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 5/9 glibc: Perform rseq(2) registration at C startup and thread creation (v17)
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.
---
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.
--- NEWS | 12 +- 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 + .../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 | 25 +++ 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 | 186 ++++++++++++++++++ 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, 728 insertions(+), 5 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 +11 -1
@@ -9,7 +9,17 @@ Version 2.32 Major new features: - * New locale added: ckb_IQ (Kurdish/Sorani spoken in Iraq) +* New locale added: ckb_IQ (Kurdish/Sorani spoken in Iraq) + +* 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:
elf/libc_early_init.c +3 -0
@@ -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
@@ -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
@@ -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
@@ -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_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/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 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
@@ -177,6 +177,9 @@ libc { GLIBC_2.30 { getdents64; gettid; tgkill; } + GLIBC_2.32 { + __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
@@ -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_2.32 __rseq_abi T 0x20
sysdeps/unix/sysv/linux/alpha/libc.abilist +1 -0
@@ -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_2.32 __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
@@ -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_2.4 _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
@@ -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_2.4 _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
@@ -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
@@ -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_2.32 __rseq_abi T 0x20
sysdeps/unix/sysv/linux/hppa/libc.abilist +1 -0
@@ -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_2.32 __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
@@ -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_2.32 __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
@@ -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_2.32 __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/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_2.4 _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
@@ -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_2.32 __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
@@ -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_2.32 __rseq_abi T 0x20
sysdeps/unix/sysv/linux/microblaze/le/libc.abilist +1 -0
@@ -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_2.32 __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
@@ -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_2.32 __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
@@ -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_2.32 __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
@@ -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_2.32 __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
@@ -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_2.32 __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
@@ -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_2.32 __rseq_abi T 0x20
sysdeps/unix/sysv/linux/powerpc/bits/rseq.h new file +37 -0
@@ -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
@@ -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_2.32 __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
@@ -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_2.32 __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
@@ -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_2.32 __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
@@ -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_2.32 __rseq_abi T 0x20
sysdeps/unix/sysv/linux/riscv/rv64/libc.abilist +1 -0
@@ -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_2.32 __rseq_abi T 0x20
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 +25 -0
@@ -0,0 +1,25 @@ +/* Restartable Sequences exported symbols. Linux Implementation. +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/>. */ + +#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) 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
@@ -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_2.32 __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
@@ -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_2.32 __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
@@ -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_2.32 __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
@@ -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_2.32 __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
@@ -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_2.32 __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
@@ -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_2.32 __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 +186 -0
186 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) 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
@@ -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_2.32 __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
@@ -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_2.32 __rseq_abi T 0x20
Signature · 1 line
2.17.1
Re: [PATCH glibc 5/9] glibc: Perform rseq(2) registration at C startup and thread creation (v17)
* Mathieu Desnoyers via Libc-alpha:
NEWS Version 2.32 · line 22
Quoted 9 lines
NEWS +9
-0+* 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.
This should refer documentation in the glibc manual.
(It is currently a glibc project requirement to add documentation for new Linux interfaces, something that I do not necessarily agree with.)
Quoted 11 lines
Deprecated and removed features, and other changes affecting compatibility:
elf/libc_early_init.c +0
-0@@ -18,10 +18,13 @@#include <ctype.h> #include <libc-early-init.h>
elf/libc_early_init.c line 30
Quoted 10 lines
elf/libc_early_init.c +3
-0+#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 (); }
The cast to void should be removed (see below the comment about the return type.
Quoted 35 lines
manual/threads.texi +18
-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@@ -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.
manual/threads.texi Behaves like @code{pthread_timedjoin_np} except that the absolute time in · line 909
Quoted 10 lines
manual/threads.texi +10
-0+@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
This should say Linux, not GNU as the standards source, given that the interface is not added to the GNU ABI.
Is this sufficient documentation of this feature? What do others think about this?
misc/rseq-internal.h line 25
Quoted 12 lines
misc/rseq-internal.h new file +5
-0@@ -0,0 +1,33 @@ +static inline int +rseq_register_current_thread (void ) +{ +return -1 ; +}
Nothing checks the return value of this function as far as I can see, so it can return void.
misc/rseq-internal.h line 31
Quoted 5 lines
misc/rseq-internal.h new file +5
-0+static inline int +rseq_unregister_current_thread (void ) +{ +return -1 ; +}
This function is unused. This also applies to the full version. I believe we switched to implicit deregistration on thread exit, so I think you can just remove it.
Quoted 13 lines
nptl/pthread_create.c +1
-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>
nptl/pthread_create.c START_THREAD_DEFN · line 390
Quoted 6 lines
nptl/pthread_create.c +3
-0@@ -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 (); +
The cast can go.
Quoted 8 lines
nptl/pthread_create.c +1
-0#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
nptl/pthread_create.c START_THREAD_DEFN · line 595
Quoted 10 lines
nptl/pthread_create.c +7
-0+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 ();
Some of these lines are too long. Also two spaces after . at the end of sentences.
Quoted 13 lines
sysdeps/unix/sysv/linux/rseq-internal.h new file +7
-0+static inline int +rseq_register_current_thread (void ) +{ +int rc, ret =0 ; + +if (__rseq_abi.cpu_id ==RSEQ_CPU_ID_REGISTRATION_FAILED ) +return -1 ;
sysdeps/unix/sysv/linux/rseq-internal.h line 44
Quoted 10 lines
sysdeps/unix/sysv/linux/rseq-internal.h new file +10
-0+ rc =INTERNAL_SYSCALL_CALL (rseq, &__rseq_abi,sizeof (struct rseq ), +0 ,RSEQ_SIG ); +if (!rc) + goto end; +if (INTERNAL_SYSCALL_ERRNO (rc) !=EBUSY ) + __rseq_abi.cpu_id =RSEQ_CPU_ID_REGISTRATION_FAILED ; + ret =-1 ; +end: +return ret; +}
This does not seem to use INTERNAL_SYSCALL_CALL correctly. I think you need to use INTERNAL_SYSCALL_ERROR_P on the result to check for an error, and only then use INTERNAL_SYSCALL_ERRNO to extract the error code.
