEXEC	Set the close-on-exec flag (since Linux 2.6.23)	v4
O_CREAT	Create file if it doesn't already exist	v3
O_DIRECTORY	Fail if pathname is not a directory	v4
O_EXCL	With O_CREAT: create file exclusively	v3
O_LARGEFILE	Used on 32-bit systems to open large files
O_NOCTTY	Don't let pathname become the controlling terminal	v3
O_NOFOLLOW	Don't dereference symbolic links	v4
O_TRUNC	Truncate existing file to zero length	v3
O_APPEND	Writes are always appended to end of file	v3
O_ASYNC	Generate a signal when I/O is possible
O_DIRECT	File I/O bypasses buffer cache
O_DSYNC	Provide synchronized I/O data integrity (since Linux 2.6.33)	v3
O_NOATIME	Don't update file last access time on read() (since Linux 2.6.8)
O_NONBLOCK	Open in nonblocking mode	v3
O_SYNC	Make file writes synchronous	v3

Explanation:

As noted on page 93, O_DIRECT and O_NOATIME are open file status flags that can be retrieved and modified using fcntl(). These flags were accidentally misplaced in the file creation flags section of Table 4-3.

Reported by Madhavan Kasthurirangan.

2011-09-04

3

This is a note to the text, rather than an erratum.

Since Linux 2.6.39, the open() system call supports a new file creation flag, O_PATH, that can be used to obtain a file descriptor with specific, limited uses. The returned file descriptor can be closed, duplicated, or passed via a UNIX domain socket (see Section 61.13.3); the file descriptor can't be read from or written to. O_PATH can be used to obtain file descriptors for symbolic links. O_PATH file descriptors are provided for use as the directory file descriptor argument in the *at() system calls (see Section 18.11).

2011-06-26

WONTFIX

In the first bullet point at the top of the page, change:

They can be retrieved using the fcntl() F_GETFL operation (Section 5.3).

to:

Only one of these values should be specified in flags. The access mode can be retrieved using the fcntl() F_GETFL operation (Section 5.3).

Explanation:

This is just a clarification that reiterates a point made on page 72.

Suggested from a conversation with Krzysztof Żelechowski.

2010-12-16

2

In the second and third bullet points at the top of the page, change:

• File creation flags: These are the flags shown in the second part of Table 4-3. They control various aspects of the behavior of the open() call, as well as options for subsequent I/O operations. These flags can’t be retrieved or changed.
• Open file status flags: These are the remaining flags in Table 4-3. They can be retrieved and modified using the fcntl() F_GETFL and F_SETFL operations (Section 5.3).

to:

• File creation flags: These are the flags shown in the second part of Table 4-3. They control various aspects of the behavior of the open() call itself. These flags can’t be retrieved or changed.
• Open file status flags: These are the remaining flags in Table 4-3. They affect the semantics of subsequent I/O system calls and can be retrieved and modified using the fcntl() F_GETFL and F_SETFL operations (Section 5.3).

Reported by 徐锋.

2020-06-12

12

In the third line of the O_CLOEXEC paragraph at the bottom of the page, change:

additional fcntl() F_SETFD and F_SETFD operations

to:

additional fcntl() F_GETFD and F_SETFD operations

Reported by Subu Rama.

2010-11-30

2

In the prototype box for write(), change:

void *buffer

to:

const void *buffer

2014-11-17

6

This is a note to the text, rather than an erratum.

This note is also relevant for the discussion of file holes on page 83.

Linux 3.1 adds two new operations for the lseek() system call: SEEK_HOLE and SEEK_DATA. These operations can be used to determine the start and end of holes in a file. SEEK_HOLE seeks to the offset of the next file hole whose location is greater than or equal to the current file offset. SEEK_DATA seeks to the offset of the next non-hole region whose location is greater than or equal to the current file offset.

2011-07-29

WONTFIX

In the middle of the last paragraph (small-font note), change:

This allows an application to be sure that a later write() to the file won't fail because disk space is exhausted (which could otherwise occur if a hole in the file was filled in, or some other application consumed space on the disk).

to:

This allows an application to be sure that a later write() to the file won't fail because disk space is exhausted (which could otherwise occur if a hole in the file was filled in, or some other application consumed space on the disk before the full extent of the file was written).

Explanation:

The parenthetical clause describes two reasons why the write() might fail. The added words simply clarify the second reason a little.

Reported by Kiju Kim.

2011-03-19

2

This is a note to the text, rather than an erratum.

Linux 2.6.38 added FALLOC_FL_PUNCH_HOLE, a new operation for the fallocate() system call. This operation creates a hole in a file in the byte range indicated by the system call's offset and len arguments. (The file data in the specified range is lost.) Among the file systems that already support this operation are XFS and (since Linux 3.0) ext4. Btrfs is capable of supporting the operation, and support is likely to be added in the future.

2011-06-30

WONTFIX

Near the start of the main() function in Listing 4-3 (fileio/seek_io.c) change:

    int fd, ap, j;
    char *buf;
    ssize_t numRead, numWritten;

to:

    int fd, ap, j;
    unsigned char *buf;
    ssize_t numRead, numWritten;

Explanation:

See the erratum for page 85.

Reported by Philippe Brodeur.

2019-12-15

12

In Listing 4-3 (fileio/seek_io.c), about one third of the way down the page, change:

                for (j = 0; j < numRead; j++) {
                    if (argv[ap][0] == 'r')
                        printf("%c", isprint((unsigned char) buf[j]) ?
                                                buf[j] : '?');
                    else
                        printf("%02x ", (unsigned int) buf[j]);
                }

to:

                for (j = 0; j < numRead; j++) {
                   if (argv[ap][0] == 'r')
                        printf("%c", isprint(buf[j]) ?  buf[j] : '?');
                   else
                        printf("%02x ", buf[j]);
               }

Explanation:

The fundamental issue was a sign-extension problem that occurred with the "R" operation when displaying bytes whose most significant bit was set. A simple fix would have been to just change the (unsigned int) cast in the printf() to (unsigned char). However, fixing the type of buf is tidier, since it allows us to remove the casts from both printf() calls.

Reported by Philippe Brodeur, with a fix (s/byte/bit/) to the explanation on 2024-10-15 after a note from Dave Chupreev.

2019-12-15

12

In the first paragraph of Section 5.1, change:

All system calls are executed atomically. By this, we mean that the kernel guarantees that all of the steps in a system call are completed as a single operation, without being interrupted by another process or thread.

to:

Various system call operations are executed atomically. By this, we mean that the kernel guarantees that all of the steps in the operation are completed without being interrupted by another process or thread.

Explanation:

While the rest of this section makes the idea of atomicity clear, these two sentences as originally given were confusing. Most obviously, blocking system calls are not atomic.

Reported by Christopher Morris.

2013-11-22

5

At the beginning of Listing 5-1, change:

fd = open(argv[1], O_WRONLY);       /* Open 1: check if file exists */
    if (fd != -1) {                     /* Open succeeded */

to:

    fd = open(argv[1], O_WRONLY);       /* Open 1: check if file exists */
    if (fd != -1) {                     /* Open succeeded */

2013-09-16

5

In the prototype box for fcntl() at the top of the page, change:

Return on success depends on cmd, or –1 on error

to:

Return value on success depends on cmd; returns –1 on error

Explanation:

A minor wording improvement.

2011-06-16

3

In the first sentence of the second-to-last paragraph, change:

The shell achieves the redirection of standard error by duplicating file descriptor 2 so that it refers to the same open file description as file descriptor 1

to:

The shell redirects standard error by making file descriptor 2 a duplicate of file descriptor 1, so that it refers to the same open file description

Explanation:

The intent of the original wording was probably clear, but was potentially confusing.

Reported by Jeffrey Thompson.

2012-12-18

4

About one third of the way down the page (in the paragraph starting "Assuming the normal situation"), change:

dup() will create the duplicate of descriptor 1 using file 3.

to:

dup() will create the duplicate of descriptor 1 using descriptor 3.

Reported by Simon Durrant.

2011-08-07

3

At the end of the paragraph following the prototype box for dup2(), change:

(Any error that occurs during this close is silently ignored; safer programming practice is to explicitly close() newfd if it is open before the call to dup2().)

to:

(Any error that occurs during this close is silently ignored.) The closing and reuse of newfd are performed atomically, which avoids the possibility that newfd is reused between the two steps in a signal handler or a parallel thread that allocates a file descriptor.

Explanation:

The text completely omitted to mention a key aspect of the operation of dup2(), and the suggestion in the deleted text to close() newfd before calling dup2() allows the possibility of the very race condition that the atomicity of the call was designed to avoid. For further details, see the (recently amended) dup2() man page.

2014-05-30

5

In the paragraph at the bottom of the page, change:

The integer count specifies

to:

The integer iovcnt specifies

Reported by Subu Rama.

2010-11-30

2

This is a note to the text, rather than an erratum.

The details described in the second paragraph (starting "However, the glibc wrapper functions for readv() and writev()) in the small-font note at the top of the page have no longer been true for quite a while. Since glibc 2.10, if the wrapper functions for readv() and writev() detect that the kernel version is 2.6.18 or later, they always only invoke the raw system calls. And since glibc 2.20, where the minimum Linux kernel version required by the library is 2.6.32, the wrapper functions always just directly invoke the system calls. For further details, see the preadv(2) manual page.

2015-01-20

WONTFIX

At the end of the paragraph second from the bottom of the page, add a sentence:

The preadv() and pwritev() system calls perform the same task as readv() and writev(), but perform the I/O at the file location specified by offset (like pread() and pwrite()). These system calls don't change the file offset.

Explanation:

The reader may have been able to draw this implication from the text (since it is explained that these system calls are like readv() and writev()), but it's better to make the point explicit.

Reported by Sun Jian.

2011-06-15

3

In the small-font at the top of the page, change:

The main difference was that a nonblocking write() on System V returned 0 if a write() could not be completed or if no input was available to satisfy a read().

to:

The main difference was that a nonblocking write() on System V returned 0 if a write() could not be completed and a nonblocking read() returned 0 if no input was available.

Reported by Sun Jian.

2011-06-15

3

In the first line of the small-font note toward the bottom of the page, change:

(e.g., Alpha, IA-64)

to:

(e.g., x86-64, Alpha, and IA-64)

Explanation:

It was an oversight to omit the most common 64-bit architecture here.

2011-12-11

3

This erratum consists of two related changes.

In the first line of the second paragraph of Section 5.11, change:

Opening one of the files in the /dev/fd directory is equivalent to duplicating the corresponding file descriptor.

to:

On some systems (but not Linux), opening one of the files in the /dev/fd directory is equivalent to duplicating the corresponding file descriptor.

Just above the small-font at the foot of the page, insert a new normal-size paragraph:

On Linux, opening one of the files in /dev/fd is equivalent to reopening the original file; that is, the new file descriptor is associated with a new open file description (and thus has distinct file status flags and file offset).

Explanation:

On at least Solaris and FreeBSD, opening /dev/fd/n is equivalent to duplicating file descriptor n. However, this is not the case on Linux. In common practical use (i.e., the use of /dev/stdin and /dev/stderr in shell pipelines used to connect filters), this difference in semantics is normally invisible.

2013-09-20

5

In the last sentence of Section 5.11, change:

As a convenience, the names /dev/stdin, /dev/stdout, and /dev/stderr are provided as symbolic links to, respectively, /dev/fd/0, /dev/fd/1, and /dev/fd/2.

to:

As a convenience, the names /dev/stdin, /dev/stdout, and /dev/stderr are provided as symbolic links to, respectively, /proc/self/fd/0, /proc/self/fd/1, and /proc/self/fd/2 (equivalent to /dev/fd/0, /dev/fd/1, and /dev/fd/2).

Reported by Ghe Rivero.

2020-07-02

12

In the first line of Exercise 5-1, change:

Modify the program in Listing 5-3

to:

If you have access to a 32-bit Linux system, modify the program in Listing 5-3

Explanation:

From Section 5.10, the reader can deduce that this exercise applies only to 32-bit systems. However, it is of course better to make that assumption explicit.

Reported by Sandipan Razzaque.

2011-12-11

3

(Exercise 5-3) In the first sentence of the last paragraph on the page, change:

This file should open

to:

This program should open

Reported by Justin Pryzby.

2010-12-28

2

In Exercise 5-6, change the line:

    write(fd2, "world", 6);

to:

    write(fd2, " world", 6);

(i.e., add a space before world).

Reported by John Wiersba. (I initially wrote up this erratum incorrectly; fixed on 2010-12-07, after a note from Geoff Clare.)

2010-11-30

2

In the last paragraph of Section 6.2, change:

the Ppid field

to:

the PPid field

(i.e., make the second "P" upper case).

Reported by Subu Rama.

2010-12-20

2

At the end of the first paragraph in the small-font note in the second half of this page, remove the following sentence:

Sometimes, the term section is used instead of segment, since section is more consistent with the terminology used in the now ubiquitous ELF specification for executable file formats.

Explanation:

That sentence wasn't correct. An ELF file can contain sections and segments, but the two terms are not synonymous.

Reported by Martin McClure.

2020-11-24

13

About one third of the way down the page, change:

    extern char etext, edata, end;
            /* For example, &etext gives the address of the end
               of the program text / start of initialized data */

to:

    extern char etext, edata, end;
            /* For example, &etext gives the address of the next byte past
               the end of the program text / start of initialized data */

Explanation:

The original wording was intended (the details are explained in the text), but the rewording is more precise and avoids possibilities for confusion.

Reported by Sun Jian.

2011-02-16

2

In the second paragraph, change:

An example of this technique is provided by the gzip(1), gunzip(1), and zcat(1) commands, all of which are links to the same executable file.

to:

An example of this technique is provided by the gzip(1), gunzip(1), and zcat(1) commands: on some distributions, these are all links to the same executable file.

Explanation:

Yongzhi Pan pointed out that the observation about the three commands linking to the same file is not true on some distributions.

Reported by Yongzhi Pan.

2013-12-02

5

In the last line on this page, change:

Returns 0 on success, or a nonzero on error

to:

Returns 0 on success, or nonzero on error

Reported by Yaakov Eisenberg.

2023-02-09

15

In the last paragraph on this page, change:

When we compile the program in Listing 6-6 normally, we see the expected output:

to:

When we compile the program in Listing 6-6 without optimization, we see the expected output:

Explanation:

By default (i.e., "normally"), gcc compiles without optimization. This change makes that point more explicit.

Reported by Bill McConnaughey.

2011-06-28

3

This is a note to the text, rather than an erratum.

Regarding the discussion on pages 135-137 of longjmp() clobbering variables (and especially the example shown a bit more than half way down page 137), things seem to have changed a little since the book was published. In more recent versions of gcc, the "warning: variable `nvar' might be clobbered by `longjmp' or `vfork'" warning is generated in cases where compiler optimization is disabled, and is not generated when optimization is enabled. Presumably, gcc is warning about the cases where clobbering can occur, and in more recent versions of the compiler, that is the case when optimization is disabled. I've yet to determine in which gcc version the behavior changed, or the rationale for the change.

Reported by Henri van de Water.

2015-03-24

WONTFIX

In the paragraph below the prototype box for malloc(), change:

aligned on a byte boundary suitable for any type of C data structure.

to:

aligned on a byte boundary suitable for efficient access to any type of C data structure.

Explanation:

This change makes the intended idea behind the explanation clearer.

Reported by Sun Jian.

2011-04-07

3

Add a line at the top of Listing 7-1 (memalloc/free_and_sbrk.c):

#define _BSD_SOURCE
#include "tlpi_hdr.h"

#define MAX_ALLOCS 1000000

Explanation:

As noted on page 140, a feature test macro definition is needed for sbrk(). Without it, gcc complains ("implicit declaration of function 'sbrk'") when invoked with (for example) -std=c99.

I missed this problem because of the combination of two reasons: (1) the feature test macro requirements for sbrk() changed in the version of glibc (2.12) that was released not long before the book went to press (see the man page for details), and (2) there was a breakage in my main Makefile (now fixed in the latest version of the source code tarball).

Reported by Lei Yang.

2011-07-05

3

In the shell session shown at the bottom of the page, change:

$ ./free_and_sbrk 1000 10240 1 500 1000
Initial program break:           0x804a6bc
Allocating 1000*10240 bytes
Program break is now:            0x8a13000
Freeing blocks from 500 to 1000 in steps of 1
After free(), program break is:  0x852b000

to:

$ ./free_and_sbrk 1000 10240 1 501 1000
Initial program break:           0x804a6bc
Allocating 1000*10240 bytes
Program break is now:            0x8a13000
Freeing blocks from 501 to 1000 in steps of 1
After free(), program break is:  0x852b000

Explanation:

An off-by-one error in the demonstration...

Reported by Joaquín Moreno.

2024-01-09

15

In the first line of the example code near the top of this page, change:

    struct { /* Some field definitions */ } myStruct;

to:

    struct myStruct { /* Some field definitions */ };

Reported by Sangman Lee.

2011-04-13

3

In the bullet point at the top of the page, change:

The address of the allocated memory is returned in memptr.

to:

The address of the allocated memory is returned in *memptr.

Reported by Yaakov Eisenberg.

2023-02-09

15

In the second bullet point, change:

In this case, the password field in /etc/passwd conventionally contains the letter x (although any nonempty character string may appear), and the encrypted password is instead stored in the shadow password file (Section 8.2).

to:

In this case, the password field in /etc/passwd contains the letter x, and the encrypted password is instead stored in the shadow password file (Section 8.2).

2012-02-08

3

In the last paragraph of Section 8.2, change the sentence:

SUSv3 doesn't specify shadow passwords, and not all UNIX implementations provide this feature.

to:

SUSv3 doesn't specify shadow passwords. Not all UNIX implementations provide this feature, and on implementations where it is provided the details of file locations and APIs vary.

Explanation:

It's worth emphasizing that even on systems where shadow passwords are provided, the details are unstandardized and therefore vary across systems.

Reported by John Wiersba.

2010-12-13

2

In the last paragraph of Section 8.2, change:

Not all UNIX implementations provide this feature, and on implementations where it is provided the details of locations and APIs vary.

to:

Not all UNIX implementations provide this feature, and on implementations where it is provided the details of the file locations and APIs vary.

Explanation:

This change fixes an error that was introduced in the seventh print run.

Reported by Yaakov Eisenberg.

2023-02-09

15

In the bullet point for Encrypted password ID about one third of the way down the page, change:

the passwd command

to:

the gpasswd command

Reported by Justin Pryzby.

2010-12-28

2

In the bullet point for Group ID about half way down the page, change the sentence:

There is normally one group defined with the group ID 0, named root (like the /etc/passwd record with user ID of 0).

to:

There is normally one group defined with the group ID 0, named root (like the /etc/passwd record with user ID of 0, but unlike the user ID 0, this group has no special privileges).

Explanation:

The added detail prevents possible misunderstandings about the significance of GID 0.

Reported by John Wiersba.

