glibc: sched_getcpu(): use rseq cpu_id TLS on Linux (v7)
Subsystems (3)
A message that is not in the archive (sent off-list or not yet received): [email protected]
PATCH 1/3 glibc: Perform rseq registration at C startup and thread creation (v19)
Register rseq 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
Those are based on glibc master branch commit 6f0baacf0f89. The rseq system call was merged into Linux 4.18.
---
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.
Changes since v15:
- Remove __ASSUME_RSEQ from kernel features.
- rseq internal: remove assume rseq
- remove assume rseq and struct rseq def from sysdeps/unix/sysv/linux/rseq-sym.c
- sys/rseq.h: detect rseq header, implement fallback
- sysdeps/unix/sysv/linux/sys/rseq.h include cdefs.h, add _Static_assert to validate struct rseq and struct rseq_cs alignment.
- sys/rseq.h: document that posix_memalign should be used rather than malloc if allocating struct rseq or struct rseq_cs on the heap. This is required to guarantee 32-byte alignement.
Changes since v16:
- Move rseq NEWS entry under 2.32.
- Move new __rseq_abi symbol to GLIBC_2.32.
Changes since v17:
- Change copyright year to 2020.
- Refer to GNU C Library manual rather than rseq manpage in NEWS.
- Use "initial" parameter from __libc_early_init ().
- Manual: rseq is Linux rather than GNU std.
- Remove rseq_unregister_current_thread () (unused).
- rseq_register_current_thread () returns void.
- Coding style fixes.
- sys/rseq.h: use "32" for alignment.
- Change http:// for https:// in comments.
- Add const struct rseq_cs * field to rseq_cs union.
Changes since v18:
- NEWS update.
- Manual update.
- Move misc/rseq-internal.h to sysdeps/generic/rseq-internal.h.
- Fix coding style in sysdeps/unix/sysv/linux/sys/rseq.h.
- Abort libc if __rseq_abi is already initialized in rseq_register_current_thread (), which is unexpected.
- Abort libc in rseq_register_current_thread () on all errno except EPERM (seccomp), EACCES (seccomp), and ENOSYS (not implemented).
--- NEWS | 10 + elf/libc_early_init.c | 4 + manual/threads.texi | 66 +++++- nptl/pthread_create.c | 13 ++ sysdeps/generic/rseq-internal.h | 26 +++ sysdeps/unix/sysv/linux/Makefile | 5 +- sysdeps/unix/sysv/linux/Versions | 1 + 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 + .../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 | 73 ++++++ sysdeps/unix/sysv/linux/rseq-sym.c | 26 +++ 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 | 207 ++++++++++++++++++ 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 + 47 files changed, 778 insertions(+), 4 deletions(-) create mode 100644 sysdeps/generic/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
@@ -23,6 +23,16 @@ Major new features: toolchains. It is recommended to use GCC 8 or newer when testing this option. +* Support for automatically registering threads with the Linux rseq + 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. + The GNU C Library manual has details on integration of Restartable + Sequences. + Deprecated and removed features, and other changes affecting compatibility: * The deprecated <sys/sysctl.h> header and the sysctl function have been
elf/libc_early_init.c +4 -0
@@ -18,10 +18,14 @@#include <ctype.h> #include <libc-early-init.h> +#include <rseq-internal.h> void __libc_early_init (_Bool initial) {/* Initialize ctype data. */ __ctype_init (); +/* Register rseq ABI to the kernel for the main program's libc. */ +if (initial) +rseq_register_current_thread (); }
manual/threads.texi +64 -2
75 lines not on this page · raw message
nptl/pthread_create.c +13 -0
@@ -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. */ +rseq_register_current_thread (); +#ifndef __ASSUME_SET_ROBUST_LIST if (__set_robust_list_avail >= 0)#endif
@@ -578,6 +582,15 @@ START_THREAD_DEFNprocess 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/generic/rseq-internal.h new file +26 -0
@@ -0,0 +1,26 @@ +/* Restartable Sequences internal API. Stub version. +Copyright (C) 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 +<https://www.gnu.org/licenses/>. */ + +#ifndef RSEQ_INTERNAL_H +#define RSEQ_INTERNAL_H + +static inline void +rseq_register_current_thread (void ) +{ +} + +#endif /* rseq-internal.h */
sysdeps/unix/sysv/linux/Makefile +3 -2
@@ -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 \ bits/termios-c_lflag.h bits/termios-tcflow.h \ bits/termios-misc.h \ bits/types/struct_semid_ds.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 +1 -0
@@ -178,6 +178,7 @@ libc { getdents64; gettid; tgkill; } GLIBC_2.32 { + __rseq_abi; } GLIBC_PRIVATE { # functions used in other libraries
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
@@ -2146,4 +2146,5 @@ GLIBC_2.30 getdents64 F GLIBC_2.30 gettid F GLIBC_2.30 tgkill F GLIBC_2.30 twalk_r F +GLIBC_2.32 __rseq_abi T 0x20 GLIBC_2.32 pthread_sigmask F
sysdeps/unix/sysv/linux/alpha/libc.abilist +1 -0
@@ -2226,6 +2226,7 @@ GLIBC_2.30 getdents64 F GLIBC_2.30 gettid F GLIBC_2.30 tgkill F GLIBC_2.30 twalk_r F +GLIBC_2.32 __rseq_abi T 0x20 GLIBC_2.32 pthread_sigmask F GLIBC_2.4 _IO_fprintf F GLIBC_2.4 _IO_printf F
sysdeps/unix/sysv/linux/arm/be/libc.abilist +1 -0
@@ -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_2.32 __rseq_abi T 0x20 GLIBC_2.32 pthread_sigmask F GLIBC_2.4 _Exit F GLIBC_2.4 _IO_2_1_stderr_ 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
@@ -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_2.32 __rseq_abi T 0x20 GLIBC_2.32 pthread_sigmask F GLIBC_2.4 _Exit F GLIBC_2.4 _IO_2_1_stderr_ D 0xa0
sysdeps/unix/sysv/linux/bits/rseq.h new file +29 -0
@@ -0,0 +1,29 @@ +/* Restartable Sequences architecture header. Stub version. +Copyright (C) 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 +<https://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
@@ -2090,4 +2090,5 @@ GLIBC_2.30 getdents64 F GLIBC_2.30 gettid F GLIBC_2.30 tgkill F GLIBC_2.30 twalk_r F +GLIBC_2.32 __rseq_abi T 0x20 GLIBC_2.32 pthread_sigmask F
sysdeps/unix/sysv/linux/hppa/libc.abilist +1 -0
@@ -2047,6 +2047,7 @@ GLIBC_2.30 getdents64 F GLIBC_2.30 gettid F GLIBC_2.30 tgkill F GLIBC_2.30 twalk_r F +GLIBC_2.32 __rseq_abi T 0x20 GLIBC_2.32 pthread_sigmask F GLIBC_2.4 __confstr_chk F GLIBC_2.4 __fgets_chk F