Quoted 5 lines
sysdeps/unix/sysv/linux/rseq-sym.c new file +0
-0
sysdeps/unix/sysv/linux/rseq-sym.c line 25
Quoted 9 lines
sysdeps/unix/sysv/linux/rseq-sym.c new file +8
-0+#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 , +};
{ should go onto its own line. I'd also add attribute_tls_model_ie, also it's implied by the declaration in the header.
sysdeps/unix/sysv/linux/sys/rseq.h
Quoted 6 lines
sysdeps/unix/sysv/linux/sys/rseq.h new file +0
-0@@ -0,0 +1,186 @@
I think there is some value in making this header compatible with inclusion from the assembler (including constants for the relevant struct offsets), but that can be a later change.
sysdeps/unix/sysv/linux/sys/rseq.h line 22
Quoted 5 lines
sysdeps/unix/sysv/linux/sys/rseq.h new file +5
-0+#ifndef _SYS_RSEQ_H +#define _SYS_RSEQ_H 1 + +/* Architecture-specific rseq signature. */ +#include <bits/rseq.h>
Maybe add a newline between the above and the following, to make clear the comment only applies to the first #include.
Quoted 54 lines
sysdeps/unix/sysv/linux/sys/rseq.h new file +54
-0+#include <stdint.h> +#include <sys/cdefs.h> + +#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 +/* 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. */ + +#include <asm/byteorder.h> + +enum rseq_cpu_id_state + { +RSEQ_CPU_ID_UNINITIALIZED =-1 , +RSEQ_CPU_ID_REGISTRATION_FAILED =-2 , + }; + +enum rseq_flags + { +RSEQ_FLAG_UNREGISTER = (1 <<0 ), + }; + +enum rseq_cs_flags_bit + { +RSEQ_CS_FLAG_NO_RESTART_ON_PREEMPT_BIT =0 , +RSEQ_CS_FLAG_NO_RESTART_ON_SIGNAL_BIT =1 , +RSEQ_CS_FLAG_NO_RESTART_ON_MIGRATE_BIT =2 , + }; + +enum rseq_cs_flags + { +RSEQ_CS_FLAG_NO_RESTART_ON_PREEMPT = + (1U <<RSEQ_CS_FLAG_NO_RESTART_ON_PREEMPT_BIT ), +RSEQ_CS_FLAG_NO_RESTART_ON_SIGNAL = + (1U <<RSEQ_CS_FLAG_NO_RESTART_ON_SIGNAL_BIT ), +RSEQ_CS_FLAG_NO_RESTART_ON_MIGRATE = + (1U <<RSEQ_CS_FLAG_NO_RESTART_ON_MIGRATE_BIT ), + }; + +/* struct rseq_cs is aligned on 4 * 8 bytes to ensure it is always +contained within a single cache-line. It is usually declared as +link-time constant data. */ +struct rseq_cs
sysdeps/unix/sysv/linux/sys/rseq.h line 86
Quoted 10 lines
sysdeps/unix/sysv/linux/sys/rseq.h new file +10
-0+ { +/* Version of this structure. */ +uint32_t 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; +} __attribute__((aligned(4 * sizeof(uint64_t))));
The comment is wrong. 32-byte alignment does not put struct rseq_cs on its own cache line on many (most?) CPUs. Not using the constant 32 looks like unnecessary obfuscation to me.
I still think we should avoid the alignment. The _ip fields should perhaps be _pc (IP is more or less specific to x86).
{ and } are not aligned. Please do not forget to add spaces before opening parentheses, and two spaces after the . and the end of sentences. The opening { should always be on its own line. (This also applies to the definition of struct rseq below.)
sysdeps/unix/sysv/linux/sys/rseq.h line 97
Quoted 11 lines
sysdeps/unix/sysv/linux/sys/rseq.h new file +11
-0+ +/* struct rseq is aligned on 4 * 8 bytes to ensure it is always +contained within a single cache-line. + +A single struct rseq per thread is allowed. */ +struct rseq + { +/* Restartable sequences cpu_id_start field. Updated by the +kernel. Read by user-space with single-copy atomicity +semantics. This field should only be read by the thread which +registered this data structure. Aligned on 32-bit. Always
What does “Aligned on 32-bit” mean in this context? Do you mean to reference 32-*byte* alignment here?
Quoted 22 lines
sysdeps/unix/sysv/linux/sys/rseq.h new file +22
-0+/* Restartable sequences rseq_cs field. + +Contains NULL when no critical section is active for the current +thread, or holds a pointer to the currently active struct rseq_cs. + +Updated by user-space, which sets the address of the currently +active rseq_cs at the beginning of assembly instruction sequence +block, and set to NULL by the kernel when it restarts an assembly +instruction sequence block, as well as when the kernel detects that +it is preempting or delivering a signal outside of the range +targeted by the rseq_cs. Also needs to be set to NULL by user-space +before reclaiming memory that contains the targeted struct rseq_cs. + +Read and set by the kernel. Set by user-space with single-copy +atomicity semantics. This field should only be updated by the +thread which registered this data structure. Aligned on 64-bit. */ + union { + uint64_t ptr64; +#ifdef __LP64__ + uint64_t ptr; +#else + struct {
sysdeps/unix/sysv/linux/sys/rseq.h line 149
Quoted 10 lines
sysdeps/unix/sysv/linux/sys/rseq.h new file +10
-0+#if (defined(__BYTE_ORDER) && (__BYTE_ORDER == __BIG_ENDIAN)) || defined(__BIG_ENDIAN) +uint32_t padding;/* Initialized to zero. */ +uint32_t ptr32; +#else /* LITTLE */ +uint32_t ptr32; +uint32_t padding;/* Initialized to zero. */ +#endif /* ENDIAN */ + } ptr; +#endif + } rseq_cs;
Are these conditionals correct for x32? Shouldn't there be a member of type const struct rseq_cs * somewhere?