2010-12-13

2

This is a note to the text, rather than an erratum.

About half-way down the page, there is the following piece of code in Listing 8-2 (users_groups/check_password.c):

    spwd = getspnam(username);
    if (spwd == NULL && errno == EACCES)
        fatal("no permission to read shadow password file");

    if (spwd != NULL)           /* If there is a shadow password record */
        pwd->pw_passwd = spwd->sp_pwdp;     /* Use the shadow password */

However, depending on your system configuration, getspnam() may not return the error EACCES if the user doesn't have permission to read the shadow password file. In particular, it may be that in your system's /etc/nsswitch.conf file, you see systemd listed:

shadow:     files   systemd

(It is likely that you will see this if you have a Debian-based or Arch-based distribution.)

In the case where the Name Server Switch falls back to using systemd, and systemd determines that the user does not have sufficient permission to access the shadow password file, the getspnam() call simply returns NULL without setting errno. In that case, the program will fall through to a later check that prints a message that the password is incorrect:

    authOk = strcmp(encrypted, pwd->pw_passwd) == 0;
    if (!authOk) {
        printf("Incorrect password\n");
        exit(EXIT_FAILURE);
    }

---

Reported by Joaquín Moreno.

2024-01-09

WONTFIX

In Exercise 8-1, completely replace the text:

8-1. When we execute the following code, we find that it displays the same number twice, even though the two users have different IDs in the password file. Why is this?

    printf("%ld %ld\n", (long) (getpwnam("avr")->pw_uid),
                        (long) (getpwnam("tsr")->pw_uid));

with the following text:

8-1. When we execute the following code, which attempts to display the usernames for two different user IDs, we find that it displays the same username twice. Why is this?

    printf("%s %s\n", getpwuid(uid1)->pw_name,
                      getpwuid(uid2)->pw_name);

Explanation:

The exercise was intended to demonstrate the nonreentrant nature of getpwnam() (and related functions, such as getpwuid()), but the experiment I proposed to do so was wrong. (And unfortunately it looks as though I didn't test this exact example.) As René Thomsen pointed out, the code of the original exercise wouldn't yield the results described in the text "since after each argument a copy of the value in the passwd structure for pw_uid is passed, and since this a direct value (not a reference) there should be no trouble [with] the next call overwriting this value, since it still has the copy."

The revised exercise correctly shows the point I wanted to make. A complete program that can be used to demonstrate the point is shown below.

#include <stdlib.h>
#include <stdio.h>
#include <errno.h>
#include <string.h>
#include <pwd.h>

int
main(int argc, char *argv[])
{
    if (argc != 3 || strcmp(argv[1], "--help") == 0) {
        fprintf(stderr, "Usage: %s uid1 uid2\n", argv[0]);
        exit(EXIT_FAILURE);
    }

    printf("%s %s\n", getpwuid(atoi(argv[1]))->pw_name,
                      getpwuid(atoi(argv[2]))->pw_name);

    exit(EXIT_SUCCESS);
}

Reported by René Thomsen.

2011-07-14

3

In the last sentence of the first paragraph of Section 9.3, a small formatting fix to change:

(The terms set-user-ID program and set-group-ID program are sometimes abbreviated as set-UID program and set-GID program.)

to:

(The terms set-user-ID program and set-group-ID program are sometimes abbreviated as set-UID program and set-GID program.)

Reported by Yaakov Eisenberg.

2023-02-09

15

Completely replace the second small-font paragraph at the top of this page, changing:

In all versions of glibc (including modern ones), setegid(egid) is implemented as setregid(–1, egid). As with seteuid(), this means that we can specify egid as the same value as the current effective group ID, although this behavior is not specified in SUSv3. It also means that setegid() changes the saved set-group-ID if the effective group ID is set to a value other than the current real group ID. (A similar remark applies for the older implementation of seteuid() using setreuid().) Again, this behavior is not specified in SUSv3.

to:

Analogous remarks apply for the glibc implementation of setegid(egid), except that the switch in implementation from setregid(–1, egid) to setresgid(–1, egid, –1) occurred in glibc 2.2 or 2.3 (the precise version depends on the hardware architecture).

Explanation:

The old text was based on ancient knowledge that for a certain period (i.e., glibc 2.1 to 2.2), the implementations of seteuid() and setegid() were not analogous. Somehow, I overlooked the fact that the two implementations later came back into sync in glibc 2.2/2.3.

The new text drops some details of the effect of the old seteuid() and setegid() implementations on the saved set IDs. However, this is probably not too important: the information applies only for rather old versions of glibc and can anyway be deduced from the description of setreuid() and setregid().

Reported by Junjiro Okajima.

2012-06-30

4

About two thirds of the way down the page, change:

Alternatively, an application can make a call to getgroups() specifying gidtsetsize as 0.

to:

Alternatively, an application can make a call to getgroups() specifying gidsetsize as 0.

Reported by Simon Durrant.

2011-08-12

3

In the fifth line from the top of the page, change:

IDs by scanning /etc/groups

to:

IDs by scanning /etc/group

Reported by Yang Yang.

2011-11-30

3

In the third paragraph, change:

Instead, it defines the initial real user ID, effective user ID, and saved set-user-ID of the login shell.

to:

Instead, it defines the initial real group ID, effective group ID, and saved set-group-ID of the login shell.

Reported by Justin Pryzby.

2011-01-03

2

Replace the first bullet point on this page:

The glibc implementations of seteuid() (as setresuid(–1, e, –1)) and setegid() (as setregid(–1, e)) also allow the effective ID to be set to the same value it already has, but this is not specified in SUSv3. The setegid() implementation also changes the saved set-group-ID if the effective user ID is set to a value other than that of the current real user ID. (SUSv3 doesn't specify that setegid() makes changes to the saved set-group-ID.)

with:

The glibc implementations of seteuid() and setegid() also allow the effective ID to be set to the same value it already has, but this is not specified in SUSv3.

Explanation:

See the erratum for page 175 for some background to this erratum.

This erratum renders obsolete an earlier (2011-01-03) erratum by Justin Pryzby. That erratum changed:

The setegid() implementation also changes the saved set-group-ID if the effective user ID is set to a value other than that of the current real user ID.

to:

The setegid() implementation also changes the saved set-group-ID if the effective group ID is set to a value other than that of the current real group ID.

Reported by Junjiro Okajima.

2012-07-02

4

In the first paragraph of Section 10.1, change:

Universal Coordinated Time

to:

Coordinated Universal Time

2022-08-17

14

At the end of the last paragraph on this page, add a sentence:

This argument is now obsolete and should always be specified as NULL. (SUSv4 marks gettimeofday() obsolete, presumably in favor of the POSIX clocks API described in Section 23.5.)

Explanation:

The fact that SUSv4 marks gettimeofday() obsolescent was noted in page 16, but it should also be mentioned here.

2011-08-12

3

In the small-font note at the top of the page, change:

The tz_minuteswest field indicates the number of minutes that must be added to times in this zone to match UTC, with a negative value indicating that an adjustment of minutes to the east of UTC

to:

The tz_minuteswest field indicates the number of minutes that must be added to times in this zone to match UTC, with a negative value indicating an adjustment of minutes to the east of UTC

Reported by Yaakov Eisenberg.

2023-02-09

15

In the third line of the small-font note about two thirds of the way down the page, change:

4.3BSD added the more precise

to:

4.2BSD added the more precise

2010-12-26

2

At the end of the page, add a small-font note as follows:

SUSv4 marks ctime() and asctime() obsolete, because they do not return localized strings (and they are nonreentrant).

Explanation:

This point was noted on page 16, but it bears repeating here, where these functions are described in detail.

2011-08-12

3

In the second to last paragraph on page 189, change:

The tm_sec field can be up to 60 (rather than 59) to account for the leap seconds that are occasionally applied to adjust human calendars to the astronomically exact (the so-called tropical) year.

to:

The tm_sec field can be up to 60 (rather than 59) to account for the leap seconds that are occasionally applied to account for the difference between the precise International Atomic Time (TAI) and the observed solar time, which is less precise because of factors such as geological events and the gradual slowdown in the earth's rotation.

Reported by Yaakov Eisenberg.

2022-08-22

14

In the final paragraph, change the opening piece of the first sentence:

The std and dst components are strings identifying the standard and DST timezones;

to:

The std and dst components are strings that define names for the standard and DST timezones;

Explanation:

This is a clarification to remove the ambiguity of the word "identifying" (does "identifying" mean "defining a name for" or "selecting"?).

Suggested from a conversation with Subu Rama.

2010-12-20

2

About a quarter of the way down the page, add a closing quote at the end of the first code snippet, so that it reads:

    TZ="CET-1:00:00CEST-2:00:00,M3.5.0,M10.5.0"

Reported by Subu Rama.

2010-12-20

2

In the last sentence of the paragraph that precedes Section 10.4, change:

the preceding form is less readable than the Linux-specific near equivalent:

to:

the preceding form is less readable than the near equivalent:

Explanation:

As described at the start of Section 10.3, a timezone string that starts with a colon (as shown in the line following the the quoted text) is standardized in SUSv3. It is only the form without a preceding colon that is nonstandard.

Reported by Junjiro Okajima.

2012-07-02

4

In the last line of the page, change "4.3BSD" to "4.2BSD".

Reported by Roger M. Levasseur.

2010-12-26

2

In the second sentence of exercise 1, change:

Assuming that the clock_t value returned by times() is an unsigned 32-bit integer

to:

Assuming that the clock_t value returned by times() is a signed 32-bit integer

Reported by Olivier Blin.

2015-09-19

7

Starting in the sixth line of the small-font note about one third of the way down the page, change:

which is the per-user limit on the number of processes that may created by this process

to:

which is the per-user limit on the number of processes that may be created by this process

Reported by Yongzhi Pan.

2011-09-20

3

This is a note to the text, rather than an erratum.

In the first paragraph of this page, I refer to /proc as a "virtual file system". It would have been worth noting two points in relation to this term:

• Some other sources use the term "pseudofilesystem" synonymously.
• The Virtual File System described in Section 14,5, is a term used to refer to a specific component of the Linux kernel. This usage is unrelated to the term as used when describing the /proc file system.

However, there isn't sufficient space on the page to add this information.

Reported by Bill McConnaughey.

2011-07-18

WONTFIX

In the third line of the third paragraph from the bottom of the page, change:

/proc/1968/1 is a symbolic link

to:

/proc/1968/fd/1 is a symbolic link

Reported by Junjiro Okajima.

2012-09-27

4

Toward the end of Listing 12-1 (sysinfo/procfs_pidmax.c), change:

    if (argc > 1) {
        if (write(fd, argv[1], strlen(argv[1])) != strlen(argv[1]))
            fatal("write() failed");

        system("echo /proc/sys/kernel/pid_max now contains "
               "`cat /proc/sys/kernel/pid_max`");
    }

to:

    if (argc > 1) {
        if (lseek(fd, 0, SEEK_SET) == -1)
            errExit("lseek");
 
        if (write(fd, argv[1], strlen(argv[1])) != strlen(argv[1]))
            fatal("write() failed");

        system("echo /proc/sys/kernel/pid_max now contains "
               "`cat /proc/sys/kernel/pid_max`");
    }

Explanation:

The code as originally given takes advantage of the fact that the file offset is ignored when writing new values into /proc/sys files. However, this approach isn't very pedagogically correct. Furthermore, with the default value of 0 in the /proc/sys/kernel/sysctl_strict_writes file that was added in Linux 3.16 (see the proc(5) man page), it is now actively discouraged. Therefore, it's better to explicitly set the file offset before performing a write().

Reported by Haiyan Meng.

2015-05-09

6

At the end of the first small-font paragraph at the top of the page, change:

the Linux-specific /proc/hostname file

to:

the Linux-specific /proc/sys/kernel/hostname file

Reported by Ursache Vladimir

2012-04-26

4

At the end of the second small-font paragraph at the top of the page, change:

the Linux-specific /proc/domainname file

to:

the Linux-specific /proc/sys/kernel/domainname file

Reported by Ursache Vladimir

2012-04-26

4

At the top of Listing 12-2 (sysinfo/t_uname.c), change:

#define _GNU_SOURCE

to:

#ifdef __linux__
#define _GNU_SOURCE
#endif

Explanation:

This change improves portability to non-Linux systems, by ensuring that the program doesn't attempt to use the Linux-specific uts.domainname field on those systems.

2013-03-11

4

In the first sentence after the heading in the middle of the page, change:

The kernel performs the same number of disk accesses, regardless of whether we perform 1000 writes of a single byte or a single write of a 1000 bytes.

to:

The kernel performs the same number of disk accesses, regardless of whether we perform 1000 writes of a single byte or a single write of 1000 bytes.

Reported by Yaakov Eisenberg.

2023-02-25

15

In the sentence following the fflush() prototype box, change:

If stream is NULL, fflush() flushes all stdio buffers.

to:

If stream is NULL, fflush() flushes all stdio buffers that are associated with output streams.

2015-01-28

6

In the fifth line from the top of the page, change:

synchronized I/O data completion

to:

synchronized I/O data integrity completion

Reported by Junjiro Okajima.

2012-01-16

3

This is a note to the text, rather than an erratum.

Since Linux 2.6.39, the new syncfs() system call can be used to flush all kernel buffers containing updated file information for a specified file system (i.e., a per-file-system sync()). The file system is specified using a file descriptor that refers to an open file in the file system.

2011-06-26

WONTFIX

This is a note to the text, rather than an erratum.

The discussion of the pdflush kernel thread in the two small-font paragraphs near the bottom of this page is obsolete. (It was already obsolete at the time the book went to press, when kernel 2.6.35 was the current kernel version, but I missed the change.) Although the general principles described in the text remain accurate, in Linux 2.6.32 the pdflush was replaced by the so-called per-BDI flush mechanism, described in the LWN.net article, "Flushing out pdflush".

2012-08-21

WONTFIX

Two changes in the third paragraph.

In the fourth line, change:

synchronized I/O data integrity (i.e., prior to performing the read…

to:

synchronized I/O data integrity completion (i.e., prior to performing the read…

In the seventh and eighth lines, change:

synchronized I/O file integrity (i.e., prior to performing the read…

to:

synchronized I/O file integrity completion (i.e., prior to performing the read…

Reported by Junjiro Okajima.

2012-01-16

3

In the last sentence of the paragraph explaining POSIX_FADV_SEQUENTIAL, change:

to the twice the default size

to:

to twice the default size

Reported by Simon Durrant.

2011-08-23

3

In Listing 13-1 (filebuff/direct_read.c), change one line in the comment in the code:

    we ensure that 'buf' is aligned on a non-power-of-two multiple of

to:

    we ensure that 'buf' is aligned on an odd multiple of

Explanation:

An odd multiple is sufficient, and is what the code actually does.

Reported by Yaakov Eisenberg.

2023-02-25

15

In the second bullet point half-way down the page, change:

Examples of block devices include disks and tape drives.

to:

Disks are a common example of block devices.

Explanation:

Tapes are normally character devices.

Reported by Junjiro Okajima.

2012-01-16

3

Change the last two sentences of the second-to-last paragraph:

Tracks themselves are divided into a number of sectors, each of which consists of a series of physical blocks. Physical blocks are typically 512 bytes (or some multiple thereof) in size, and represent the smallest unit of information that the drive can read or write.

to:

Tracks themselves are divided into a number of sectors, which are typically 512 bytes (or some multiple thereof) in size, and represent the smallest unit of information that the drive can read or write.

Reported by Yaakov Eisenberg.

2023-02-25

15

In the second-to-last paragraph, change:

For even larger files, the fourteenth pointer (numbered 13 in the diagram) is a double indirect pointer—it points to blocks of pointers that in turn point to blocks of pointers that in turn point to data blocks of the file.

to:

For even larger files, the fourteenth pointer (numbered 13 in the diagram) is a double indirect pointer—it points to a block pointers that in turn point to blocks of pointers that in turn point to data blocks of the file.

Reported by Yaakov Eisenberg.

2023-02-25

15

In the "Key" box in the upper right hand corner of Figure 14-2, change:

2IPB = Double IBP

to:

2IPB = Double IPB

Reported by Yongzhi Pan.

2011-09-20

3

This report applies only for the PDF version of the book, and involves two fixes to Figure 14-2.

In print run 3, a fix was made to Figure 14-2, but the fix was was accidentally reverted in print run 6. The same fix must be reapplied.

In the "Key" box in the upper right hand corner of Figure 14-2, change:

2IPB = Double IBP

to:

2IPB = Double IPB

In addition, a formatting glitch appeared in print run 6, such that the bottom edge of a box in the diagram has been lost. This will be fixed in the next PDF version.

Reported by Yaakov Eisenberg.

2023-02-25

15

In Figure 14-4, the arrow pointing from Mount points to the windows icon should point instead to the C icon.

Reported by Dave Walker.

2010-11-28

2

Two changes in Table 14-1.

In the line for the flag MS_REC, change:

Recursive mount (since Linux 2.6.20)

to:

Recursive mount (since Linux 2.4.11)

In the line for the flag MS_RELATIME, change:

Update last access time only if older than last modification time or last status change time (since Linux 2.4.11)

to:

Update last access time only if older than last modification time or last status change time (since Linux 2.6.20)

Explanation:

The version numbers for these two flags were inadvertently swapped. The main text on page 266 showed the correct version numbers for the flags.

Reported by Damien Grassart.

2013-07-17

4

In the paragraph near the top of the page describing MS_BIND, change:

If this flag is specified, then the fstype, mountflags, and data arguments are ignored.

to:

If this flag is specified, then the fstype and data arguments are ignored, as are flags in mountflags other than MS_REC (described below).

2011-04-02

3

In the paragraph in the middle of the page describing MS_MOVE, change the last sentence:

When this flag is specified, the fstype, mountflags, and data arguments are ignored.

to:

When this flag is specified, the fstype and data arguments are ignored, as are the remaining flags in mountflags.

Reported by Sangman Lee.

2011-04-02

3

In the paragraph at the bottom of the page describing MS_REMOUNT, change:

When using this flag, the source and target arguments should be the same as for the original mount() call, and the fstype argument is ignored.

to:

When using this flag, the target argument should be the same as for the original mount() call, and the source and fstype arguments are ignored.

Reported by Junjiro Okajima.

2012-07-04

4

In the second sentence of the first paragraph of Section 14.9.1, change:

From kernel 2.4 onward, a file system can be mounted at multiple locations within the file system.

to:

From kernel 2.4 onward, a file system can be mounted at multiple locations within the single directory hierarchy.

Explanation:

The rewording is more precise, and coherent with Figure 14-4 and the text in Section 14-7.

Reported by Guilherme Maciel Ferreira.

2013-07-21

4

In the sentence immediately preceding Section 14.9.5, change:

we can simply create bind mounts for these directories (possibly mounted read-only) within the jail

to:

we can simply create bind mounts for these directories within the jail

Explanation:

It isn't possible to make read-only bind mounts to read-write directories. As noted on page 265, mountflags other than MS_REC are ignored when MS_BIND is specified.

