Skip to contentSkip to conversation

sigaction.2: Document SA_EXPOSE_TAGBITS and the flag support detection protocol

PATCH Superseded Peter Collingbourne 2 participants 3 messages Jul 13, 2021 – Jul 15, 2021

Subsystems (2)

v3 of 4 Superseded Latest is v4 →

lore
Share & tools
lore
#0 lore raw

PATCH v3 1/1 sigaction.2: Document SA_EXPOSE_TAGBITS and the flag support detection protocol

---
This feature landed back in 5.11, but the manpage update seems to have fallen through the cracks. Here's a v3 with the introducing version specified and with one formatting nit fixed.

v3:
- s/5.x/5.11/g
- s/.IR/.I/ in one location

v2:
- fix formatting
- address feedback from Dave

 man2/sigaction.2 | 125 +++++++++++++++++++++++++++++++++++++++++++++++
 1 file changed, 125 insertions(+)
man2/sigaction.2 +125 -0
index 57ad6418c..6b90982d4 100644
@@ -261,6 +261,44 @@ This flag is meaningful only when establishing a signal handler.
 .\" .I sa_sigaction
 .\" field was added in Linux 2.1.86.)
 .\"
+.TP
+.BR SA_UNSUPPORTED
+Used to dynamically probe for flag bit support.
+.IP
+If an attempt to register a handler succeeds with this flag set in
+.I act->sa_flags
+alongside other flags that are potentially unsupported by the kernel,
+and an immediately subsequent
+.BR sigaction ()
+call specifying the same signal number n and with non-NULL
+.I oldact
+yields
+.B SA_UNSUPPORTED
+.I clear
+in
+.IR oldact->sa_flags ,
+then
+.I oldact->sa_flags
+may be used as a bitmask
+describing which of the potentially unsupported flags are,
+in fact, supported.
+See the section "Dynamically probing for flag bit support"
+below for more details.
+.TP
+.BR SA_EXPOSE_TAGBITS " (since Linux 5.11)"
+Normally, when delivering a signal,
+an architecture-specific set of tag bits are cleared from the
+.I si_addr
+field of
+.IR siginfo_t .
+If this flag is set,
+an architecture-specific subset of the tag bits will be preserved in
+.IR si_addr .
+.IP
+Programs that need to be compatible with Linux versions older than 5.11
+must use
+.B SA_UNSUPPORTED
+to probe for support.
 .SS The siginfo_t argument to a SA_SIGINFO handler
 When the
 .B SA_SIGINFO
@@ -846,6 +884,93 @@ Triggered by a
 .BR seccomp (2)
 filter rule.
 .RE
+.SS Dynamically probing for flag bit support
+The
+.BR sigaction ()
+call on Linux accepts unknown bits set in
+.I act->sa_flags
+without error.
+The behavior of the kernel starting with Linux 5.11 is that a second
+.BR sigaction ()
+will clear unknown bits from
+.IR oldact->sa_flags .
+However, historically, a second
+.BR sigaction ()
+call would typically leave those bits set in
+.IR oldact->sa_flags .
+.PP
+This means that support for new flags cannot be detected
+simply by testing for a flag in
+.IR sa_flags ,
+and a program must test that
+.B SA_UNSUPPORTED
+has been cleared before relying on the contents of
+.IR sa_flags .
+.PP
+Since the behavior of the signal handler cannot be guaranteed
+unless the check passes,
+it is wise to either block the affected signal
+while registering the handler and performing the check in this case,
+or where this is not possible,
+for example if the signal is synchronous, to issue the second
+.BR sigaction ()
+in the signal handler itself.
+.PP
+In kernels that do not support a specific flag,
+the kernel's behavior is as if the flag was not set,
+even if the flag was set in
+.IR act->sa_flags .
+.PP
+The flags
+.BR SA_NOCLDSTOP ,
+.BR SA_NOCLDWAIT ,
+.BR SA_SIGINFO ,
+.BR SA_ONSTACK ,
+.BR SA_RESTART ,
+.BR SA_NODEFER ,
+.BR SA_RESETHAND ,
+and, if defined by the architecture,
+.B SA_RESTORER
+may not be reliably probed for using this mechanism,
+because they were introduced before Linux 5.11.
+However, in general, programs may assume that these flags are supported,
+since they have all been supported since Linux 2.6,
+which was released in the year 2003.
+.PP
+The following example program exits with status 0 if
+.B SA_EXPOSE_TAGBITS
+is determined to be supported, and 1 otherwise.
+.PP
+.EX
+#include <signal.h>
+#include <stdio.h>
+#include <unistd.h>
+
+void handler(int signo, siginfo_t *info, void *context) {
+  struct sigaction oldact;
+  if (sigaction(SIGSEGV, 0, &oldact) == 0 &&
+      !(oldact.sa_flags & SA_UNSUPPORTED) &&
+      (oldact.sa_flags & SA_EXPOSE_TAGBITS)) {
+    _exit(0);
+  } else {
+    _exit(1);
+  }
+}
+
+int main(void) {
+  struct sigaction act = {};
+  act.sa_flags = SA_SIGINFO | SA_UNSUPPORTED | SA_EXPOSE_TAGBITS;
+  act.sa_sigaction = handler;
+  if (sigaction(SIGSEGV, &act, 0) != 0) {
+    perror("sigaction");
+    return 1;
+  }
+
+  /* Force a SIGSEGV. */
+  *(volatile int *)0 = 0;
+  return 1;
+}
+.EE
 .SH RETURN VALUE
 .BR sigaction ()
 returns 0 on success; on error, \-1 is returned, and
Signature · 1 line
2.32.0.93.g670b81a890-goog
#1 lore raw