sysdeps/unix/sysv/linux/x86/bits/rseq.h line 2
Quoted 8 lines
sysdeps/unix/sysv/linux/x86/bits/rseq.h new file +2
-0@@ -0,0 +1,30 @@ +/* Restartable Sequences Linux x86 architecture header. +Copyright (C) 2019-2020 Free Software Foundation, Inc.
Please make sure that none of the new files reference the 2019 year. It should be 2020, per GNU policy.
The patch needs some rebasing on top of current master.
Re: [PATCH glibc 5/9] glibc: Perform rseq(2) registration at C startup and thread creation (v17)
Quoted 12 lines · ----- On Apr 27, 2020, at 5:11 AM, Florian Weimer [email protected] wrote:
* Mathieu Desnoyers via Libc-alpha:
NEWS Version 2.32 · line 22
Quoted 9 lines
NEWS +9
-0+* 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.
Quoted 4 lines
This should refer documentation in the glibc manual.
(It is currently a glibc project requirement to add documentation for new Linux interfaces, something that I do not necessarily agree with.)
A related issue here is that editing of the rseq(2) man pages has been stalled since March 2019. I have been waiting for reply from Michael Kerrisk, and I suspect he might have be side-tracked by other projects. I just bumped that thread.
Ref.: https://lore.kernel.org/r/[email protected]
So as of today, "man 2 rseq" does not exist in the kernel man pages, so I suggest we remove that sentence. Would the following change be OK with you ?
- See 'man 2 rseq' for the details of the ABI shared between libc and the - kernel. + The GNU C Library manual has details on integration of Restartable + Sequences.
Quoted 14 lines
Quoted 14 lines
Deprecated and removed features, and other changes affecting compatibility:
elf/libc_early_init.c +1
-0@@ -18,10 +18,13 @@#include <ctype.h> #include <libc-early-init.h> +#include <rseq-internal.h>
Quoted 11 lines
elf/libc_early_init.c line 30
Quoted 8 lines
elf/libc_early_init.c +2
-0void __libc_early_init (void ) {/* Initialize ctype data. */ __ctype_init (); +/* Register rseq ABI to the kernel. */ + (void )rseq_register_current_thread (); }The cast to void should be removed (see below the comment about the return type.
OK
Quoted 42 lines
Quoted 42 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@@ -881,3 +883,27 @@ Behaves like @code{pthread_timedjoin_np} except that theabsolute time in
@cpthread_spin_unlock
@cpthread_testcancel
@cpthread_yieldmanual/threads.texi +19
-0+ +@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
Quoted 8 lines
manual/threads.texi Behaves like @code{pthread_timedjoin_np} except that the absolute time in · line 909
Quoted 5 lines
+@file{sys/rseq.h} which contains a signature. That signature is expected to be
manual/threads.texi +4
-0+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 deftypevrThis should say Linux, not GNU as the standards source, given that the interface is not added to the GNU ABI.
Good point, done.
Quoted 3 lines
Is this sufficient documentation of this feature? What do others think about this?
I'll leave this question to others.
Quoted 8 lines
Quoted 7 lines
misc/rseq-internal.h new file +0
-0@@ -0,0 +1,33 @@
Quoted 8 lines
misc/rseq-internal.h line 25
Quoted 5 lines
misc/rseq-internal.h new file +5
-0+static inline int +rseq_register_current_thread (void ) +{ +return -1 ; +}Nothing checks the return value of this function as far as I can see, so it can return void.
OK
Quoted 10 lines
misc/rseq-internal.h line 31
Quoted 6 lines
misc/rseq-internal.h new file +5
-0+static inline int +rseq_unregister_current_thread (void ) +{ +return -1 ; +}This function is unused. This also applies to the full version. I believe we switched to implicit deregistration on thread exit, so I think you can just remove it.
OK
Quoted 17 lines
Quoted 17 lines
nptl/pthread_create.c +1
-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 ();
Quoted 5 lines
nptl/pthread_create.c START_THREAD_DEFN · line 390
Quoted 3 lines
nptl/pthread_create.c +3
-0+/* Register rseq TLS to the kernel. */ + (void )rseq_register_current_thread (); +The cast can go.
OK
Quoted 17 lines
Quoted 17 lines
nptl/pthread_create.c +1
-0#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 implicitunregistration
+ performed by the kernel on thread teardown. This is not a problem because
thenptl/pthread_create.c +3
-0+ 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 TLSarea
+ memory. It is NOT sufficient to block signals because the kernel may write
Quoted 8 lines
nptl/pthread_create.c START_THREAD_DEFN · line 595
Quoted 5 lines
nptl/pthread_create.c +2
-0+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 ();Some of these lines are too long. Also two spaces after . at the end of sentences.