Reported by Junjiro Okajima.

2012-01-16

3

In the bullet point at the bottom of the page, change:

A tmpfs file system mounted at /dev/shm

to:

A tmpfs file system mounted at /dev/shm (/run/shm on some systems)

Explanation:

See the erratum for page 1108.

2013-09-18

5

In the first sentence of the bullet point at the bottom of the page, change:

For most file Linux systems

to:

For most Linux file systems

Reported by Subu Rama.

2011-01-04

2

In the third bullet point from the top of the page, change:

The f_fsid field is used on some UNIX implementations to return a unique identifier for the file system–for example, a value based on the identifier of the device on which the file system resides. For most Linux file systems, this field contains 0.

to:

The f_fsid field is used to return a unique identifier for the file system–for example, a value based on the identifier of the device on which the file system resides. On older Linux kernels, many types of file systems return 0 in this field.

Explanation:

A few years ago, f_fsid was unsupported on most types of file systems. I overlooked the fact that over time most Linux file systems have added support for this field.

Reported by Junjiro Okajima.

2012-07-04

4

In the third line from the bottom of the page, change:

(x000001, x000001, x0000002, and so on)

to:

(x000000, x000001, x0000002, and so on)

Reported by Junjiro Okajima.

2012-01-16

3

In the third line of the small-font note at the bottom of the page, change:

None of other fields

to:

None of the other fields

Reported by Simon Durrant.

2011-08-23

3

Near the end of the small-font note at the bottom of the page, change:

(On all major contemporary UNIX implementations, symbolic links are implemented as i-nodes.)

to:

(On all major contemporary UNIX implementations, symbolic links are implemented as i-nodes. See Section 18.2 for further details.)

Explanation:

Adding this forward reference to the relevant section is helpful for the reader.

Reported by Justin Pryzby.

2010-12-28

2

At the end of the small-font note at the bottom of the page (after the sentence beginning "On Linux, lstat() returns…"), add the following sentence:

SUSv4 tightens the requirements on an implementation, requiring lstat() to return valid information in all fields of the stat structure except the permission bits of st_mode.

2011-09-30

3

In the last line of the third paragraph, change:

On Linux, they are exposed by <sys/types.h> if the _BSD_SOURCE macro is defined.

to:

On Linux, they are exposed by <sys/sysmacros.h>.

Explanation:

The use of <sys/types.h> is not wrong, but <sys/sysmacros.h> is the preferred method to get the definitions.

2016-04-20

7

At the start of Listing 15-1 (files/t_stat.h), change:

<#define _BSD_SOURCE     /* Get major() and minor() from <sys/types.h> */
#include <sys/types.h>
#include <sys/stat.h>

to:

#include <sys/sysmacros.h>
#include <sys/stat.h>

Explanation:

See the erratum for page 281.

2016-04-20

7

In Table 15-2, in the entry for truncate(), change:

Same for ftruncate(); timestamps change only if file size changes

to:

Same for ftruncate()

Explanation:

The original text did not completely match Linux behavior, but the story here is complicated. On Linux, ftruncate() always changes the file timestamps, even if the file size is not changed. However, Linux truncate() only changes the timestamps if the file size changes. This more or less mirrors the specifications in SUSv3. SUSv3 makes no mention of file size changes when specifying that ftruncate() changes the file timestamps. By contrast, SUSv3 says for truncate() that if the file size is changed, the file timestamps are updated, implying that if the file size is unchanged, the timestamps should not be changed. A planned fix to SUSv4 would require that truncate() behave like ftruncate().

2012-02-14

3

In the sentence immediately preceding heading 15.2.1, change:

can be accessed using field names such st_atim.tv_nsec

to:

can be accessed using field names such as st_atim.tv_nsec

Reported by Yongzhi Pan.

2011-09-20

3

In the prototype box for futimens() at the bottom of the page change:

#include _GNU_SOURCE

to:

#define _GNU_SOURCE

Reported by Junjiro Okajima.

2012-01-16

3

In the middle of this page, change:

An unprivileged process can use chown() to change the group ID of a file that it owns (i.e., the process’s effective user ID matches the user ID of the file) to any of the groups of which they are a member.

to:

An unprivileged process can use chown() to change the group ID of a file that it owns (i.e., the process’s effective user ID matches the user ID of the file) to any of the groups of which it is a member.

Reported by Yaakov Eisenberg.

2023-04-30

15

At the bottom of the page, delete this entire sentence:

(However, if we use the chown(1) command under a root login to change the ownership of a file, then, after calling chown(2), the chown command uses the chmod() system call to reenable the set-user-ID and set-group-ID bits.)

Explanation:

I recall testing this point years ago, but unfortunately I didn't record a log of my test. Junjiro pointed out that the chown command on his system doesn't behave as was described in the deleted sentence, and on my current systems I can't reproduce the behavior either. Furthermore, a web search turns up no description of this point. Thus, either things have changed or my earlier testing was incorrect. At this point, it seems safest to remove that text, but if anyone can confirm the behavior that I thought I saw, please let me know.

Update: Justin Pryzby kindly pointed me at Debian Bug 112597, which shows that the behavior described in the now-deleted text did indeed occur in the past, but was subsequently changed.

Reported by Junjiro Okajima.

2012-07-05

4

In the fourth line of code from the top of this page, change:

            fatal("No group user (%s)", argv[1]);

to:

            fatal("No group user (%s)", argv[2]);

2018-06-19

9

In the fourth line of code from the top of this page, change:

            fatal("No group user (%s)", argv[1]);

to:

            fatal("No such group (%s)", argv[2]);

Reported by Yaakov Eisenberg.

2023-02-26

15

Add some words to the last sentence in the small-font note at the top of the page:

To ensure that we are using an unadulterated ls, we can specify the full pathname of the command (/bin/ls) or precede the ls command with a backslash to prevent alias substitution.

Reported by Junjiro Okajima.

2012-01-16

3

This is a note to the text, rather than an erratum.

The example shell session at the top of this page is technically correct, but could have been better chosen. As explained earlier in Section 15.4.5, the sticky bit no longer serves any purpose for regular (executable) files, but it does serve a purpose for directories. Therefore, a better example would have been to demonstrate manipulating the sticky bit of a directory, rather than a file.

Reported by Simon Durrant.

2011-08-23

WONTFIX

In the shell session near the top of the page, change:

    $ chattr +ai myfile             Turn on Append Only and Immutable flags

to:

    $ sudo chattr +ai myfile        Turn on Append Only and Immutable flags

Explanation:

As shown in Table 15-8 and explained on page 306, privilege is required to set the Append Only and Immutable attributes.

Reported by Liu Jiaming.

2013-06-05

4

In the first line on the page, change:

The various FL_* flags and their meanings are as follows:

to:

The various FS_* flags and their meanings are as follows:

Reported by Douglas Luu.

2011-11-12

3

In the last line of the paragraph for FS_NOTAIL_FL, change:

mount -notail

to:

mount -o notail

Reported by Justin Pryzby.

2010-12-28

2

This is a note to the text, rather than an erratum.

The first bullet point on this page notes that various filesystems require that the filesystem be mounted with the mount –o user_xattr option in order to employ user EAs. Among the listed filesystems is ext4. However, since Linux 2.6.39, this option is no longer required for ext4: user EAs are enabled by default.

2015-01-22

WONTFIX

In the first sentence of the paragraph that precedes the bulleted list in Section 16.2, change:

It is only possible to place user EAs on files and directories.

to:

It is only possible to place user EAs on regular files and directories.

Explanation:

From the bulleted list, the reader can deduce that this sentence is referring to regular files, but it's better to make that explicit.

Reported by Junjiro Okajima.

2012-01-16

3

This is a note to the text, rather than an erratum.

The second bullet point on this page notes that on JFS there is a 128 kB limit on the total bytes used by the names and values of all EAs on a file. More recent kernels have removed this limit, so that it is possible to create much larger sets of EAs on a file. I haven't determined what the limiting factors are now, or when the changes occurred, but on a Linux 3.0 system I was able to attach 4 MB of EAs to a file on JFS.

Reported by Junjiro Okajima.

2012-07-06

WONTFIX

At the bottom of this page (in Listing 16-1, xattr/xattr_view.c), change:

    if (optind >= argc + 2)
        usageError(argv[0]);

to:

    if (optind >= argc)
        usageError(argv[0]);

Reported by Liu Jiaming.

2013-06-05

4

Towards the end of Listing 16-1 (xattr/xattr_view.c), fix a casting error by changing:

                printf("value=");
                for (k = 0; k < valueLen; k++)
                    printf("%02x ", (unsigned int) value[k]);

to:

                printf("value=");
                for (k = 0; k < valueLen; k++)
                    printf("%02x ", (unsigned char) value[k]);

2018-05-03

9

This is a note to the text, rather than an erratum.

The second paragraph of the small-font note on this page notes that various filesystems require that the filesystem be mounted with the mount –o acl option in order to user ACLs. Among the listed filesystems is ext4. However, since Linux 2.6.39, this option is no longer required for ext4: ACLs are enabled by default.

2015-01-22

WONTFIX

In the first bullet point, about half way down the page, change:

Simply modifying the ACL_GROUP_OBJ and ACL_USER_OBJ entries

to:

Simply modifying the ACL_GROUP_OBJ and ACL_OTHER entries

Reported by Jorge Merlino.

2011-03-08

2

At the bottom of this page, change:

    chmod(pathname, 751);

to:

    chmod(pathname, 0751);

Reported by Junjiro Okajima.

2020-09-02

13

In the box in the bottom left-hand corner of Figure 17-2, change "acl_type_t" to "acl_tag_t".

(This erratum was modified slightly on 2021-08-17, after a note from Claudio Fornaro.)

Reported by Andrew McKinlay.

2020-08-11

13

In the second paragraph under the heading Retrieving entries from an in-memory ACL, change the last sentence:

Thus, we can loop through all of the entries in an ACL by specifying type as ACL_FIRST_ENTRY in the first call to acl_get_entry() and specifying type as ACL_NEXT_ENTRY in subsequent calls.

to:

Thus, we can loop through all of the entries in an ACL by specifying entry_id as ACL_FIRST_ENTRY in the first call to acl_get_entry() and specifying entry_id as ACL_NEXT_ENTRY in subsequent calls.

Reported by Yongzhi Pan.

2011-09-20

3

In the last sentence on this page, change:

The tag_type argument has the type acl_type_t

to:

The tag_type argument has the type acl_tag_t

Reported by Andrew McKinlay.

2020-08-11

13

In the first sentence of the small-font note near the top of the page, change:

The acl_check() and acl_error() functions (the latter is a Linux extension)

to:

The acl_check() and acl_error() functions (both are Linux extensions)

Reported by Junjiro Okajima.

2012-01-16

3

About one third of the way down the page, change:

        /* Retrieve and display optional tag qualifier */

        if (tag == ACL_USER) {
            uidp = acl_get_qualifier(entry);
            if (uidp == NULL)
                errExit("acl_get_qualifier");

            name = groupNameFromId(*uidp);

to:

        /* Retrieve and display optional tag qualifier */

        if (tag == ACL_USER) {
            uidp = acl_get_qualifier(entry);
            if (uidp == NULL)
                errExit("acl_get_qualifier");

            name = userNameFromId(*uidp);

Reported by René Thomsen.

2011-08-11

3

The bottommost rectangle in the I-node table should be gray, as are the other two areas between explicitly shown i-nodes further up in the table.

2013-11-14

5

This is a note to the text, rather than an erratum.

This note relates to the section headed Interpretation of symbolic links by system calls. Linux 3.6 added some restrictions on the operation of symbolic links; these restrictions can cause system calls that dereference symbolic links to fail with the error EACCES if they are violated. The goal of these restrictions is to prevent a class of security issues caused by time-of-check, time-of-use races when accessing symbolic links. For more information, see this LWN.net article and the documentation of /proc/sys/fs/protected_symlinks in the proc(5) man page.

2012-10-15

WONTFIX

This is a note to the text, rather than an erratum.

Linux 3.6 added some restrictions on the creation of hard links; these restrictions can cause link() to return with an EPERM error if they are violated. The goal of the changes is to prevent a class of security issues caused by hard-link-based time-of-check, time-of-use races, most commonly seen in world-writable directories such as /tmp. For more information, see the documentation of /proc/sys/fs/protected_hardlinks in the proc(5) man page.

2012-10-15

WONTFIX

This is a note to the text, rather than an erratum.

This note relates to the shell session shown at the top of the page. Depending on factors including the file system type, the size of file created, and system scheduling decisions, you may not always see the behavior shown in this example. Specifically, by the time that the second df command is executed (via system()) the blocks of the closed file may not yet have been freed, so that the output of the second df shows no change in the amount of disk space used in the file system.

As a workaround, adding a sufficiently long sleep() call immediately after the close() call in dirs_links/t_unlink.c should ensure that the file blocks have been freed; in all cases that I tested (including very large file sizes), sleep(1) was sufficient to ensure that the file space was freed by the time of the second df command. Note that calling sync() after the close() call does not suffice to eliminate this behavior. Presumably, this is because sync() flushes dirty blocks in the buffer cache, whereas the file blocks in question here are being (asynchronously) deallocated by the kernel.

Reported by Yang Yang.

2011-12-04

WONTFIX

In the last sentence of the third paragraph, change:

Alternatively, we can size pathname using the PATH_MAX constant

to:

Alternatively, we can size buffer using the PATH_MAX constant

Reported by Olivier Blin.

2015-09-21

7

Below the prototype box for readdir() (about half way down the page), change:

Each call to readdir() reads the next directory from the directory stream

to:

Each call to readdir() reads the next directory entry from the directory stream

Reported by Subu Rama.

2011-01-12

2

In the second line, change:

stat() on the pathname

to:

lstat() (or stat(), if a symbolic link should be dereferenced) on the pathname

Suggested by a report by Justin Pryzby.

2010-12-28

2

At the end of the small-font note near the top of the page, add a sentence:

Although not specified in SUSv3, scandir() is provided on most UNIX implementations. SUSv4 added a specification for scandir().

Reported by Geoff Clare.

2010-12-31

2

This is a note to the text, rather than an erratum.

Since glibc 2.24, the readdir_r() function has been marked as deprecated, for the reasons discussed in Austin Group bug report 696, this glibc bug report, and in the readdir_r() manual page.

This erratum was updated on 15 Aug 2017 to amend the glibc version where readdir)r() was deprecated from 2.23 to 2.24.

2016-02-29

WONTFIX

In the last line of the page, change:

(say 10 or more)

to:

(say 20 or more)

Explanation:

A recommendation of "10 or more" isn't wrong as such. However, one case where we might use nftw() is to traverse the entire file system under the root directory, and commonly the directory tree for that case would be deeper than 10 levels (for example, on one of my personal systems, the file system tree runs to a depth of 19).

Reported by Junjiro Okajima.

2012-07-09

4

Near the top of the page, change:

The fourth argument to func, ftwbuf, is pointer to a structure defined as follows:

to:

The fourth argument to func, ftwbuf, is a pointer to a structure defined as follows:

Reported by Subu Rama.

2011-01-12

2

In Listing 18-3 (dirs_links/nftw_dir_tree.c), change:

    switch (sbuf->st_mode & S_IFMT) {      /* Print file type */
    case S_IFREG:  printf("-"); break;
    case S_IFDIR:  printf("d"); break;
    case S_IFCHR:  printf("c"); break;
    case S_IFBLK:  printf("b"); break;
    case S_IFLNK:  printf("l"); break;
    case S_IFIFO:  printf("p"); break;
    case S_IFSOCK: printf("s"); break;
    default:       printf("?"); break;     /* Should never happen (on Linux) */
    }
 
    printf(" %s  ", (type == FTW_D)  ? "D  " : (type == FTW_DNR) ? "DNR" :
            (type == FTW_DP) ? "DP " : (type == FTW_F)   ? "F  " :
            (type == FTW_SL) ? "SL " : (type == FTW_SLN) ? "SLN" :
            (type == FTW_NS) ? "NS " : "  ");
 
    if (type != FTW_NS)
        printf("%7ld ", (long) sbuf->st_ino);
    else
        printf("        ");
 
    printf(" %*s", 4 * ftwb->level, "");        /* Indent suitably */
    printf("%s\n",  &pathname[ftwb->base]);     /* Print basename */
    return 0;                                   /* Tell nftw() to continue */
}

to:

    if (type == FTW_NS) {                  /* Could not stat() file */
        printf("?");
    } else {
        switch (sbuf->st_mode & S_IFMT) {  /* Print file type */
        case S_IFREG:  printf("-"); break;
        case S_IFDIR:  printf("d"); break;
        case S_IFCHR:  printf("c"); break;
        case S_IFBLK:  printf("b"); break;
        case S_IFLNK:  printf("l"); break;
        case S_IFIFO:  printf("p"); break;
        case S_IFSOCK: printf("s"); break;
        default:       printf("?"); break; /* Should never happen (on Linux) */
        }
    }

    printf(" %s  ", (type == FTW_D)  ? "D  " : (type == FTW_DNR) ? "DNR" :
            (type == FTW_DP) ? "DP " : (type == FTW_F)   ? "F  " :
            (type == FTW_SL) ? "SL " : (type == FTW_SLN) ? "SLN" :
            (type == FTW_NS) ? "NS " : "  ");
    if (type != FTW_NS)
        printf("%7ld ", (long) sbuf->st_ino);
    else
        printf("        ");
    printf(" %*s", 4 * ftwb->level, "");        /* Indent suitably */
    printf("%s\n",  &pathname[ftwb->base]);     /* Print basename */
    return 0;                                   /* Tell nftw() to continue */
}

Explanation:

Before inspecting sbuf->st_mode, we must check if the file was was stat()-able. (There are additional white-space changes to ensure that the listing will still fit on the printed page.)

Reported by Raam sri.

2017-11-25

9

This is a note to the text, rather than an erratum.

Because of a regression that took place in glibc 2.4, the example shell session demonstrating the use of the nftw_dir_tree.c program may (if you are using a glibc version that does not have a fix for the regression) produce different results than are shown on this page.

In particular, in more or less the middle of the page there is the following line of output:

    1 SLN  2327987       dsl

Compare this line with an identical line shown further down the page where the nftw_dir_tree.c program is employed with the –p option, which specifies the use of the FTW_PHYS flag, meaning display information about symbolic links themselves, rather than the files they refer to.

The fact that the same i-node number is shown for the dangling symbolic link, dsl, both with and without the use of the FTW_PHYS flag indicates that even when the FTW_PHYS flag is not specified, if nftw() encounters a dangling symbolic link, then it uses lstat() to obtain a stat structure describing the symbolic link and that structure is passed to the callback function.

In glibc 2.4, there was a change in behavior that appears to have been unintended. Following that change, in the case where the FTW_PHYS flag is not specified, then when a dangling symbolic link is encountered, the contents of the stat structure that is passed to the callback function are left undefined. (In practice, this typically means that the last stat structure that was previously fetched for a file will be supplied to the callback function.)

