Skip to contentSkip to conversation

glibc: sched_getcpu(): use rseq cpu_id TLS on Linux (v6)

RFC PATCH In review Mathieu Desnoyers 2 participants 14 messages Mar 19, 2020 – Mar 20, 2020 1 message not in the archive

Subsystems (3)
lore
Share & tools
lore

A message that is not in the archive (sent off-list or not yet received): [email protected]

#1 lore raw

RFC PATCH 4/8 glibc: Perform rseq(2) registration at C startup and thread creation (v15)

Register rseq(2) TLS for each thread (including main), and unregister for each thread (excluding main). "rseq" stands for Restartable Sequences.

See the rseq(2) man page proposed here:
https://lkml.org/lkml/2018/9/19/647

This patch depends on three patches from Florian Weimer:

- Introduce <elf_machine_sym_no_match.h>
- Implement __libc_early_init
- nptl: Start new threads with all signals blocked [BZ #25098]

those are based on glibc master branch commit 1fabdb99084df004f7f4cdc7068d1be209a258be. The rseq(2) system call was merged into Linux 4.18.

* NEWS: Add Restartable Sequences feature description.
* elf/libc_early_init.c: Perform rseq(2) registration at C startup. startup for shared libc.
* nptl/pthread_create.c: Perform rseq(2) registration at thread creation.
* manual/threads.texi: Document __rseq_abi, RSEQ_SIG, sys/rseq.h.
* sysdeps/unix/sysv/linux/kernel-features.h: Add __ASSUME_RSEQ.
* sysdeps/unix/sysv/linux/Makefile: Add rseq-sym, sys/rseq.h, bits/rseq.h.
* sysdeps/unix/sysv/linux/Versions: Export __rseq_abi from libc.
* sysdeps/unix/sysv/linux/aarch64/libc.abilist: Likewise.
* sysdeps/unix/sysv/linux/alpha/libc.abilist: Likewise.
* sysdeps/unix/sysv/linux/arm/be/libc.abilist: Likewise.
* sysdeps/unix/sysv/linux/arm/le/libc.abilist: Likewise.
* sysdeps/unix/sysv/linux/csky/libc.abilist: Likewise.
* sysdeps/unix/sysv/linux/hppa/libc.abilist: Likewise.
* sysdeps/unix/sysv/linux/i386/libc.abilist: Likewise.
* sysdeps/unix/sysv/linux/ia64/libc.abilist: Likewise.
* sysdeps/unix/sysv/linux/m68k/coldfire/libc.abilist: Likewise.
* sysdeps/unix/sysv/linux/m68k/m680x0/libc.abilist: Likewise.
* sysdeps/unix/sysv/linux/microblaze/be/libc.abilist: Likewise.
* sysdeps/unix/sysv/linux/microblaze/le/libc.abilist: Likewise.
* sysdeps/unix/sysv/linux/mips/mips32/fpu/libc.abilist: Likewise.
* sysdeps/unix/sysv/linux/mips/mips32/nofpu/libc.abilist: Likewise.
* sysdeps/unix/sysv/linux/mips/mips64/n32/libc.abilist: Likewise.
* sysdeps/unix/sysv/linux/mips/mips64/n64/libc.abilist: Likewise.
* sysdeps/unix/sysv/linux/nios2/libc.abilist: Likewise.
* sysdeps/unix/sysv/linux/powerpc/powerpc32/fpu/libc.abilist: Likewise.
* sysdeps/unix/sysv/linux/powerpc/powerpc32/nofpu/libc.abilist: Likewise.
* sysdeps/unix/sysv/linux/powerpc/powerpc64/be/libc.abilist: Likewise.
* sysdeps/unix/sysv/linux/powerpc/powerpc64/le/libc.abilist: Likewise.
* sysdeps/unix/sysv/linux/riscv/rv64/libc.abilist: Likewise.
* sysdeps/unix/sysv/linux/s390/s390-32/libc.abilist: Likewise.
* sysdeps/unix/sysv/linux/s390/s390-64/libc.abilist: Likewise.
* sysdeps/unix/sysv/linux/sh/be/libc.abilist: Likewise.
* sysdeps/unix/sysv/linux/sh/le/libc.abilist: Likewise.
* sysdeps/unix/sysv/linux/sparc/sparc32/libc.abilist: Likewise.
* sysdeps/unix/sysv/linux/sparc/sparc64/libc.abilist: Likewise.
* sysdeps/unix/sysv/linux/x86_64/64/libc.abilist: Likewise.
* sysdeps/unix/sysv/linux/x86_64/x32/libc.abilist: Likewise.
* misc/rseq-internal.h: New file.
* sysdeps/unix/sysv/linux/rseq-internal.h: Likewise.
* sysdeps/unix/sysv/linux/rseq-sym.c: Likewise.
* sysdeps/unix/sysv/linux/sys/rseq.h: Likewise.
* sysdeps/unix/sysv/linux/bits/rseq.h: Likewise.
* sysdeps/unix/sysv/linux/aarch64/bits/rseq.h: Likewise.
* sysdeps/unix/sysv/linux/arm/bits/rseq.h: Likewise.
* sysdeps/unix/sysv/linux/mips/bits/rseq.h: Likewise.
* sysdeps/unix/sysv/linux/powerpc/bits/rseq.h: Likewise.
* sysdeps/unix/sysv/linux/s390/bits/rseq.h: Likewise.
* sysdeps/unix/sysv/linux/x86/bits/rseq.h: Likewise.

CC Carlos O'Donell <[email protected]>
CC Florian Weimer <[email protected]>
CC Joseph Myers <[email protected]>
CC Szabolcs Nagy <[email protected]>
CC Thomas Gleixner <[email protected]>
CC Ben Maurer <[email protected]>
CC Peter Zijlstra <[email protected]>
CC "Paul E. McKenney" <[email protected]>
CC Boqun Feng <[email protected]>
CC Will Deacon <[email protected]>
CC Dave Watson <[email protected]>
CC Paul Turner <[email protected]>
CC Rich Felker <[email protected]>

---
Changes since v1:
- Move __rseq_refcount to an extra field at the end of __rseq_abi to eliminate one symbol.

All libraries/programs which try to register rseq (glibc, early-adopter applications, early-adopter libraries) should use the rseq refcount. It becomes part of the ABI within a user-space process, but it's not part of the ABI shared with the kernel per se.

- Restructure how this code is organized so glibc keeps building on non-Linux targets.

- Use non-weak symbol for __rseq_abi.

- Move rseq registration/unregistration implementation into its own nptl/rseq.c compile unit.

- Move __rseq_abi symbol under GLIBC_2.29.

Changes since v2:
- Move __rseq_refcount to its own symbol, which is less ugly than trying to play tricks with the rseq uapi.
- Move __rseq_abi from nptl to csu (C start up), so it can be used across glibc, including memory allocator and sched_getcpu(). The __rseq_refcount symbol is kept in nptl, because there is no reason to use it elsewhere in glibc.

Changes since v3:
- Set __rseq_refcount TLS to 1 on register/set to 0 on unregister because glibc is the first/last user.
- Unconditionally register/unregister rseq at thread start/exit, because glibc is the first/last user.
- Add missing abilist items.
- Rebase on glibc master commit a502c5294.
- Add NEWS entry.

Changes since v4:
- Do not use "weak" symbols for __rseq_abi and __rseq_refcount. Based on "System V Application Binary Interface", weak only affects the link editor, not the dynamic linker.
- Install a new sys/rseq.h system header on Linux, which contains the RSEQ_SIG definition, __rseq_abi declaration and __rseq_refcount declaration. Move those definition/declarations from rseq-internal.h to the installed sys/rseq.h header.
- Considering that rseq is only available on Linux, move csu/rseq.c to sysdeps/unix/sysv/linux/rseq-sym.c.
- Move __rseq_refcount from nptl/rseq.c to sysdeps/unix/sysv/linux/rseq-sym.c, so it is only defined on Linux.
- Move both ABI definitions for __rseq_abi and __rseq_refcount to sysdeps/unix/sysv/linux/Versions, so they only appear on Linux.
- Document __rseq_abi and __rseq_refcount volatile.
- Document the RSEQ_SIG signature define.
- Move registration functions from rseq.c to rseq-internal.h static inline functions. Introduce empty stubs in misc/rseq-internal.h, which can be overridden by architecture code in sysdeps/unix/sysv/linux/rseq-internal.h.
- Rename __rseq_register_current_thread and __rseq_unregister_current_thread to rseq_register_current_thread and rseq_unregister_current_thread, now that those are only visible as internal static inline functions.
- Invoke rseq_register_current_thread() from libc-start.c LIBC_START_MAIN rather than nptl init, so applications not linked against libpthread.so have rseq registered for their main() thread. Note that it is invoked separately for SHARED and !SHARED builds.

Changes since v5:
- Replace __rseq_refcount by __rseq_lib_abi, which contains two
uint32_t: register_state and refcount. The "register_state" field allows inhibiting rseq registration from signal handlers nested on top of glibc registration and occuring after rseq unregistration by glibc.
- Introduce enum rseq_register_state, which contains the states allowed for the struct rseq_lib_abi register_state field.

Changes since v6:
- Introduce bits/rseq.h to define RSEQ_SIG for each architecture. The generic bits/rseq.h does not define RSEQ_SIG, meaning that each architecture implementing rseq needs to implement bits/rseq.h.
- Rename enum item RSEQ_REGISTER_NESTED to RSEQ_REGISTER_ONGOING.
- Port to glibc-2.29.

Changes since v7:
- Remove __rseq_lib_abi symbol, including refcount and register_state fields.
- Remove reference counting and nested signals handling from registration/unregistration functions.
- Introduce new __rseq_handled exported symbol, which is set to 1 by glibc on C startup when it handles restartable sequences. This allows glibc to coexist with early adopter libraries and applications wishing to register restartable sequences when it is not handled by glibc.
- Introduce rseq_init (), which sets __rseq_handled to 1 from C startup.
- Update NEWS entry.
- Update comments at the beginning of new files.
- Registration depends on both __NR_rseq and RSEQ_SIG.
- Remove ARM, powerpc, MIPS RSEQ_SIG until we agree with maintainers on the signature choice.
- Update x86, s390 RSEQ_SIG based on discussion with arch maintainers.
- Remove rseq-internal.h from headers list of misc/Makefile, so it it not installed by make install.

Changes since v8:
- Introduce RSEQ_SIG_CODE and RSEQ_SIG_DATA on aarch64 to handle compiling with -mbig-endian.

Changes since v9:
- Update Changelog.
- Remove unneeded new file comment header newlines.