Hi Peter,

Quoted 4 lines · Peter Collingbourne

---
This feature landed back in 5.11, but the manpage
update seems to have fallen through the cracks.

Sorry, I don't remember what happened :/
Thanks for the ping.

Quoted 2 lines

Here's a v3 with the introducing version specified and with one formatting nit fixed.

Please See some comments below (especially in the example code).

BTW, could you please link to the thread in the kernel mailing list that added this feature?

Thanks,

Alex

Quoted 17 lines

v3:
- s/5.x/5.11/g
- s/.IR/.I/ in one location

v2:
- fix formatting
- address feedback from Dave

 man2/sigaction.2 | 125 +++++++++++++++++++++++++++++++++++++++++++++++
 1 file changed, 125 insertions(+)
man2/sigaction.2 +0 -0
index 57ad6418c..6b90982d4 100644

man2/sigaction.2 This flag is meaningful only when establishing a signal handler. · line 265

Quoted 5 lines
man2/sigaction.2 +2 -0
@@ -261,6 +261,44 @@ This flag is meaningful only when establishing a signal handler.
 .\" .I sa_sigaction
 .\" field was added in Linux 2.1.86.)
 .\"
+.TP
+.BR SA_UNSUPPORTED

s/.BR/.B/

Quoted 96 lines
man2/sigaction.2 +89 -0
+Used to dynamically probe for flag bit support.
+.IP
+If an attempt to register a handler succeeds with this flag set in
+.I act->sa_flags
+alongside other flags that are potentially unsupported by the kernel,
+and an immediately subsequent
+.BR sigaction ()
+call specifying the same signal number n and with non-NULL
+.I oldact
+yields
+.B SA_UNSUPPORTED
+.I clear
+in
+.IR oldact->sa_flags ,
+then
+.I oldact->sa_flags
+may be used as a bitmask
+describing which of the potentially unsupported flags are,
+in fact, supported.
+See the section "Dynamically probing for flag bit support"
+below for more details.
+.TP
+.BR SA_EXPOSE_TAGBITS " (since Linux 5.11)"
+Normally, when delivering a signal,
+an architecture-specific set of tag bits are cleared from the
+.I si_addr
+field of
+.IR siginfo_t .
+If this flag is set,
+an architecture-specific subset of the tag bits will be preserved in
+.IR si_addr .
+.IP
+Programs that need to be compatible with Linux versions older than 5.11
+must use
+.B SA_UNSUPPORTED
+to probe for support.
 .SS The siginfo_t argument to a SA_SIGINFO handler
 When the
 .B SA_SIGINFO
@@ -846,6 +884,93 @@ Triggered by a
 .BR seccomp (2)
 filter rule.
 .RE
+.SS Dynamically probing for flag bit support
+The
+.BR sigaction ()
+call on Linux accepts unknown bits set in
+.I act->sa_flags
+without error.
+The behavior of the kernel starting with Linux 5.11 is that a second
+.BR sigaction ()
+will clear unknown bits from
+.IR oldact->sa_flags .
+However, historically, a second
+.BR sigaction ()
+call would typically leave those bits set in
+.IR oldact->sa_flags .
+.PP
+This means that support for new flags cannot be detected
+simply by testing for a flag in
+.IR sa_flags ,
+and a program must test that
+.B SA_UNSUPPORTED
+has been cleared before relying on the contents of
+.IR sa_flags .
+.PP
+Since the behavior of the signal handler cannot be guaranteed
+unless the check passes,
+it is wise to either block the affected signal
+while registering the handler and performing the check in this case,
+or where this is not possible,
+for example if the signal is synchronous, to issue the second
+.BR sigaction ()
+in the signal handler itself.
+.PP
+In kernels that do not support a specific flag,
+the kernel's behavior is as if the flag was not set,
+even if the flag was set in
+.IR act->sa_flags .
+.PP
+The flags
+.BR SA_NOCLDSTOP ,
+.BR SA_NOCLDWAIT ,
+.BR SA_SIGINFO ,
+.BR SA_ONSTACK ,
+.BR SA_RESTART ,
+.BR SA_NODEFER ,
+.BR SA_RESETHAND ,
+and, if defined by the architecture,
+.B SA_RESTORER
+may not be reliably probed for using this mechanism,
+because they were introduced before Linux 5.11.
+However, in general, programs may assume that these flags are supported,
+since they have all been supported since Linux 2.6,
+which was released in the year 2003.
+.PP