sysdeps/unix/sysv/linux/i386/libc.abilist +1 -0
@@ -2213,6 +2213,7 @@ GLIBC_2.30 getdents64 F GLIBC_2.30 gettid F GLIBC_2.30 tgkill F GLIBC_2.30 twalk_r F +GLIBC_2.32 __rseq_abi T 0x20 GLIBC_2.32 pthread_sigmask F GLIBC_2.4 __confstr_chk F GLIBC_2.4 __fgets_chk F
sysdeps/unix/sysv/linux/ia64/libc.abilist +1 -0
@@ -2079,6 +2079,7 @@ GLIBC_2.30 getdents64 F GLIBC_2.30 gettid F GLIBC_2.30 tgkill F GLIBC_2.30 twalk_r F +GLIBC_2.32 __rseq_abi T 0x20 GLIBC_2.32 pthread_sigmask F GLIBC_2.4 __confstr_chk F GLIBC_2.4 __fgets_chk F
sysdeps/unix/sysv/linux/m68k/coldfire/libc.abilist +1 -0
@@ -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_2.32 __rseq_abi T 0x20 GLIBC_2.32 pthread_sigmask F GLIBC_2.4 _Exit F GLIBC_2.4 _IO_2_1_stderr_ D 0x98
sysdeps/unix/sysv/linux/m68k/m680x0/libc.abilist +1 -0
@@ -2159,6 +2159,7 @@ GLIBC_2.30 twalk_r F GLIBC_2.31 msgctl F GLIBC_2.31 semctl F GLIBC_2.31 shmctl F +GLIBC_2.32 __rseq_abi T 0x20 GLIBC_2.32 pthread_sigmask F GLIBC_2.4 __confstr_chk F GLIBC_2.4 __fgets_chk F
sysdeps/unix/sysv/linux/microblaze/be/libc.abilist +1 -0
@@ -2141,4 +2141,5 @@ GLIBC_2.30 twalk_r F GLIBC_2.31 msgctl F GLIBC_2.31 semctl F GLIBC_2.31 shmctl F +GLIBC_2.32 __rseq_abi T 0x20 GLIBC_2.32 pthread_sigmask F
sysdeps/unix/sysv/linux/microblaze/le/libc.abilist +1 -0
@@ -2138,4 +2138,5 @@ GLIBC_2.30 getdents64 F GLIBC_2.30 gettid F GLIBC_2.30 tgkill F GLIBC_2.30 twalk_r F +GLIBC_2.32 __rseq_abi T 0x20 GLIBC_2.32 pthread_sigmask F
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
@@ -2130,6 +2130,7 @@ GLIBC_2.30 getdents64 F GLIBC_2.30 gettid F GLIBC_2.30 tgkill F GLIBC_2.30 twalk_r F +GLIBC_2.32 __rseq_abi T 0x20 GLIBC_2.32 pthread_sigmask F GLIBC_2.4 __confstr_chk F GLIBC_2.4 __fgets_chk F
sysdeps/unix/sysv/linux/mips/mips32/nofpu/libc.abilist +1 -0
@@ -2128,6 +2128,7 @@ GLIBC_2.30 getdents64 F GLIBC_2.30 gettid F GLIBC_2.30 tgkill F GLIBC_2.30 twalk_r F +GLIBC_2.32 __rseq_abi T 0x20 GLIBC_2.32 pthread_sigmask F GLIBC_2.4 __confstr_chk F GLIBC_2.4 __fgets_chk F
sysdeps/unix/sysv/linux/mips/mips64/n32/libc.abilist +1 -0
@@ -2136,6 +2136,7 @@ GLIBC_2.30 getdents64 F GLIBC_2.30 gettid F GLIBC_2.30 tgkill F GLIBC_2.30 twalk_r F +GLIBC_2.32 __rseq_abi T 0x20 GLIBC_2.32 pthread_sigmask F GLIBC_2.4 __confstr_chk F GLIBC_2.4 __fgets_chk F
sysdeps/unix/sysv/linux/mips/mips64/n64/libc.abilist +1 -0
@@ -2130,6 +2130,7 @@ GLIBC_2.30 getdents64 F GLIBC_2.30 gettid F GLIBC_2.30 tgkill F GLIBC_2.30 twalk_r F +GLIBC_2.32 __rseq_abi T 0x20 GLIBC_2.32 pthread_sigmask F GLIBC_2.4 __confstr_chk F GLIBC_2.4 __fgets_chk F
sysdeps/unix/sysv/linux/nios2/libc.abilist +1 -0
@@ -2179,4 +2179,5 @@ GLIBC_2.30 getdents64 F GLIBC_2.30 gettid F GLIBC_2.30 tgkill F GLIBC_2.30 twalk_r F +GLIBC_2.32 __rseq_abi T 0x20 GLIBC_2.32 pthread_sigmask F
sysdeps/unix/sysv/linux/powerpc/bits/rseq.h new file +37 -0
@@ -0,0 +1,37 @@ +/* Restartable Sequences Linux powerpc architecture header. +Copyright (C) 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 +<https://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
@@ -2186,6 +2186,7 @@ GLIBC_2.30 getdents64 F GLIBC_2.30 gettid F GLIBC_2.30 tgkill F GLIBC_2.30 twalk_r F +GLIBC_2.32 __rseq_abi T 0x20 GLIBC_2.32 pthread_sigmask F GLIBC_2.4 _IO_fprintf F GLIBC_2.4 _IO_printf F
sysdeps/unix/sysv/linux/powerpc/powerpc32/nofpu/libc.abilist +1 -0
@@ -2219,6 +2219,7 @@ GLIBC_2.30 getdents64 F GLIBC_2.30 gettid F GLIBC_2.30 tgkill F GLIBC_2.30 twalk_r F +GLIBC_2.32 __rseq_abi T 0x20 GLIBC_2.32 pthread_sigmask F GLIBC_2.4 _IO_fprintf F GLIBC_2.4 _IO_printf F
sysdeps/unix/sysv/linux/powerpc/powerpc64/be/libc.abilist +1 -0
@@ -2049,6 +2049,7 @@ GLIBC_2.30 getdents64 F GLIBC_2.30 gettid F GLIBC_2.30 tgkill F GLIBC_2.30 twalk_r F +GLIBC_2.32 __rseq_abi T 0x20 GLIBC_2.32 pthread_sigmask F GLIBC_2.4 _IO_fprintf F GLIBC_2.4 _IO_printf F
sysdeps/unix/sysv/linux/powerpc/powerpc64/le/libc.abilist +1 -0
@@ -2288,6 +2288,7 @@ GLIBC_2.32 __qecvtieee128_r F GLIBC_2.32 __qfcvtieee128 F GLIBC_2.32 __qfcvtieee128_r F GLIBC_2.32 __qgcvtieee128 F +GLIBC_2.32 __rseq_abi T 0x20 GLIBC_2.32 __scanfieee128 F GLIBC_2.32 __snprintf_chkieee128 F GLIBC_2.32 __snprintfieee128 F
sysdeps/unix/sysv/linux/riscv/rv64/libc.abilist +1 -0
@@ -2108,4 +2108,5 @@ GLIBC_2.30 getdents64 F GLIBC_2.30 gettid F GLIBC_2.30 tgkill F GLIBC_2.30 twalk_r F +GLIBC_2.32 __rseq_abi T 0x20 GLIBC_2.32 pthread_sigmask F
sysdeps/unix/sysv/linux/rseq-internal.h new file +73 -0
73 lines not on this page · raw message
sysdeps/unix/sysv/linux/rseq-sym.c new file +26 -0
@@ -0,0 +1,26 @@ +/* Restartable Sequences exported symbols. Linux Implementation. +Copyright (C) 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 +<https://www.gnu.org/licenses/>. */ + +#include <sys/syscall.h> +#include <stdint.h> +#include <kernel-features.h> +#include <sys/rseq.h> + +__threadstruct rseq __rseq_abi = + { + .cpu_id =RSEQ_CPU_ID_UNINITIALIZED , + };
sysdeps/unix/sysv/linux/s390/bits/rseq.h new file +37 -0
@@ -0,0 +1,37 @@ +/* Restartable Sequences Linux s390 architecture header. +Copyright (C) 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 +<https://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
@@ -2184,6 +2184,7 @@ GLIBC_2.30 twalk_r F GLIBC_2.31 msgctl F GLIBC_2.31 semctl F GLIBC_2.31 shmctl F +GLIBC_2.32 __rseq_abi T 0x20 GLIBC_2.32 pthread_sigmask F GLIBC_2.4 _IO_fprintf F GLIBC_2.4 _IO_printf F
sysdeps/unix/sysv/linux/s390/s390-64/libc.abilist +1 -0
@@ -2085,6 +2085,7 @@ GLIBC_2.30 getdents64 F GLIBC_2.30 gettid F GLIBC_2.30 tgkill F GLIBC_2.30 twalk_r F +GLIBC_2.32 __rseq_abi T 0x20 GLIBC_2.32 pthread_sigmask F GLIBC_2.4 _IO_fprintf F GLIBC_2.4 _IO_printf F
sysdeps/unix/sysv/linux/sh/be/libc.abilist +1 -0
@@ -2054,6 +2054,7 @@ GLIBC_2.30 twalk_r F GLIBC_2.31 msgctl F GLIBC_2.31 semctl F GLIBC_2.31 shmctl F +GLIBC_2.32 __rseq_abi T 0x20 GLIBC_2.32 pthread_sigmask F GLIBC_2.4 __confstr_chk F GLIBC_2.4 __fgets_chk F
sysdeps/unix/sysv/linux/sh/le/libc.abilist +1 -0
@@ -2051,6 +2051,7 @@ GLIBC_2.30 getdents64 F GLIBC_2.30 gettid F GLIBC_2.30 tgkill F GLIBC_2.30 twalk_r F +GLIBC_2.32 __rseq_abi T 0x20 GLIBC_2.32 pthread_sigmask F GLIBC_2.4 __confstr_chk F GLIBC_2.4 __fgets_chk F
sysdeps/unix/sysv/linux/sparc/sparc32/libc.abilist +1 -0