Changes since v10:
- Remove volatile from __rseq_abi declaration.
- Document that __rseq_handled is about library managing rseq registration, independently of whether rseq is available or not.
- Move __rseq_handled symbol to ld.so, initialize this symbol within the dynamic linker initialization for both shared (rtld.c) and static (dl-support.c) builds.
- Only register the rseq TLS on initialization once in multiple-libc scenarios. Use rtld_active () for this purpose.
- In the static libc case, register the rseq TLS after LD_PRELOAD constructors are run, so it matches the order of this initialization vs LD_PRELOAD contructors execution for the shared libc.
- Agreed on signature choice with powerpc and MIPS maintainers, re-adding those signatures,
- The main architecture still left out signature-wise is ARM32.

Changes since v11:
- Rebase on glibc 2.30.
- Re-introduce ARM RSEQ_SIG following feedback from Will Deacon.

Changes since v12:
- Remove __rseq_handled,
- Rely on OS implicit rseq unregistration on thread teardown,
- Register main thread in __libc_early_init ().
- Add Restartable Sequences entry to threads manual.

Changes since v13:
- Update following be/le abilist split for arm, microblaze, and sh.
- Update manual to add the __rseq_abi variable and RSEQ_SIG macro to generate manual index entries, and add missing "Restartable Sequences" menu entry to the threads chapter.

Changes since v14:
- Update copyright range to include 2020.
- Introduce __ASSUME_RSEQ defined for --enable-kernel=4.18.0 and higher.
- Use ifdef __ASSUME_RSEQ rather than ifdef __NR_rseq to discover rseq availability. This is necessary now that the system call numbers are integrated within glibc.

CC Carlos O'Donell <[email protected]>
CC Florian Weimer <[email protected]>
CC Joseph Myers <[email protected]>
CC Szabolcs Nagy <[email protected]>
CC Thomas Gleixner <[email protected]>
CC Ben Maurer <[email protected]>
CC Peter Zijlstra <[email protected]>
CC "Paul E. McKenney" <[email protected]>
CC Boqun Feng <[email protected]>
CC Will Deacon <[email protected]>
CC Dave Watson <[email protected]>
CC Paul Turner <[email protected]>
CC Rich Felker <[email protected]>
---
 NEWS                                          | 10 +++
 elf/libc_early_init.c                         |  3 +
 manual/threads.texi                           | 30 ++++++-
 misc/rseq-internal.h                          | 33 ++++++++
 nptl/pthread_create.c                         | 12 +++
 sysdeps/unix/sysv/linux/Makefile              |  5 +-
 sysdeps/unix/sysv/linux/Versions              |  3 +
 sysdeps/unix/sysv/linux/aarch64/bits/rseq.h   | 43 ++++++++++
 sysdeps/unix/sysv/linux/aarch64/libc.abilist  |  1 +
 sysdeps/unix/sysv/linux/alpha/libc.abilist    |  1 +
 sysdeps/unix/sysv/linux/arm/be/libc.abilist   |  1 +
 sysdeps/unix/sysv/linux/arm/bits/rseq.h       | 83 +++++++++++++++++++
 sysdeps/unix/sysv/linux/arm/le/libc.abilist   |  1 +
 sysdeps/unix/sysv/linux/bits/rseq.h           | 29 +++++++
 sysdeps/unix/sysv/linux/csky/libc.abilist     |  1 +
 sysdeps/unix/sysv/linux/hppa/libc.abilist     |  1 +
 sysdeps/unix/sysv/linux/i386/libc.abilist     |  1 +
 sysdeps/unix/sysv/linux/ia64/libc.abilist     |  1 +
 sysdeps/unix/sysv/linux/kernel-features.h     |  5 ++
 .../sysv/linux/m68k/coldfire/libc.abilist     |  1 +
 .../unix/sysv/linux/m68k/m680x0/libc.abilist  |  1 +
 .../sysv/linux/microblaze/be/libc.abilist     |  1 +
 .../sysv/linux/microblaze/le/libc.abilist     |  1 +
 sysdeps/unix/sysv/linux/mips/bits/rseq.h      | 62 ++++++++++++++
 .../sysv/linux/mips/mips32/fpu/libc.abilist   |  1 +
 .../sysv/linux/mips/mips32/nofpu/libc.abilist |  1 +
 .../sysv/linux/mips/mips64/n32/libc.abilist   |  1 +
 .../sysv/linux/mips/mips64/n64/libc.abilist   |  1 +
 sysdeps/unix/sysv/linux/nios2/libc.abilist    |  1 +
 sysdeps/unix/sysv/linux/powerpc/bits/rseq.h   | 37 +++++++++
 .../linux/powerpc/powerpc32/fpu/libc.abilist  |  1 +
 .../powerpc/powerpc32/nofpu/libc.abilist      |  1 +
 .../linux/powerpc/powerpc64/be/libc.abilist   |  1 +
 .../linux/powerpc/powerpc64/le/libc.abilist   |  1 +
 .../unix/sysv/linux/riscv/rv64/libc.abilist   |  1 +
 sysdeps/unix/sysv/linux/rseq-internal.h       | 76 +++++++++++++++++
 sysdeps/unix/sysv/linux/rseq-sym.c            | 44 ++++++++++
 sysdeps/unix/sysv/linux/s390/bits/rseq.h      | 37 +++++++++
 .../unix/sysv/linux/s390/s390-32/libc.abilist |  1 +
 .../unix/sysv/linux/s390/s390-64/libc.abilist |  1 +
 sysdeps/unix/sysv/linux/sh/be/libc.abilist    |  1 +
 sysdeps/unix/sysv/linux/sh/le/libc.abilist    |  1 +
 .../sysv/linux/sparc/sparc32/libc.abilist     |  1 +
 .../sysv/linux/sparc/sparc64/libc.abilist     |  1 +
 sysdeps/unix/sysv/linux/sys/rseq.h            | 30 +++++++
 sysdeps/unix/sysv/linux/x86/bits/rseq.h       | 30 +++++++
 .../unix/sysv/linux/x86_64/64/libc.abilist    |  1 +
 .../unix/sysv/linux/x86_64/x32/libc.abilist   |  1 +
 48 files changed, 598 insertions(+), 4 deletions(-)
 create mode 100644 misc/rseq-internal.h
 create mode 100644 sysdeps/unix/sysv/linux/aarch64/bits/rseq.h
 create mode 100644 sysdeps/unix/sysv/linux/arm/bits/rseq.h
 create mode 100644 sysdeps/unix/sysv/linux/bits/rseq.h
 create mode 100644 sysdeps/unix/sysv/linux/mips/bits/rseq.h
 create mode 100644 sysdeps/unix/sysv/linux/powerpc/bits/rseq.h
 create mode 100644 sysdeps/unix/sysv/linux/rseq-internal.h
 create mode 100644 sysdeps/unix/sysv/linux/rseq-sym.c
 create mode 100644 sysdeps/unix/sysv/linux/s390/bits/rseq.h
 create mode 100644 sysdeps/unix/sysv/linux/sys/rseq.h
 create mode 100644 sysdeps/unix/sysv/linux/x86/bits/rseq.h
NEWS +10 -0
index e0379fc53c..abc62ae5f5 100644
@@ -61,6 +61,16 @@ Major new features:
   responses, indicating a lack of DNSSEC validation.  (Therefore, the name
   servers and the network path to them are treated as untrusted.)
 
+* Support for automatically registering threads with the Linux rseq(2)
+  system call has been added.  This system call is implemented starting
+  from Linux 4.18.  The Restartable Sequences ABI accelerates user-space
+  operations on per-cpu data.  It allows user-space to perform updates
+  on per-cpu data without requiring heavy-weight atomic operations.
+  Automatically registering threads allows all libraries, including libc,
+  to make immediate use of the rseq(2) support by using the documented ABI.
+  See 'man 2 rseq' for the details of the ABI shared between libc and the
+  kernel.
+
 Deprecated and removed features, and other changes affecting compatibility:
 
 * The totalorder and totalordermag functions, and the corresponding
elf/libc_early_init.c +3 -0
index 1ac66d895d..30466afea0 100644
@@ -18,10 +18,13 @@
 
 #include <ctype.h>
 #include <libc-early-init.h>
+#include <rseq-internal.h>
 
 void
 __libc_early_init (void)
 {
   /* Initialize ctype data.  */
   __ctype_init ();
+  /* Register rseq ABI to the kernel.   */
+  (void) rseq_register_current_thread ();
 }
manual/threads.texi +28 -2
index 0858ef8f92..59f634e432 100644
@@ -9,8 +9,10 @@ This chapter describes functions used for managing threads.
 POSIX threads.
 
 @menu
-* ISO C Threads::	Threads based on the ISO C specification.
-* POSIX Threads::	Threads based on the POSIX specification.
+* ISO C Threads::		Threads based on the ISO C specification.
+* POSIX Threads::		Threads based on the POSIX specification.
+* Restartable Sequences::	Linux-specific Restartable Sequences
+				integration.
 @end menu
 
 
@@ -881,3 +883,27 @@ Behaves like @code{pthread_timedjoin_np} except that the absolute time in
 @c pthread_spin_unlock
 @c pthread_testcancel
 @c pthread_yield
+
+@node Restartable Sequences
+@section Restartable Sequences
+@cindex rseq
+
+This section describes @theglibc{} Restartable Sequences integration.
+
+@deftypevar {struct rseq} __rseq_abi
+@standards{GNU, sys/rseq.h}
+@Theglibc{} implements a @code{__rseq_abi} TLS symbol to interact with the
+Restartable Sequences system call (Linux-specific).  The layout of this
+structure is defined by the Linux kernel @file{linux/rseq.h} UAPI.
+Registration of each thread's @code{__rseq_abi} is performed by
+@theglibc{} at libc initialization and pthread creation.
+@end deftypevar
+
+@deftypevr Macro int RSEQ_SIG
+@standards{GNU, sys/rseq.h}
+Each supported architecture provide a @code{RSEQ_SIG} macro in
+@file{sys/rseq.h} which contains a signature.  That signature is expected to be
+present in the code before each Restartable Sequences abort handler.  Failure
+to provide the expected signature may terminate the process with a Segmentation
+fault.
+@end deftypevr
misc/rseq-internal.h new file +33 -0
new file mode 100644
index 0000000000..d564cf1bc3
@@ -0,0 +1,33 @@
+/* Restartable Sequences internal API. Stub version.
+   Copyright (C) 2019-2020 Free Software Foundation, Inc.
+
+   The GNU C Library is free software; you can redistribute it and/or
+   modify it under the terms of the GNU Lesser General Public
+   License as published by the Free Software Foundation; either
+   version 2.1 of the License, or (at your option) any later version.
+
+   The GNU C Library is distributed in the hope that it will be useful,
+   but WITHOUT ANY WARRANTY; without even the implied warranty of
+   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+   Lesser General Public License for more details.
+
+   You should have received a copy of the GNU Lesser General Public
+   License along with the GNU C Library; if not, see
+   <http://www.gnu.org/licenses/>.  */
+
+#ifndef RSEQ_INTERNAL_H
+#define RSEQ_INTERNAL_H
+
+static inline int
+rseq_register_current_thread (void)
+{
+  return -1;
+}
+
+static inline int
+rseq_unregister_current_thread (void)
+{
+  return -1;
+}
+
+#endif /* rseq-internal.h */
nptl/pthread_create.c +12 -0
index afd379e89a..1ff248042e 100644
@@ -33,6 +33,7 @@
 #include <default-sched.h>
 #include <futex-internal.h>
 #include <tls-setup.h>
+#include <rseq-internal.h>
 #include "libioP.h"
 
 #include <shlib-compat.h>
@@ -384,6 +385,9 @@ START_THREAD_DEFN
   /* Initialize pointers to locale data.  */
   __ctype_init ();
 
+  /* Register rseq TLS to the kernel. */
+  (void) rseq_register_current_thread ();
+
 #ifndef __ASSUME_SET_ROBUST_LIST
   if (__set_robust_list_avail >= 0)
 #endif
@@ -578,6 +582,14 @@ START_THREAD_DEFN
      process is really dead since 'clone' got passed the CLONE_CHILD_CLEARTID
      flag.  The 'tid' field in the TCB will be set to zero.
 