This changed behaviour of the glibc nftw() implementation differed from every other implementation that I've tested or examined (Solaris, OpenBSD, FreeBSD, musl) and also my reading of POSIX.1, which notes that in the case where FTW_NS is passed in the typeflag argument of the callback function, the contents of the stat structure are undefined, but makes no similar statement regarding the case where FTW_SLN is passed in the typeflag argument (with the implication that the contents of the stat buffer are defined). The older (pre-2.4) glibc implementation of nftw() behaved in the same way as other current implementations.

I was already aware that there was an issue regarding the difference between the current glibc nftw() implementation and the behavior of nftw() on other systems, but LaoK's bug report against my book reminded me that I had myself earlier determined that the glibc implementation used to behave the same as other nftw() implementations.

This regression led to some bug reports against glibc (Red Hat bugzilla report 1422736 and glibc bug 23501) and Austin bug report 1121 against the POSIX.1 standard. Ultimately, it was resolved that this was a glibc bug, and the regression was fixed in glibc 2.30 (and possibly backported to some earlier glibc versions).

This erratum was updated on 2020-07-02 to reflect the fact that the glibc bug has been fixed.

Reported by LaoK.

2017-06-11

WONTFIX

The same typo, where a "1" should be an "l", appears twice on this page.

About half way down the page, change:

In the above output, we can see that the symbolic link s1 was resolved.

to:

In the above output, we can see that the symbolic link sl was resolved.

Further down the page, change:

From the above output, we can see that the symbolic link s1 was not resolved.

to:

From the above output, we can see that the symbolic link sl was not resolved.

Reported by Junjiro Okajima.

2013-04-16

4

This is a note to the text, rather than an erratum.

Since Section 18.11 was written, some new AT_* flags have been added to various system calls mentioned in this section.

Linux 2.6.38 added the AT_NO_AUTOMOUNT flag for fstatat(). This flag can be used to prevent automounting of the terminal component of a pathname given to fstatat().

Linux 2.6.39 added the AT_EMPTY_PATH flag for linkat(), fchownat(), fstatat(), and name_to_handle_at(). (An explanation of the last of these system calls can be found on LWN.net.) This flag allows empty relative pathnames to be specified for these system calls, in which case the calls operate on the directory file descriptor.

2011-06-26

WONTFIX

Just over half way down the page, change:

The classic example of the use of chroot() is in the ftp program. As a security measure, when a user logs in anonymously under FTP, the ftp program uses chroot() to set the root directory for the new process to the directory specifically reserved for anonymous logins.

to:

The classic example of the use of chroot() is in the ftpd program (the FTP server daemon). As a security measure, when a user logs in anonymously under FTP, the ftpd program uses chroot() to set the root directory for the new process to the directory specifically reserved for anonymous logins.

Reported by Junjiro Okajima.

2012-07-09

4

Two changes associated with the description of inotify_rm_watch().

1. In the prototype box at the top of the page, change the line:

int inotify_rm_watch(int fd, uint32_t wd);

to:

int inotify_rm_watch(int fd, int wd);

2. In the paragraph below the prototype box, remove the sentence:

(The uint32_t data type is an unsigned 32-bit integer.)

Explanation:

The calling signature of this function changed in glibc 2.10.

2010-10-25

2

This is a note to the text, rather than an erratum.

Linux 2.6.36 adds a new input flag for inotify, IN_EXCL_UNLINK, that prevents children of a watched directory from generating events for a directory after they have been unlinked from that directory. This flag would slot into Table 19-1, just below IN_DONT_FOLLOW. Further information can be found in the inotify(7) manual page.

2011-05-08

WONTFIX

In the discussion of the IN_ATTRIB event in the first bullet point, it is worth noting that the "link count changed" case also includes deletion of a file. However, there isn't sufficient space on the page to add this information.

Reported by Bill McConnaughey.

2011-07-18

WONTFIX

At the end of the first of the two bullet points at the bottom of this page, there is the following sentence:

An IN_IGNORED event is not generated when a watch that was established with IN_ONESHOT is automatically removed because an event was triggered.

The semantics described in that sentence were by design, The idea being that if an application employed the IN_ONESHOT flag, then it knew that on the next event the watch would be dropped, and did not need the "noise" of an event to be informed of that fact.

However, starting with kernel 2.6.36, an IN_ONESHOT event is generated in this case. This change in semantics was almost certainly an unintended consequence of code reworking that went on during the development of the fanotify API.

2014-03-28

WONTFIX

In the second bullet point at the top of the page, change:

We describe the IN_Q_OVERFLOW in Section 19.5,

to:

We describe the IN_Q_OVERFLOW event in Section 19.5,

Reported by Yaakov Eisenberg.

2023-04-06

15

In the third line of the paragraph near the top of the page that starts "The cookie field…", change:

and then an IN_MOVED_TO is

to:

and then an IN_MOVED_TO event is

2011-12-08

3

Add a sentence at the end of the paragraph near the top of the page that starts "The cookie field…":

These two events will have the same unique value in their cookie field, thus allowing the application to associate them. For all other types of event, the cookie field is set to 0.

Explanation:

This point was made implicitly in the program in Listing 19-1 (where the displayInotifyEvent() function contains the check if (i->cookie > 0)), but should of course have been made explicit in the text.

Reported by Yang Yang.

2011-12-06

3

In the third paragraph from the bottom of the page (starting "Using a larger buffer size"), change:

A read() from an inotify file descriptor returns the minimum of the number of events that are available and the number of events that will fit in the supplied buffer.

to:

A read() from an inotify file descriptor reads the minimum of the number of events that are available and the number of events that will fit in the supplied buffer.

Explanation:

This change is made to prevent the possible misinterpretation that the return value of read() is the number of items read. As shown in Figure 19-2, a read() from an inotify file descriptor returns the number of bytes read (as with a read() from other types of file descriptors).

Reported by Junjiro Okajima.

2012-01-16

3

In Listing 19-1 (inotify/demo_inotify.c), near the top of the page, change:

int
main(int argc, char *argv[])
{
    int inotifyFd, wd, j;
    char buf[BUF_LEN];

to:

int
main(int argc, char *argv[])
{
    int inotifyFd, wd, j;
    char buf[BUF_LEN] __attribute__((aligned(8)));

Explanation:

This change is required because further down in the program there is the following assignment, which casts p (which was initialized to point to buf):

event = (struct inotify_event *) p;

The code as originally given will work on many architectures, since the compiler will most likely align buf to some natural boundary (e.g., 4-byte) on the machine. Furthermore, on x86, things will work even if buf is aligned on a boundary other than 4-byte, since x86 tends to be very forgiving about alignments.

However, the C standards do not guarantee that compilers will align variables in this fashion and do not guarantee that one can cast possibly nonaligned integer fields as I've shown in the original code (i.e., an architecture may require that the integer fields in struct inotify are aligned on a suitable boundary). Therefore, one should force buf to be suitably aligned. The gcc aligned attribute shown above forces buf to be aligned on a 8-byte boundary, which provides correct alignment even on ILP64 platforms, where the size of int is 64 bits.

A more precise (albeit long-winded) way of ensuring the correct alignment would be to declare buf as follows:

    char buf[BUF_LEN]
        __attribute__ ((aligned(__alignof__(struct inotify_event))));

(The gcc __alignof__ keyword yields the alignment of an object.)

Yet another approach that is a little less explicit, but does not depend on gcc features, is to declare buf as an array of struct inotify_event, relying on the fact that C aligns arrays of structures to the alignment requirements of the base structure. Then, some casts are needed for the subsequent pointer operations. The required changes to the code are as follows:

@@ -32,13 +32,11 @@
         printf("        name = %s\n", i->name);
 }

-#define BUF_LEN (10 * (sizeof(struct inotify_event) + NAME_MAX + 1))
-
 int
 main(int argc, char *argv[])
 {
     int inotifyFd, wd, j;
-    char buf[BUF_LEN] __attribute__ ((aligned(8)));
+    struct inotify_event buf[1000];
     ssize_t numRead;
     char *p;
     struct inotify_event *event;
@@ -59,7 +57,7 @@
     }

     for (;;) {                                  /* Read events forever */
-        numRead = read(inotifyFd, buf, BUF_LEN);
+        numRead = read(inotifyFd, buf, sizeof(buf));
         if (numRead == 0)
             fatal("read() from inotify fd returned 0!");

@@ -70,7 +68,7 @@

         /* Process all of the events in buffer returned by read() */

-        for (p = buf; p < buf + numRead; ) {
+        for (p = (char *) buf; p < (char *) buf + numRead; ) {
             event = (struct inotify_event *) p;
             displayInotifyEvent(event);

Note, however, that is approach will cause gcc -Wpedantic to give a warning "invalid use of structure with flexible array member" because of the field definition char name[] at the end of the inotify_event structure.

This erratum was originally composed on 2013-11-18, but then modified (and the explanation considerably expanded) after a conversation with Heinrich Schuchardt.

Reported by Matt Wojciak.

2014-05-21

5

In the first shell session shown at the top of the page, change:

$ mkdir dir2/ddd
Read 32 bytes from inotify fd
    wd = 1; mask = IN_CREATE IN_ISDIR
        name = ddd

to:

$ mkdir dir2/ddd
Read 32 bytes from inotify fd
    wd = 2; mask = IN_CREATE IN_ISDIR
        name = ddd

Reported by Eric Dickinson.

2015-01-13

6

This is a note to the text, rather than an erratum.

At the end of Chapter 19, rather more that could have been said about the limitations of the inotify API. In the meantime, I have made some substantial revisions to the inotify(7) man page to discuss limitations and pitfalls that one should be aware of when using the API. Having read Chapter 19, it's strongly advisable to read that manual page.

2014-04-06

WONTFIX

Change the opening sentence on this page:

This chapter and next two chapters discuss signals.

to:

This chapter and the next two chapters discuss signals.

2011-01-03

2

In the fourth bullet point, change:

the use of a process signal mask

to:

the use of the process signal mask

Explanation:

Each process has a single signal mask, so using the definite article "the" is better than the indefinite article "a".

2011-01-03

2

In the third line of the paragraph describing SIGLOST, change:

if the NSF client fails

to:

if the NFS client fails

Reported by Suse Shi.

2011-04-22

3

In the third line of the paragraph describing SIGLOST, change:

by the those processes

to:

by those processes

Reported by Simon Durrant.

2012-12-12

4

A 2012-07-02 erratum that was incorporated into the fourth print run unfortunately added a technical error into the text. That erratum read:

In the final bullet point on the page, replace the third sentence:

Excluding the effective user ID of the target from the check serves a complementary purpose: it prevents one user from sending signals to another user's process that is running a set-user-ID program belonging to the user trying to send the signal.

with:

Excluding the receiver's effective user ID from the check serves a complementary purpose: it allows a set-user-ID program to retain privileges while preventing the delivery of signals from the user that started the program (by using setresuid() to change the saved set-user-ID to the same value as the real user ID).

The words "from the user that started the program" should be "from the user that owns the executable". However, an even better replacement text is the following:

Furthermore, on Linux and other systems that provide the setresuid() system call, a set-user-ID program can take advantage of this rule to prevent itself being sent signals by the user that owns the executable, by using setresuid() to make its saved set-user-ID the same as the real user ID.

Thus the entire text of the bullet point should now read:

An unprivileged process can send a signal to another process if the real or effective user ID of the sending process matches the real user ID or saved set-user-ID of the receiving process, as shown in Figure 20-2. This rule allows users to send signals to set-user-ID programs that they have started, regardless of the current setting of the target process's effective user ID. Furthermore, on Linux and other systems that provide the setresuid() system call, a set-user-ID program can take advantage of this rule to prevent itself being sent signals by the user that owns the executable, by using setresuid() to make its saved set-user-ID the same as the real user ID. (SUSv3 mandates the rules shown in Figure 20-2, but Linux followed slightly different rules in kernel versions before 2.0, as described in the kill(2) manual page.)

Reported by Lorenzo Monti

2014-01-05

5

In the paragraph just above the heading for Section 20-7, change:

This program takes two command-line arguments, a signal number and a process ID,

to:

This program takes two command-line arguments, a process ID and a signal number,

Explanation:

This makes the order of the arguments given in the text match the order of the arguments for Listing 20-3.

2011-05-18

3

In Listing 20-3 (signals/t_kill.c), change:

    if (argc != 3 || strcmp(argv[1], "--help") == 0)
        usageErr("%s sig-num pid\n", argv[0]);

to:

    if (argc != 3 || strcmp(argv[1], "--help") == 0)
        usageErr("%s pid sig-num\n", argv[0]);

Explanation:

See also the erratum on page 404.

Reported by Madhavan Kasthurirangan.

2011-05-18

3

This is a note to the text, rather than an erratum.

Starting with version 2.32 (August 2020), the sys_siglist variable described in Section 20.8 is no longer exported by glibc.

2020-10-10

WONTFIX

In the first sentence below the prototype box for sigemptyset() and sigfillset(), change:

One of sigemptyset() or sigaddset() must be used to initialize a signal set.

to:

One of sigemptyset() or sigfillset() must be used to initialize a signal set.

Reported by Bill McConnaughey.

2011-07-18

3

In the second prototype box on this page, change:

int sigaddset(sigset_t *set, int sig);
int sigdelset(sigset_t *set, int sig);

to:

int sigaddset(const sigset_t *set, int sig);
int sigdelset(const sigset_t *set, int sig);

Reported by Robert P. J. Day.

2013-09-24

5

At the bottom of this page, change the text in the prototype box for sigismember() and the sentence immediately following the prototype box as follows:

Returns 1 if sig is a member of set, otherwise 0

The sigismember() function returns 1 (true) if sig is a member of set, and 0 (false) otherwise

to:

Returns 1 if sig is a member of set, 0 if it is not, or –1 on error

The sigismember() function returns 1 (true) if sig is a member of set, 0 (false) if it is not a member, or –1 on error (e.g., sig is not a valid signal number)

2015-06-15

6

In the last line of the prototype box at the top of the page, change:

Returns 1 if sig is empty, otherwise 0

to:

Returns 1 if set is empty, otherwise 0

Reported by Junjiro Okajima.

2012-01-16

3

In the prototype box at the top of the page, change the line:

int sigandset(sigset_t *set, sigset_t *left, sigset_t *right);

to:

int sigandset(sigset_t *dest, sigset_t *left, sigset_t *right);

Reported by Matias Virsu.

2010-12-22

2

Change the text of Exercise 20-4:

Implement siginterrupt() using sigaction().

to:

Implement the siginterrupt() function described in Section 21.5 using sigaction().

Explanation:

This exercise should have been at the end of Chapter 21. However, as an exercise in the use of sigaction(), it works fine in Chapter 20 as well.

Reported by Subu Rama.

2011-01-14

2

in Listing 21-1 (signals/nonreentrant.c), change:

static char *str2;              /* Set from argv[2] */
static int handled = 0;         /* Counts number of calls to handler */

to:

static char *str2;                  /* Set from argv[2] */
static volatile int handled = 0;    /* Counts number of calls to handler */

Explanation:

Since the variable handled is accessed both from the main program and asynchronously from the signal handler, it should me marked as volatile in order to prevent the optimizer aliasing the variable to a register.

Reported by Joaquín Moreno.

2024-03-11

15

This is a note to the text, rather than an erratum.

Regarding Table 21-1, POSIX.1-2008 Technical Corrigendum 1 (POSIX.1-2013) adds the following functions to the list of async-signal-safe functions: fchdir(), pthread_kill(), pthread_self(), and pthread_sigmask().

POSIX.1-2008 Technical Corrigendum 2 (POSIX.1-2016), adds even more functions to the list of async-signal-safe functions, including longjmp() and siglongjmp(), as well as a host of functions that deal with strings, wide-character strings, and memory comparisons and copying. For details, see the new signal-safety(7) manual page.

Explanation:

This note was updated on 12 Nov 2016 to reflect the POSIX.1-2016 changes.

2015-09-11

WONTFIX

In the first paragraph of Section 21.1.3, in the third line from end of the paragraph, change:

using the volatile attribute

to:

using the volatile keyword

Explanation:

Tightening up the terminology, so as to avoid possible confusion with the gcc attribute feature.

Reported by Yang Yang.

2011-12-09

3

This is a note to the text, rather than an erratum.

A small-font note near the bottom of this page says

If we define the _BSD_SOURCE feature test macro when compiling a program, then (the glibc) setjmp() follows the BSD semantics.

Since the deprecation of _BSD_SOURCE in glibc 2.19, _BSD_SOURCE no longer has this effect for setjmp(). For further details, see the setjmp(3) manual page.

2017-08-15

WONTFIX

In the shell session about one quarter of the way down the page, change the second line:

    $ ./sigmask_siglongjmp x

to:

    $ ./sigmask_siglongjmp

Explanation:

The "x" argument is superfluous (it was used in an earlier version of the program).

Reported by Subu Rama.

2011-01-14

2

Near the top of the page, in the paragraph under si_code, change:

as shown in Table 21-1.

to:

as shown in Table 21-2.

Reported by Subu Rama.

2011-01-14

2

Near the top of of this table, change the entry in the second column for the row SI_USER from:

A user process via kill() or raise()

to:

A user process via kill()

Explanation:

Once upon a time, raise() did produce an si_code of SI_USER, but modern glibc versions implement raise() raise using tgkill(), which yields an si_code of SI_TKILL. The simplest fix here is to remove mention of raise() in this table. I made some recent changes to the raise() man page to allow programmers to deduce the details.

2012-04-20

4

This is a note to the text, rather than an erratum.

In Section 21.5, in the first paragraph following the numbered list, the error EINTR is described in parentheses as "Interrupted function". This text is the SUSv3 description of this error. However, glibc associates this error with a slightly different text: "Interrupted system call".

Reported by Junjiro Okajima.

2012-01-16

WONTFIX

In the first bullet point on this page, change the second sentence:

The I/O system calls are interruptible, and hence automatically restarted by SA_RESTART only when operating on a "slow" device.

to:

The I/O system calls are interruptible, and hence automatically restarted by SA_RESTART, only when operating on a "slow" device.

Reported by Ulf (Ulfalizer) Magnusson.

2015-02-01

6

A bit more than one third of the way down this page, delete a sentence:

The sem_wait() and sem_timedwait() functions used to decrement a POSIX semaphore. (On some UNIX implementations, sem_wait() is restarted if the SA_RESTART flag is specified.)

to:

The sem_wait() and sem_timedwait() functions used to decrement a POSIX semaphore.

Reported by Yaakov Eisenberg.

2023-04-30

15

This is a note to the text, rather than an erratum.

A bullet point at the bottom of this page notes that a read from an inotify file descriptor is never restarted, regardless of the setting of the SA_RESTART flag. However, starting with Linux 3.8, the SA_RESTART flag does cause a read from an inotify file descriptor to be restarted when interrupted by a signal handler.