@@ -2175,6 +2175,7 @@ GLIBC_2.30 getdents64 F GLIBC_2.30 gettid F GLIBC_2.30 tgkill F GLIBC_2.30 twalk_r F +GLIBC_2.32 __rseq_abi T 0x20 GLIBC_2.32 pthread_sigmask F GLIBC_2.4 _IO_fprintf F GLIBC_2.4 _IO_printf F
sysdeps/unix/sysv/linux/sparc/sparc64/libc.abilist +1 -0
@@ -2102,6 +2102,7 @@ GLIBC_2.30 getdents64 F GLIBC_2.30 gettid F GLIBC_2.30 tgkill F GLIBC_2.30 twalk_r F +GLIBC_2.32 __rseq_abi T 0x20 GLIBC_2.32 pthread_sigmask F GLIBC_2.4 __confstr_chk F GLIBC_2.4 __fgets_chk F
sysdeps/unix/sysv/linux/sys/rseq.h new file +207 -0
207 lines not on this page · raw message
sysdeps/unix/sysv/linux/x86/bits/rseq.h new file +30 -0
@@ -0,0 +1,30 @@ +/* Restartable Sequences Linux x86 architecture header. +Copyright (C) 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 +<https://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
@@ -2060,6 +2060,7 @@ GLIBC_2.30 getdents64 F GLIBC_2.30 gettid F GLIBC_2.30 tgkill F GLIBC_2.30 twalk_r F +GLIBC_2.32 __rseq_abi T 0x20 GLIBC_2.32 pthread_sigmask F GLIBC_2.4 __confstr_chk F GLIBC_2.4 __fgets_chk F
sysdeps/unix/sysv/linux/x86_64/x32/libc.abilist +1 -0
@@ -2159,4 +2159,5 @@ GLIBC_2.30 getdents64 F GLIBC_2.30 gettid F GLIBC_2.30 tgkill F GLIBC_2.30 twalk_r F +GLIBC_2.32 __rseq_abi T 0x20 GLIBC_2.32 pthread_sigmask F
Signature · 1 line
2.17.1
Re: [PATCH glibc 1/3] glibc: Perform rseq registration at C startup and thread creation (v19)
* Mathieu Desnoyers via Libc-alpha:
Quoted 7 lines
NEWS +0
-0@@ -23,6 +23,16 @@ Major new features: toolchains. It is recommended to use GCC 8 or newer when testing this option.
NEWS Major new features: · line 34
Quoted 10 lines
NEWS +9
-0+* Support for automatically registering threads with the Linux rseq + 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. + The GNU C Library manual has details on integration of Restartable + Sequences.
“rseq” instead “rseq(2)”.
Quoted 9 lines
elf/libc_early_init.c +1
-0@@ -18,10 +18,14 @@#include <ctype.h> #include <libc-early-init.h> +#include <rseq-internal.h>
elf/libc_early_init.c line 31
Quoted 10 lines
elf/libc_early_init.c +3
-0void __libc_early_init (_Bool initial) {/* Initialize ctype data. */ __ctype_init (); +/* Register rseq ABI to the kernel for the main program's libc. */ +if (initial) +rseq_register_current_thread (); }
Okay.
Quoted 5 lines
manual/threads.texi +0
-0
manual/threads.texi This chapter describes functions used for managing threads. · line 16
Quoted 10 lines
manual/threads.texi +4
-2@@ -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
This should go into the extensions menu (@node Non-POSIX Extensions).
General comment: Please wrap the lines around 72 or so characters. Thanks.
manual/threads.texi Behaves like @code{pthread_timedjoin_np} except that the absolute time in · line 892
Quoted 11 lines
manual/threads.texi +7
-0@@ -881,3 +883,63 @@ 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 Restartable Sequences + +This section describes Restartable Sequences integration for +@theglibc{}.
“This functionality is only available on Linux.” (The @standards parts are not visible to readers.)
manual/threads.texi Behaves like @code{pthread_timedjoin_np} except that the absolute time in · line 900
Quoted 8 lines
manual/threads.texi +8
-0+ +@deftypevar {struct rseq} __rseq_abi +@standards{Linux, 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 @file{sys/rseq.h} header. Registration of each +thread's @code{__rseq_abi} is performed by @theglibc{} at library +initialization and thread creation.
Can drop “(Linux-specific)” here.
sysdeps/unix/sysv/linux/aarch64/bits/rseq.h line 41
Quoted 11 lines
sysdeps/unix/sysv/linux/aarch64/bits/rseq.h new file +5
-0+#ifdef __AARCH64EB__ +#define RSEQ_SIG_DATA 0x00bc28d4/* BRK #0x45E0. */ +#else +#define RSEQ_SIG_DATA RSEQ_SIG_CODE +#endif
Missing indentation on the #defines (sorry!).
sysdeps/unix/sysv/linux/arm/bits/rseq.h line 83
Quoted 11 lines
sysdeps/unix/sysv/linux/arm/bits/rseq.h new file +5
-0+#ifdef __ARMEB__ +#define RSEQ_SIG 0xf3def5e7/* udf #24035 ; 0x5de3 (ARMv6+) */ +#else +#define RSEQ_SIG 0xe7f5def3/* udf #24035 ; 0x5de3 */ +#endif
Likewise, missing indentation.
Quoted 6 lines
sysdeps/unix/sysv/linux/bits/rseq.h new file +0
-0
sysdeps/unix/sysv/linux/bits/rseq.h line 29
Quoted 9 lines
sysdeps/unix/sysv/linux/bits/rseq.h new file +8
-0@@ -0,0 +1,29 @@ +/* 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. */
I wonder if we should say something somewhere that the correct way to check for compile-time rseq support in glibc is something like this?
#if __has_include (<sys/rseq.h>)
# include <sys/rseq.h> #endif #ifdef RSEQ_SIG … #endif
Or perhaps we should suppress installation of the headers if we only have support for the stub.
(I think this fine tuning can be deferred to later patch.)
Quoted 14 lines
sysdeps/unix/sysv/linux/mips/bits/rseq.h new file +7
-0@@ -0,0 +1,62 @@ +#if defined(__nanomips__) +# ifdef __MIPSEL__ +# define RSEQ_SIG 0x03500010 +# else +# define RSEQ_SIG 0x00100350 +# endif +#elif defined(__mips_micromips)
sysdeps/unix/sysv/linux/mips/bits/rseq.h line 62
Quoted 10 lines
sysdeps/unix/sysv/linux/mips/bits/rseq.h new file +10
-0+# ifdef __MIPSEL__ +# define RSEQ_SIG 0xd4070000 +# else +# define RSEQ_SIG 0x0000d407 +# endif +#elif defined(__mips__) +# define RSEQ_SIG 0x0350000d +#else +/* Unknown MIPS architecture. */ +#endif
Please use “defined (”, with a space.
sysdeps/unix/sysv/linux/rseq-internal.h line 64
Quoted 13 lines
sysdeps/unix/sysv/linux/rseq-internal.h new file +6
-0+#ifdef RSEQ_SIG +static inline void +rseq_register_current_thread (void) + if (msg) + __libc_fatal (msg); + }
“if (msg != NULL)”, please.
sysdeps/unix/sysv/linux/rseq-internal.h line 71
Quoted 6 lines
sysdeps/unix/sysv/linux/rseq-internal.h new file +6
-0+#else +static inline void +rseq_register_current_thread (void ) +{ +} +#endif
Maybe add /* RSEQ_SIG */ comments to #else/#endif here as well.
Quoted 6 lines
sysdeps/unix/sysv/linux/sys/rseq.h new file +0
-0
sysdeps/unix/sysv/linux/sys/rseq.h line 36
Quoted 10 lines
sysdeps/unix/sysv/linux/sys/rseq.h new file +10
-0+#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
Too much indentation on the define line, I think.