+     rseq TLS is still registered at this point. Rely on implicit unregistration
+     performed by the kernel on thread teardown. This is not a problem because the
+     rseq TLS lives on the stack, and the stack outlives the thread. If TCB
+     allocation is ever changed, additional steps may be required, such as
+     performing explicit rseq unregistration before reclaiming the rseq TLS area
+     memory. It is NOT sufficient to block signals because the kernel may write
+     to the rseq area even without signals.
+
      The exit code is zero since in case all threads exit by calling
      'pthread_exit' the exit status must be 0 (zero).  */
   __exit_thread ();
sysdeps/unix/sysv/linux/Makefile +3 -2
index 60dc5cf9e5..6c6f669d21 100644
@@ -41,7 +41,7 @@ update-syscall-lists: arch-syscall.h
 endif
 
 ifeq ($(subdir),csu)
-sysdep_routines += errno-loc
+sysdep_routines += errno-loc rseq-sym
 endif
 
 ifeq ($(subdir),assert)
@@ -92,7 +92,8 @@ sysdep_headers += sys/mount.h sys/acct.h sys/sysctl.h \
 		  bits/termios-baud.h bits/termios-c_cflag.h \
 		  bits/termios-c_lflag.h bits/termios-tcflow.h \
 		  bits/termios-misc.h \
-		  bits/ipc-perm.h
+		  bits/ipc-perm.h \
+		  sys/rseq.h bits/rseq.h
 
 tests += tst-clone tst-clone2 tst-clone3 tst-fanotify tst-personality \
 	 tst-quota tst-sync_file_range tst-sysconf-iov_max tst-ttyname \
sysdeps/unix/sysv/linux/Versions +3 -0
index d385085c61..7f0da50580 100644
@@ -177,6 +177,9 @@ libc {
   GLIBC_2.30 {
     getdents64; gettid; tgkill;
   }
+  GLIBC_2.31 {
+    __rseq_abi;
+  }
   GLIBC_PRIVATE {
     # functions used in other libraries
     __syscall_rt_sigqueueinfo;
sysdeps/unix/sysv/linux/aarch64/bits/rseq.h new file +43 -0

43 lines not on this page · raw message

sysdeps/unix/sysv/linux/aarch64/libc.abilist +1 -0
index a4c31932cb..6784f13c09 100644
@@ -2145,3 +2145,4 @@ GLIBC_2.30 getdents64 F
 GLIBC_2.30 gettid F
 GLIBC_2.30 tgkill F
 GLIBC_2.30 twalk_r F
+GLIBC_2.31 __rseq_abi T 0x20
sysdeps/unix/sysv/linux/alpha/libc.abilist +1 -0
index e7f2174ac2..71db8422a2 100644
@@ -2225,6 +2225,7 @@ GLIBC_2.30 getdents64 F
 GLIBC_2.30 gettid F
 GLIBC_2.30 tgkill F
 GLIBC_2.30 twalk_r F
+GLIBC_2.31 __rseq_abi T 0x20
 GLIBC_2.4 _IO_fprintf F
 GLIBC_2.4 _IO_printf F
 GLIBC_2.4 _IO_sprintf F
sysdeps/unix/sysv/linux/arm/be/libc.abilist +1 -0
index b152c0e24a..eecebc908f 100644
@@ -133,6 +133,7 @@ GLIBC_2.30 twalk_r F
 GLIBC_2.31 msgctl F
 GLIBC_2.31 semctl F
 GLIBC_2.31 shmctl F
+GLIBC_2.31 __rseq_abi T 0x20
 GLIBC_2.4 _Exit F
 GLIBC_2.4 _IO_2_1_stderr_ D 0xa0
 GLIBC_2.4 _IO_2_1_stdin_ D 0xa0
sysdeps/unix/sysv/linux/arm/bits/rseq.h new file +83 -0

83 lines not on this page · raw message

sysdeps/unix/sysv/linux/arm/le/libc.abilist +1 -0
index 9371927927..3e7434d3d5 100644
@@ -130,6 +130,7 @@ GLIBC_2.30 getdents64 F
 GLIBC_2.30 gettid F
 GLIBC_2.30 tgkill F
 GLIBC_2.30 twalk_r F
+GLIBC_2.31 __rseq_abi T 0x20
 GLIBC_2.4 _Exit F
 GLIBC_2.4 _IO_2_1_stderr_ D 0xa0
 GLIBC_2.4 _IO_2_1_stdin_ D 0xa0
sysdeps/unix/sysv/linux/bits/rseq.h new file +29 -0
new file mode 100644
index 0000000000..b8a63ed26d
@@ -0,0 +1,29 @@
+/* Restartable Sequences architecture header. Stub version.
+   Copyright (C) 2019-2020 Free Software Foundation, Inc.
+
+   The GNU C Library is free software; you can redistribute it and/or
+   modify it under the terms of the GNU Lesser General Public
+   License as published by the Free Software Foundation; either
+   version 2.1 of the License, or (at your option) any later version.
+
+   The GNU C Library is distributed in the hope that it will be useful,
+   but WITHOUT ANY WARRANTY; without even the implied warranty of
+   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+   Lesser General Public License for more details.
+
+   You should have received a copy of the GNU Lesser General Public
+   License along with the GNU C Library; if not, see
+   <http://www.gnu.org/licenses/>.  */
+
+#ifndef _SYS_RSEQ_H
+# error "Never use <bits/rseq.h> directly; include <sys/rseq.h> instead."
+#endif
+
+/* RSEQ_SIG is a signature required before each abort handler code.
+
+   It is a 32-bit value that maps to actual architecture code compiled
+   into applications and libraries. It needs to be defined for each
+   architecture. When choosing this value, it needs to be taken into
+   account that generating invalid instructions may have ill effects on
+   tools like objdump, and may also have impact on the CPU speculative
+   execution efficiency in some cases.  */
sysdeps/unix/sysv/linux/csky/libc.abilist +1 -0
index 9b3cee65bb..b7ed346b1c 100644
@@ -2089,3 +2089,4 @@ GLIBC_2.30 getdents64 F
 GLIBC_2.30 gettid F
 GLIBC_2.30 tgkill F
 GLIBC_2.30 twalk_r F
+GLIBC_2.31 __rseq_abi T 0x20
sysdeps/unix/sysv/linux/hppa/libc.abilist +1 -0
index df6d96fbae..d55b153c48 100644
@@ -2046,6 +2046,7 @@ GLIBC_2.30 getdents64 F
 GLIBC_2.30 gettid F
 GLIBC_2.30 tgkill F
 GLIBC_2.30 twalk_r F
+GLIBC_2.31 __rseq_abi T 0x20
 GLIBC_2.4 __confstr_chk F
 GLIBC_2.4 __fgets_chk F
 GLIBC_2.4 __fgets_unlocked_chk F
sysdeps/unix/sysv/linux/i386/libc.abilist +1 -0
index fcb625b6bf..c9061600f6 100644
@@ -2212,6 +2212,7 @@ GLIBC_2.30 getdents64 F
 GLIBC_2.30 gettid F
 GLIBC_2.30 tgkill F
 GLIBC_2.30 twalk_r F
+GLIBC_2.31 __rseq_abi T 0x20
 GLIBC_2.4 __confstr_chk F
 GLIBC_2.4 __fgets_chk F
 GLIBC_2.4 __fgets_unlocked_chk F
sysdeps/unix/sysv/linux/ia64/libc.abilist +1 -0
index cb556c5998..f794303f0e 100644
@@ -2078,6 +2078,7 @@ GLIBC_2.30 getdents64 F
 GLIBC_2.30 gettid F
 GLIBC_2.30 tgkill F
 GLIBC_2.30 twalk_r F
+GLIBC_2.31 __rseq_abi T 0x20
 GLIBC_2.4 __confstr_chk F
 GLIBC_2.4 __fgets_chk F
 GLIBC_2.4 __fgets_unlocked_chk F
sysdeps/unix/sysv/linux/kernel-features.h +5 -0
index 32533e94cf..2c15da9389 100644
@@ -121,6 +121,11 @@
 # define __ASSUME_STATX 1
 #endif
 
+/* Support for rseq was added in kernel 4.18.  */
+#if __LINUX_KERNEL_VERSION >= 0x041200
+# define __ASSUME_RSEQ 1
+#endif
+
 /* Support for clone call used on fork.  The signature varies across the
    architectures with current 4 different variants:
 
sysdeps/unix/sysv/linux/m68k/coldfire/libc.abilist +1 -0
index 5e3cdea246..e5e545f3af 100644
@@ -134,6 +134,7 @@ GLIBC_2.30 twalk_r F
 GLIBC_2.31 msgctl F
 GLIBC_2.31 semctl F
 GLIBC_2.31 shmctl F
+GLIBC_2.31 __rseq_abi T 0x20
 GLIBC_2.4 _Exit F
 GLIBC_2.4 _IO_2_1_stderr_ D 0x98
 GLIBC_2.4 _IO_2_1_stdin_ D 0x98
sysdeps/unix/sysv/linux/m68k/m680x0/libc.abilist +1 -0
index ea5e7a41af..3fc1223e2c 100644
@@ -2158,6 +2158,7 @@ GLIBC_2.30 twalk_r F
 GLIBC_2.31 msgctl F
 GLIBC_2.31 semctl F
 GLIBC_2.31 shmctl F
+GLIBC_2.31 __rseq_abi T 0x20
 GLIBC_2.4 __confstr_chk F
 GLIBC_2.4 __fgets_chk F
 GLIBC_2.4 __fgets_unlocked_chk F
sysdeps/unix/sysv/linux/microblaze/be/libc.abilist +1 -0
index ac55b0acd7..81db30b543 100644
@@ -2140,3 +2140,4 @@ GLIBC_2.30 twalk_r F
 GLIBC_2.31 msgctl F
 GLIBC_2.31 semctl F
 GLIBC_2.31 shmctl F
+GLIBC_2.31 __rseq_abi T 0x20
sysdeps/unix/sysv/linux/microblaze/le/libc.abilist +1 -0
index f7ced487f7..a2ce147dde 100644
@@ -2137,3 +2137,4 @@ GLIBC_2.30 getdents64 F
 GLIBC_2.30 gettid F
 GLIBC_2.30 tgkill F
 GLIBC_2.30 twalk_r F
+GLIBC_2.31 __rseq_abi T 0x20
sysdeps/unix/sysv/linux/mips/bits/rseq.h new file +62 -0