2014-12-11

WONTFIX

This is a note to the text, rather than an erratum.

Under the heading Unhandled stop signals can generate EINTR for some Linux system calls, it is noted that certain system calls can produce the EINTR error even when there is no signal handler. Among the operations listed here are reads from inotify file descriptors. However, starting with Linux 3.8, reads from inotify file descriptors no longer demonstrate this behavior.

2014-12-11

WONTFIX

About half way down this page, change:

In kernels before 2.6.24, poll() also exhibited this behavior, as did sem_wait(), sem_timedwait(), futex(FUTEX_WAIT), in kernels before 2.6.22, msgrcv() and msgsnd() in kernels before 2.6.9, and nanosleep() in Linux 2.4 and earlier.

to:

In kernels before 2.6.24, poll() also exhibited this behavior, as did sem_wait(), sem_timedwait(), and futex(FUTEX_WAIT) in kernels before 2.6.22, msgrcv() and msgsnd() in kernels before 2.6.9, and nanosleep() in Linux 2.4 and earlier.

Explanation:

This small edit makes the sentence easier to parse.

Reported by Yaakov Eisenberg.

2023-04-30

15

Remove the fifth bullet point on this page:

interruption of system calls by signal handlers, and how to automatically restart interrupted system calls;

Explanation:

This topic was covered in the previous chapter, in Section 21-5 (and indeed, there's a very similar bullet point on the first page of Chapter 21). At a certain point during some chapter reorganizations, I overlooked this superfluous piece of text at the start of Chapter 22.

Reported by Trevor Woerner.

2012-09-28

4

In the second-to-last bullet point on the page, change:

the use of signalfd() to receive a signal via file descriptor; and

to:

the use of signalfd() to receive a signal via a file descriptor; and

Reported by Murray McAllister.

2012-05-16

4

This is a note to the text, rather than an erratum.

With respect to Table 22-1, Linux 3.0 added a new specifier, %E. This specifier is replaced by the pathname of the executable file, with slashes (/) substituted by exclamation marks (!). (The substitution is required because slashes can't form part of a filename—they delimit the filename components of a pathname.)

Reported by Junjiro Okajima.

2012-01-16

WONTFIX

This is a note to the text, rather than an erratum.

With respect to Table 22-1, Linux 3.7 added a new specifier, %P. This specifier is replaced by the PID of the process as seen in the initial PID namespace (whereas the existing %p specifier is replaced by the PID in the PID namespace where the process resides).

2014-03-14

WONTFIX

This is a note to the text, rather than an erratum.

With respect to Table 22-1, Linux 3.12 added a new specifier, %d. This specifier is replaced by the "dumpable" mode of the process (the same value as is returned by the prctl(2) PR_GET_DUMPABLE operation).

2014-03-14

WONTFIX

In the first paragraph, change:

If, at the time it was execed, a program finds that the disposition of a terminal-generated signals

to:

If, at the time it was execed, a program finds that the disposition of a terminal-generated signal

Reported by Vince Manapat.

2014-01-05

5

In the second and third lines of the paragraph starting "SUSv3 requires..." that follows the two bullet points at the top of the page, change:

The Linux kernel defines 32 different realtime signals, numbered from 32 to 63.

to:

The Linux kernel defines 33 different realtime signals, numbered from 32 to 64.

2015-02-18

6

In the third line of the paragraph starting "SUSv3 requires..." that follows the two bullet points at the top of the page, change:

The <signal.h> header file

to:

The <limits.h> header file

Reported by Trevor Woerner.

2012-09-30

4

In the second line of the last paragraph on this page, change:

that may be queued to a process

to:

that may be queued by a process

2011-12-12

3

Replace the first two paragraphs on this page:

On Linux, this call returns –1. The reason for this is that Linux employs a different model for limiting the number of realtime signals that may be queued to a process. In Linux versions up to and including 2.6.7, the kernel enforces a system-wide limit on the total number of realtime signals that may be queued to all processes. This limit can be viewed and (with privilege) changed via the Linux-specific /proc/sys/kernel/rtsig-max file. The default value in this file is 1024. The number of currently queued realtime signals can be found in the Linux-specific /proc/sys/kernel/rtsig-nr file.

Starting with Linux 2.6.8, this model was changed, and the aforementioned /proc interfaces were removed. Under the new model, the RLIMIT_SIGPENDING soft resource limit defines a limit on the number of signals that can be queued to all processes owned by a particular real user ID. We describe this limit further in Section 36.3.

with:

In systems with glibc versions before 2.4, this call returns –1. Since glibc 2.4, the return value depends on the kernel version. Before Linux 2.6.8, the call returns the value in the Linux-specific /proc/sys/kernel/rtsig-max file. This file defines a system-wide limit on the number of realtime signals that may be queued to all processes. The default value is 1024, but a privileged process can change it. The Linux-specific /proc/sys/kernel/rtsig-nr file shows the number of currently queued realtime signals.

Starting with Linux 2.6.8, these /proc files disappear. In their place, the RLIMIT_SIGPENDING resource limit (Section 36.3) limits the number of signals that can be queued to all processes owned by a particular real user ID. Since glibc 2.10, the sysconf() call returns the RLIMIT_SIGPENDING limit. (The SigQ field of the Linux-specific /proc/PID/status file displays the number of realtime signals pending for a process.)

Explanation:

Once upon a time, sysconf(_SC_SIGQUEUE_MAX) always returned –1. Although the other points in the text relating to the /proc files and the RLIMIT_SIGPENDING limit were all correct, I missed the fact that the glibc behavior changed, and to best capture the implications of that fact required reworking the entire text.

Reported by Yang Yang.

2011-12-12

3

This is a note to the text, rather than an erratum.

Although I refer to sigqueue() twice on this page (and once on page 477) as a system call, strictly speaking, it is a library function (layered on top of the Linux rt_sigqueueinfo(2) system call). In September 2011, I made changes to the Linux man pages to more accurately reflect that fact, documenting the underlying system call and moving the sigqueue() manual page from Section 2 (System Calls) to Section 3 (Library Functions).

2011-09-30

WONTFIX

In the second paragraph from the top of the page, change the second sentence:

This program takes up to four arguments, of which the first three are mandatory: a signal number, a target process ID, and an integer value to accompany the realtime signal.

to:

This program takes up to four arguments, of which the first three are mandatory: a target process ID, a signal number, and an integer value to accompany the realtime signal.

Explanation:

This change makes the order of the text match the order of the arguments in the program in Listing 22-2.

Reported by Sangman Lee.

2011-04-02

3

Near the start of listing 22-3 (signals/catch_rtsigs.c), change:

static volatile int handlerSleepTime;
static volatile int sigCnt = 0;         /* Number of signals received */
static volatile int allDone = 0;

to:

static volatile int handlerSleepTime;
static volatile int sigCnt = 0;         /* Number of signals received */
static volatile sig_atomic_t allDone = 0;

Explanation:

As was already shown in various other programs and described in Section 21.1.3 of TLPI, a flag variable that is shared between the main program and a signal handler should be declared with the type sig_atomic_t. Note that the two other global variables that are accessed in the signal handler, sigCnt and handlerSleepTime don't need to be declared in the same way, because the logic of the code is such that the signal handler can never interrupt the main program while it is accessing those variables.

Reported by Alexander Pozdneev.

2019-11-21

11

At the end of Listing 22-3 (signals/catch_rtsigs.c), change:

    while (!allDone)                    /* Wait for incoming signals */
        pause();
}

to:

    while (!allDone)                    /* Wait for incoming signals */
        pause();

    printf("Caught %d signals\n", sigCnt);
    exit(EXIT_SUCCESS);
}

Explanation:

These lines were accidentally deleted from the published version of the program. Without them, the line of output "Caught 6 signals" will not appear in the shell session shown on page 462.

Reported by Yang Yang.

2011-12-12

3

Near the top of the page, change:

Uses sigsuspend() to unblock SIGINT and SIGQUIT ...

to:

Use sigsuspend() to unblock SIGINT and SIGQUIT ...

Reported by Yaakov Eisenberg.

2023-04-30

15

In the prototype box, change:

Returns number of delivered signal on success

to:

Returns signal number on success

Explanation:

The new wording is a little more concise, and avoids the possible misreading "number of delivered signals". (See also the erratum below about rewording "is delivered" to "becomes pending".)

2011-01-03

2

In the paragraph immediately following the sigwaitinfo() prototype box, change:

The sigwaitinfo() system call suspends execution of the process until one of the signals in the signal set pointed to by set is delivered. If one of the signals in set is already pending at the time of the call, sigwaitinfo() returns immediately. The delivered signal is removed from the process's list of pending signals, and the signal number is returned as the function result

to:

The sigwaitinfo() system call suspends execution of the process until one of the signals in the signal set pointed to by set becomes pending. If one of the signals in set is already pending at the time of the call, sigwaitinfo() returns immediately. One of the signals is removed from the process's list of pending signals, and the signal number is returned as the function result

Explanation:

The sigwaitinfo() system call does not "deliver" signals. Rather, it removes a signal from the caller's set of pending signals, blocking if necessary until a signal becomes pending. The use of "delivered" in this context was a little sloppy, since it may mislead the reader into thinking of a signal being delivered to a signal handler (and indeed, in normal usage, signals are blocked while calling sigwaitinfo() so that they can't be caught by a signal handler).

Reported by Daniel Zingaro.

2012-07-20

4

At the end of the second-to-last paragraph, change:

(see Exercise 22-3.)

to:

(see Exercise 22-3)

2011-01-04

2

This is a note to the text, rather than an erratum.

Ideally, the following information regarding portability should have been included at the bottom of this page:

On Linux, a blocked signal whose disposition is ignore is added to the process's set of pending signals if it is generated. However, according to SUSv3, an implementation may discard the signal in this case. (SUSv3 XSH 2.4.1: "If the action associated with a blocked signal is to ignore the signal and if that signal is generated for the process, it is unspecified whether the signal is discarded immediately upon generation or remains pending.") Thus, in addition to blocking the signal, a portable application that uses sigwaitinfo() to accept a signal must ensure that the signal is not ignored, by either installing a handler for the signal or ensuring that the signal's disposition is SIG_DFL (unless the default action is to ignore the signal).

However, there isn't sufficient space on the page to include this information.

Explanation:

This note was originally composed on 2011-12-08, but then expanded on 2012-02-14.

2012-02-14

WONTFIX

At the top of the page, insert a small-font note as follows:

According to SUSv3, calling sigwaitinfo() without blocking the signals in set results in undefined behavior.

2011-12-08

3

In the prototype box near the top of the page, change:

Returns number of delivered signal on success

to:

Returns signal number on success

Explanation:

The new wording is a little more concise, and avoids the possible misreading "number of delivered signals". (See also the erratum for page 468 about rewording "is delivered" to "becomes pending".)

2011-01-03

2

In the definition of the struct signalfd_siginfo structure, change:

    int32_t ssi_fd;        /* File descriptor (SIGPOLL/SIGIO) */
    uint32_t ssi_tid;      /* Kernel timer ID (POSIX timers) */
    uint32_t ssi_band;     /* Band event (SIGPOLL/SIGIO) */
    uint32_t ssi_tid;      /* (Kernel-internal) timer ID (POSIX timers) */
    uint32_t ssi_overrun;  /* Overrun count (POSIX timers) */

to:

    int32_t ssi_fd;        /* File descriptor (SIGPOLL/SIGIO) */
    uint32_t ssi_tid;      /* (Kernel-internal) timer ID (POSIX timers) */
    uint32_t ssi_band;     /* Band event (SIGPOLL/SIGIO) */
    uint32_t ssi_overrun;  /* Overrun count (POSIX timers) */

Explanation:

The definition of the ssi_tid field was accidentally duplicated, with two different accompanying comments. The fix removes the duplicate while retaining the slightly more detailed comment.

Reported by Sangman Lee.

2011-05-12

3

In the prototype box for sigvec(), change:

int sigvec(int sig, struct sigvec *vec, struct sigvec *ovec);

to:

int sigvec(int sig, const struct sigvec *vec, struct sigvec *ovec);

2014-05-21

5

In the definition of the sigvec structure shown in the middle of the page, change:

    void (*sv_handler)();

to:

    void (*sv_handler)(int);

Reported by Li Wenjun.

2012-09-06

4

In the prototype box at the bottom of the page, change:

int sigmask(sig);

to:

int sigmask(int sig);

Reported by Junjiro Okajima.

2012-01-16

3

In the second paragraph on this page, change:

Note that this function is defined with different calling signatures in the System V and BSD APIs.

to:

Note that although a function with this name exists in both the System V and BSD APIs, the int argument is interpreted differently in the two APIs.

Reported by Yaakov Eisenberg.

2023-05-05

15

In the first line of Section 22.14, change:

This file contains

to:

Core dumps contain

Explanation:

This is a rewording that clarifies the text a little.

Reported by Justin Pryzby.

2010-12-28

2

In the second sentence of the second paragraph of Section 22.14, change:

Asynchronous generation occurs when a signal is sent a process by the kernel

to:

Asynchronous generation occurs when a signal is sent to a process by the kernel

Reported by Murray McAllister.

2012-05-16

4

In the second line of the third paragraph from the bottom of the page, change:

suspend execution until a signal arrives, The atomicity of

to:

suspend execution until a signal arrives. The atomicity of

(i.e., change a comma to a period.)

Reported by Murray McAllister.

2012-05-16

4

In the last line of Exercise 22-2, change:

period time

to:

period of time

2012-09-29

4

In the paragraph in the middle of the page beginning "The default disposition", change:

we must to establish

to:

we must establish

Reported by Eric Arora.

2011-01-05

2

In the sixth line from the bottom of the page, change:

the timer is expires

to:

the timer expires

Reported by Eric Arora.

2011-01-05

2

In the third paragraph, change:

When the timer reaches 0, the corresponding signal is sent to the process, and then, if the interval (it_interval) is nonzero, the timer value (it_value) is reloaded, and counting down toward 0 recommences.

to:

When the timer reaches 0, the corresponding signal is sent to the process, and then, if the interval (it_interval) is nonzero, the timer value (it_value) is reloaded with the interval value, and counting down toward 0 recommences.

Explanation:

This change helps avoid reader misunderstanding by reiterating a point already made (in different words) at the foot of page 480.

Reported by Sangman Lee.

2011-04-02

3

In the fourth line from the top of the page, change:

non-async-signal-functions

to:

non-async-signal-safe functions

Reported by Kiju Kim.

2012-02-15

3

In the main() function of Listing 23-1 (timers/real_timer.c), change:

    if (argc > 1 && strcmp(argv[1], "--help") == 0)
        usageErr("%s [secs [usecs [int-secs [int-usecs]]]]\n", argv[0]);
 
    sigCnt = 0;

to:

    if (argc > 1 && strcmp(argv[1], "--help") == 0)
        usageErr("%s [secs [usecs [int-secs [int-usecs]]]]\n", argv[0]);

Explanation:

The sigCnt variable is also initialized to zero lower down in the main() function.

Reported by Cyrus Ramavarapu.

2019-02-15

11

This fix applies to a segment of code that spans the bottom of page 483 and the top of page 484. The change moves a block of code to a location lower down in the program (but otherwise leaves the lines of code unchanged). Change:

    sigemptyset(&sa.sa_mask);
    sa.sa_flags = 0;
    sa.sa_handler = sigalrmHandler;
    if (sigaction(SIGALRM, &sa, NULL) == -1)
        errExit("sigaction");

    /* Exit after 3 signals, or on first signal if interval is 0 */

    maxSigs = (itv.it_interval.tv_sec == 0 &&
                itv.it_interval.tv_usec == 0) ? 1 : 3;

    displayTimes("START:", FALSE);

    /* Set timer from the command-line arguments */

    itv.it_value.tv_sec = (argc > 1) ? getLong(argv[1], 0, "secs") : 2;
    itv.it_value.tv_usec = (argc > 2) ? getLong(argv[2], 0, "usecs") : 0;
    itv.it_interval.tv_sec = (argc > 3) ? getLong(argv[3], 0, "int-secs") : 0;
    itv.it_interval.tv_usec = (argc > 4) ? getLong(argv[4], 0, "int-usecs") : 0;

    if (setitimer(ITIMER_REAL, &itv, 0) == -1)
        errExit("setitimer");

to:

    sigemptyset(&sa.sa_mask);
    sa.sa_flags = 0;
    sa.sa_handler = sigalrmHandler;
    if (sigaction(SIGALRM, &sa, NULL) == -1)
        errExit("sigaction");

    /* Set timer from the command-line arguments */

    itv.it_value.tv_sec = (argc > 1) ? getLong(argv[1], 0, "secs") : 2;
    itv.it_value.tv_usec = (argc > 2) ? getLong(argv[2], 0, "usecs") : 0;
    itv.it_interval.tv_sec = (argc > 3) ? getLong(argv[3], 0, "int-secs") : 0;
    itv.it_interval.tv_usec = (argc > 4) ? getLong(argv[4], 0, "int-usecs") : 0;

    /* Exit after 3 signals, or on first signal if interval is 0 */

    maxSigs = (itv.it_interval.tv_sec == 0 &&
                itv.it_interval.tv_usec == 0) ? 1 : 3;

    displayTimes("START:", FALSE);
    if (setitimer(ITIMER_REAL, &itv, 0) == -1)
        errExit("setitimer");

Explanation:

The initialization of itv produces a value used in the initialization of maxSigs. Therefore the initializations must be in that order. (When I originally wrote the program the initializations were in the correct order, but during the production phase of the book I accidentally made a change that introduced the error.)

Reported by Subu Rama.

2011-01-17

2

In Listing 23-1 (timers/real_timer.c) in the line marked with a circled 4, change:

    if (setitimer(ITIMER_REAL, &itv, 0) == -1)

to:

    if (setitimer(ITIMER_REAL, &itv, NULL) == -1)

Explanation:

Strictly speaking, using 0 for the last argument of setitimer() is exactly equivalent to using NULL, since the C standard guarantees that 0 is treated as NULL in locations where a pointer expected. However, using NULL more clearly indicates that this is a pointer, and is what I do in all of the rest of the code in the book.

Reported by Trevor Woerner.

2012-09-30

4

In Listing 23-2 (timers/timed_read.c), in the line of code fifth from the top of the page, change:

    numRead = read(STDIN_FILENO, buf, BUF_SIZE - 1);

to:

    numRead = read(STDIN_FILENO, buf, BUF_SIZE);

Explanation:

A small error that crept in as a thinko or via cut-and-paste. Since there's no need to tack a terminating null byte onto the end of this buffer, it's okay to allow read() to use the complete buffer.

2014-06-30

5

In the first line of Section 23.4.1, change:

The sleep() function suspends execution of the calling process

to:

The sleep() function suspends execution of the calling thread

2014-04-05

5

In the last sentence of the small-font note at the bottom of the page, change:

and the time spent in the sleep state

to:

and the time spent in the stopped state

Reported by Alexander Jaoshvili.