Still missing: #ifdef __ASSEMBLER__. .S files should be able to include <sys/rseq.h> to get the definition of RSEQ_SIG. But I think this can be deferred to a follow-up.
sysdeps/unix/sysv/linux/sys/rseq.h line 42
Quoted 5 lines
sysdeps/unix/sysv/linux/sys/rseq.h new file +5
-0+#ifdef __GLIBC_HAVE_KERNEL_RSEQ +/* We use the structures declarations from the kernel headers. */ +# include <linux/rseq.h> +#else +/* We use a copy of the include/uapi/linux/rseq.h kernel header. */
This comment is not true, the kernel headers do not have uptr support.
If we revert the uptr change, we also need to update the manual, I think.
sysdeps/unix/sysv/linux/sys/rseq.h line 197
Quoted 3 lines
sysdeps/unix/sysv/linux/sys/rseq.h new file +3
-0+/* Ensure the compiler supports __attribute__ ((aligned)). */ +_Static_assert (__alignof__ (struct rseq_cs ) >=32 ,"alignment" ); +_Static_assert (__alignof__ (struct rseq ) >=32 ,"alignment" );
This needs #ifndef __cplusplus or something like that. I'm surprised that this passes the installed header tests.
sysdeps/unix/sysv/linux/sys/rseq.h line 205
Quoted 7 lines
sysdeps/unix/sysv/linux/sys/rseq.h new file +7
-0+/* 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. */ + +extern __thread struct rseq __rseq_abi + __attribute__ ((tls_model ("initial-exec")));
Should be __tls_model__.
We're getting really close now. 8-)
Thanks,
Florian
Re: [PATCH glibc 1/3] glibc: Perform rseq registration at C startup and thread creation (v19)
Quoted 10 lines · ----- On May 20, 2020, at 7:40 AM, Florian Weimer [email protected] wrote:
* Mathieu Desnoyers via Libc-alpha:
Quoted 8 lines
NEWS +0
-0@@ -23,6 +23,16 @@ Major new features: toolchains. It is recommended to use GCC 8 or newer when testing this option.
Quoted 11 lines
NEWS Major new features: · line 34
Quoted 9 lines
NEWS +9
-0+* Support for automatically registering threads with the Linux rseq + 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. + The GNU C Library manual has details on integration of Restartable + Sequences.“rseq” instead “rseq(2)”.
OK
[...]
Quoted 8 lines
Quoted 8 lines
manual/threads.texi +0
-0@@ -9,8 +9,10 @@ This chapter describes functions used for managing threads. POSIX threads.
Quoted 10 lines
manual/threads.texi This chapter describes functions used for managing threads. · line 16
Quoted 8 lines
manual/threads.texi +4
-2@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 menuThis should go into the extensions menu (@node Non-POSIX Extensions).
OK
Quoted 3 lines
General comment: Please wrap the lines around 72 or so characters. Thanks.
OK
Quoted 16 lines
manual/threads.texi Behaves like @code{pthread_timedjoin_np} except that the absolute time in · line 892
Quoted 13 lines
manual/threads.texi +0
-0@@ -881,3 +883,63 @@ Behaves like @code{pthread_timedjoin_np} except that theabsolute time in
@cpthread_spin_unlock
@cpthread_testcancel
@cpthread_yieldmanual/threads.texi +7
-0+ +@node Restartable Sequences +@section Restartable Sequences +@cindex Restartable Sequences + +This section describes Restartable Sequences integration for +@theglibc{}.“This functionality is only available on Linux.” (The @standards parts are not visible to readers.)
OK
Quoted 11 lines
manual/threads.texi Behaves like @code{pthread_timedjoin_np} except that the absolute time in · line 900
Quoted 9 lines
manual/threads.texi +8
-0+ +@deftypevar {struct rseq} __rseq_abi +@standards{Linux, 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 @file{sys/rseq.h} header. Registration of each +thread's @code{__rseq_abi} is performed by @theglibc{} at library +initialization and thread creation.Can drop “(Linux-specific)” here.
OK
Quoted 8 lines
Quoted 7 lines
sysdeps/unix/sysv/linux/aarch64/bits/rseq.h new file +0
-0
Quoted 7 lines
sysdeps/unix/sysv/linux/aarch64/bits/rseq.h line 41
Quoted 5 lines
sysdeps/unix/sysv/linux/aarch64/bits/rseq.h new file +5
-0+#ifdef __AARCH64EB__ +#define RSEQ_SIG_DATA 0x00bc28d4/* BRK #0x45E0. */ +#else +#define RSEQ_SIG_DATA RSEQ_SIG_CODE +#endif Missing indentation on the #defines (sorry!).
Sorry! My bad!
Quoted 8 lines
Quoted 7 lines
sysdeps/unix/sysv/linux/arm/bits/rseq.h new file +0
-0
Quoted 7 lines
sysdeps/unix/sysv/linux/arm/bits/rseq.h line 83
Quoted 5 lines
sysdeps/unix/sysv/linux/arm/bits/rseq.h new file +5
-0+#ifdef __ARMEB__ +#define RSEQ_SIG 0xf3def5e7/* udf #24035 ; 0x5de3 (ARMv6+) */ +#else +#define RSEQ_SIG 0xe7f5def3/* udf #24035 ; 0x5de3 */ +#endif Likewise, missing indentation.
At least I was consistently wrong. ;-)
Quoted 28 lines
Quoted 8 lines
sysdeps/unix/sysv/linux/bits/rseq.h new file +0
-0sysdeps/unix/sysv/linux/bits/rseq.h line 29
Quoted 9 lines
sysdeps/unix/sysv/linux/bits/rseq.h new file +8
-0@@ -0,0 +1,29 @@ +/* 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. */ I wonder if we should say something somewhere that the correct way to check for compile-time rseq support in glibc is something like this?
#if
__has_include(<sys/rseq.h>)# include <sys/rseq.h> #endif #ifdef RSEQ_SIG … #endif
Quoted 4 lines
Or perhaps we should suppress installation of the headers if we only have support for the stub.
(I think this fine tuning can be deferred to later patch.)
I would prefer we keep this for a later patch.
That being said, I notice that gcc prior to 4.9 has trouble compiling __has_include. Therefore, I would tend to recommend the following as a backward-compatible way of using the rseq header from external projects:
- At configure time, check for availability of sys/rseq.h to ensure the glibc version is recent enough. In the autotools world, this is usually done with a small test program which will fail to build if the header is not there.
- At preprocessing time, include <sys/rseq.h> and check whether RSEQ_SIG is defined.
If a project only aims at building with gcc 4.9 or newer, then it can use __has_include.
Quoted 20 lines
Quoted 20 lines
sysdeps/unix/sysv/linux/mips/bits/rseq.h new file +11
-0@@ -0,0 +1,62 @@ +#if defined(__nanomips__) +# ifdef __MIPSEL__ +# define RSEQ_SIG 0x03500010 +# else +# define RSEQ_SIG 0x00100350 +# endif +#elif defined(__mips_micromips) +# ifdef __MIPSEL__ +# define RSEQ_SIG 0xd4070000 +# else +# define RSEQ_SIG 0x0000d407
Quoted 8 lines
sysdeps/unix/sysv/linux/mips/bits/rseq.h line 62
Quoted 6 lines
sysdeps/unix/sysv/linux/mips/bits/rseq.h new file +6
-0+# endif +#elif defined(__mips__) +# define RSEQ_SIG 0x0350000d +#else +/* Unknown MIPS architecture. */ +#endif Please use “defined (”, with a space.
OK
Quoted 12 lines
Quoted 11 lines
sysdeps/unix/sysv/linux/rseq-internal.h new file +3
-0+#ifdef RSEQ_SIG +static inline void +rseq_register_current_thread (void )
Quoted 5 lines
sysdeps/unix/sysv/linux/rseq-internal.h line 64
Quoted 3 lines
sysdeps/unix/sysv/linux/rseq-internal.h new file +3
-0+if (msg) +__libc_fatal (msg); + }“if (msg != NULL)”, please.
OK
Quoted 9 lines
sysdeps/unix/sysv/linux/rseq-internal.h line 71
Quoted 7 lines
sysdeps/unix/sysv/linux/rseq-internal.h new file +6
-0+#else +static inline void +rseq_register_current_thread (void ) +{ +} +#endif Maybe add /*
RSEQ_SIG*/ comments to #else/#endif here as well.