62 lines not on this page · raw message

sysdeps/unix/sysv/linux/mips/mips32/fpu/libc.abilist +1 -0
index 06c2e64edd..c0040ddd4e 100644
@@ -2129,6 +2129,7 @@ GLIBC_2.30 getdents64 F
 GLIBC_2.30 gettid F
 GLIBC_2.30 tgkill F
 GLIBC_2.30 twalk_r F
+GLIBC_2.31 __rseq_abi T 0x20
 GLIBC_2.4 __confstr_chk F
 GLIBC_2.4 __fgets_chk F
 GLIBC_2.4 __fgets_unlocked_chk F
sysdeps/unix/sysv/linux/mips/mips32/nofpu/libc.abilist +1 -0
index bdfd073b86..61f19076eb 100644
@@ -2127,6 +2127,7 @@ GLIBC_2.30 getdents64 F
 GLIBC_2.30 gettid F
 GLIBC_2.30 tgkill F
 GLIBC_2.30 twalk_r F
+GLIBC_2.31 __rseq_abi T 0x20
 GLIBC_2.4 __confstr_chk F
 GLIBC_2.4 __fgets_chk F
 GLIBC_2.4 __fgets_unlocked_chk F
sysdeps/unix/sysv/linux/mips/mips64/n32/libc.abilist +1 -0
index 3d61d4974a..df4f3a3c04 100644
@@ -2135,6 +2135,7 @@ GLIBC_2.30 getdents64 F
 GLIBC_2.30 gettid F
 GLIBC_2.30 tgkill F
 GLIBC_2.30 twalk_r F
+GLIBC_2.31 __rseq_abi T 0x20
 GLIBC_2.4 __confstr_chk F
 GLIBC_2.4 __fgets_chk F
 GLIBC_2.4 __fgets_unlocked_chk F
sysdeps/unix/sysv/linux/mips/mips64/n64/libc.abilist +1 -0
index 675acca5db..a96de2e467 100644
@@ -2129,6 +2129,7 @@ GLIBC_2.30 getdents64 F
 GLIBC_2.30 gettid F
 GLIBC_2.30 tgkill F
 GLIBC_2.30 twalk_r F
+GLIBC_2.31 __rseq_abi T 0x20
 GLIBC_2.4 __confstr_chk F
 GLIBC_2.4 __fgets_chk F
 GLIBC_2.4 __fgets_unlocked_chk F
sysdeps/unix/sysv/linux/nios2/libc.abilist +1 -0
index 7fec0c9670..7b2ccbe953 100644
@@ -2178,3 +2178,4 @@ GLIBC_2.30 getdents64 F
 GLIBC_2.30 gettid F
 GLIBC_2.30 tgkill F
 GLIBC_2.30 twalk_r F
+GLIBC_2.31 __rseq_abi T 0x20
sysdeps/unix/sysv/linux/powerpc/bits/rseq.h new file +37 -0
new file mode 100644
index 0000000000..9d10000a6e
@@ -0,0 +1,37 @@
+/* Restartable Sequences Linux powerpc architecture header.
+   Copyright (C) 2019-2020 Free Software Foundation, Inc.
+
+   The GNU C Library is free software; you can redistribute it and/or
+   modify it under the terms of the GNU Lesser General Public
+   License as published by the Free Software Foundation; either
+   version 2.1 of the License, or (at your option) any later version.
+
+   The GNU C Library is distributed in the hope that it will be useful,
+   but WITHOUT ANY WARRANTY; without even the implied warranty of
+   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+   Lesser General Public License for more details.
+
+   You should have received a copy of the GNU Lesser General Public
+   License along with the GNU C Library; if not, see
+   <http://www.gnu.org/licenses/>.  */
+
+#ifndef _SYS_RSEQ_H
+# error "Never use <bits/rseq.h> directly; include <sys/rseq.h> instead."
+#endif
+
+/* RSEQ_SIG is a signature required before each abort handler code.
+
+   It is a 32-bit value that maps to actual architecture code compiled
+   into applications and libraries. It needs to be defined for each
+   architecture. When choosing this value, it needs to be taken into
+   account that generating invalid instructions may have ill effects on
+   tools like objdump, and may also have impact on the CPU speculative
+   execution efficiency in some cases.
+
+   RSEQ_SIG uses the following trap instruction:
+
+   powerpc-be:    0f e5 00 0b           twui   r5,11
+   powerpc64-le:  0b 00 e5 0f           twui   r5,11
+   powerpc64-be:  0f e5 00 0b           twui   r5,11  */
+
+#define RSEQ_SIG	0x0fe5000b
sysdeps/unix/sysv/linux/powerpc/powerpc32/fpu/libc.abilist +1 -0
index 1e8ff6f83e..6f4c6515dc 100644
@@ -2185,6 +2185,7 @@ GLIBC_2.30 getdents64 F
 GLIBC_2.30 gettid F
 GLIBC_2.30 tgkill F
 GLIBC_2.30 twalk_r F
+GLIBC_2.31 __rseq_abi T 0x20
 GLIBC_2.4 _IO_fprintf F
 GLIBC_2.4 _IO_printf F
 GLIBC_2.4 _IO_sprintf F
sysdeps/unix/sysv/linux/powerpc/powerpc32/nofpu/libc.abilist +1 -0
index b5a0751d90..f9875b4e22 100644
@@ -2218,6 +2218,7 @@ GLIBC_2.30 getdents64 F
 GLIBC_2.30 gettid F
 GLIBC_2.30 tgkill F
 GLIBC_2.30 twalk_r F
+GLIBC_2.31 __rseq_abi T 0x20
 GLIBC_2.4 _IO_fprintf F
 GLIBC_2.4 _IO_printf F
 GLIBC_2.4 _IO_sprintf F
sysdeps/unix/sysv/linux/powerpc/powerpc64/be/libc.abilist +1 -0
index 0c86217fc6..db06080db8 100644
@@ -2048,6 +2048,7 @@ GLIBC_2.30 getdents64 F
 GLIBC_2.30 gettid F
 GLIBC_2.30 tgkill F
 GLIBC_2.30 twalk_r F
+GLIBC_2.31 __rseq_abi T 0x20
 GLIBC_2.4 _IO_fprintf F
 GLIBC_2.4 _IO_printf F
 GLIBC_2.4 _IO_sprintf F
sysdeps/unix/sysv/linux/powerpc/powerpc64/le/libc.abilist +1 -0
index 2229a1dcc0..608ad49593 100644
@@ -2247,3 +2247,4 @@ GLIBC_2.30 getdents64 F
 GLIBC_2.30 gettid F
 GLIBC_2.30 tgkill F
 GLIBC_2.30 twalk_r F
+GLIBC_2.31 __rseq_abi T 0x20
sysdeps/unix/sysv/linux/riscv/rv64/libc.abilist +1 -0
index 31010e6cf7..c7657ce7f6 100644
@@ -2107,3 +2107,4 @@ GLIBC_2.30 getdents64 F
 GLIBC_2.30 gettid F
 GLIBC_2.30 tgkill F
 GLIBC_2.30 twalk_r F
+GLIBC_2.31 __rseq_abi T 0x20
sysdeps/unix/sysv/linux/rseq-internal.h new file +76 -0