man2/sigaction.2 Triggered by a · line 949

Quoted 10 lines
man2/sigaction.2 +10 -0
+The following example program exits with status 0 if
+.B SA_EXPOSE_TAGBITS
+is determined to be supported, and 1 otherwise.
+.PP
+.EX
+#include <signal.h>
+#include <stdio.h>
+#include <unistd.h>
+
+void handler(int signo, siginfo_t *info, void *context) {

Use 4-space indents.

man2/sigaction.2 Triggered by a · line 961

Quoted 12 lines
man2/sigaction.2 +12 -0
+  struct sigaction oldact;
+  if (sigaction(SIGSEGV, 0, &oldact) == 0 &&
+      !(oldact.sa_flags & SA_UNSUPPORTED) &&
+      (oldact.sa_flags & SA_EXPOSE_TAGBITS)) {
+    _exit(0);
+  } else {
+    _exit(1);
+  }
+}
+
+int main(void) {
+  struct sigaction act = {};

The {} initializer is a GCC extension. Please use the C99-conforming {0} zero initializer. See
<https://stackoverflow.com/questions/11152160/initializing-a-struct-to-0>.

man2/sigaction.2 Triggered by a · line 970

Quoted 9 lines
man2/sigaction.2 +9 -0
+  act.sa_flags = SA_SIGINFO | SA_UNSUPPORTED | SA_EXPOSE_TAGBITS;
+  act.sa_sigaction = handler;
+  if (sigaction(SIGSEGV, &act, 0) != 0) {
+    perror("sigaction");
+    return 1;
+  }
+
+  /* Force a SIGSEGV. */
+  *(volatile int *)0 = 0;

Isn't this equivalent to the following (which wouldn't need a comment)?:

raise(SIGSEGV);

Quoted 7 lines
man2/sigaction.2 +3 -0
+  return 1;
+}
+.EE
 .SH RETURN VALUE
 .BR sigaction ()
 returns 0 on success; on error, \-1 is returned, and
Signature · 3 lines
Alejandro Colomar
Linux man-pages comaintainer; https://www.kernel.org/doc/man-pages/
http://www.alejandro-colomar.es/
#2 lore raw
Quoted 9 lines · Alejandro Colomar (man-pages)

Hi Peter,

Quoted 4 lines · Peter Collingbourne

---
This feature landed back in 5.11, but the manpage
update seems to have fallen through the cracks.

Quoted 7 lines

Sorry, I don't remember what happened :/
Thanks for the ping.

Quoted 2 lines

Here's a v3 with the introducing version specified and with one formatting nit fixed.

Please See some comments below (especially in the example code).

Addressed all comments in v4.

Quoted 2 lines

BTW, could you please link to the thread in the kernel mailing list that added this feature?

There were several threads (the patch went through several revisions). But you can find your way to most of the threads starting from here:
https://lore.kernel.org/linux-arm-kernel/?q=pcc+far_el1

Peter