OK
Quoted 8 lines
Quoted 7 lines
sysdeps/unix/sysv/linux/sys/rseq.h new file +0
-0
Quoted 12 lines
sysdeps/unix/sysv/linux/sys/rseq.h line 36
Quoted 10 lines
sysdeps/unix/sysv/linux/sys/rseq.h new file +10
-0+#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 Too much indentation on the define line, I think.
OK
Quoted 4 lines
Still missing: #ifdef
__ASSEMBLER__. .S files should be able to include <sys/rseq.h> to get the definition ofRSEQ_SIG. But I think this can be deferred to a follow-up.
Would the plan be to only do the #include <bits/rseq.h> bits in #ifdef __ASSEMBLER__ and skip all the rest ?
Quoted 11 lines
sysdeps/unix/sysv/linux/sys/rseq.h line 42
Quoted 6 lines
sysdeps/unix/sysv/linux/sys/rseq.h new file +5
-0+#ifdef __GLIBC_HAVE_KERNEL_RSEQ +/* We use the structures declarations from the kernel headers. */ +# include <linux/rseq.h> +#else +/* We use a copy of the include/uapi/linux/rseq.h kernel header. */ This comment is not true, the kernel headers do not have uptr support.
If we revert the uptr change, we also need to update the manual, I think.
Ah, I get to your uptr concern now! Good :)
The larger question here is: considering that we re-implement the entire uapi header within glibc (which includes the uptr addition), do we still care about using the header provided by the Linux kernel ?
Having different definitions depending on whether a kernel header is installed or not when including a glibc header seems rather unexpected.
*If* we want to use the uapi header, I think something is semantically missing. Here is the scheme I envision. We could rely on the kernel header version.h to figure out which of glibc or kernel uapi header is more recent. Any new concept we try to integrate into glibc (e.g. uptr) should go into the upstream Linux uapi header first.
For the coming glibc e.g. 2.32, we use the kernel uapi header if kernel version is >= 4.18.0. Within glibc, the fallback implements exactly the API exposed by the kernel rseq.h header.
As we eventually introduce the uptr change into the Linux kernel, and say it gets merged for Linux 5.9.0, we mirror this change into glibc (e.g. release 2.33), and bump the Linux kernel version cutoff to 5.9.0. So starting from that version, we use the Linux kernel header only if version >= 5.9.0, else we fallback on glibc's own implementation.
This is clearly something we need to get right.
Quoted 7 lines
sysdeps/unix/sysv/linux/sys/rseq.h line 197
Quoted 4 lines
sysdeps/unix/sysv/linux/sys/rseq.h new file +3
-0+/* Ensure the compiler supports __attribute__ ((aligned)). */ +_Static_assert (__alignof__ (struct rseq_cs ) >=32 ,"alignment" ); +_Static_assert (__alignof__ (struct rseq ) >=32 ,"alignment" );This needs #ifndef
__cplusplusor something like that. I'm surprised that this passes the installed header tests.
Would the following be ok ?
#ifdef __cplusplus
#define rseq_static_assert static_assert
#else
#define rseq_static_assert _Static_assert
#endif
/* Ensure the compiler supports __attribute__ ((aligned)). */ rseq_static_assert (__alignof__ (struct rseq_cs) >= 32, "alignment"); rseq_static_assert (__alignof__ (struct rseq) >= 32, "alignment");
Quoted 10 lines
sysdeps/unix/sysv/linux/sys/rseq.h line 205
Quoted 8 lines
sysdeps/unix/sysv/linux/sys/rseq.h new file +7
-0+/* 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. */ + +extern __thread struct rseq __rseq_abi + __attribute__ ((tls_model ("initial-exec")));Should be
__tls_model__.
OK
Quoted 2 lines
We're getting really close now. 8-)
Great!
Thanks for the feedback!
Mathieu
Quoted 3 lines
Thanks,
Florian
Signature · 3 lines
Mathieu Desnoyers EfficiOS Inc. http://www.efficios.com
Re: [PATCH glibc 1/3] glibc: Perform rseq registration at C startup and thread creation (v19)
* Mathieu Desnoyers:
Quoted 3 lines
The larger question here is: considering that we re-implement the entire uapi header within glibc (which includes the uptr addition), do we still care about using the header provided by the Linux kernel ?
We don't care, but our users do. Eventually, they want to include <sys/rseq.h> and <linux/rseq.h> to get new constants that are not yet known to glibc.
Quoted 2 lines
Having different definitions depending on whether a kernel header is installed or not when including a glibc header seems rather unexpected.
Indeed.
Quoted 5 lines
*If* we want to use the uapi header, I think something is semantically missing. Here is the scheme I envision. We could rely on the kernel header
version.hto figure out which of glibc or kernel uapi header is more recent. Any new concept we try to integrate into glibc (e.g. uptr) should go into the upstream Linux uapi header first.
I think we should always prefer the uapi header. The Linux version check does not tell you anything about backports.
Quoted 3 lines
For the coming glibc e.g. 2.32, we use the kernel uapi header if kernel version is >= 4.18.0. Within glibc, the fallback implements exactly the API exposed by the kernel
rseq.hheader.
Agreed.
Quoted 5 lines
As we eventually introduce the uptr change into the Linux kernel, and say it gets merged for Linux 5.9.0, we mirror this change into glibc (e.g. release 2.33), and bump the Linux kernel version cutoff to 5.9.0. So starting from that version, we use the Linux kernel header only if version >= 5.9.0, else we fallback on glibc's own implementation.
Fortunately, we don't need to settle this today. 8-)
Let's stick to the 4.18 definitions for the fallback for now, and discuss the incorporation of future changes later.
Quoted 15 lines
Quoted 6 lines
sysdeps/unix/sysv/linux/sys/rseq.h line 197
Quoted 3 lines
sysdeps/unix/sysv/linux/sys/rseq.h new file +3
-0+/* Ensure the compiler supports __attribute__ ((aligned)). */ +_Static_assert (__alignof__ (struct rseq_cs ) >=32 ,"alignment" ); +_Static_assert (__alignof__ (struct rseq ) >=32 ,"alignment" );This needs #ifndef
__cplusplusor something like that. I'm surprised that this passes the installed header tests.Would the following be ok ?
#ifdef
__cplusplus
#definerseq_static_assertstatic_assert
#else
#definerseq_static_assert_Static_assert
#endif
Quoted 3 lines
/* Ensure the compiler supports
__attribute__((aligned)). */rseq_static_assert(__alignof__(structrseq_cs) >= 32, "alignment");rseq_static_assert(__alignof__(struct rseq) >= 32, "alignment");
Seems reasonable, yes. __alignof__ is still a GCC extension. C++11 has alignof, C11 has _Alignof. So you could use something like this (perhaps without indentation for the kernel header version):
#ifdef __cplusplus
# if __cplusplus >= 201103L # define rseq_static_assert(x) static_assert x; # define rseq_alignof alignof # endif #elif __STDC_VERSION__ >= 201112L # define rseq_static_assert(x) _Static_assert x; # define rseq_alignof _Alignof #endif #ifndef rseq_static_assert # define rseq_static_assert /* nothing */ #endif rseq_static_assert ((rseq_alignof__ (struct rseq_cs) >= 32, "alignment")) rseq_static_assert ((rseq_alignof (struct rseq) >= 32, "alignment"))
And something similar for _Alignas/attribute aligned, with an error for older standards and !__GNUC__ compilers (because neither the type nor __thread can be represented there).
Thanks,
Florian
Re: [PATCH glibc 1/3] glibc: Perform rseq registration at C startup and thread creation (v19)
Quoted 9 lines · ----- On May 25, 2020, at 11:20 AM, Florian Weimer [email protected] wrote:
* Mathieu Desnoyers:
Quoted 3 lines
The larger question here is: considering that we re-implement the entire uapi header within glibc (which includes the uptr addition), do we still care about using the header provided by the Linux kernel ?
We don't care, but our users do. Eventually, they want to include <
sys/rseq.h> and <linux/rseq.h> to get new constants that are not yet known to glibc.
Good point!
Quoted 6 lines
Quoted 3 lines
Having different definitions depending on whether a kernel header is installed or not when including a glibc header seems rather unexpected.
Indeed.