2016-08-03

7

About two thirds of the way down the page, in the paragraph beginning "The time value", change the last sentence in the paragraph:

The clock_getres() system call returns a pointer to a timespec structure containing the resolution of the clock specified in clockid.

to:

The clock_getres() system call returns, via the argument res, a pointer to a timespec structure containing the resolution of the clock specified in clockid.

Explanation:

The reader could probably deduce this information from the declaration in the preceding prototype box, but this change removes any ambiguity about whether the information is returned as the function result or via the argument res.

Reported by Bill McConnaughey.

2011-08-12

3

About two thirds of the way down the page, in the paragraph beginning "The time value", change the last sentence in the paragraph:

The clock_getres() system call returns via the argument res, a pointer to a timespec structure containing the resolution of the clock specified in clockid.

to:

The clock_getres() system call returns the resolution of the clock specified in clockid in the timespec structure pointed to by res.

Explanation:

A refinement to an earlier fix that was made in the third print run.

Reported by Yaakov Eisenberg.

2023-05-05

15

This is a note to the text, rather than an erratum.

The second paragraph of Section 23.5 says:

On Linux, programs using this API must be compiled with the –lrt option, in order to link against the librt (realtime) library.

Starting with glibc 2.17, this is no longer true. This change was requested in glibc bug 14743, so that single-threaded programs using the clock_* functions do not end up incurring the cost of dragging in libpthread. The change affects clock_getcpuclockid(), clock_getres(), clock_gettime(), clock_nanosleep(), and clock_settime().

2012-10-29

WONTFIX

This is a note to the text, rather than an erratum.

Since Section 23.5 was written, some new clocks have been added to the POSIX clocks API.

Linux 2.6.39 adds CLOCK_BOOTTIME. This clock is the same as CLOCK_MONOTONIC, except that it includes time that the system has spent in suspended state. The new clock is intended for applications that want a monotonically increasing clock and also want to be aware of time the system has been suspended.

Linux 3.0 adds two clocks to the POSIX clocks API: CLOCK_BOOTTIME_ALARM and CLOCK_REALTIME_ALARM. These clocks behave identically to CLOCK_REALTIME and CLOCK_BOOTTIME, but the _ALARM suffixed clocks will wake the system if it is suspended. A new CAP_WAKE_ALARM capability governs the use of these clocks.

2011-06-26

WONTFIX

In the second of the two small-font paragraphs in the middle of the page, change the first sentence:

Linux 2.6.32 adds two more new clocks to those listed in Table 23-1: CLOCK_REALTIME_COARSE and CLOCK_MONOTIC_COARSE.

to:

Linux 2.6.32 adds two more new clocks to those listed in Table 23-1: CLOCK_REALTIME_COARSE and CLOCK_MONOTONIC_COARSE.

Reported by Junjiro Okajima.

2012-08-07

4

This is a note to the text, rather than an erratum.

Relevant to Section 23.5.2: since Linux 2.6.39, the new clock_adjtime() system call provides an interface to adjust the value of a POSIX clock (if that clock supports the interface) in the same manner as the adjtimex(2) system call.

2011-06-26

WONTFIX

In the definition of the sigevent structure in the middle of the page, change:

        pid_t      _tid;      /* ID of thread to be signaled /

to:

        pid_t      _tid;      /* ID of thread to be signaled */

Reported by Junjiro Okajima.

2012-01-16

3

Add a sentence under the description of SIGEV_THREAD_ID; change:

(With SIGEV_SIGNAL notification, a signal is queued to the process as a whole, and, if there are multiple threads in the process, the signal will be delivered to an arbitrarily selected thread in the process.)

to:

(With SIGEV_SIGNAL notification, a signal is queued to the process as a whole, and, if there are multiple threads in the process, the signal will be delivered to an arbitrarily selected thread in the process. See Section 33.2 for a discussion of the interaction of threads and signals.)

2011-07-19

3

In the second paragraph of Section 23.6.3, change:

The curr_value.it_value field returns the time until next timer expiration, even if this timer was established as an absolute timer using TIMER_ABSTIME.

to:

The curr_value.it_value field returns the time until the next timer expiration, even if this timer was established as an absolute timer using TIMER_ABSTIME.

Reported by Yaakov Eisenberg.

2023-05-06

15

Remove an unneeded line from the top of Listing 23-6 (timers/itimerspec_from_str.c). Change:

#define _POSIX_C_SOURCE 199309
#include <string.h>
#include <stdlib.h>

to:

#include <string.h>
#include <stdlib.h>

2013-07-11

4

In Listing 23-6 (timers/itimerspec_from_str.c), change:

void
itimerspecFromStr(char *str, struct itimerspec *tsp)
{
    char *cptr, *sptr;

    cptr = strchr(str, ':');
    if (cptr != NULL)
        *cptr = '\0';
    sptr = strchr(str, '/');
    if (sptr != NULL)
        *sptr = '\0';

    tsp->it_value.tv_sec = atoi(str);

to:

void
itimerspecFromStr(char *str, struct itimerspec *tsp)
{
    char *dupstr, *cptr, *sptr;

    dupstr = strdup(str);
 
    cptr = strchr(dupstr, ':');
    if (cptr != NULL)
        *cptr = '\0';
    sptr = strchr(dupstr, '/');
    if (sptr != NULL)
        *sptr = '\0';

    tsp->it_value.tv_sec = atoi(dupstr);

Explanation:

With this change, the itimerspecFromStr() function no longer modifies its string argument. Without this fix, the programs shown in Listing 23-5 (timers/ptmr_sigev_signal.c) and Listing 23-7 (timers/ptmr_sigev_thread.c) will not produce exactly the output shown on pages 503 and 505. In particular, the lines near the top of each of the shell sessions, which show the timer ID and the corresponding command-line argument, will not correctly echo the command-line argument strings. (This error occurred because the text and source code tarballs inadvertently included an older version of the source code.)

Reported by Liu Jiaming.

2013-07-11

4

At the end of Listing 23-6 (timers/itimerspec_from_str.c), change:

        tsp->it_interval.tv_sec = atoi(cptr + 1);
        tsp->it_interval.tv_nsec = (sptr != NULL) ? atoi(sptr + 1) : 0;
    }
}

to:

        tsp->it_interval.tv_sec = atoi(cptr + 1);
        tsp->it_interval.tv_nsec = (sptr != NULL) ? atoi(sptr + 1) : 0;
    }
    free(dupstr);
}

Explanation:

See the explanation for the erratum on page 502.

Reported by Liu Jiaming.

2013-07-11

4

This is a correction for the PDF version of the book only. In the second sentence of Section 23.6.7, the text "30" should hyperlink to Chapter 30 in the PDF. However, this link is broken and needs repairing to point to the correct location.

Reported by Yaakov Eisenberg.

2023-05-19

15

In the shell session in the bottom half of the page change five instances of:

    main(): count =

to:

    main(): expireCnt =

Explanation:

After a late change to the source code, I missed refreshing the shell session log.

Reported by Trevor Woerner.

2012-10-02

4

This is a note to the text, rather than an erratum.

Since Linux 3.0, the timerfd_settime() system call supports a further flag, TFD_TIMER_CANCEL_ON_SET. If this flag is set for a CLOCK_REALTIME absolute (TFD_TIMER_ABSTIME) timer, then the timer is expired if the clock is reset.

2011-06-26

WONTFIX

In the first paragraph of this page, change:

In this and the next three chapters, we look at how a process is created and terminates, and how a process can execute a new program. This chapter covers process creation. However, before diving into that subject, we present a short overview of the main system calls covered in these four chapters.

to:

In this and the next four chapters, we look at how a process is created and terminates, and how a process can execute a new program. This chapter covers process creation. However, before diving into that subject, we present a short overview of the main system calls covered in these chapters.

Explanation:

Late in the writing work, there was a chapter split: Chapters 25 and 26 were once one chapter. During the split, this text at start of the Chapter 24 did not get fixed.

Reported by Trevor Woerner.

2012-10-03

4

In the small-font note at the top of this page, change:

generally only one of the parent and child terminate by calling exit();

to:

generally only one of the parent and child terminates by calling exit();

Reported by Yaakov Eisenberg.

2023-05-19

15

In the the first bullet point at the top of the page, fix a formatting error by changing:

Second, the termination status of the child is returned in the status argument of wait().

to:

Second, the termination status of the child is returned in the status argument of wait().

Reported by Yaakov Eisenberg.

2023-05-19

15

In the second sentence of the second paragraph, change:

are initially exact duplicates of the corresponding parts the parent's memory

to:

are initially exact duplicates of the corresponding parts of the parent's memory

Reported by Murray McAllister.

2012-05-19

4

In the second paragraph from the bottom of the page, change:

that it inherits during the during the fork().

to:

that it inherits during the fork().

Reported by Bill McConnaughey.

2011-07-18

3

In the paragraph at the bottom of the page, change:

These duplicates are made in the manner of dup(), which means that corresponding descriptors in the parent and the child refer to the same open file description.

to:

The duplicated file descriptors in the child refer to the same open file descriptions as the corresponding descriptors in the parent.

Explanation:

The old text could potentially mislead the reader about the treatment of file descriptor flags during fork(), in particular, the close-on-exec flag discussed on pages 576-578. When a file descriptor is duplicated by dup(), the close-on-exec flag is cleared for the duplicated descriptor. By contrast (and as described on page 613), when a file descriptor is duplicated in the child during fork(), the value of the close-on-exec flag is made a copy of the value for the corresponding file descriptor in the parent.

Reported by Conor O'Rourke.

2019-01-09

11

This is a note to the text, rather than an erratum.

The first paragraph below the listing on this page starts with the following text:

Sharing of open file attributes between the parent and child processes is frequently useful. For example, if the parent and child are both writing to a file, sharing the file offset ensures that the two processes don't overwrite each other's output.

Implicit in that text is the notion that file I/O operations (e.g., read() and write()) are atomic with respect to updates to the file offset. POSIX.1-2008/SUSv4 Section XSI 2.9.7 ("Thread Interactions with Regular File Operations") is explicit on this point:

All of the following functions shall be atomic with respect to each other in the effects specified in POSIX.1-2008 when they operate on regular files or symbolic links: … read(), readv(). … write(), writev().

Among the effects specified by POSIX for various I/O functions is of course updating the current file offset.

However, a longstanding bug meant that Linux did not conform to the POSIX requirements (and was not consistent with other implementations such as FreeBSD and Solaris). In short, when performing file I/O system calls, the I/O operation plus accompanying file offset change were not performed atomically on Linux. Consequently, if two process (or threads) shared an open file description (and thus the current file offset, as described in Section 5.4 of TLPI), and performed simultaneous reads or writes, then there was a chance that the blocks of data transferred by those I/O system calls would (incorrectly) overlap with one another. This problem was fixed in Linux 3.14.

I made the problem report that led to the fix after Yongzhi Pan reported puzzling results when running my disc_SIGHUP.c program (Listing 34-4) as described on pages 713 and 714 of TLPI. Yongzhi noted that some of the expected output was not appearing and when I dug deeper, I discovered the nonatomicity issue. The reason that I had not discovered the (kernel) problem when I originally wrote the program was that the problem behavior appears much less frequently on single-CPU systems, and I happened to have tested the program on a single-CPU system as I wrote that chapter of the book. After Yongzhi's report, I was immediately (and easily) able to reproduce the behavior he reported on a multi-CPU system.

Reported by Yongzhi Pan.

2014-05-04

WONTFIX

In the first line of the last paragraph on the page, change:

If sharing of file descriptors

to:

If sharing of open file attributes

Reported by Simon Durrant.

2012-12-12

4

In the last line of the first bullet point at the top of the page, change:

virtual memory page frames

to:

physical memory page frames

Reported by Madhavan Kasthurirangan.

2011-09-04

3

In the fourth line from the end of the second bullet point, change:

This new page copy is assigned to the faulting process, and the corresponding page-table entry for the child is adjusted appropriately.

to:

This new page copy is assigned to the faulting process, and the corresponding page-table entry for the other process is adjusted appropriately.

Reported by Trevor Woerner.

2012-10-16

4

In Exercise 24-3, change:

at a given moment in time

to:

at a given point in the program

Explanation:

The revised wording is more precise.

Reported by Bill McConnaughey.

2011-07-18

3

In the first normal-size paragraph at the top of the page, change:

2,147,482,647

to:

2,147,483,647

Reported by Junjiro Okajima.

2012-01-16

3

Fix a page reflow error that first appeared in print run 5, so that the code snippet at the bottom of the page does not split badly onto the top of page 536.

Reported by Yaakov Eisenberg.

2023-05-19

15

In the second bullet point, half way down the page, change:

in an application that creates child processes, typically only one of the processes (most often the parent) should terminate via exit(), while the other processes should terminate via _exit().

to:

in an application that creates child processes that don't exec new programs, typically only one of the processes (most often the parent) should terminate via exit(), while the other processes should terminate via _exit().

Explanation:

This change clarifies the intended meaning of the explanation by adding an important detail.

Suggested by a conversation with Justin Pryzby.

2011-01-03

2

In Exercise 25-1, change:

If a child process makes the call exit(-1), what exit status (as returned by WEXITSTATUS()) will be seen by the parent?

to:

If a child process makes the call exit(-1), what exit status will be seen by the parent?

Explanation:

Mention of WEXITSTATUS()), which is only covered in the next chapter, isn't needed to solve this exercise.

2013-10-24

5

This is a note to the text, rather than an erratum.

The first paragraph on this page includes the following text:

(A future corrigendum to SUSv4 is likely to add a requirement that si_pid and si_signo are zeroed in this case.)

This requirement was indeed added in the 2013 Technical Corrigendum 1 to SUSv4.

Reported by Ильдар Низамов (Ildar Nizamov)

2017-09-14

WONTFIX

This is a note to the text, rather than an erratum.

Section 26.2 discusses the traditional semantics for adoption of orphaned children by the init process. Linux 3.4 added the PR_SET_CHILD_SUBREAPER prctl() operation, which allows a "service manager" process to mark itself as a sort of 'sub-init', so that it adopts all orphaned processes created by the services it starts. All SIGCHLD signals will be delivered to the service manager. There is a corresponding PR_GET_CHILD_SUBREAPER prctl() operation.

More recent init frameworks employ the PR_SET_CHILD_SUBREAPER feature to create a dedicated subreaper process that becomes the parent of all orphaned children. On such frameworks, which include systemd (dependent on start-up options) and (the now obsolete) upstart, getppid() will report the new parent PID as the PID of the subreaper process (i.e., a value other than 1) when a process becomes orphaned.

Explanation:

This errata updated on 2017-11-02, after a note from Matthew Meffan.

2012-04-24

WONTFIX

In Listing 26-4 (procexec/make_zombie.c), change:

    default:    /* Parent */
        sleep(3);               /* Give child a chance to start and exit */
        snprintf(cmd, CMD_SIZE, "ps | grep %s", basename(argv[0]));
        cmd[CMD_SIZE - 1] = '\0';       /* Ensure string is null-terminated */
        system(cmd);            /* View zombie child */

to:

    default:    /* Parent */
        sleep(3);               /* Give child a chance to start and exit */
        snprintf(cmd, CMD_SIZE, "ps | grep %s", basename(argv[0]));
        system(cmd);            /* View zombie child */

Explanation:

While writing a few code examples, it looks like I forgot that snprintf() always terminates its argument, even when it completely fills the buffer. (This was probably me confusing with the behavior of strncpy(). See page 793 of the book.)

Reported by Madhavan Kasthurirangan.

2011-09-02

3

Change the second and third sentences of the first paragraph of Section 26.3:

A parent can't predict when one of its child will terminate. (Even if the parent sends a SIGKILL signal to the child, the exact time of termination is still dependent on when the child is next scheduled for use of a CPU.)

to:

A parent can't predict when one of its children will terminate. (Even if the parent sends a SIGKILL signal to a child, the exact time of termination is still dependent on when the child is next scheduled for use of a CPU.)

Reported by Murray McAllister.

2012-05-31

4

This is a note to the text, rather than an erratum.

Exercise 26-1 asks the reader to verify that after a process becomes orphaned, its parent PID (PPID, as returned by getppid()) becomes 1. Historically, this was true. But on some modern init frameworks such as systemd, a dedicated process (with PID other than 1) takes on the role of adopting orphaned children. Thus, in code that solves the exercise, the PPID of the process will change as the child becomes orphaned, but the change with be to a value other than 1. For further details, see the erratum for page 553.

A similar qualification also applies for exercise 26-2.

Explanation:

This erratum was updated on 2019-11-19 include mention of exercise 26-2.

Reported by Matthew Meffan.

2017-11-02

WONTFIX

In the explanation of ETXTBSY, change:

The file referred to by pathname is open for writing by another process

to:

The file referred to by pathname is open for writing by one or more processes

Explanation:

The ETXTBSY error could of course occur if the caller of execve() has itself opened pathname for writing.

Reported by Junjiro Okajima.

2012-01-16

3

Towards the end of Listing 27-4 (procexec/t_execle.c), change:

    execle(argv[1], filename, "hello world", (char *) NULL, envVec);

to:

    execle(argv[1], filename, "hello world", "goodbye", (char *) NULL, envVec);

Explanation:

This makes the program do the same thing as the program in Listing 27-1, as stated in the text on page 570.

Reported by Tobin Harding.

2016-04-01

7

This is a note to the text, rather than an erratum.

The last paragraph on the page should have noted that:

The fexecve() function is not specified in SUSv3, but is specified in SUSv4.

Explanation:

At various other places in the book (pages 15 and 426), I noted that SUSv4 specified fexecve(). That fact should also have been mentioned in Section 27.2.4. However, there isn't sufficient space on the page to add this sentence.

2010-12-31

WONTFIX

In the fifth line of the first paragraph of the small-font note on this page, change:

(which is given as an argument to the script,

to:

(which is given as an argument to the interpreter,

Reported by Junjiro Okajima.

2012-01-16

3

In the fourth line of the second paragraph of the small-font note on this page, change:

(and these are passed as separate words to the script)

to:

(and these are passed as separate words to the interpreter)

Reported by Junjiro Okajima.

2012-01-16

3

In the last line of the small-font note at the top of the page (starting "The Linux 2.2 kernel…"), change:

would show just the value echo.

to:

would show just the value necho.

Reported by Subu Rama.

2011-01-27

2

In the third of the bullet points at the top of the page, change:

If all system calls succeed, then system() returns the termination status of the child shell used to execute command. (The termination status of a shell is the termination status of the last command it executes.)

to:

If all system calls succeed, then system() returns the termination status of the child shell used to execute command. The termination status of a shell is the exit status of the last command it executes; if that command is killed by a signal, most shells exit with the value 128+n, where n is the signal number. (If the child shell is itself killed by a signal, the termination status is as described in Section 26.1.3.)

Explanation:

There are two clarifications here. The first of these reiterates a point made in a small-font note near the top of page 532: when a command is terminated by a signal, the shell indicates that fact with a status of 128 plus the signal number.