OK
Quoted 22 lines
Quoted 22 lines
sysdeps/unix/sysv/linux/rseq-internal.h new file +14
-0+static inline int +rseq_register_current_thread (void ) +{ +int rc, ret =0 ; + +if (__rseq_abi.cpu_id ==RSEQ_CPU_ID_REGISTRATION_FAILED ) +return -1 ; + rc =INTERNAL_SYSCALL_CALL (rseq, &__rseq_abi,sizeof (struct rseq ), +0 ,RSEQ_SIG ); +if (!rc) + goto end; +if (INTERNAL_SYSCALL_ERRNO (rc) !=EBUSY ) + __rseq_abi.cpu_id =RSEQ_CPU_ID_REGISTRATION_FAILED ; + ret =-1 ;
Quoted 8 lines
sysdeps/unix/sysv/linux/rseq-internal.h line 44
Quoted 3 lines
sysdeps/unix/sysv/linux/rseq-internal.h new file +3
-0+end: +return ret; +}This does not seem to use
INTERNAL_SYSCALL_CALLcorrectly. I think you need to useINTERNAL_SYSCALL_ERROR_Pon the result to check for an error, and only then useINTERNAL_SYSCALL_ERRNOto extract the error code.
OK
Quoted 8 lines
Quoted 7 lines
sysdeps/unix/sysv/linux/rseq-sym.c new file +0
-0
Quoted 10 lines
sysdeps/unix/sysv/linux/rseq-sym.c line 25
Quoted 8 lines
sysdeps/unix/sysv/linux/rseq-sym.c new file +8
-0+#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 , +};{ should go onto its own line.
OK
Quoted 2 lines
I'd also add
attribute_tls_model_ie,
also it's implied by the declaration in the header.
This contradicts feedback I received from Szabolcs Nagy in September 2019:
https://public-inbox.org/libc-alpha/[email protected]/
"note that libpthread.so is built with -ftls-model=initial-exec
(and if it wasn't then you'd want to put the attribute on the declaration in the internal header file, not on the definition, so the actual tls accesses generate the right code)"
In the context of his feedback, __rseq_abi was defined within nptl/pthread_create.c. It is now defined in sysdeps/unix/sysv/linux/rseq-sym.c, which is built into the csu which AFAIU ends up in libc.so. His comment still applies though, because libc.so is also built with -ftls-model=initial-exec.
So should I apply the "initial-exec" TLS model only to the __rseq_abi declaration, or is it preferred to apply it to both the declaration and the definition ?
Quoted 12 lines
sysdeps/unix/sysv/linux/sys/rseq.h
Quoted 8 lines
sysdeps/unix/sysv/linux/sys/rseq.h new file +0
-0@@ -0,0 +1,186 @@I think there is some value in making this header compatible with inclusion from the assembler (including constants for the relevant struct offsets), but that can be a later change.
Agreed. By "later", do you mean before merging the patch, between merge of the patch and next glibc release, or for a subsequent glibc release ?
Quoted 9 lines
sysdeps/unix/sysv/linux/sys/rseq.h line 22
Quoted 6 lines
sysdeps/unix/sysv/linux/sys/rseq.h new file +5
-0+#ifndef _SYS_RSEQ_H +#define _SYS_RSEQ_H 1 + +/* Architecture-specific rseq signature. */ +#include <bits/rseq.h> Maybe add a newline between the above and the following, to make clear the comment only applies to the first #include.
OK
Quoted 61 lines
Quoted 61 lines
sysdeps/unix/sysv/linux/sys/rseq.h new file +60
-0+#include <stdint.h> +#include <sys/cdefs.h> + +#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 +/* 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. */ + +#include <asm/byteorder.h> + +enum rseq_cpu_id_state + { + RSEQ_CPU_ID_UNINITIALIZED = -1, + RSEQ_CPU_ID_REGISTRATION_FAILED = -2, + }; + +enum rseq_flags + { + RSEQ_FLAG_UNREGISTER = (1 << 0), + }; + +enum rseq_cs_flags_bit + { + RSEQ_CS_FLAG_NO_RESTART_ON_PREEMPT_BIT = 0, + RSEQ_CS_FLAG_NO_RESTART_ON_SIGNAL_BIT = 1, + RSEQ_CS_FLAG_NO_RESTART_ON_MIGRATE_BIT = 2, + }; + +enum rseq_cs_flags + { + RSEQ_CS_FLAG_NO_RESTART_ON_PREEMPT = + (1U << RSEQ_CS_FLAG_NO_RESTART_ON_PREEMPT_BIT), + RSEQ_CS_FLAG_NO_RESTART_ON_SIGNAL = + (1U << RSEQ_CS_FLAG_NO_RESTART_ON_SIGNAL_BIT), + RSEQ_CS_FLAG_NO_RESTART_ON_MIGRATE = + (1U << RSEQ_CS_FLAG_NO_RESTART_ON_MIGRATE_BIT), + }; + +/* struct rseq_cs is aligned on 4 * 8 bytes to ensure it is always + contained within a single cache-line. It is usually declared as + link-time constant data. */ +struct rseq_cs + { + /* Version of this structure. */ + uint32_t version; + /* enum rseq_cs_flags. */ + uint32_t flags; + uint64_t start_ip;
Quoted 8 lines
sysdeps/unix/sysv/linux/sys/rseq.h line 86
Quoted 4 lines
sysdeps/unix/sysv/linux/sys/rseq.h new file +4
-0+/* Offset from start_ip. */ +uint64_t post_commit_offset; +uint64_t abort_ip; +} __attribute__((aligned(4 * sizeof(uint64_t))));The comment is wrong. 32-byte alignment does not put struct
rseq_cson its own cache line on many (most?) CPUs. Not using the constant 32 looks like unnecessary obfuscation to me.
There is a difference between "being contained within a single cache-line" and "being the only structure in a cache-line". The objective here is the former.