Quoted 8 lines
Quoted 5 lines
*If* we want to use the uapi header, I think something is semantically missing. Here is the scheme I envision. We could rely on the kernel header
version.hto figure out which of glibc or kernel uapi header is more recent. Any new concept we try to integrate into glibc (e.g. uptr) should go into the upstream Linux uapi header first.I think we should always prefer the uapi header. The Linux version check does not tell you anything about backports.
Fair enough.
Quoted 13 lines
Quoted 4 lines
For the coming glibc e.g. 2.32, we use the kernel uapi header if kernel version is >= 4.18.0. Within glibc, the fallback implements exactly the API exposed by the kernel
rseq.hheader.Agreed.
Quoted 5 lines
As we eventually introduce the uptr change into the Linux kernel, and say it gets merged for Linux 5.9.0, we mirror this change into glibc (e.g. release 2.33), and bump the Linux kernel version cutoff to 5.9.0. So starting from that version, we use the Linux kernel header only if version >= 5.9.0, else we fallback on glibc's own implementation.
Quoted 4 lines
Fortunately, we don't need to settle this today. 8-)
Let's stick to the 4.18 definitions for the fallback for now, and discuss the incorporation of future changes later.
OK
Quoted 25 lines
Quoted 16 lines
Quoted 7 lines
sysdeps/unix/sysv/linux/sys/rseq.h line 197
Quoted 4 lines
sysdeps/unix/sysv/linux/sys/rseq.h new file +3
-0+/* Ensure the compiler supports __attribute__ ((aligned)). */ +_Static_assert (__alignof__ (struct rseq_cs ) >=32 ,"alignment" ); +_Static_assert (__alignof__ (struct rseq ) >=32 ,"alignment" );This needs #ifndef
__cplusplusor something like that. I'm surprised that this passes the installed header tests.Would the following be ok ?
#ifdef
__cplusplus
#definerseq_static_assertstatic_assert
#else
#definerseq_static_assert_Static_assert
#endifQuoted 3 lines
/* Ensure the compiler supports
__attribute__((aligned)). */rseq_static_assert(__alignof__(structrseq_cs) >= 32, "alignment");rseq_static_assert(__alignof__(struct rseq) >= 32, "alignment");Seems reasonable, yes.
__alignof__is still a GCC extension. C++11 has alignof, C11 has_Alignof. So you could use something like this (perhaps without indentation for the kernel header version):#ifdef
__cplusplus
Quoted 13 lines
# if __cplusplus >= 201103L # define rseq_static_assert(x) static_assert x; # define rseq_alignof alignof # endif #elif __STDC_VERSION__ >= 201112L # define rseq_static_assert(x) _Static_assert x; # define rseq_alignof _Alignof #endif #ifndef rseq_static_assert # define rseq_static_assert /* nothing */ #endif rseq_static_assert ((rseq_alignof__ (struct rseq_cs) >= 32, "alignment")) rseq_static_assert ((rseq_alignof (struct rseq) >= 32, "alignment"))
Something like this ?
#ifdef __cplusplus
# if __cplusplus >= 201103L # define rseq_static_assert (expr, diagnostic) static_assert (expr, diagnostic) # define rseq_alignof alignof # endif #elif __STDC_VERSION__ >= 201112L # define rseq_static_assert (expr, diagnostic) _Static_assert (expr, diagnostic) # define rseq_alignof _Alignof #endif
#ifndef rseq_static_assert
# define rseq_static_assert (expr, diagnostic) /* nothing */ #endif
/* Ensure the compiler supports __attribute__ ((aligned)). */ rseq_static_assert ((rseq_alignof (struct rseq_cs) >= 32, "alignment")); rseq_static_assert ((rseq_alignof (struct rseq) >= 32, "alignment"));
Quoted 1 line
And something similar for _Alignas/attribute aligned,
I don't see where _Alignas is needed here ?
For attribute aligned, what would be the oldest supported C and C++ standards ?
Quoted 3 lines
with an error for
older standards and !__GNUC__ compilers (because neither the type nor__threadcan be represented there).
By "type" you mean "struct rseq" here ? What does it contain that requires a __GNUC__ compiler ?
About __thread, I recall other compilers have other means to declare it. In liburcu, I end up with the following:
#if defined (__cplusplus) && (__cplusplus >= 201103L)
# define URCU_TLS_STORAGE_CLASS thread_local #elif defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 201112L) # define URCU_TLS_STORAGE_CLASS _Thread_local #elif defined (_MSC_VER) # define URCU_TLS_STORAGE_CLASS __declspec(thread) #else # define URCU_TLS_STORAGE_CLASS __thread #endif
Would something along those lines be OK for libc ?
Thanks,
Mathieu
Signature · 3 lines
Mathieu Desnoyers EfficiOS Inc. http://www.efficios.com
Re: [PATCH glibc 1/3] glibc: Perform rseq registration at C startup and thread creation (v19)
* Mathieu Desnoyers:
Quoted 12 lines
Something like this ?
#ifdef
__cplusplus# if __cplusplus >= 201103L # define rseq_static_assert (expr, diagnostic) static_assert (expr, diagnostic) # define rseq_alignof alignof # endif #elif __STDC_VERSION__ >= 201112L # define rseq_static_assert (expr, diagnostic) _Static_assert (expr, diagnostic) # define rseq_alignof _Alignof #endif
Quoted 3 lines
#ifndef
rseq_static_assert# define rseq_static_assert (expr, diagnostic) /* nothing */ #endif
You can't have a space in #defines like that, no matter what GNU style says. 8-)
Quoted 3 lines
/* Ensure the compiler supports
__attribute__((aligned)). */rseq_static_assert((rseq_alignof(structrseq_cs) >= 32, "alignment"));rseq_static_assert((rseq_alignof(struct rseq) >= 32, "alignment"));
You need to move the ; into rseq_static_assert. And if you use explicit arguments, you can't use double parentheses.
Quoted 6 lines
Quoted 1 line
And something similar for _Alignas/attribute aligned,
I don't see where
_Alignasis needed here ?For attribute aligned, what would be the oldest supported C and C++ standards ?
There are no standardized attributes for C, there is only _Alignas. C++11 has an alignas specifier; it's not an attribute either. I think these are syntactically similar.
Quoted 6 lines
Quoted 3 lines
with an error for
older standards and !__GNUC__ compilers (because neither the type nor__threadcan be represented there).By "type" you mean "struct rseq" here ? What does it contain that requires a
__GNUC__compiler ?
__attribute__ and __thread support.
Quoted 13 lines
About
__thread, I recall other compilers have other means to declare it. In liburcu, I end up with the following:#if defined (
__cplusplus) && (__cplusplus>= 201103L)# define URCU_TLS_STORAGE_CLASS thread_local #elif defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 201112L) # define URCU_TLS_STORAGE_CLASS _Thread_local #elif defined (_MSC_VER) # define URCU_TLS_STORAGE_CLASS __declspec(thread) #else # define URCU_TLS_STORAGE_CLASS __thread #endif
Quoted 1 line
Would something along those lines be OK for libc ?
Yes, it would be okay (minus the Visual C++ part). This part does not have to go into UAPI headers first. A fallback definition of __thread should be okay. Outside glibc, the TLS model declaration is optional, I think. The glibc *definition* ensures that the variable is initial-exec.
Thanks,
Florian
Re: [PATCH glibc 1/3] glibc: Perform rseq registration at C startup and thread creation (v19)
Quoted 16 lines · ----- On May 26, 2020, at 8:41 AM, Florian Weimer [email protected] wrote:
* Mathieu Desnoyers:
Quoted 14 lines
Something like this ?
#ifdef
__cplusplus# if __cplusplus >= 201103L # define rseq_static_assert (expr, diagnostic) static_assert (expr, diagnostic) # define rseq_alignof alignof # endif #elif __STDC_VERSION__ >= 201112L # define rseq_static_assert (expr, diagnostic) _Static_assert (expr, diagnostic) # define rseq_alignof _Alignof #endif
Quoted 6 lines
Quoted 3 lines
#ifndef
rseq_static_assert# define rseq_static_assert (expr, diagnostic) /* nothing */ #endifYou can't have a space in #defines like that, no matter what GNU style says. 8-)
Yes, I noticed when failing to build this ;)
Quoted 7 lines
Quoted 4 lines
/* Ensure the compiler supports
__attribute__((aligned)). */rseq_static_assert((rseq_alignof(structrseq_cs) >= 32, "alignment"));rseq_static_assert((rseq_alignof(struct rseq) >= 32, "alignment"));You need to move the ; into
rseq_static_assert. And if you use explicit arguments, you can't use double parentheses.