76 lines not on this page · raw message

sysdeps/unix/sysv/linux/rseq-sym.c new file +44 -0

44 lines not on this page · raw message

sysdeps/unix/sysv/linux/s390/bits/rseq.h new file +37 -0
new file mode 100644
index 0000000000..c25ee67ee7
@@ -0,0 +1,37 @@
+/* Restartable Sequences Linux s390 architecture header.
+   Copyright (C) 2019-2020 Free Software Foundation, Inc.
+
+   The GNU C Library is free software; you can redistribute it and/or
+   modify it under the terms of the GNU Lesser General Public
+   License as published by the Free Software Foundation; either
+   version 2.1 of the License, or (at your option) any later version.
+
+   The GNU C Library is distributed in the hope that it will be useful,
+   but WITHOUT ANY WARRANTY; without even the implied warranty of
+   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+   Lesser General Public License for more details.
+
+   You should have received a copy of the GNU Lesser General Public
+   License along with the GNU C Library; if not, see
+   <http://www.gnu.org/licenses/>.  */
+
+#ifndef _SYS_RSEQ_H
+# error "Never use <bits/rseq.h> directly; include <sys/rseq.h> instead."
+#endif
+
+/* RSEQ_SIG is a signature required before each abort handler code.
+
+   It is a 32-bit value that maps to actual architecture code compiled
+   into applications and libraries. It needs to be defined for each
+   architecture. When choosing this value, it needs to be taken into
+   account that generating invalid instructions may have ill effects on
+   tools like objdump, and may also have impact on the CPU speculative
+   execution efficiency in some cases.
+
+   RSEQ_SIG uses the trap4 instruction. As Linux does not make use of the
+   access-register mode nor the linkage stack this instruction will always
+   cause a special-operation exception (the trap-enabled bit in the DUCT
+   is and will stay 0). The instruction pattern is
+       b2 ff 0f ff        trap4   4095(%r0)  */
+
+#define RSEQ_SIG	0xB2FF0FFF
sysdeps/unix/sysv/linux/s390/s390-32/libc.abilist +1 -0
index 4feca641b0..9c29ec0d2d 100644
@@ -2183,6 +2183,7 @@ GLIBC_2.30 twalk_r F
 GLIBC_2.31 msgctl F
 GLIBC_2.31 semctl F
 GLIBC_2.31 shmctl F
+GLIBC_2.31 __rseq_abi T 0x20
 GLIBC_2.4 _IO_fprintf F
 GLIBC_2.4 _IO_printf F
 GLIBC_2.4 _IO_sprintf F
sysdeps/unix/sysv/linux/s390/s390-64/libc.abilist +1 -0
index efe588a072..99424ceac9 100644
@@ -2084,6 +2084,7 @@ GLIBC_2.30 getdents64 F
 GLIBC_2.30 gettid F
 GLIBC_2.30 tgkill F
 GLIBC_2.30 twalk_r F
+GLIBC_2.31 __rseq_abi T 0x20
 GLIBC_2.4 _IO_fprintf F
 GLIBC_2.4 _IO_printf F
 GLIBC_2.4 _IO_sprintf F
sysdeps/unix/sysv/linux/sh/be/libc.abilist +1 -0
index 6bfc2b7439..80198de5d6 100644
@@ -2053,6 +2053,7 @@ GLIBC_2.30 twalk_r F
 GLIBC_2.31 msgctl F
 GLIBC_2.31 semctl F
 GLIBC_2.31 shmctl F
+GLIBC_2.31 __rseq_abi T 0x20
 GLIBC_2.4 __confstr_chk F
 GLIBC_2.4 __fgets_chk F
 GLIBC_2.4 __fgets_unlocked_chk F
sysdeps/unix/sysv/linux/sh/le/libc.abilist +1 -0
index 4b057bf4a2..916aa0b7f0 100644
@@ -2050,6 +2050,7 @@ GLIBC_2.30 getdents64 F
 GLIBC_2.30 gettid F
 GLIBC_2.30 tgkill F
 GLIBC_2.30 twalk_r F
+GLIBC_2.31 __rseq_abi T 0x20
 GLIBC_2.4 __confstr_chk F
 GLIBC_2.4 __fgets_chk F
 GLIBC_2.4 __fgets_unlocked_chk F
sysdeps/unix/sysv/linux/sparc/sparc32/libc.abilist +1 -0
index 49cd597fd6..9a27df8e43 100644
@@ -2174,6 +2174,7 @@ GLIBC_2.30 getdents64 F
 GLIBC_2.30 gettid F
 GLIBC_2.30 tgkill F
 GLIBC_2.30 twalk_r F
+GLIBC_2.31 __rseq_abi T 0x20
 GLIBC_2.4 _IO_fprintf F
 GLIBC_2.4 _IO_printf F
 GLIBC_2.4 _IO_sprintf F
sysdeps/unix/sysv/linux/sparc/sparc64/libc.abilist +1 -0
index 95e68e0ba1..32908666c4 100644
@@ -2101,6 +2101,7 @@ GLIBC_2.30 getdents64 F
 GLIBC_2.30 gettid F
 GLIBC_2.30 tgkill F
 GLIBC_2.30 twalk_r F
+GLIBC_2.31 __rseq_abi T 0x20
 GLIBC_2.4 __confstr_chk F
 GLIBC_2.4 __fgets_chk F
 GLIBC_2.4 __fgets_unlocked_chk F
sysdeps/unix/sysv/linux/sys/rseq.h new file +30 -0
new file mode 100644
index 0000000000..eb64fd41d7
@@ -0,0 +1,30 @@
+/* Restartable Sequences exported symbols. Linux header.
+   Copyright (C) 2019-2020 Free Software Foundation, Inc.
+
+   The GNU C Library is free software; you can redistribute it and/or
+   modify it under the terms of the GNU Lesser General Public
+   License as published by the Free Software Foundation; either
+   version 2.1 of the License, or (at your option) any later version.
+
+   The GNU C Library is distributed in the hope that it will be useful,
+   but WITHOUT ANY WARRANTY; without even the implied warranty of
+   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+   Lesser General Public License for more details.
+
+   You should have received a copy of the GNU Lesser General Public
+   License along with the GNU C Library; if not, see
+   <http://www.gnu.org/licenses/>.  */
+
+#ifndef _SYS_RSEQ_H
+#define _SYS_RSEQ_H	1
+
+/* We use the structures declarations from the kernel headers.  */
+#include <linux/rseq.h>
+/* Architecture-specific rseq signature.  */
+#include <bits/rseq.h>
+#include <stdint.h>
+
+extern __thread struct rseq __rseq_abi
+__attribute__ ((tls_model ("initial-exec")));
+
+#endif /* sys/rseq.h */
sysdeps/unix/sysv/linux/x86/bits/rseq.h new file +30 -0
new file mode 100644
index 0000000000..75f52d9788
@@ -0,0 +1,30 @@
+/* Restartable Sequences Linux x86 architecture header.
+   Copyright (C) 2019-2020 Free Software Foundation, Inc.
+
+   The GNU C Library is free software; you can redistribute it and/or
+   modify it under the terms of the GNU Lesser General Public
+   License as published by the Free Software Foundation; either
+   version 2.1 of the License, or (at your option) any later version.
+
+   The GNU C Library is distributed in the hope that it will be useful,
+   but WITHOUT ANY WARRANTY; without even the implied warranty of
+   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+   Lesser General Public License for more details.
+
+   You should have received a copy of the GNU Lesser General Public
+   License along with the GNU C Library; if not, see
+   <http://www.gnu.org/licenses/>.  */
+
+#ifndef _SYS_RSEQ_H
+# error "Never use <bits/rseq.h> directly; include <sys/rseq.h> instead."
+#endif
+
+/* RSEQ_SIG is a signature required before each abort handler code.
+
+   RSEQ_SIG is used with the following reserved undefined instructions, which
+   trap in user-space:
+
+   x86-32:    0f b9 3d 53 30 05 53      ud1    0x53053053,%edi
+   x86-64:    0f b9 3d 53 30 05 53      ud1    0x53053053(%rip),%edi  */
+
+#define RSEQ_SIG	0x53053053
sysdeps/unix/sysv/linux/x86_64/64/libc.abilist +1 -0
index 1f2dbd1451..7366565608 100644
@@ -2059,6 +2059,7 @@ GLIBC_2.30 getdents64 F
 GLIBC_2.30 gettid F
 GLIBC_2.30 tgkill F
 GLIBC_2.30 twalk_r F
+GLIBC_2.31 __rseq_abi T 0x20
 GLIBC_2.4 __confstr_chk F
 GLIBC_2.4 __fgets_chk F
 GLIBC_2.4 __fgets_unlocked_chk F
sysdeps/unix/sysv/linux/x86_64/x32/libc.abilist +1 -0
index 59da85a5d8..c1aa86f06e 100644
@@ -2158,3 +2158,4 @@ GLIBC_2.30 getdents64 F
 GLIBC_2.30 gettid F
 GLIBC_2.30 tgkill F
 GLIBC_2.30 twalk_r F
+GLIBC_2.31 __rseq_abi T 0x20
Signature · 1 line
2.17.1
#2 lore raw

Re: [RFC PATCH glibc 4/8] glibc: Perform rseq(2) registration at C startup and thread creation (v15)

* Mathieu Desnoyers via Libc-alpha:

Quoted 6 lines

Changes since v14:
- Update copyright range to include 2020.
- Introduce __ASSUME_RSEQ defined for --enable-kernel=4.18.0 and higher.
- Use ifdef __ASSUME_RSEQ rather than ifdef __NR_rseq to discover rseq availability. This is necessary now that the system call numbers are integrated within glibc.

It's not quite clear to me why you need __ASSUME_RSEQ.

Can you use __has_include in <sys/rseq.h>, with a copy of the kernel definitions if the kernel header is not available?

#3 lore raw

Re: [RFC PATCH glibc 4/8] glibc: Perform rseq(2) registration at C startup and thread creation (v15)

Quoted 10 lines · ----- On Mar 19, 2020, at 10:53 AM, Florian Weimer [email protected] wrote:

* Mathieu Desnoyers via Libc-alpha:

Quoted 6 lines

Changes since v14:
- Update copyright range to include 2020.
- Introduce __ASSUME_RSEQ defined for --enable-kernel=4.18.0 and higher.
- Use ifdef __ASSUME_RSEQ rather than ifdef __NR_rseq to discover rseq availability. This is necessary now that the system call numbers are integrated within glibc.

It's not quite clear to me why you need __ASSUME_RSEQ.

This is how I noticed other Linux features were being integrated, so I figured this is how I should do it. Monkey see, monkey do. ;-)

Quoted 2 lines

Can you use __has_include in <sys/rseq.h>, with a copy of the kernel definitions if the kernel header is not available?

Sure. Should I pull a verbatim copy of uapi linux/rseq.h into glibc ? If so, where should I put it ?

Thanks,

Mathieu

Signature · 3 lines
Mathieu Desnoyers
EfficiOS Inc.
http://www.efficios.com
#4 lore raw

Re: [RFC PATCH glibc 4/8] glibc: Perform rseq(2) registration at C startup and thread creation (v15)

* Mathieu Desnoyers:

Quoted 5 lines
Quoted 2 lines

Can you use __has_include in <sys/rseq.h>, with a copy of the kernel definitions if the kernel header is not available?

Sure. Should I pull a verbatim copy of uapi linux/rseq.h into glibc ? If so, where should I put it ?

Probably into <sys/rseq.h>, perhaps with a construct like this (untested):

#ifdef __has_include

# if __has_include ("linux/rseq.h")
#   define __GLIBC_HAVE_KERNEL_RSEQ
# endif
#else
# include <linux/version.h>
# if LINUX_VERSION_CODE >= KERNEL_VERSION (4, 18, 0)
#   define __GLIBC_HAVE_KERNEL_RSEQ
# endif
#endif

#ifdef __GLIBC_HAVE_KERNEL_RSEQ

# include <linux/rseq.h>
#else

… (fallback goes here)
#endif

We have an ongoing debate whether the fallback definition should use __u64 or uint64_t.

You also need to add an assert that the compiler supports __attribute__ ((aligned)) because ignoring it produces an ABI-incompatible header. The struct rseq/struct rseq_cs definitions are broken, they should not try to change the alignment.

PS: I have Internet connection trouble. Nobody should be worried if I drop off the net for a while. I understand this is quite a bad time for that. 8-(

#5 lore raw

Re: [RFC PATCH glibc 4/8] glibc: Perform rseq(2) registration at C startup and thread creation (v15)

Quoted 22 lines · ----- On Mar 19, 2020, at 12:03 PM, Florian Weimer [email protected] wrote:

* Mathieu Desnoyers:

Quoted 5 lines
Quoted 2 lines

Can you use __has_include in <sys/rseq.h>, with a copy of the kernel definitions if the kernel header is not available?

Sure. Should I pull a verbatim copy of uapi linux/rseq.h into glibc ? If so, where should I put it ?

Probably into <sys/rseq.h>, perhaps with a construct like this (untested):

#ifdef __has_include

# if __has_include ("linux/rseq.h")
#   define __GLIBC_HAVE_KERNEL_RSEQ
# endif
#else
# include <linux/version.h>
# if LINUX_VERSION_CODE >= KERNEL_VERSION (4, 18, 0)
#   define __GLIBC_HAVE_KERNEL_RSEQ
# endif
#endif
Quoted 6 lines

#ifdef __GLIBC_HAVE_KERNEL_RSEQ

# include <linux/rseq.h>
#else

… (fallback goes here)
#endif

OK will do.

Quoted 3 lines

We have an ongoing debate whether the fallback definition should use __u64 or uint64_t.

Then I'll keep including <linux/types.h> in the fallback and use __u{32,64} for now. If this proves to be an issue we can change it later. This is the minimal change from the uapi header.

Quoted 4 lines

You also need to add an assert that the compiler supports __attribute__ ((aligned)) because ignoring it produces an ABI-incompatible header.