For instance, if we do not enforce any minimum alignment, the compiler could decide to align that structure on __alignof__(uint64_t) which happens to be 4 bytes on some architectures. This can cause this frequently accessed structure to be sitting across 2 cache-lines, thus requiring the CPU to load two cache lines rather than one very frequently.
I think what you have in mind is "being the only structure in a cache-line", which is useful to eliminate false-sharing. However, considering that this is a TLS variable, we don't care about false-sharing, because it is never meant to be updated concurrently by many threads or CPUs.
So I think my comment is correct.
I agree that the constant 32 may be clearer here. I will change to align(32).
Quoted 2 lines
I still think we should avoid the alignment. The
_ipfields should perhaps be_pc(IP is more or less specific to x86).
I am concerned that removing an alignment attribute which is exposed in a public Linux UAPI header can be an ABI breakage.
I am also concerned about changing field names for fields already exposed in a public Linux UAPI header, especially if the change is only for cosmetic reasons.
Quoted 5 lines
{ and } are not aligned. Please do not forget to add spaces before opening parentheses, and two spaces after the . and the end of sentences. The opening { should always be on its own line. (This also applies to the definition of struct rseq below.)
OK. Old coding style habits die hard ;)
Quoted 15 lines
sysdeps/unix/sysv/linux/sys/rseq.h line 97
Quoted 12 lines
sysdeps/unix/sysv/linux/sys/rseq.h new file +11
-0+ +/* struct rseq is aligned on 4 * 8 bytes to ensure it is always +contained within a single cache-line. + +A single struct rseq per thread is allowed. */ +struct rseq + { +/* Restartable sequences cpu_id_start field. Updated by the +kernel. Read by user-space with single-copy atomicity +semantics. This field should only be read by the thread which +registered this data structure. Aligned on 32-bit. Always What does “Aligned on 32-bit” mean in this context? Do you mean to reference 32-*byte* alignment here?
No. I really mean 32-bit (4-byte). Being aligned on 32-byte guarantees that this field is aligned at least on 4-byte. This is required by single-copy atomicity semantics.
Should I update this comment to state "Aligned on 4-byte" instead ?
Quoted 28 lines
Quoted 28 lines
sysdeps/unix/sysv/linux/sys/rseq.h new file +23
-0+/* Restartable sequences rseq_cs field. + +Contains NULL when no critical section is active for the current +thread, or holds a pointer to the currently active struct rseq_cs. + +Updated by user-space, which sets the address of the currently +active rseq_cs at the beginning of assembly instruction sequence +block, and set to NULL by the kernel when it restarts an assembly +instruction sequence block, as well as when the kernel detects that +it is preempting or delivering a signal outside of the range +targeted by the rseq_cs. Also needs to be set to NULL by user-space +before reclaiming memory that contains the targeted struct rseq_cs. + +Read and set by the kernel. Set by user-space with single-copy +atomicity semantics. This field should only be updated by the +thread which registered this data structure. Aligned on 64-bit. */ + union { + uint64_t ptr64; +#ifdef __LP64__ + uint64_t ptr; +#else + struct { +#if (defined(__BYTE_ORDER) && (__BYTE_ORDER == __BIG_ENDIAN)) ||
defined(__BIG_ENDIAN)sysdeps/unix/sysv/linux/sys/rseq.h new file +3
-0+uint32_t padding;/* Initialized to zero. */ +uint32_t ptr32; +#else /* LITTLE */
Quoted 8 lines
sysdeps/unix/sysv/linux/sys/rseq.h line 149
Quoted 6 lines
sysdeps/unix/sysv/linux/sys/rseq.h new file +6
-0+uint32_t ptr32; +uint32_t padding;/* Initialized to zero. */ +#endif /* ENDIAN */ + } ptr; +#endif + } rseq_cs;Are these conditionals correct for x32?
Let's see. With x86 gcc:
-m64: (__x86_64__ && __LP64__)
-m32: (__i386__)
-mx32: (__x86_64__ && __ILP32__)
So with "#ifdef __LP64__" we specifically target 64-bit pointers. The rest falls into the "else" case, which expects 32-bit pointers. Considering that x32 has 32-bit pointers, I don't see any issue here.
Quoted 1 line
Shouldn't there be a member of type const struct
rseq_cs* somewhere?
Having pointers within structures in kernel UAPI headers is frowned upon. Indeed here having it in the union could possibly make some use-cases easier in user-space, so I'm open to it. It basically depends on how much we want the Linux UAPI header and the glibc header to stay in sync, and if other kernel maintainers are open to this addition.
We don't mind that user-space uses that pointer, but we never want the kernel to touch that pointer rather than the 32/64-bit-aware fields. One possibility would be to do:
union
{
uint64_t ptr64;
#ifdef __LP64__
uint64_t ptr;
#else
struct
{#if (defined (__BYTE_ORDER) && (__BYTE_ORDER == __BIG_ENDIAN)) || defined (__BIG_ENDIAN)
uint32_t padding; /* Initialized to zero. */
uint32_t ptr32;
#else /* LITTLE */
uint32_t ptr32;
uint32_t padding; /* Initialized to zero. */
#endif /* ENDIAN */
} ptr;#endif
#ifndef __KERNEL__
const struct rseq_cs *uptr;
#endif
} rseq_cs;in the union, so only user-space can see that field. Thoughts ?
Quoted 13 lines
sysdeps/unix/sysv/linux/x86/bits/rseq.h line 2
Quoted 10 lines
sysdeps/unix/sysv/linux/x86/bits/rseq.h new file +2
-0@@ -0,0 +1,30 @@ +/* Restartable Sequences Linux x86 architecture header. +Copyright (C) 2019-2020 Free Software Foundation, Inc. Please make sure that none of the new files reference the 2019 year. It should be 2020, per GNU policy.