Why move the ";" into the macro ?
AFAIU, the only gain here would be to make sure we don't emit useless ";" in the "/* nothing */" case. But does it matter ?
Examples I can find of "static_assert" explicitly have the ";" at the end, so I find it weird to integrate it into the rseq_static_assert macro, which makes it different from static_assert.
Agreed on the need to remove the double-parentheses.
Quoted 11 lines
Quoted 7 lines
Quoted 2 lines
And something similar for _Alignas/attribute aligned,
I don't see where
_Alignasis needed here ?For attribute aligned, what would be the oldest supported C and C++ standards ?
There are no standardized attributes for C, there is only
_Alignas. C++11 has an alignas specifier; it's not an attribute either. I think these are syntactically similar.
There appears to be an interesting difference between attribute aligned and alignas. It seems like alignas cannot be used on a structure declaration, only on fields, e.g.:
struct blah {
int a;
} _Alignas (16);o.c:3:1: warning: useless ‘_Alignas’ in empty declaration
} _Alignas (16);
But
struct blah {
int _Alignas (16) a;
};is OK. So if I change e.g. struct rseq_cs to align the first field:
struct rseq_cs
{
/* Version of this structure. */
uint32_t rseq_align (32) version;
/* enum rseq_cs_flags. */
uint32_t flags;
uint64_t start_ip;
/* Offset from start_ip. */
uint64_t post_commit_offset;
uint64_t abort_ip;};
It should work.
Quoted 9 lines
Quoted 7 lines
Quoted 4 lines
with an error for
older standards and !__GNUC__ compilers (because neither the type nor__threadcan be represented there).By "type" you mean "struct rseq" here ? What does it contain that requires a
__GNUC__compiler ?
__attribute__and__threadsupport.
OK
Quoted 14 lines
Quoted 14 lines
About
__thread, I recall other compilers have other means to declare it. In liburcu, I end up with the following:#if defined (
__cplusplus) && (__cplusplus>= 201103L)# define URCU_TLS_STORAGE_CLASS thread_local #elif defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 201112L) # define URCU_TLS_STORAGE_CLASS _Thread_local #elif defined (_MSC_VER) # define URCU_TLS_STORAGE_CLASS __declspec(thread) #else # define URCU_TLS_STORAGE_CLASS __thread #endif
Quoted 7 lines
Quoted 1 line
Would something along those lines be OK for libc ?
Yes, it would be okay (minus the Visual C++ part). This part does not have to go into UAPI headers first. A fallback definition of
__threadshould be okay. Outside glibc, the TLS model declaration is optional, I think. The glibc *definition* ensures that the variable is initial-exec.
AFAIU you are technically correct when stating that the tls model on the declaration is optional, but I think it's a good thing to have it there rather than only at the definition. It makes it clear to all users of this variable that its model is IE. Especially in scenarios where early-adopter libraries and applications can define their own __rseq_abi symbol, I think it's good to explicitly keep the IE tls model attribute in the header.
I end up with the following:
#ifdef __cplusplus
# if __cplusplus >= 201103L # define rseq_static_assert(expr, diagnostic) static_assert (expr, diagnostic) # define rseq_alignof(type) alignof (type) # define rseq_alignas(x) alignas (x) # define rseq_tls_storage_class thread_local # endif #elif (defined __STDC_VERSION__ ? __STDC_VERSION__ : 0) >= 201112L # define rseq_static_assert(expr, diagnostic) _Static_assert (expr, diagnostic) # define rseq_alignof(type) _Alignof (type) # define rseq_alignas(x) _Alignas (x) # define rseq_tls_storage_class _Thread_local #endif
#ifndef rseq_static_assert
/* Try to use _Static_assert macro from sys/cdefs.h. */
# ifdef _Static_assert # define rseq_static_assert(expr, diagnostic) _Static_assert (expr, diagnostic) # else # define rseq_static_assert(expr, diagnostic) /* Nothing. */ # endif #endif
/* Rely on GNU extensions for older standards and tls model. */ #ifdef __GNUC__
# ifndef rseq_alignof
# define rseq_alignof(x) __alignof__ (x)
# endif
# ifndef rseq_alignas
# define rseq_alignas(x) __attribute__ ((aligned (x)))
# endif
# define rseq_tls_model_ie __attribute__ ((__tls_model__ ("initial-exec")))
#else
/* Specifying the TLS model on the declaration is optional. */
# define rseq_tls_model_ie /* Nothing. */
#endif/* Fall back to __thread for TLS storage class. */ #ifndef rseq_tls_storage_class
# define rseq_tls_storage_class __thread #endif
[...]
/* Ensure the compiler supports rseq_align. */ rseq_static_assert (rseq_alignof (struct rseq_cs) >= 32, "alignment"); rseq_static_assert (rseq_alignof (struct rseq) >= 32, "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. */
extern rseq_tls_storage_class struct rseq __rseq_abi rseq_tls_model_ie;
Thanks,
Mathieu
Signature · 3 lines
Mathieu Desnoyers EfficiOS Inc. http://www.efficios.com
Re: [PATCH glibc 1/3] glibc: Perform rseq registration at C startup and thread creation (v19)
* Mathieu Desnoyers:
Quoted 2 lines
AFAIU, the only gain here would be to make sure we don't emit useless ";" in the "/* nothing */" case. But does it matter ?
I don't think C allows empty constructs like this at the top level.
Quoted 19 lines
Quoted 10 lines
Quoted 6 lines
Quoted 1 line
And something similar for _Alignas/attribute aligned,
I don't see where
_Alignasis needed here ?For attribute aligned, what would be the oldest supported C and C++ standards ?
There are no standardized attributes for C, there is only
_Alignas. C++11 has an alignas specifier; it's not an attribute either. I think these are syntactically similar.There appears to be an interesting difference between attribute aligned and alignas. It seems like alignas cannot be used on a structure declaration, only on fields, e.g.:
struct blah { int a; } _Alignas (16);
Quoted 8 lines
o.c:3:1: warning: useless ‘_Alignas’ in empty declaration
}_Alignas(16);But
struct blah { int _Alignas (16) a; };
Like the attribute, it needs to come right after the struct keyword, I think. (Trailing attributes can be ambiguous, but not in this case.)
Quoted 13 lines
is OK. So if I change e.g. struct
rseq_csto align the first field:struct
rseq_cs{ /* Version of this structure. */ uint32_t rseq_align (32) version; /* enum rseq_cs_flags. */ uint32_t flags; uint64_t start_ip; /* Offset from start_ip. */ uint64_t post_commit_offset; uint64_t abort_ip;
Quoted 3 lines
};
It should work.
Indeed.
Quoted 14 lines
/* Rely on GNU extensions for older standards and tls model. */ #ifdef
__GNUC__# ifndef rseq_alignof # define rseq_alignof(x) __alignof__ (x) # endif # ifndef rseq_alignas # define rseq_alignas(x) __attribute__ ((aligned (x))) # endif # define rseq_tls_model_ie __attribute__ ((__tls_model__ ("initial-exec"))) #else /* Specifying the TLS model on the declaration is optional. */ # define rseq_tls_model_ie /* Nothing. */ #endif
Quoted 4 lines
/* Fall back to
__threadfor TLS storage class. */ #ifndefrseq_tls_storage_class# define rseq_tls_storage_class __thread #endif
If they are only used in the glibc headers, they should have __rseq prefixes, so that application code doesn't start using them (in case we have to change/fix them, or move the into <sys/cdefs.h> later).
Rest looks fine.
Thanks,
Florian
Re: [PATCH glibc 1/3] glibc: Perform rseq registration at C startup and thread creation (v19)
Quoted 29 lines · ----- On May 26, 2020, at 10:38 AM, Florian Weimer [email protected] wrote:
* Mathieu Desnoyers:
Quoted 2 lines
AFAIU, the only gain here would be to make sure we don't emit useless ";" in the "/* nothing */" case. But does it matter ?
I don't think C allows empty constructs like this at the top level.