Are you aware of some helper macro I should use to do this, or is it done elsewhere in glibc ?

Quoted 2 lines

The struct rseq/struct rseq_cs definitions
are broken, they should not try to change the alignment.

AFAIU, this means we should ideally not have used __attribute__((aligned)) in the uapi headers in the first place. Why is it broken ? However, now that it is in the wild, it's a bit late to change that.

Quoted 3 lines

PS: I have Internet connection trouble. Nobody should be worried if I drop off the net for a while. I understand this is quite a bad time for that. 8-(

Allright, thanks for the heads up! Stay safe!

Thanks,

Mathieu

Signature · 3 lines
Mathieu Desnoyers
EfficiOS Inc.
http://www.efficios.com
#6 lore raw

Re: [RFC PATCH glibc 4/8] glibc: Perform rseq(2) registration at C startup and thread creation (v15)

* Mathieu Desnoyers:

Quoted 6 lines
Quoted 3 lines

You also need to add an assert that the compiler supports __attribute__ ((aligned)) because ignoring it produces an ABI-incompatible header.

Are you aware of some helper macro I should use to do this, or is it done elsewhere in glibc ?

I don't think we have any such GCC-only types yet. max_align_t is provided by GCC itself.

Quoted 5 lines
Quoted 2 lines

The struct rseq/struct rseq_cs definitions
are broken, they should not try to change the alignment.

AFAIU, this means we should ideally not have used __attribute__((aligned)) in the uapi headers in the first place. Why is it broken ?

Compilers which are not sufficiently GCC-compatible define __attribute__(X) as the empty expansion, so you silently get a different ABI.

There is really no need to specify 32-byte alignment here. Is not even the size of a standard cache line. It can result in crashes if these structs are heap-allocated using malloc, when optimizing for AVX2.

For example, clang turns

void
clear (struct rseq *p)
{
memset (p, 0, sizeof (*p));
}

into:

	vxorps	%xmm0, %xmm0, %xmm0
	vmovaps	%ymm0, (%rdi)
	vzeroupper
	retq

My understanding is that vmovaps will trap if the pointer is misaligned (“When the source or destination operand is a memory operand, the operand must be aligned on a 32-byte boundary or a general-protection exception (#GP) will be generated.”).

Quoted 1 line

However, now that it is in the wild, it's a bit late to change that.

I had forgotten about the alignment crashes. I think we should seriously consider changing the types. 8-(

#7 lore raw

Re: [RFC PATCH glibc 4/8] glibc: Perform rseq(2) registration at C startup and thread creation (v15)

Quoted 11 lines · ----- On Mar 19, 2020, at 2:16 PM, Florian Weimer [email protected] wrote:

* Mathieu Desnoyers:

Quoted 6 lines
Quoted 3 lines

You also need to add an assert that the compiler supports __attribute__ ((aligned)) because ignoring it produces an ABI-incompatible header.

Are you aware of some helper macro I should use to do this, or is it done elsewhere in glibc ?

I don't think we have any such GCC-only types yet. max_align_t is provided by GCC itself.

I was thinking of adding the following to

sysdeps/unix/sysv/linux/rseq-internal.h: rseq_register_current_thread()

+ /* Ensure the compiler supports __attribute__ ((aligned)). */
+ _Static_assert (__alignof__ (struct rseq_cs) >= 4 * sizeof(uint64_t),
+ "alignment");
+ _Static_assert (__alignof__ (struct rseq) >= 4 * sizeof(uint64_t),
+ "alignment");
+

Quoted 9 lines
Quoted 5 lines
Quoted 2 lines

The struct rseq/struct rseq_cs definitions
are broken, they should not try to change the alignment.

AFAIU, this means we should ideally not have used __attribute__((aligned)) in the uapi headers in the first place. Why is it broken ?

Compilers which are not sufficiently GCC-compatible define __attribute__(X) as the empty expansion, so you silently get a different ABI.

It is worth noting that rseq.h is not the only Linux uapi header which uses __attribute__ ((aligned)), so this ABI problem exists today anyway for those compilers.

Quoted 5 lines

There is really no need to specify 32-byte alignment here. Is not even the size of a standard cache line. It can result in crashes if these structs are heap-allocated using malloc, when optimizing for AVX2.

Why would it be valid to allocate those with malloc ? Isn't it the purpose of posix_memalign() ?

Quoted 21 lines

For example, clang turns

void
clear (struct rseq *p)
{
memset (p, 0, sizeof (*p));
}

into:

	vxorps	%xmm0, %xmm0, %xmm0
	vmovaps	%ymm0, (%rdi)
	vzeroupper
	retq

My understanding is that vmovaps will trap if the pointer is misaligned (“When the source or destination operand is a memory operand, the operand must be aligned on a 32-byte boundary or a general-protection exception (#GP) will be generated.”).

Quoted 4 lines
Quoted 1 line

However, now that it is in the wild, it's a bit late to change that.

I had forgotten about the alignment crashes. I think we should seriously consider changing the types. 8-(

I don't think this is an option at this stage given that it is part of the Linux kernel UAPI. I am not convinced that it is valid at all to allocate struct rseq or struct rseq_cs with malloc(), because it does not guarantee any alignment.

Thanks,

Mathieu

Signature · 3 lines
Mathieu Desnoyers
EfficiOS Inc.
http://www.efficios.com
#8 lore raw

Re: [RFC PATCH glibc 4/8] glibc: Perform rseq(2) registration at C startup and thread creation (v15)

* Mathieu Desnoyers:

Quoted 18 lines
Quoted 11 lines · ----- On Mar 19, 2020, at 2:16 PM, Florian Weimer [email protected] wrote:

* Mathieu Desnoyers:

Quoted 6 lines
Quoted 3 lines

You also need to add an assert that the compiler supports __attribute__ ((aligned)) because ignoring it produces an ABI-incompatible header.

Are you aware of some helper macro I should use to do this, or is it done elsewhere in glibc ?

I don't think we have any such GCC-only types yet. max_align_t is provided by GCC itself.

I was thinking of adding the following to

sysdeps/unix/sysv/linux/rseq-internal.h: rseq_register_current_thread()

Review

Quoted 6 lines
(quoted diff) +6 -0
+  /* Ensure the compiler supports __attribute__ ((aligned)).  */
+  _Static_assert (__alignof__ (struct rseq_cs) >= 4 * sizeof(uint64_t),
+                 "alignment");
+  _Static_assert (__alignof__ (struct rseq) >= 4 * sizeof(uint64_t),
+                 "alignment");
+

Something like it would have to go into the *public* header.

Inside glibc, you can assume __attribute__ support.

Quoted 6 lines
Quoted 6 lines
Quoted 5 lines
Quoted 2 lines

The struct rseq/struct rseq_cs definitions
are broken, they should not try to change the alignment.

AFAIU, this means we should ideally not have used __attribute__((aligned)) in the uapi headers in the first place. Why is it broken ?

Quoted 7 lines
Quoted 3 lines

Compilers which are not sufficiently GCC-compatible define __attribute__(X) as the empty expansion, so you silently get a different ABI.

It is worth noting that rseq.h is not the only Linux uapi header which uses __attribute__ ((aligned)), so this ABI problem exists today anyway for those compilers.

Yuck. Even with larger-than-16 alignment?

Quoted 7 lines
Quoted 4 lines

There is really no need to specify 32-byte alignment here. Is not even the size of a standard cache line. It can result in crashes if these structs are heap-allocated using malloc, when optimizing for AVX2.

Why would it be valid to allocate those with malloc ? Isn't it the purpose of posix_memalign() ?

It would not be valid, but I don't think we have diagnostics for C like we have them for C++'s operator new.

Quoted 9 lines
Quoted 4 lines
Quoted 1 line

However, now that it is in the wild, it's a bit late to change that.

I had forgotten about the alignment crashes. I think we should seriously consider changing the types. 8-(

I don't think this is an option at this stage given that it is part of the Linux kernel UAPI. I am not convinced that it is valid at all to allocate struct rseq or struct rseq_cs with malloc(), because it does not guarantee any alignment.

The kernel ABI doesn't change. The kernel cannot use the alignment information anyway. Userspace struct layout may change in subtle ways, though.

#9 lore raw

Re: [RFC PATCH glibc 4/8] glibc: Perform rseq(2) registration at C startup and thread creation (v15)

Quoted 27 lines · ----- On Mar 19, 2020, at 2:34 PM, Florian Weimer [email protected] wrote:

* Mathieu Desnoyers:

Quoted 18 lines
Quoted 11 lines · ----- On Mar 19, 2020, at 2:16 PM, Florian Weimer [email protected] wrote:

* Mathieu Desnoyers:

Quoted 6 lines
Quoted 3 lines

You also need to add an assert that the compiler supports __attribute__ ((aligned)) because ignoring it produces an ABI-incompatible header.

Are you aware of some helper macro I should use to do this, or is it done elsewhere in glibc ?

I don't think we have any such GCC-only types yet. max_align_t is provided by GCC itself.

I was thinking of adding the following to

sysdeps/unix/sysv/linux/rseq-internal.h: rseq_register_current_thread()

Review

Quoted 6 lines
(quoted diff) +6 -0
+  /* Ensure the compiler supports __attribute__ ((aligned)).  */
+  _Static_assert (__alignof__ (struct rseq_cs) >= 4 * sizeof(uint64_t),
+                 "alignment");
+  _Static_assert (__alignof__ (struct rseq) >= 4 * sizeof(uint64_t),
+                 "alignment");
+
Quoted 3 lines

Something like it would have to go into the *public* header.

Inside glibc, you can assume __attribute__ support.

OK, so the _Static_assert () could sit in sys/rseq.h

Quoted 11 lines
Quoted 11 lines
Quoted 10 lines
Quoted 6 lines
Quoted 3 lines

The struct rseq/struct rseq_cs definitions
are broken, they should not try to change the alignment.

AFAIU, this means we should ideally not have used __attribute__((aligned)) in the uapi headers in the first place. Why is it broken ?

Compilers which are not sufficiently GCC-compatible define __attribute__(X) as the empty expansion, so you silently get a different ABI.

Quoted 5 lines
Quoted 3 lines

It is worth noting that rseq.h is not the only Linux uapi header which uses __attribute__ ((aligned)), so this ABI problem exists today anyway for those compilers.

Yuck. Even with larger-than-16 alignment?