OK
Quoted 2 lines
The patch needs some rebasing on top of current master.
Done.
Thanks for the review!
Mathieu
Signature · 3 lines
Mathieu Desnoyers EfficiOS Inc. http://www.efficios.com
Re: [PATCH glibc 5/9] glibc: Perform rseq(2) registration at C startup and thread creation (v17)
* Mathieu Desnoyers:
Quoted 26 lines
Quoted 10 lines
sysdeps/unix/sysv/linux/rseq-sym.c line 25
Quoted 8 lines
sysdeps/unix/sysv/linux/rseq-sym.c new file +8
-0+#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 , +};{ should go onto its own line.
OK
Quoted 2 lines
I'd also add
attribute_tls_model_ie,
also it's implied by the declaration in the header.This contradicts feedback I received from Szabolcs Nagy in September 2019:
https://public-inbox.org/libc-alpha/[email protected]/
"note that libpthread.so is built with -ftls-model=initial-exec
(and if it wasn't then you'd want to put the attribute on the declaration in the internal header file, not on the definition, so the actual tls accesses generate the right code)"
Quoted 8 lines
In the context of his feedback,
__rseq_abiwas defined withinnptl/pthread_create.c. It is now defined insysdeps/unix/sysv/linux/rseq-sym.c, which is built into the csu which AFAIU ends up in libc.so. His comment still applies though, because libc.so is also built with -ftls-model=initial-exec.So should I apply the "initial-exec" TLS model only to the
__rseq_abideclaration, or is it preferred to apply it to both the declaration and the definition ?
I do not have a strong preference here. Technically, the declaration in the header file should be enough.
Quoted 12 lines
Quoted 11 lines
sysdeps/unix/sysv/linux/sys/rseq.h
Quoted 7 lines
sysdeps/unix/sysv/linux/sys/rseq.h new file +0
-0@@ -0,0 +1,186 @@I think there is some value in making this header compatible with inclusion from the assembler (including constants for the relevant struct offsets), but that can be a later change.
Quoted 3 lines
Agreed. By "later", do you mean before merging the patch, between merge of the patch and next glibc release, or for a subsequent glibc release ?
It can happen some time after merging the patch, preferably for this release. But I don't think it's release-critical.
Quoted 19 lines
Quoted 10 lines
Quoted 10 lines
sysdeps/unix/sysv/linux/sys/rseq.h new file +10
-0+/* struct rseq_cs is aligned on 4 * 8 bytes to ensure it is always +contained within a single cache-line. It is usually declared as +link-time constant data. */ +struct rseq_cs + { +/* Version of this structure. */ +uint32_t version; +/* enum rseq_cs_flags. */ +uint32_t flags; +uint64_t start_ip;Quoted 8 lines
sysdeps/unix/sysv/linux/sys/rseq.h line 86
Quoted 4 lines
sysdeps/unix/sysv/linux/sys/rseq.h new file +4
-0+/* Offset from start_ip. */ +uint64_t post_commit_offset; +uint64_t abort_ip; +} __attribute__((aligned(4 * sizeof(uint64_t))));The comment is wrong. 32-byte alignment does not put struct
rseq_cson its own cache line on many (most?) CPUs. Not using the constant 32 looks like unnecessary obfuscation to me.
Quoted 3 lines
There is a difference between "being contained within a single cache-line" and "being the only structure in a cache-line". The objective here is the former.
Fair enough.
Quoted 1 line
I agree that the constant 32 may be clearer here. I will change to align(32).
With a space, please. 8-)
Quoted 14 lines
Quoted 5 lines
Quoted 5 lines
sysdeps/unix/sysv/linux/sys/rseq.h new file +5
-0+/* struct rseq is aligned on 4 * 8 bytes to ensure it is always +contained within a single cache-line. + +A single struct rseq per thread is allowed. */ +struct rseq Quoted 8 lines
sysdeps/unix/sysv/linux/sys/rseq.h line 97
Quoted 5 lines
sysdeps/unix/sysv/linux/sys/rseq.h new file +5
-0+ { +/* Restartable sequences cpu_id_start field. Updated by the +kernel. Read by user-space with single-copy atomicity +semantics. This field should only be read by the thread which +registered this data structure. Aligned on 32-bit. Always What does “Aligned on 32-bit” mean in this context? Do you mean to reference 32-*byte* alignment here?
Quoted 5 lines
No. I really mean 32-bit (4-byte). Being aligned on 32-byte guarantees that this field is aligned at least on 4-byte. This is required by single-copy atomicity semantics.
Should I update this comment to state "Aligned on 4-byte" instead ?
I think this is implied by all Linux ABIs. And the explicit alignment specification for struct rseq makes the alignment 32 bytes.
Quoted 42 lines
Quoted 27 lines
Quoted 27 lines
sysdeps/unix/sysv/linux/sys/rseq.h new file +23
-0+/* Restartable sequences rseq_cs field. + +Contains NULL when no critical section is active for the current +thread, or holds a pointer to the currently active struct rseq_cs. + +Updated by user-space, which sets the address of the currently +active rseq_cs at the beginning of assembly instruction sequence +block, and set to NULL by the kernel when it restarts an assembly +instruction sequence block, as well as when the kernel detects that +it is preempting or delivering a signal outside of the range +targeted by the rseq_cs. Also needs to be set to NULL by user-space +before reclaiming memory that contains the targeted struct rseq_cs. + +Read and set by the kernel. Set by user-space with single-copy +atomicity semantics. This field should only be updated by the +thread which registered this data structure. Aligned on 64-bit. */ + union { + uint64_t ptr64; +#ifdef __LP64__ + uint64_t ptr; +#else + struct { +#if (defined(__BYTE_ORDER) && (__BYTE_ORDER == __BIG_ENDIAN)) ||
defined(__BIG_ENDIAN)sysdeps/unix/sysv/linux/sys/rseq.h new file +3
-0+uint32_t padding;/* Initialized to zero. */ +uint32_t ptr32; +#else /* LITTLE */ Quoted 8 lines
sysdeps/unix/sysv/linux/sys/rseq.h line 149
Quoted 6 lines
sysdeps/unix/sysv/linux/sys/rseq.h new file +6
-0+uint32_t ptr32; +uint32_t padding;/* Initialized to zero. */ +#endif /* ENDIAN */ + } ptr; +#endif + } rseq_cs;Are these conditionals correct for x32?