The second, more significant fix (regarding what happens if the child shell is killed by a signal) results from Yang Yang's report. Yang Yang noted that on his Ubuntu system, the output in the last few lines of the shell session is different from that shown in the book. In the book, the following output is shown near the end of the shell session:

    $ fg
    ./t_system
    system() returned: status=0x000f (0,15)
    child killed by signal 15 (Terminated)

By contrast, Yang Yang obtained output of the form:

    $ fg
    ./t_system
    system() returned: status=0x8f00 (143,0)
    child exited, status=143

The differing output is a consequence of differences in the shells executed by system() on the two systems. A small-font note at the bottom of page 583 notes that some shells, when executing a simple command with sh -c, directly execute the command, rather than forking a new child shell to execute the command. On shells that perform this optimization, when system() is asked to execute the sleep 100 command, only a single process is created as a consequence of the call:

    system()    [process calls system("sleep 100")]
        |
    fork()
      :  \
      :   exec("sh")                  [Child process execs a shell]
      :     |
      :   exec("sleep")               [Child shell directly execs "sleep"]
      :     |
      :   [killed by SIGTERM]
      :  /
     waitpid(..., &status, ...)       [Sees "Killed by signal 15"]

By contrast, some shells don't perform this optimization, and the call to system() creates two processes, yielding the scenario shown in Figure 27-2 on page 584:

    system()    [process calls system("sleep 100")]
        |
    fork()
      :  \
      :   exec("sh")                  [Child process execs a shell]
      :     |
      :   fork()                      [Shell creates a new child]
      :     : \
      :     :  exec("sleep")          [New child execs "sleep"]
      :     :    |
      :     :  [killed by SIGTERM]
      :     : /
      :   waitpid(..., &status, ...)  [Sees "Killed by signal 15"]
      :     |
      :   exit(128+15)
      :  /
    waitpid(..., &status, ...)        [Sees "Exited normally, with status 143"]

The output in the book corresponds to the first of the scenarios described above. The shell executed by system() is bash, which does perform the optimization. In this case, the process running sleep is the child shell created by system(), and when that process is killed the status reported back by system() is as for a process killed by a signal.

The output for Yang Yang's case corresponds to the second scenario. The shell executed by system() is one that does not perform the optimization (dash, the default shell used to execute scripts on Ubuntu since version 6.10). In this case, the process running sleep is a child of the child shell created by system(). When that process is killed by a signal, the child shell observes the termination, and itself terminates with a status of the form 128+signum, which is in turn returned to the caller of system().

Reported by Yang Yang.

2011-12-23

3

In the second-to-last line of the first bulleted paragraph on the page, change:

executable it if it is

to:

able to be executed if it is

Explanation:

There are two fixes here. One rewords a slightly cumbersome double use of the word executable, and the other removes a superfluous word, it.

Reported by Justin Pryzby.

2010-12-28

2

In the last sentence of the small-font note near the top of the page, change:

In this case, if a signal was delivered to the child in the small time interval before the call to execl(), then the handler would be invoked in the child, after the signal was unblocked by sigprocmask().

to:

In this case, if a signal was delivered to the child in the small time interval before the call to execl(), then the handler would be invoked in the child. (By contrast, in the implementation shown in Listing 27-9, there is a small window of time before the execl() call where these signals will be ignored rather than being handled according to the dispositions that were in effect in the caller.)

Explanation:

The deleted text removes an inaccuracy (the SIGINT and SIGQUIT signals were not blocked). The added text clarifies a detail that might occur to curious readers.

Reported by Romain Vimont.

2018-01-08

9

In the last line on this page (in Listing 28-1, procexec/acct_on.c), change:

        usageErr("%s [file]\n");

to:

        usageErr("%s [file]\n", argv[0]);

Reported by Liu Jiaming.

2013-09-13

5

In the CLONE_NEWNET row of Table 28-2, change:

(2.4.24 onward)

to:

(2.6.24 onward)

Reported by Renato Westphal.

2012-06-05

4

In the CLONE_NEWUTS row of Table 28-2, change:

Child gets new UTS (utsname()) namespace

to:

Child gets new UTS (uname()) namespace

Reported by Yongzhi Pan.

2013-12-20

5

In the second bullet point on this page, at the end of the first sentence ("If CHILD_SIG is nonzero…"), add a circled "5" as a reference to the corresponding line in Listing 28-3.

Reported by Junjiro Okajima.

2012-01-16

3

In the second sentence of the first paragraph under the heading for CLONE_THREAD, change:

If this flag not set

to:

If this flag is not set

Reported by Murray McAllister.

2012-05-31

4

In the first sentence of the second paragraph under the heading for CLONE_THREAD, change:

Threads groups were introduced in Linux 2.4

to:

Thread groups were introduced in Linux 2.4

Reported by Murray McAllister.

2012-05-31

4

This is a note to the text, rather than an erratum.

At the foot of this page, there is a brief discussion of various CLONE_NEW* that are used with containers. For more information on these flags, see this article series on namespaces that I wrote for LWN.net.

2013-03-07

WONTFIX

In the entry for prctl(PR_SET_PDEATHSIG), the exec() column should say "See notes". In the fourth column, the text should say something like the following:

If a set-user-ID or set-group-ID program is being executed, the parent death signal value is cleared, otherwise it is preserved.

However, there isn't sufficient space on the page to add this information.

2012-10-27

WONTFIX

In the last line of the page, change:

case of a multithreaded processes

to:

case of a multithreaded process

Reported by Subu Rama.

2011-01-27

2

In the second paragraph on the page (starting "The threads in a process"), change the second sentence:

On a multiprocessor system, multiple threads can execute parallel.

to:

On a multiprocessor system, multiple threads can execute in parallel.

Reported by Eric Arora.

2011-01-05

2

In the second paragraph on the page (starting "The threads in a process"), change:

(Although it sometimes useful

to:

(Although it is sometimes useful

Reported by Subu Rama.

2011-01-27

2

In the label of the upward pointing arrow on the left-hand side of Figure 29-1, change:

increasing virtual addesses

to:

increasing virtual addresses

Reported by Giannis Tsaraias.

2016-12-26

8

In the second paragraph, change:

On Linux, a thread-specific errno is achieved in a similar manner to most other UNIX implementations: errno is defined as a macro that expands into a function call returning a modifiable lvalue that is distinct for each thread.

to:

On Linux, a thread-specific errno is achieved in a similar manner to most other UNIX implementations: errno is defined as a macro that produces a modifiable lvalue by dereferencing the pointer returned by a function that returns a pointer to a thread-specific data area.

Explanation:

See also the erratum for page 53.

2022-12-20

15

In the code segment at the bottom of the page, change:

    pthread_t *thread;

to:

    pthread_t thread;

Reported by Subu Rama.

2011-02-17

2

Inside the prototype box for pthread_exit(), change:

include <pthread.h>

to:

#include <pthread.h>

Reported by Fabien Galand.

2011-07-13

3

Inside the prototype boxes for pthread_self() and pthread_equal() (i.e., two changes!), change:

include <pthread.h>

to:

#include <pthread.h>

Reported by Fabien Galand.

2011-07-13

3

In the sentence just above the box containing the prototype for pthread_equal(), change:

The pthread_equal() function allows us check

to:

The pthread_equal() function allows us to check

Reported by Jens Thoms Toerring.

2011-02-27

2

In the prototype box for pthread_equal() near the bottom of the page, change:

Returns nonzero value

to:

Returns a nonzero value

Reported by Yaakov Eisenberg.

2023-06-09

15

Inside the prototype box for pthread_join(), change:

include <pthread.h>

to:

#include <pthread.h>

Reported by Fabien Galand.

2011-07-13

3

In listing 29-1 (threads/simple_thread.c), change:

static void *
threadFunc(void *arg)
{
    char *s = (char *) arg;

to:

static void *
threadFunc(void *arg)
{
    char *s = arg;

Explanation:

No cast is needed when assigning a void pointer to another pointer type.

Reported by Alejandro Colomar.

2021-09-17

14

This is a note to the text, rather than an erratum.

In Section 29.7, I could have noted that calling pthread_detach() more than once on the same thread results in unspecified behavior.

Explanation:

It would have been desirable to note this point in the text. (It is noted in the pthread_detach() manual page that I wrote.) However, there isn't sufficient space on the page to add this sentence.

Reported by Jens Thoms Toerring.

2011-02-28

WONTFIX

Near the top of Listing 30-1 (threads/thread_incr.c), change:

static int glob = 0;

static void *                   /* Loop 'arg' times incrementing 'glob' */
threadFunc(void *arg)

to:

static volatile int glob = 0;   /* "volatile" prevents compiler optimizations
                                   of arithmetic operations on 'glob' */
static void *                   /* Loop 'arg' times incrementing 'glob' */
threadFunc(void *arg)

Explanation:

Making glob volatile means that the program more easily produces the "incorrect" output described on pages 633 and 634, even in the face of compiler optimization. Yang Yang's email report summarized the issue well, so I will just reproduce a lightly edited version of his comments:

I think failing to declare "glob" as volatile makes the program less "incorrect", because in this case, the output is determined by both the kernel's CPU scheduling decisions (as described on page 634) and the compiler's optimization decisions.

I've compiled Listing 30-1 with –O[0123] (GCC 4.5.2):

1. –O2 and –O3 results in the same assembly code, which ably simplifies the entire loop to a single addition "glob += n". It's very hard to observe an incorrect output even with a huge n (overflow doesn't count), although the race condition does exist.
2. –O1 uses a register to save the intermediate sum in each iteration, and writes the result into "glob" only once when the loop ends. Thus the output will be either n (more likely with a big n) or 2*n (more likely with a small n, as the first output on page 633).
3. –O0 faithfully executes the load-increment-store routine (as if volatile were added), and its behavior matches both outputs on page 633.

With volatile, of course, the program's behavior will always be the same "incorrect" as in the text.

Because it is a common misunderstanding, it's perhaps worth pointing out that it is not generally necessary to employ the volatile qualifier on global variables in multithreaded programs. For further information on this point, see, for example, various Stack Overflow questions: When to use volatile with multi threading?, Should ALL global variables be volatile-qualified?, and Is 'volatile' needed in this multi-threaded C++ code?.

Reported by Yang Yang.

2011-12-25

3

In the sentence that immediately precedes Section 30.1.1, change:

In order to safely handle shared variables, all threads must cooperate in their use of a mutex, abiding by the locking rules it enforces.

to:

In order to safely handle shared variables, all threads must cooperate in their use of a mutex, abiding by its locking rules.

Explanation:

The use of the word "enforces" in the original text was a little confusing, suggesting a meaning that runs counter to the point that mutex locking is advisory.

Reported by Bill McConnaughey.

2011-07-18

3

This is a note to the text, rather than an erratum.

Section 30.1.1 discusses statically initialized mutexes, stating that PTHREAD_MUTEX_INITIALIZER can be used only for statically allocated mutexes. This follows the specification in POSIX.1-2008. However, the subsequent Austin Group Defect Report 70 noted that this restriction imposed by the standard arose from the constraints of C89, which was the standard that was current when that detail of the specification was originally written. The defect report notes that because C99 allows structure copying, PTHREAD_MUTEX_INITIALIZER can be used generally to initialize a mutex, regardless of whether it is statically allocated or not. This detail was subsequently clarified in POSIX.1-2013, which removed the restriction that PTHREAD_MUTEX_INITIALIZER can be used only with statically allocated mutexes. The use of pthread_mutex_init() is required only to initialize a mutex with nondefault attributes.

This change means that the various statements in TLPI about statically allocated mutexes either no longer hold true or need slight amendment. The affected pages are at least: 635, 639, 640, and 641.

2017-10-14

WONTFIX

Near the top of Listing 30-2 (threads/thread_incr_mutex.c), change:

static int glob = 0;

to:

static volatile int glob = 0;

Explanation:

See also the notes for the erratum on page 632.

Reported by Arnaud Frugier.

2013-12-04

5

This is a note to the text, rather than an erratum.

A word of caution about the timing statistics for mutexes, file locks, and semaphores that are presented in Section 30.1.3. These statistics were obtained on an x86-32 system. The (relative) performance of the various synchronization methods depends on the hardware architecture and the implementation of each method on that architecture. In cases where you suspect that performance of a synchronization method may be critical to the performance of your application, there is no substitute for application-specific benchmarking on the target architecture.

Reported by Stefan Puiu.

2013-07-18

WONTFIX

In the second-to-last line on the page, remove the extra space between the words "early" and "futex".

Reported by Fabien Galand.

2011-11-26

3

In the fourth line from the top of the page, change:

    s = pthread_mutex_init(mtx, &mtxAttr);

to:

    s = pthread_mutex_init(&mtx, &mtxAttr);

Reported by Gary Hu.

2011-01-05

2

This is a note to the text, rather than an erratum.

Section 30.1.1 discusses statically initialized condition variables, stating that PTHREAD_COND_INITIALIZER can be used only for statically allocated condition variables. This follows the specification in POSIX.1-2008. However, the subsequent Austin Group Defect Report 70 noted that this restriction imposed by the standard arose from the constraints of C89, which was the standard that was current when that detail of the specification was originally written. The defect report notes that because C99 allows structure copying, PTHREAD_COND_INITIALIZER can be used generally to initialize a condition variable, regardless of whether it is statically allocated or not. This detail was subsequently clarified in POSIX.1-2013, which removed the restriction that PTHREAD_COND_INITIALIZER can be used only with statically allocated condition variables. The use of pthread_cond_init() is required only to initialize a condition variable with nondefault attributes.

This change means that the various statements in TLPI about statically allocated mutexes either no longer hold true or need slight amendment. The affected pages are at least: 643, 651, and 652.

2017-10-14

WONTFIX

In the seventh line of the second paragraph, change:

threads are designed to perform the exactly same task

to:

threads are designed to perform exactly the same task

Reported by Simon Durrant.

2011-09-26

3

In the third paragraph from the bottom of the page, change:

By contrast, pthread_cond_broadcast() handles the case where the waiting threads are designed to perform different tasks (in which case they probably have different predicates associated with the condition variable).

to:

By contrast, pthread_cond_broadcast() handles the case where the waiting threads are designed to perform different tasks. In this case, the threads probably have different predicates associated with the condition variable (see Section 30.2.3).

Reported by Yaakov Eisenberg.

2023-08-03

15

At the bottom of the page, add the following small-font note:

Chapter 32, which describes thread cancellation, notes that pthread_cond_wait() is a cancellation point. If a thread is canceled while blocked in a call to pthread_cond_wait(), then the mutex is relocked before the first cancellation cleanup handler is called. This means a cleanup handler can safely unlock the mutex (as is done, for example, in Listing 32-2).

Explanation:

Page 636 notes that it is an error to call pthread_mutex_unlock() on a mutex that is unlocked. Page 646 notes that pthread_cond_wait() unlocks its mutex before blocking. Chia Hao Lo pointed out that lacking any further details, the example in Listing 32-2 seemed to be at odds with those statements: if the thread is canceled while blocked in pthread_cond_wait(), wouldn't the cleanup handler then be unlocking a mutex that had already been unlocked by pthread_cond_wait()? The text added by the erratum resolves this question.

Reported by Chia Hao Lo.

2012-03-07

4

Near the top of Listing 30-4 (threads/thread_multijoin), change:

static pthread_cond_t threadDied = PTHREAD_COND_INITIALIZER;
static pthread_mutex_t threadMutex = PTHREAD_MUTEX_INITIALIZER;
                /* Protects all of the following global variables */

static int totThreads = 0;      /* Total number of threads created */
static int numLive = 0;         /* Total number of threads still alive or
                                   terminated but not yet joined */

to:

static int totThreads = 0;      /* Total number of threads created */
static int numLive = 0;         /* Total number of threads still alive or
                                   terminated but not yet joined */
static pthread_cond_t threadDied = PTHREAD_COND_INITIALIZER;
static pthread_mutex_t threadMutex = PTHREAD_MUTEX_INITIALIZER;
                /* Protects all of the following global variables */

Explanation:

The comment saying that the threadMutex mutex protects all of the following variables was not accurate. It does not (and does not need to) protect accesses to totThreads and numLive, because those variables are accessed only by the main thread. The simplest way to address this inaccuracy was to reorder the declarations. (The mutex is also not used to protect accesses to the sleepTime field of the threads array, because the timing of these accesses does not require mutex protection.)

Reported by Yaakov Eisenberg.

2023-08-05

15

This erratum makes changes on pages 649 and 650 to fix a single problem in Listing 30-4 (threads/thread_multijoin.c).

On page 649, in the second line from the bottom of the page, change:

    int idx = *((int *) arg);

to:

    int idx = (int) arg;

On page 650, about two thirds of the way down the page, change:

        s = pthread_create(&thread[idx].tid, NULL, threadFunc, &idx);

to:

        s = pthread_create(&thread[idx].tid, NULL, threadFunc, (void *) idx);

Explanation:

There was a race condition in the program. In the main program, the value &idx was passed as the final argument of the call to pthread_create(), which is located within a for loop indexed on idx. In the new thread, the passed pointer value was dereferenced to obtain the value for the local variable idx used within the thread. However, this will only succeed if the new thread dereferences the pointer before the main thread increments idx inside the for loop.

The changes given in the erratum correct the problem, and should generally work on any platform, but, as noted toward the bottom of page 622, casting an int back and forth to a pointer type is not strictly standards-conformant. A standards-conformant solution would be to apply a patch to the original program something like the following:

@@ -16,19 +16,22 @@
     TS_JOINED                   /* Thread terminated, and joined */
 };

-static struct {                 /* Info about each thread */
+struct tinfo {                  /* Info about each thread */
     pthread_t tid;              /* ID of this thread */
     enum tstate state;          /* Thread state (TS_* constants above) */
     int sleepTime;              /* Number seconds to live before terminating */
-} *thread;
+};
+
+static struct tinfo *thread;

 static void *                   /* Start function for thread */
 threadFunc(void *arg)
 {
-    int idx = *((int *) arg);
+    struct tinfo *tptr = arg;
+    int idx = tptr - thread;    /* Obtain index in 'thread' array */
     int s;

-    sleep(thread[idx].sleepTime);       /* Simulate doing some work */
+    sleep(tptr->sleepTime);     /* Simulate doing some work */
     printf("Thread %d terminating\n", idx);

     s = pthread_mutex_lock(&threadMutex);
@@ -36,7 +39,7 @@
         errExitEN(s, "pthread_mutex_lock");

     numUnjoined++;
-    thread[idx].state = TS_TERMINATED;
+    tptr->state = TS_TERMINATED;

     s = pthread_mutex_unlock(&threadMutex);
     if (s != 0)
@@ -65,7 +68,7 @@
     for (idx = 0; idx < argc - 1; idx++) {
         thread[idx].sleepTime = getInt(argv[idx + 1], GN_NONNEG, NULL);
         thread[idx].state = TS_ALIVE;
-        s = pthread_create(&thread[idx].tid, NULL, threadFunc, &idx);
+        s = pthread_create(&thread[idx].tid, NULL, threadFunc, &thread[idx]);
         if (s != 0)
             errExitEN(s, "pthread_create");
     }

2012-10-15 follow-up: Ideally, instead of using int in the modified code, it would have been preferable to use uintptr_t:

    uintptr_t idx = (uintptr_t) arg;

The problem with using int is that the cast generates a compiler warning ("cast from pointer to integer of different size") when compiling on 64-bit systems. However, using uintptr_t would have required more substantial changes to the code: an adjustment to the format specifier used in the printf() call in threadFunc(), and a corresponding change to the declaration of idx in main(). But, there's not much point in trying to make a perfect solution here, since the correct fix is really something like the longer patch above.