There are two:

target_core_user.h
45:#define ALIGN_SIZE 64 /* Should be enough for most CPUs */
58: __u32 cmd_tail __attribute__((__aligned__(ALIGN_SIZE)));

netfilter_bridge/ebtables.h:90: char data[0] __attribute__ ((aligned (__alignof__(struct ebt_replace))));
netfilter_bridge/ebtables.h:132: unsigned char data[0] __attribute__ ((aligned (__alignof__(struct ebt_replace))));
netfilter_bridge/ebtables.h:145: unsigned char data[0] __attribute__ ((aligned (__alignof__(struct ebt_replace))));
netfilter_bridge/ebtables.h:158: unsigned char data[0] __attribute__ ((aligned (__alignof__(struct ebt_replace))));
netfilter_bridge/ebtables.h:191: unsigned char elems[0] __attribute__ ((aligned (__alignof__(struct ebt_replace))));

Quoted 11 lines
Quoted 8 lines
Quoted 5 lines

There is really no need to specify 32-byte alignment here. Is not even the size of a standard cache line. It can result in crashes if these structs are heap-allocated using malloc, when optimizing for AVX2.

Why would it be valid to allocate those with malloc ? Isn't it the purpose of posix_memalign() ?

It would not be valid, but I don't think we have diagnostics for C like we have them for C++'s operator new.

We could at least make an effort to let people know that alignment is required here when allocating struct rseq and struct rseq_cs on the heap by adding some comments to that effect in linux/rseq.h ?

Quoted 6 lines
Quoted 6 lines
Quoted 5 lines
Quoted 2 lines

However, now that it is in the wild, it's a bit late to change that.

I had forgotten about the alignment crashes. I think we should seriously consider changing the types. 8-(

Quoted 8 lines
Quoted 4 lines

I don't think this is an option at this stage given that it is part of the Linux kernel UAPI. I am not convinced that it is valid at all to allocate struct rseq or struct rseq_cs with malloc(), because it does not guarantee any alignment.

The kernel ABI doesn't change. The kernel cannot use the alignment information anyway. Userspace struct layout may change in subtle ways, though.

Considering the amount of pain this can cause in user-space, and because it can break userspace, this is not a UAPI change I am willing to consider. I'm not sure why we are even discussing the possibility of breaking a Linux UAPI considering that those are set in stone.

Thanks,

Mathieu

Signature · 3 lines
Mathieu Desnoyers
EfficiOS Inc.
http://www.efficios.com
#10 lore raw

Re: [RFC PATCH glibc 4/8] glibc: Perform rseq(2) registration at C startup and thread creation (v15)

* Mathieu Desnoyers:

Quoted 3 lines
Quoted 1 line

Inside glibc, you can assume __attribute__ support.

OK, so the _Static_assert () could sit in sys/rseq.h

It requires a C11 compiler. In this case, you could use _Alignas.

Quoted 16 lines
Quoted 16 lines
Quoted 15 lines
Quoted 11 lines
Quoted 7 lines
Quoted 4 lines

The struct rseq/struct rseq_cs definitions
are broken, they should not try to change the alignment.

AFAIU, this means we should ideally not have used __attribute__((aligned)) in the uapi headers in the first place. Why is it broken ?

Compilers which are not sufficiently GCC-compatible define __attribute__(X) as the empty expansion, so you silently get a different ABI.

It is worth noting that rseq.h is not the only Linux uapi header which uses __attribute__ ((aligned)), so this ABI problem exists today anyway for those compilers.

Quoted 7 lines
Quoted 1 line

Yuck. Even with larger-than-16 alignment?

There are two:

target_core_user.h
45:#define ALIGN_SIZE 64 /* Should be enough for most CPUs */
58: __u32 cmd_tail __attribute__((__aligned__(ALIGN_SIZE)));

That one is tough to figure out:

struct tcmu_mailbox {
        __u16 version;
        __u16 flags;
        __u32 cmdr_off;
        __u32 cmdr_size;

        __u32 cmd_head;

        /* Updated by user. On its own cacheline */
        __u32 cmd_tail __attribute__((__aligned__(ALIGN_SIZE)));

} __attribute__((packed));

Apparently, the expectation is that the compiler ignores __attribute__ ((packed) in this context. Ugh.

Quoted 5 lines

netfilter_bridge/ebtables.h:90: char data[0] __attribute__ ((aligned (__alignof__(struct ebt_replace))));
netfilter_bridge/ebtables.h:132: unsigned char data[0] __attribute__ ((aligned (__alignof__(struct ebt_replace))));
netfilter_bridge/ebtables.h:145: unsigned char data[0] __attribute__ ((aligned (__alignof__(struct ebt_replace))));
netfilter_bridge/ebtables.h:158: unsigned char data[0] __attribute__ ((aligned (__alignof__(struct ebt_replace))));
netfilter_bridge/ebtables.h:191: unsigned char elems[0] __attribute__ ((aligned (__alignof__(struct ebt_replace))));

I think these values are lower than max_align_t, so uncritical.

Quoted 8 lines
Quoted 8 lines
Quoted 7 lines
Quoted 4 lines

There is really no need to specify 32-byte alignment here. Is not even the size of a standard cache line. It can result in crashes if these structs are heap-allocated using malloc, when optimizing for AVX2.

Why would it be valid to allocate those with malloc ? Isn't it the purpose of posix_memalign() ?

Quoted 6 lines
Quoted 2 lines

It would not be valid, but I don't think we have diagnostics for C like we have them for C++'s operator new.

We could at least make an effort to let people know that alignment is required here when allocating struct rseq and struct rseq_cs on the heap by adding some comments to that effect in linux/rseq.h ?

We could use different types on the glibc side, then no special programmer action will be needed.

Quoted 10 lines
Quoted 10 lines
Quoted 9 lines
Quoted 4 lines
Quoted 1 line

However, now that it is in the wild, it's a bit late to change that.

I had forgotten about the alignment crashes. I think we should seriously consider changing the types. 8-(

I don't think this is an option at this stage given that it is part of the Linux kernel UAPI. I am not convinced that it is valid at all to allocate struct rseq or struct rseq_cs with malloc(), because it does not guarantee any alignment.

Quoted 8 lines
Quoted 3 lines

The kernel ABI doesn't change. The kernel cannot use the alignment information anyway. Userspace struct layout may change in subtle ways, though.

Considering the amount of pain this can cause in user-space, and because it can break userspace, this is not a UAPI change I am willing to consider. I'm not sure why we are even discussing the possibility of breaking a Linux UAPI considering that those are set in stone.

Again, the kernel interface is NOT affected. Only if the struct is used in a non-top-level fashion across an ABI boundary in userspace. I think making the change now is better than dealing with the breakage in rseq users when they are built with -mavx2.

#11 lore raw

Re: [RFC PATCH glibc 4/8] glibc: Perform rseq(2) registration at C startup and thread creation (v15)

Quoted 7 lines · ----- On Mar 19, 2020, at 3:05 PM, Florian Weimer [email protected] wrote:

* Mathieu Desnoyers:

Quoted 3 lines
Quoted 1 line

Inside glibc, you can assume __attribute__ support.

OK, so the _Static_assert () could sit in sys/rseq.h

It requires a C11 compiler. In this case, you could use _Alignas.

How would _Alignas replace:

+_Static_assert (__alignof__ (struct rseq_cs) >= 4 * sizeof(uint64_t),
+ "alignment");
+_Static_assert (__alignof__ (struct rseq) >= 4 * sizeof(uint64_t),
+ "alignment");

?

Moreover, I notice that sys/cdefs.h implements a fallback for _Static_assert for cases where it is not supported by the compiler. So I do not think it strictly depends on C11 if I include sys/cdefs.h from sys/rseq.h.

Quoted 10 lines
Quoted 10 lines
Quoted 10 lines
Quoted 10 lines
Quoted 9 lines
Quoted 5 lines
Quoted 2 lines

The struct rseq/struct rseq_cs definitions
are broken, they should not try to change the alignment.

AFAIU, this means we should ideally not have used __attribute__((aligned)) in the uapi headers in the first place. Why is it broken ?

Compilers which are not sufficiently GCC-compatible define __attribute__(X) as the empty expansion, so you silently get a different ABI.

[...]

Quoted 11 lines
Quoted 11 lines
Quoted 10 lines
Quoted 7 lines
Quoted 4 lines

There is really no need to specify 32-byte alignment here. Is not even the size of a standard cache line. It can result in crashes if these structs are heap-allocated using malloc, when optimizing for AVX2.

Why would it be valid to allocate those with malloc ? Isn't it the purpose of posix_memalign() ?

It would not be valid, but I don't think we have diagnostics for C like we have them for C++'s operator new.

Quoted 6 lines
Quoted 3 lines

We could at least make an effort to let people know that alignment is required here when allocating struct rseq and struct rseq_cs on the heap by adding some comments to that effect in linux/rseq.h ?

We could use different types on the glibc side, then no special programmer action will be needed.

Can't this lead to problems when mixing up compile units which have been compiled with linux/rseq.h with compile units compiled against sys/rseq.h ?

Let me take a step back and try to understand.

So far, there appears to be two scenarios where having a 64-byte alignment attribute on struct rseq and struct rseq_cs can cause problems:

1) A user-space programmer uses malloc() to dynamically allocate struct rseq or struct rseq_cs, which does not satisfy any of the alignment requirement of the structure. Combining this with compiler expectations that the structure needs to be aligned on 64-byte (e.g. -mavx2) breaks things.

For this first scenario, I am proposing that we document that the programmer should have used posix_memalign(), which provides the required alignment guarantees.

2) A user-space programmer mixes code compiled with compilers honouring the aligned attribute with other compile units compiled with compilers which discard those GCC extension attributes silently, embeds those into a structure, and get different struct layouts.

The _Static_assert in sys/rseq.h should detect the case where a compiler is not honouring the aligned attribute, right ?

Quoted 20 lines
Quoted 11 lines
Quoted 11 lines
Quoted 10 lines
Quoted 5 lines
Quoted 2 lines

However, now that it is in the wild, it's a bit late to change that.

I had forgotten about the alignment crashes. I think we should seriously consider changing the types. 8-(

I don't think this is an option at this stage given that it is part of the Linux kernel UAPI. I am not convinced that it is valid at all to allocate struct rseq or struct rseq_cs with malloc(), because it does not guarantee any alignment.

Quoted 8 lines
Quoted 3 lines

The kernel ABI doesn't change. The kernel cannot use the alignment information anyway. Userspace struct layout may change in subtle ways, though.

Considering the amount of pain this can cause in user-space, and because it can break userspace, this is not a UAPI change I am willing to consider. I'm not sure why we are even discussing the possibility of breaking a Linux UAPI considering that those are set in stone.

Quoted 4 lines

Again, the kernel interface is NOT affected. Only if the struct is used in a non-top-level fashion across an ABI boundary in userspace. I think making the change now is better than dealing with the breakage in rseq users when they are built with -mavx2.

What I am missing is what are the issues that persist once we add proper documentation of alignment requirements for heap allocation and a static assert to fail early when compiled with a compiler dismissing the aligned attribute ?