Let's see. With x86 gcc:
(quoted diff) +0
-3-m64: (__x86_64__ && __LP64__) -m32: (__i386__) -mx32: (__x86_64__ && __ILP32__)
Quoted 3 lines
So with "#ifdef
__LP64__" we specifically target 64-bit pointers. The rest falls into the "else" case, which expects 32-bit pointers. Considering that x32 has 32-bit pointers, I don't see any issue here.
Does the kernel have a separate 32-bit entry point for rseq on x32? If not, it will expect the 64-bit struct layout.
Quoted 20 lines
We don't mind that user-space uses that pointer, but we never want the kernel to touch that pointer rather than the 32/64-bit-aware fields. One possibility would be to do:
union { uint64_t ptr64; #ifdef __LP64__ uint64_t ptr; #else struct {#if (defined (
__BYTE_ORDER) && (__BYTE_ORDER==__BIG_ENDIAN)) || defined (__BIG_ENDIAN)
uint32_tpadding; /* Initialized to zero. */
uint32_tptr32;
#else /* LITTLE */
uint32_tptr32;
uint32_tpadding; /* Initialized to zero. */#endif /* ENDIAN */ } ptr;
Quoted 8 lines
#endif
#ifndef __KERNEL__ const struct rseq_cs *uptr; #endif } rseq_cs;in the union, so only user-space can see that field. Thoughts ?
I think this depends on where the x32 question lands.
Re: [PATCH glibc 5/9] glibc: Perform rseq(2) registration at C startup and thread creation (v17)
Quoted 34 lines · ----- On Apr 27, 2020, at 12:54 PM, Florian Weimer [email protected] wrote:
* Mathieu Desnoyers:
Quoted 32 lines
Quoted 10 lines
sysdeps/unix/sysv/linux/rseq-sym.c line 25
Quoted 8 lines
sysdeps/unix/sysv/linux/rseq-sym.c new file +8
-0+#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 , +};{ should go onto its own line.
OK
Quoted 2 lines
I'd also add
attribute_tls_model_ie,
also it's implied by the declaration in the header.This contradicts feedback I received from Szabolcs Nagy in September 2019:
https://public-inbox.org/libc-alpha/[email protected]/
"note that libpthread.so is built with -ftls-model=initial-exec
(and if it wasn't then you'd want to put the attribute on the declaration in the internal header file, not on the definition, so the actual tls accesses generate the right code)"
In the context of his feedback,
__rseq_abiwas defined withinnptl/pthread_create.c.
It is now defined insysdeps/unix/sysv/linux/rseq-sym.c, which is built into the csu which AFAIU ends up in libc.so. His comment still applies though, because libc.so is also built with -ftls-model=initial-exec.
Quoted 6 lines
Quoted 3 lines
So should I apply the "initial-exec" TLS model only to the
__rseq_abideclaration, or is it preferred to apply it to both the declaration and the definition ?I do not have a strong preference here. Technically, the declaration in the header file should be enough.
OK, so I'll just keep the attribute on the declaration in the header.
Quoted 13 lines
Quoted 13 lines
Quoted 12 lines
sysdeps/unix/sysv/linux/sys/rseq.h
Quoted 8 lines
sysdeps/unix/sysv/linux/sys/rseq.h new file +0
-0@@ -0,0 +1,186 @@I think there is some value in making this header compatible with inclusion from the assembler (including constants for the relevant struct offsets), but that can be a later change.
Quoted 6 lines
Quoted 3 lines
Agreed. By "later", do you mean before merging the patch, between merge of the patch and next glibc release, or for a subsequent glibc release ?
It can happen some time after merging the patch, preferably for this release. But I don't think it's release-critical.
OK
Quoted 15 lines
Quoted 15 lines
Quoted 14 lines
sysdeps/unix/sysv/linux/sys/rseq.h line 97
Quoted 11 lines
sysdeps/unix/sysv/linux/sys/rseq.h new file +10
-0+/* struct rseq is aligned on 4 * 8 bytes to ensure it is always +contained within a single cache-line. + +A single struct rseq per thread is allowed. */ +struct rseq + { +/* Restartable sequences cpu_id_start field. Updated by the +kernel. Read by user-space with single-copy atomicity +semantics. This field should only be read by the thread which +registered this data structure. Aligned on 32-bit. Always What does “Aligned on 32-bit” mean in this context? Do you mean to reference 32-*byte* alignment here?
Quoted 8 lines
Quoted 5 lines
No. I really mean 32-bit (4-byte). Being aligned on 32-byte guarantees that this field is aligned at least on 4-byte. This is required by single-copy atomicity semantics.
Should I update this comment to state "Aligned on 4-byte" instead ?
I think this is implied by all Linux ABIs. And the explicit alignment specification for struct rseq makes the alignment 32 bytes.
Unless a structure ends up being packed, which is of course not the case here.