Reported by Jacob Mandelson.

2011-04-02

3

A few lines from the top of the page, inside Listing 30-4 (threads/thread_multijoin.c), change:

          if (thread[idx].state == TS_TERMINATED){

to:

          if (thread[idx].state == TS_TERMINATED) {

2015-01-30

6

In the fourth function prototype, change:

Boolean lookup(char *key, void **value)

to:

Boolean lookup(tree, char *key, void **value)

Reported by Yaakov Eisenberg.

2023-06-09

15

In the second paragraph from the top of the page, change:

lock that mutex when the function is called, and unlock it when the mutex returns

to:

lock that mutex when the function is called, and unlock it when the function returns

Reported by Sun Jian.

2011-09-05

3

In the second-to-last bullet point on the page, change:

The ecvt(), fcvt(), gcvt(), gethostbyname(), and gethostbyaddr() are removed

to:

The ecvt(), fcvt(), gcvt(), gethostbyname(), and gethostbyaddr() functions are removed

Reported by Murray McAllister.

2012-08-20

4

In the last line of this page, change:

A call to pthread_setspecific() is a request to the Pthreads implementation to say

to:

A call to pthread_setspecific() is a request to the Pthreads implementation that says

Reported by Yaakov Eisenberg.

2023-06-09

15

In the last line of the prototype box toward the bottom of the page, change:

isassociated

to:

is associated

Reported by 맹규호.

2015-12-18

7

In the sentence that immediately precedes Listing 31-1, change:

points to the string Invalid operation

to:

points to the string Invalid argument

Reported by Junjiro Okajima.

2012-08-18

4

This is a note to the text, rather than an erratum.

In version 2.32, glibc removed the symbols _sys_nerr and _sys_errlist. As a consequence, the code example shown on this page (Listing 31-1, threads/strerror.c) will no longer compile on systems that have glibc 2.32 or later.

2020-10-07

WONTFIX

This is a note to the text, rather than an erratum.

In version 2.32, glibc removed the symbols _sys_nerr and _sys_errlist. As a consequence, the code example shown on this page and the following page (Listing 31-3, threads/strerror_tsd.c) will no longer compile on systems that have glibc 2.32 or later.

2020-10-07

WONTFIX

About half way down the page, change:

    static __thread buf[MAX_ERROR_LEN];

to:

    static __thread char buf[MAX_ERROR_LEN];

Reported by Sangman Lee.

2011-03-09

2

This is a note to the text, rather than an erratum.

In version 2.32, glibc removed the symbols _sys_nerr and _sys_errlist. As a consequence, the code example shown on this page (Listing 31-4, threads/strerror_tls.c) will no longer compile on systems that have glibc 2.32 or later.

2020-10-07

WONTFIX

In the second paragraph on this page, change:

Precisely what happens to the target thread, and when it happens, depends

to:

Precisely what happens to the target thread, and when it happens, depend

Reported by Yaakov Eisenberg.

2023-06-09

15

Three lines above the heading for Section 32.3, change:

When a thread calls exec(), the cancelability type and state

to:

When a thread calls exec(), the cancelability state and type

Reported by Liu Jiaming.

2013-09-13

5

In the shell session toward the bottom of the page, change:

    $ ./t_pthread_cancel

to:

    $ ./thread_cancel

Reported by Yang Yang.

2011-12-28

3

At the end of the first bullet point on this page (starting "If no command-line argument is supplied"), add the following text:

(When a pthread_cond_wait() call is canceled, the mutex is automatically relocked before the cleanup handler is invoked. This means the mutex can be safely unlocked in the cleanup handler.)

Explanation:

See the erratum for page 646.

Reported by Chia Hao Lo.

2012-03-07

4

In Listing 32-2 thread/thread_cleanup.c, calls to pthread_mutex_lock() and pthread_mutex_unlock() should surround the assignment to the global variable glob, so that the code reads:

        printf("main:    about to signal condition variable\n");
 
        s = pthread_mutex_lock(&mtx);
        if (s != 0)
            errExitEN(s, "pthread_mutex_lock");
 

        glob = 1;
 
        s = pthread_mutex_unlock(&mtx);
        if (s != 0)
            errExitEN(s, "pthread_mutex_unlock");
 

        s = pthread_cond_signal(&cond);
        if (s != 0)
            errExitEN(s, "pthread_cond_signal");

The above change has been made in the source code distribution. However, there isn't sufficient space on the page to make this change in the printed book.

Reported by Jingtian Zhang.

2014-03-09

WONTFIX

In the first sentence of Section 33.2.3, change:

to another thread in the same process

to:

to a thread in the same process as the caller

Explanation:

Writing "another" implies that this function can't be used to send a signal to the calling thread, which is of course untrue (see page 404).

Reported by Junjiro Okajima.

2012-08-18

4

In the last line of the page, change:

to another thread in the same process

to:

to a thread in the same process as the caller

Explanation:

Writing "another" implies that this function can't be used to send a signal to the calling thread, which is untrue.

Reported by Junjiro Okajima.

2012-08-18

4

Beginning in the third line of the third paragraph, change:

Similarly, parent_func and child_func are added to a list functions

to:

Similarly, parent_func and child_func are added to a list of functions

Reported by Murray McAllister.

2012-09-19

4

Under the heading Threads and exit(), change:

If any thread calls exit() or, equivalently, the main thread does a return, all threads immediately vanish

to:

If any thread calls exit() or, equivalently, the main thread does a return, all threads vanish after higher-level cleanups have been performed (e.g., calling C++ destructors)

Explanation:

Stefan demonstrated a case where a C++ destructor was blocked in a call to pthread_join() that was trying to join with another thread that was not terminating. As a consequence, the call to exit() in the main thread blocked indefinitely.

Reported by Stefan Puiu.

2013-04-08

4

In the second line of the third paragraph below the heading One-to-one (1:1) implementations (kernel-level threads), change:

are slower on a 1:1 implementations

to:

are slower on a 1:1 implementation

Reported by Murray McAllister.

2012-09-19

4

In the first bullet point below the heading LinuxThreads deviations from specified behavior, change:

Calls to getppid() reflect the fact that every thread other than the main thread is created by the process's manager thread (i.e., getppid() returns the process ID of the manager thread).

to:

Calls to getppid() in threads other than the main thread return the process ID of the manager thread, which reflects the fact that every thread other than the main thread is created by the manager thread.

Explanation:

Although the original wording was not incorrect, Yongzhi Pan pointed out that it was hard to understand, and suggested the new text.

Reported by Yongzhi Pan.

2013-12-02

5

In the second line of this page, change:

process as whole

to:

process as a whole

Reported by Yaakov Eisenberg.

2023-06-09

15

In the third bullet point from the bottom of the page, change:

Threads don't share nice value set by setpriority().

to:

Threads don't share the nice value set by setpriority().

Reported by Murray McAllister.

2012-09-19

4

In the second bullet point from the bottom of the page, change:

are not shared between the threads.

to:

are not shared between threads.

2012-09-19

4

In the third line of the second paragraph, change:

can grep though

to:

can grep through

Reported by Yongzhi Pan.

2013-12-02

5

In the last two paragraphs on the page, change:

A process group is a set of one or more processes sharing the same process group identifier (PGID). A process group ID is a number of the same type (pid_t) as a process ID. A process group has a process group leader, which is the process that creates the group and whose process ID becomes the process group ID of the group. A new process inherits its parent's process group ID.

A process group has a lifetime, which is the period of time beginning when the leader creates the group and ending when the last member process leaves the group.

to:

A process group is a set of one or more processes sharing the same process group identifier (PGID). A process group ID is a number of the same type (pid_t) as a process ID. A process group has a process group leader, which is the process that becomes the first member of the group and whose process ID becomes the process group ID of the group. A new process inherits its parent's process group ID.

A process group has a lifetime, which is the period of time beginning when the leader joins the group and ending when the last member process leaves the group.

Explanation:

As is made clear on page 702 and 703, a process, X, can become a process group leader when another process changes X's process group ID to be the same as its process ID.

This errata was added following Marek Mularczyk's report of a similar imprecision on page 730.

2014-05-02

5

In the middle of the page, change this entire paragraph:

If the pid and pgid arguments specify the same process (i.e., pgid is 0 or matches the process ID of the process specified by pid), then a new process group is created, and the specified process is made the leader of the new group (i.e., the process group ID of the process is made the same as its process ID). If the two arguments specify different values (i.e., pgid is not 0 and doesn't match the process ID of the process specified by pid), then setpgid() is being used to move a process between process groups.

to:

If setpgid() changes the process group ID of the target process (the process specified by pid) to be the same as its process ID, then the target process is made the leader of a new process group whose ID is the same as the process ID. If setpgid() changes the process group ID of the target process to a value different from its process ID, then the target process is moved to the existing process group specified by pgid. If setpgid() leaves the process group ID of the target process unchanged, then the call has no effect on the target process.

Explanation:

The original text was correct, but (1) was more complex than is needed to explain the behavior of setpgid(), and (2) didn't clearly explain the behavior if the process group ID was unchanged. The new text rectifies both of these problems.

In the sixth print run, a typo in this erratum was corrected: a mistyped "septgid()" was corrected to "setpgid()". Thanks to Ernesto Messina for reporting this in August 2015.

Reported by Jonathan Nieder.

2011-02-16

2

The following change applies only to the PDF version of the book. A correction was made in both the printed book and the PDF in print run 6, but that fix was somehow reverted in print run 7.

In the third line in the paragraph in the middle of this page, change:

If septgid() changes

to:

If setpgid() changes

Reported by Yaakov Eisenberg.

2023-06-09

15

This is a note to the text, rather than an erratum.

The paragraph immediately preceding Section 34.3 says:

If we explicitly define the _BSD_SOURCE feature test macro when compiling a program, then glibc provides the BSD-derived versions of setpgrp() and getpgrp(), instead of the default versions.

2017-08-15

WONTFIX

In the first line of the second paragraph under the heading Other (obsolete) interfaces…", change:

In the beginning, 4.2BSD provided a getprgp(pid) system call

to:

In the beginning, 4.2BSD provided a getpgrp(pid) system call

Reported by Jordi Sanfeliu.

2013-03-18

4

In the last line of the prototype box for getsid() at the bottom of the page, change:

or (pid_t) –1 on error

to:

or –1 on error

Explanation:

Such a cast is not needed when the standards require the type to be a signed integer. SUSv3 and SUSv4 incorrectly documented a cast as being needed for getsid() and a few other functions. For some more details, see this Austin bug report.

2012-01-16

3

In the last line of the prototype box for setsid() at the top of the page, change:

or (pid_t) –1 on error

to:

or –1 on error

Explanation:

See the erratum for the getsid() prototype box on page 704.

2012-01-16

3

About one third of the way down the page, change the second item in the numbered list:

The controlling terminal loses its association with the session, and can therefore be acquired as the controlling process by another session leader.

to:

The controlling terminal loses its association with the session, and can therefore be acquired as the controlling terminal by another session leader.

Reported by Liu Jiaming.

2013-09-17

5

In the code snippet toward the bottom of the page, change:

if (ioctl(fd, TIOCSCTTY) == -1)
    errExit("ioctl");

to:

if (ioctl(fd, TIOCSCTTY, 0) == -1)
    errExit("ioctl");

Explanation:

In fact, the third argument isn't accessed on FreeBSD (nor, probably, on the other BSDs), or (in the usual case) on Linux, so that the call as originally given on page 707 will work. Furthermore, the FreeBSD (and OpenBSD and NetBSD) tty(4) man page describes the third argument as void, meaning that officially no argument is required. However, scanning the source code of some utilities on BSD shows that they all seem to specify 0 or NULL as the third argument of ioctl(TIOCSCTTY). And some sources document that the third argument should be zero, so it is probably best safe practice to specify the argument as 0, even though none of the BSDs seems to need it.

On Linux, the situation is more complex. By default, the TIOCSCTTY ioctl() does not try to use a third argument, and so omitting it is (almost) safe. (See the function tiocsctty() in the Linux source file drivers/tty/tty_io.c.) But, it does access the third argument if the terminal is a controlling terminal for another process, in which case the terminal will be stolen if the third argument is 1 and the caller has the CAP_SYS_ADMIN capability. Therefore, on the unlikely chance that we are calling ioctl(TIOCSCTTY) from a privileged process and the terminal is already a controlling terminal for another process, then we better specify the third argument of ioctl() as zero, to avoid the possibility that the call randomly gets a value of 1 in the register used for the argument, and inadvertently steals another process's controlling terminal.

Summary: the BSDs invented TIOCSCTTY, but they don't need a third argument. Linux piggybacked an extra use for TIOCSCTTY, which means that the third argument should be specified as 0 in the usual case.

Reported by Liu Jiaming.

2013-11-12

5

In the second line of Section 34.5, change:

Within a session, only one process can be in the foreground

to:

Within a session, only one process group can be in the foreground

Reported by Justin Pryzby.

2010-12-28

2

In the first sentence of the paragraph just above the prototype box, change:

The tcgetpgrp() and tcsetpgrp() functions respectively retrieve and change the process group of a terminal.

to:

The tcgetpgrp() and tcsetpgrp() functions respectively retrieve and change the foreground process group of a terminal.

Reported by Yaakov Eisenberg.

2023-08-03

15

This is a note to the text, rather than an erratum.

A small-font note at the bottom of the page states:

If there is no foreground process group for this terminal, tcgetpgrp() returns a value greater than 1 that doesn't match the ID of any existing process group.

The text could have said a little more about what kind of scenarios may result in such a return value.

The most likely scenario is that (all of the processes in) the foreground process terminated, and no other process group was subsequently made the foreground process group for the terminal. In this case, tcgetpgrp() will return the process group ID of the terminated process group.

In practice, such situations are rare: normally the foreground process group consists of (child) processes created by the shell, and when the shell observes that those processes have terminated, it makes itself the foreground process group for the terminal. However, if, for example, an application program creates a child process in a new process group, makes that process group the foreground process group, and then the child process terminates, then the shell will not notice that the foreground process group has disappeared and will thus not make itself the foreground group.

Reported by Jimm Chen.

2024-02-09

WONTFIX

This is a note to the text, rather than an erratum.

Section 34.6.1 notes that when a shell is sent a SIGHUP signal (because it is the controlling process and there was a terminal hangup), it in turns sends a SIGHUP signal to each of the process groups (jobs) that it has created. It should be noted that the latter behavior is dependent on a shell option, and on many modern systems the option is disabled by default. (In bash the option is set using the command shopt -s huponexit.)

2017-10-16

WONTFIX

This is a note to the text, rather than an erratum.

A longstanding kernel bug means that if you are running the disc_SIGHUP.c program on a multi-CPU system with a Linux kernel version of 3.13 or earlier, then it is likely that you won't see the results shown at the top of this page. For further details, see the erratum note for page 519.

Reported by Yongzhi Pan.

2014-05-04

WONTFIX

In the first sentence of the last paragraph, change:

After we typed Control-Z, the shell displays the command that has been stopped in the background.

to:

After we typed Control-Z, the shell displays the command that has been stopped and moved to the background.

Explanation:

The original wording was okay, but the revised wording makes the details a little clearer.

Reported by Yaakov Eisenberg.

2023-08-03

15

In the small-font note near the top of the page, change:

The Korn and C shells provide the command stop as a shorthand for kill –stop.

to:

The Korn and C shells provide the command stop as a shorthand for kill –STOP.

Reported by Yaakov Eisenberg.

2023-08-03

15

Near the bottom of this page, change:

    for(;;)             /* Wait for signals */
        pause()

to:

    for (;;)            /* Wait for signals */
        pause()

2018-03-16

9

In the second sentence of the second paragraph, change:

We can also see that the new job is in the same session as the shell and that all of the processes are in the same process group.

to:

We can also see that the new job is in the same session as the shell and that both of its processes are in a single new process group.

Explanation:

Yaakov pointed out that the reader could misunderstand "all of the processes" to include the shell. The new wording, suggested by Yaakov, is more precise and lessens the possibility for misunderstanding.

Reported by Yaakov Eisenberg.

2023-08-03

15

In the second line from the top of the page, change:

caused one of its child to stop.

to:

caused one of its children to stop.

Reported by Yaakov Eisenberg.

2023-08-03

15

In the first sentence of the small-font note near the top of the page, change:

By definition, a session leader is in an orphaned process group.

to:

By definition, a session leader is (initially) in an orphaned process group.

Explanation:

Yaakov Eisenberg pointed out that there are scenarios whereby the process group of the session leader could subsequently become nonorphaned. For example:

1. Process A becomes a session leader.
2. Process A creates B, which moves to a new process group.
3. Process B creates process C, which begins its life in the same process group as B, but then moves to the process group of process A.

Reported by Yaakov Eisenberg.

2023-08-04

15

In the last paragraph of the page, change:

As can be seen, both children receive SIGCONT and SIGHUP after the parent exits.

to:

As can be seen, both children receive SIGCONT and SIGHUP after the parent exits. (The SIGCONT output appears first because the SIGCONT handler interrupted the SIGHUP handler before the latter handler had a chance to print its output.)

Explanation:

The additional sentence clarifies a detail that might otherwise puzzle the reader. See page 451 for a discussion of the fact that a signal handler might interrupt an already executing signal handler. See also the code shown on page 463, where all signals are handled with the sa.sa_mask field filled with a full set of signals, so as to prevent the handler for a signal from being interrupted by delivery of another signal:

    sigfillset(&sa.sa_mask);

Reported by Sergey Oskotskiy.

2017-11-10

9

Two changes in the first paragraph of Section 34.8.

1. Change:

(which is the same as the process group ID of the process group leader)

to:

(which is the same as the process ID of the process group leader)

2. Change:

(which is the same as the ID of the session leader)

to:

(which is the same as the process ID of the session leader)

The first of these changes was suggested by Justin Pryzby.

2011-01-03

2

In the first paragraph of Section 34.8, change:

A session leader is the process that created the session using setsid(). Similarly, a process group leader is the process that created the group using setpgid(). All of the members of a process group share the same process group ID (which is the same as the process ID of the process group leader), and all processes in the process groups that constitute a session have the same session ID (which is the same as the process ID of the session leader).