As you point out, changing the currently public linux/rseq.h UAPI header to remove those attributes ends up breaking user-space in scenarios of non-top-level use across ABI boundary. This is not kernel-vs-userspace ABI, but an ABI exposed by the kernel which ends up being used to coordinate user-space objects within a program. Breaking that does not appear to be any more acceptable. As I recall, the hard requirement for Linux ABIs is to do not break userspace, period. There is not mention of kernel-vs-userspace or userspace-vs-userspace. So if the end result of this change is to break user-space, it should not be changed.

Thanks,

Mathieu

Signature · 3 lines
Mathieu Desnoyers
EfficiOS Inc.
http://www.efficios.com
#12 lore raw

Re: [RFC PATCH glibc 4/8] glibc: Perform rseq(2) registration at C startup and thread creation (v15)

Quoted 112 lines · ----- On Mar 19, 2020, at 3:46 PM, Mathieu Desnoyers…
Quoted 7 lines · ----- On Mar 19, 2020, at 3:05 PM, Florian Weimer [email protected] wrote:

* Mathieu Desnoyers:

Quoted 3 lines
Quoted 1 line

Inside glibc, you can assume __attribute__ support.

OK, so the _Static_assert () could sit in sys/rseq.h

It requires a C11 compiler. In this case, you could use _Alignas.

How would _Alignas replace:

(quoted diff) +4 -0
+_Static_assert (__alignof__ (struct rseq_cs) >= 4 * sizeof(uint64_t),
+                "alignment");
+_Static_assert (__alignof__ (struct rseq) >= 4 * sizeof(uint64_t),
+                "alignment");

?

Moreover, I notice that sys/cdefs.h implements a fallback for _Static_assert for cases where it is not supported by the compiler. So I do not think it strictly depends on C11 if I include sys/cdefs.h from sys/rseq.h.

Quoted 10 lines
Quoted 10 lines
Quoted 10 lines
Quoted 10 lines
Quoted 9 lines
Quoted 5 lines
Quoted 2 lines

The struct rseq/struct rseq_cs definitions
are broken, they should not try to change the alignment.

AFAIU, this means we should ideally not have used __attribute__((aligned)) in the uapi headers in the first place. Why is it broken ?

Compilers which are not sufficiently GCC-compatible define __attribute__(X) as the empty expansion, so you silently get a different ABI.

[...]

Quoted 11 lines
Quoted 11 lines
Quoted 10 lines
Quoted 7 lines
Quoted 4 lines

There is really no need to specify 32-byte alignment here. Is not even the size of a standard cache line. It can result in crashes if these structs are heap-allocated using malloc, when optimizing for AVX2.

Why would it be valid to allocate those with malloc ? Isn't it the purpose of posix_memalign() ?

It would not be valid, but I don't think we have diagnostics for C like we have them for C++'s operator new.

Quoted 6 lines
Quoted 3 lines

We could at least make an effort to let people know that alignment is required here when allocating struct rseq and struct rseq_cs on the heap by adding some comments to that effect in linux/rseq.h ?

We could use different types on the glibc side, then no special programmer action will be needed.

Can't this lead to problems when mixing up compile units which have been compiled with linux/rseq.h with compile units compiled against sys/rseq.h ?

Let me take a step back and try to understand.

So far, there appears to be two scenarios where having a 64-byte alignment attribute on struct rseq and struct rseq_cs can cause problems:

1) A user-space programmer uses malloc() to dynamically allocate struct rseq or struct rseq_cs, which does not satisfy any of the alignment requirement of the structure. Combining this with compiler expectations that the structure needs to be aligned on 64-byte (e.g. -mavx2) breaks things.

For this first scenario, I am proposing that we document that the programmer should have used posix_memalign(), which provides the required alignment guarantees.

2) A user-space programmer mixes code compiled with compilers honouring the aligned attribute with other compile units compiled with compilers which discard those GCC extension attributes silently, embeds those into a structure, and get different struct layouts.

The _Static_assert in sys/rseq.h should detect the case where a compiler is not honouring the aligned attribute, right ?

Quoted 20 lines
Quoted 11 lines
Quoted 11 lines
Quoted 10 lines
Quoted 5 lines
Quoted 2 lines

However, now that it is in the wild, it's a bit late to change that.

I had forgotten about the alignment crashes. I think we should seriously consider changing the types. 8-(

I don't think this is an option at this stage given that it is part of the Linux kernel UAPI. I am not convinced that it is valid at all to allocate struct rseq or struct rseq_cs with malloc(), because it does not guarantee any alignment.

Quoted 8 lines
Quoted 3 lines

The kernel ABI doesn't change. The kernel cannot use the alignment information anyway. Userspace struct layout may change in subtle ways, though.

Considering the amount of pain this can cause in user-space, and because it can break userspace, this is not a UAPI change I am willing to consider. I'm not sure why we are even discussing the possibility of breaking a Linux UAPI considering that those are set in stone.

Quoted 4 lines

Again, the kernel interface is NOT affected. Only if the struct is used in a non-top-level fashion across an ABI boundary in userspace. I think making the change now is better than dealing with the breakage in rseq users when they are built with -mavx2.

What I am missing is what are the issues that persist once we add proper documentation of alignment requirements for heap allocation and a static assert to fail early when compiled with a compiler dismissing the aligned attribute ?

Quoted 9 lines

As you point out, changing the currently public linux/rseq.h UAPI header to remove those attributes ends up breaking user-space in scenarios of non-top-level use across ABI boundary. This is not kernel-vs-userspace ABI, but an ABI exposed by the kernel which ends up being used to coordinate user-space objects within a program. Breaking that does not appear to be any more acceptable. As I recall, the hard requirement for Linux ABIs is to do not break userspace, period. There is not mention of kernel-vs-userspace or userspace-vs-userspace. So if the end result of this change is to break user-space, it should not be changed.

Actually, here is an important clarification: the Linux kernel validates the struct rseq alignment on registration:

        if (!IS_ALIGNED((unsigned long)rseq, __alignof__(*rseq)) ||
            rseq_len != sizeof(*rseq))
                return -EINVAL;

So removing the aligned attribute from struct rseq is actually an ABI-breaking change, because it would be incompatible with older kernels which perform the IS_ALIGNED check expecting at least at 32 bytes alignment.

Thanks,

Mathieu

Signature · 3 lines
Mathieu Desnoyers
EfficiOS Inc.
http://www.efficios.com
#13 lore raw

Re: [RFC PATCH glibc 4/8] glibc: Perform rseq(2) registration at C startup and thread creation (v15)

----- On Mar 20, 2020, at 9:44 AM, Mathieu Desnoyers [email protected] wrote:
[...]

Quoted 11 lines

Actually, here is an important clarification: the Linux kernel validates the struct rseq alignment on registration:

       if (!IS_ALIGNED((unsigned long)rseq, __alignof__(*rseq)) ||
           rseq_len != sizeof(*rseq))
               return -EINVAL;

So removing the aligned attribute from struct rseq is actually an ABI-breaking change, because it would be incompatible with older kernels which perform the IS_ALIGNED check expecting at least at 32 bytes alignment.

So I plan to add the following to glibc's sys/rseq.h:

#include <sys/cdefs.h>

[...]

/* Ensure the compiler supports __attribute__ ((aligned)). */ _Static_assert (__alignof__ (struct rseq_cs) >= 4 * sizeof(uint64_t),

                "alignment");
_Static_assert (__alignof__ (struct rseq) >= 4 * sizeof(uint64_t),
                "alignment");

/* Allocations of struct rseq and struct rseq_cs on the heap need to
be aligned on 32 bytes. Therefore, use of malloc is discouraged because it does not guarantee alignment. posix_memalign should be used instead. */

Does it help mitigating your concerns ?

Thanks,

Mathieu

Signature · 3 lines
Mathieu Desnoyers
EfficiOS Inc.
http://www.efficios.com
#14 lore raw

RFC PATCH 5/8 glibc: sched_getcpu(): use rseq cpu_id TLS on Linux (v6)

In reply to a message not in the archive

When available, use the cpu_id field from __rseq_abi on Linux to implement sched_getcpu(). Fall-back on the vgetcpu vDSO if unavailable.

Benchmarks:

x86-64: Intel E5-2630 [email protected], 16-core, hyperthreading

glibc sched_getcpu():                     13.7 ns (baseline)
glibc sched_getcpu() using rseq:           2.5 ns (speedup:  5.5x)

inline load cpuid from __rseq_abi TLS: 0.8 ns (speedup: 17.1x)

* sysdeps/unix/sysv/linux/sched_getcpu.c: use rseq cpu_id TLS on Linux.

CC Carlos O'Donell <[email protected]>
CC Florian Weimer <[email protected]>
CC Joseph Myers <[email protected]>
CC Szabolcs Nagy <[email protected]>
CC Thomas Gleixner <[email protected]>
CC Ben Maurer <[email protected]>
CC Peter Zijlstra <[email protected]>
CC "Paul E. McKenney" <[email protected]>
CC Boqun Feng <[email protected]>
CC Will Deacon <[email protected]>
CC Dave Watson <[email protected]>
CC Paul Turner <[email protected]>

---
Changes since v1:
- rseq is only used if both __NR_rseq and RSEQ_SIG are defined.

Changes since v2:
- remove duplicated __rseq_abi extern declaration.

Changes since v3:
- update ChangeLog.

Changes since v4:
- Use atomic_load_relaxed to load the __rseq_abi.cpu_id field, a consequence of the fact that __rseq_abi is not volatile anymore.
- Include atomic.h which provides atomic_load_relaxed.

Changes since v5:
- Use __ASSUME_RSEQ to detect rseq availability.

---
 sysdeps/unix/sysv/linux/sched_getcpu.c | 30 ++++++++++++++++++++++++--
 1 file changed, 28 insertions(+), 2 deletions(-)
sysdeps/unix/sysv/linux/sched_getcpu.c +28 -2
index c019cfb3cf..4f2c5c4b1a 100644
@@ -18,10 +18,16 @@
 #include <errno.h>
 #include <sched.h>
 #include <sysdep.h>
+#include <atomic.h>
 #include <sysdep-vdso.h>
+#include <kernel-features.h>
 
-int
-sched_getcpu (void)
+#ifdef HAVE_GETCPU_VSYSCALL
+# define HAVE_VSYSCALL
+#endif
+
+static int
+vsyscall_sched_getcpu (void)
 {
   unsigned int cpu;
   int r = -1;
@@ -32,3 +38,23 @@ sched_getcpu (void)
 #endif
   return r == -1 ? r : cpu;
 }
+
+#ifdef __ASSUME_RSEQ
+#include <sys/rseq.h>
+#endif
+
+#if defined __ASSUME_RSEQ && defined RSEQ_SIG
+int
+sched_getcpu (void)
+{
+  int cpu_id = atomic_load_relaxed (&__rseq_abi.cpu_id);
+
+  return cpu_id >= 0 ? cpu_id : vsyscall_sched_getcpu ();
+}
+#else
+int
+sched_getcpu (void)
+{
+  return vsyscall_sched_getcpu ();
+}
+#endif
Signature · 1 line
2.17.1