Quoted 22 lines
Quoted 10 lines
Quoted 6 lines
Quoted 1 line
And something similar for _Alignas/attribute aligned,
I don't see where
_Alignasis needed here ?For attribute aligned, what would be the oldest supported C and C++ standards ?
There are no standardized attributes for C, there is only
_Alignas. C++11 has an alignas specifier; it's not an attribute either. I think these are syntactically similar.There appears to be an interesting difference between attribute aligned and alignas. It seems like alignas cannot be used on a structure declaration, only on fields, e.g.:
struct blah { int a; } _Alignas (16);o.c:3:1: warning: useless ‘_Alignas’ in empty declaration
}_Alignas(16);
Quoted 8 lines
Quoted 5 lines
But
struct blah { int _Alignas (16) a; };Like the attribute, it needs to come right after the struct keyword, I think. (Trailing attributes can be ambiguous, but not in this case.)
Nope. _Alignas really _is_ special :-(
struct _Alignas (16) blah {
int a;
};p.c:1:8: error: expected ‘{’ before ‘_Alignas’
struct _Alignas (16) blah {
Also:
struct blah _Alignas (16) {
int a;
};p.c:1:27: error: expected identifier or ‘(’ before ‘{’ token
struct blah _Alignas (16) {
Quoted 16 lines
Quoted 16 lines
is OK. So if I change e.g. struct
rseq_csto align the first field:struct
rseq_cs{ /* Version of this structure. */ uint32_t rseq_align (32) version; /* enum rseq_cs_flags. */ uint32_t flags; uint64_t start_ip; /* Offset from start_ip. */ uint64_t post_commit_offset; uint64_t abort_ip;};
Quoted 3 lines
Quoted 1 line
It should work.
Indeed.
OK, so let's go for that approach.
Quoted 15 lines
Quoted 15 lines
/* Rely on GNU extensions for older standards and tls model. */ #ifdef
__GNUC__# ifndef rseq_alignof # define rseq_alignof(x) __alignof__ (x) # endif # ifndef rseq_alignas # define rseq_alignas(x) __attribute__ ((aligned (x))) # endif # define rseq_tls_model_ie __attribute__ ((__tls_model__ ("initial-exec"))) #else /* Specifying the TLS model on the declaration is optional. */ # define rseq_tls_model_ie /* Nothing. */ #endif
Quoted 8 lines
Quoted 4 lines
/* Fall back to
__threadfor TLS storage class. */ #ifndefrseq_tls_storage_class# define rseq_tls_storage_class __thread #endifIf they are only used in the glibc headers, they should have
__rseqprefixes, so that application code doesn't start using them (in case we have to change/fix them, or move the into <sys/cdefs.h> later).
OK will do.
Quoted 2 lines
Rest looks fine.
One last thing I'm planning to add in sys/rseq.h to cover acessing the rseq_cs pointers with both the UAPI headers and the glibc struct rseq declarations:
/* The rseq_cs_ptr macro can be used to access the pointer to the current
rseq critical section descriptor. */
#ifdef __LP64__
# define rseq_cs_ptr(rseq) \
((const struct rseq_cs *) (rseq)->rseq_cs.ptr)
#else /* __LP64__ */
# define rseq_cs_ptr(rseq) \
((const struct rseq_cs *) (rseq)->rseq_cs.ptr.ptr32)
#endif /* __LP64__ */Does it make sense ?
Thanks,
Mathieu
Quoted 3 lines
Thanks,
Florian
Signature · 3 lines
Mathieu Desnoyers EfficiOS Inc. http://www.efficios.com
Re: [PATCH glibc 1/3] glibc: Perform rseq registration at C startup and thread creation (v19)
* Mathieu Desnoyers:
Quoted 11 lines
Quoted 2 lines
Like the attribute, it needs to come right after the struct keyword, I think. (Trailing attributes can be ambiguous, but not in this case.)
Nope.
_Alignasreally _is_ special :-(struct _Alignas (16) blah { int a; };p.c:1:8: error: expected ‘{’ before ‘_Alignas’
struct_Alignas(16) blah {
Meh, yet another unnecessary C++ incompatibility. C does not support empty structs, so I assume they didn't see the field requirement as a burden.
Quoted 7 lines
One last thing I'm planning to add in
sys/rseq.hto cover acessing therseq_cspointers with both the UAPI headers and the glibc struct rseq declarations:/* The
rseq_cs_ptrmacro can be used to access the pointer to the current
rseq critical section descriptor. */
#ifdef__LP64__
Quoted 8 lines
# define rseq_cs_ptr(rseq) \ ((const struct rseq_cs *) (rseq)->rseq_cs.ptr) #else /* __LP64__ */ # define rseq_cs_ptr(rseq) \ ((const struct rseq_cs *) (rseq)->rseq_cs.ptr.ptr32) #endif /* __LP64__ */Does it make sense ?
Written this way, it's an aliasing violation. I don't think it's very useful.
Thanks,
Florian
Re: [PATCH glibc 1/3] glibc: Perform rseq registration at C startup and thread creation (v19)
Quoted 11 lines · ----- On May 26, 2020, at 10:57 AM, Florian Weimer [email protected] wrote:
* Mathieu Desnoyers:
Quoted 9 lines
Quoted 2 lines
Like the attribute, it needs to come right after the struct keyword, I think. (Trailing attributes can be ambiguous, but not in this case.)
Nope.
_Alignasreally _is_ special :-(struct _Alignas (16) blah { int a; };
Quoted 6 lines
Quoted 2 lines
p.c:1:8: error: expected ‘{’ before ‘_Alignas’
struct_Alignas(16) blah {Meh, yet another unnecessary C++ incompatibility. C does not support empty structs, so I assume they didn't see the field requirement as a burden.
Indeed, it's weird.
Quoted 15 lines
Quoted 15 lines
One last thing I'm planning to add in
sys/rseq.hto cover acessing therseq_cspointers with both the UAPI headers and the glibc struct rseq declarations:/* The
rseq_cs_ptrmacro can be used to access the pointer to the current
rseq critical section descriptor. */
#ifdef__LP64__# define rseq_cs_ptr(rseq) \ ((const struct rseq_cs *) (rseq)->rseq_cs.ptr) #else /* __LP64__ */ # define rseq_cs_ptr(rseq) \ ((const struct rseq_cs *) (rseq)->rseq_cs.ptr.ptr32) #endif /* __LP64__ */
Quoted 4 lines
Quoted 1 line
Does it make sense ?
Written this way, it's an aliasing violation. I don't think it's very useful.
OK, I'll just remove it.
Thanks,
Mathieu
Signature · 3 lines
Mathieu Desnoyers EfficiOS Inc. http://www.efficios.com
PATCH 2/3 glibc: sched_getcpu(): use rseq cpu_id TLS on Linux (v7)
In reply to a message not in the archiveWhen 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)
---
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.
Changes since v6:
- Remove use of __ASSUME_RSEQ.
--- sysdeps/unix/sysv/linux/sched_getcpu.c | 27 ++++++++++++++++++++++++-- 1 file changed, 25 insertions(+), 2 deletions(-)
sysdeps/unix/sysv/linux/sched_getcpu.c +25 -2
@@ -18,10 +18,15 @@#include <errno.h> #include <sched.h> #include <sysdep.h> +#include <atomic.h> #include <sysdep-vdso.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 +37,21 @@ sched_getcpu (void)#endif return r ==-1 ? r : cpu; } + +#include <sys/rseq.h> + +#ifdef 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
* Mathieu Desnoyers via Libc-alpha:
Quoted 6 lines
sysdeps/unix/sysv/linux/sched_getcpu.c +0
-0@@ -18,10 +18,15 @@#include <errno.h>
sysdeps/unix/sysv/linux/sched_getcpu.c line 26
Quoted 10 lines
sysdeps/unix/sysv/linux/sched_getcpu.c +4
-2#include <sched.h> #include <sysdep.h> +#include <atomic.h> #include <sysdep-vdso.h> -int -sched_getcpu (void ) +#ifdef HAVE_GETCPU_VSYSCALL +# define HAVE_VSYSCALL +#endif
I think the #ifdef is a result of an incorrect merge of commit d0def09ff6bbf1537beec305fdfe96a21174fb31 ("linux: Fix vDSO macros build with time64 interfaces") and it should be removed.
The commit subject should probably say “Linux: Use rseq in sched_getcpu if available”.
Thanks,
Florian