I would prefer to keep the comment about 32-bit alignment requirement on the specific fields, because the motivation for alignment requirement is much more strict for fields (correctness) than the motivation for alignment of the structure (performance).
Quoted 43 lines
Quoted 43 lines
Quoted 28 lines
Quoted 28 lines
sysdeps/unix/sysv/linux/sys/rseq.h new file +23
-0+/* Restartable sequences rseq_cs field. + +Contains NULL when no critical section is active for the current +thread, or holds a pointer to the currently active struct rseq_cs. + +Updated by user-space, which sets the address of the currently +active rseq_cs at the beginning of assembly instruction sequence +block, and set to NULL by the kernel when it restarts an assembly +instruction sequence block, as well as when the kernel detects that +it is preempting or delivering a signal outside of the range +targeted by the rseq_cs. Also needs to be set to NULL by user-space +before reclaiming memory that contains the targeted struct rseq_cs. + +Read and set by the kernel. Set by user-space with single-copy +atomicity semantics. This field should only be updated by the +thread which registered this data structure. Aligned on 64-bit. */ + union { + uint64_t ptr64; +#ifdef __LP64__ + uint64_t ptr; +#else + struct { +#if (defined(__BYTE_ORDER) && (__BYTE_ORDER == __BIG_ENDIAN)) ||
defined(__BIG_ENDIAN)sysdeps/unix/sysv/linux/sys/rseq.h new file +3
-0+uint32_t padding;/* Initialized to zero. */ +uint32_t ptr32; +#else /* LITTLE */ Quoted 8 lines
sysdeps/unix/sysv/linux/sys/rseq.h line 149
Quoted 6 lines
sysdeps/unix/sysv/linux/sys/rseq.h new file +6
-0+uint32_t ptr32; +uint32_t padding;/* Initialized to zero. */ +#endif /* ENDIAN */ + } ptr; +#endif + } rseq_cs;Are these conditionals correct for x32?
Let's see. With x86 gcc:
(quoted diff) +0
-3-m64: (__x86_64__ && __LP64__) -m32: (__i386__) -mx32: (__x86_64__ && __ILP32__)
Quoted 6 lines
Quoted 3 lines
So with "#ifdef
__LP64__" we specifically target 64-bit pointers. The rest falls into the "else" case, which expects 32-bit pointers. Considering that x32 has 32-bit pointers, I don't see any issue here.Does the kernel have a separate 32-bit entry point for rseq on x32? If not, it will expect the 64-bit struct layout.
No, there is a single entry point into rseq covering all of 32-bit, 64-bit and x32. We achieve this by ensuring the layout of the linux/rseq.h structures uses the union representation for pointers. Therefore, the kernel does not care whether it reads a pointer from a 32-bit or 64-bit process. This is becoming the preferred way to design Linux kernel ABIs nowadays.
Quoted 24 lines
Quoted 24 lines
We don't mind that user-space uses that pointer, but we never want the kernel to touch that pointer rather than the 32/64-bit-aware fields. One possibility would be to do:
union { uint64_t ptr64; #ifdef __LP64__ uint64_t ptr; #else struct {#if (defined (
__BYTE_ORDER) && (__BYTE_ORDER==__BIG_ENDIAN)) || defined (__BIG_ENDIAN)
uint32_tpadding; /* Initialized to zero. */uint32_tptr32;
#else /* LITTLE */
uint32_tptr32;
uint32_tpadding; /* Initialized to zero. */#endif /* ENDIAN */ } ptr;#endif
Quoted 8 lines
Quoted 6 lines
#ifndef __KERNEL__ const struct rseq_cs *uptr; #endif } rseq_cs;in the union, so only user-space can see that field. Thoughts ?
I think this depends on where the x32 question lands.
x32 should not be an issue as explained above, so I'm very open to add this "uptr" for user-space only.
Thanks,
Mathieu
Signature · 3 lines
Mathieu Desnoyers EfficiOS Inc. http://www.efficios.com
Re: [PATCH glibc 5/9] glibc: Perform rseq(2) registration at C startup and thread creation (v17)
----- On Apr 27, 2020, at 1:26 PM, Mathieu Desnoyers [email protected] wrote:
[...]
Quoted 28 lines
Quoted 28 lines
Quoted 28 lines
We don't mind that user-space uses that pointer, but we never want the kernel to touch that pointer rather than the 32/64-bit-aware fields. One possibility would be to do:
union { uint64_t ptr64; #ifdef __LP64__ uint64_t ptr; #else struct {#if (defined (
__BYTE_ORDER) && (__BYTE_ORDER==__BIG_ENDIAN)) || defined (__BIG_ENDIAN)
uint32_tpadding; /* Initialized to zero. */uint32_tptr32;
#else /* LITTLE */
uint32_tptr32;
uint32_tpadding; /* Initialized to zero. */#endif /* ENDIAN */ } ptr;#endif
#ifndef __KERNEL__ const struct rseq_cs *uptr; #endif } rseq_cs;
Quoted 6 lines
Quoted 3 lines
Quoted 1 line
in the union, so only user-space can see that field. Thoughts ?
I think this depends on where the x32 question lands.
x32 should not be an issue as explained above, so I'm very open to add this "uptr" for user-space only.
Actually, the snippet above is broken on 32-bit. It needs to be:
union
{
uint64_t ptr64;
#ifdef __LP64__
uint64_t ptr;
# ifndef __KERNEL__
const struct rseq_cs *uptr;
# endif
#else
struct
{
#if (defined (__BYTE_ORDER) && (__BYTE_ORDER == __BIG_ENDIAN)) || defined (__BIG_ENDIAN)
uint32_t padding; /* Initialized to zero. */
uint32_t ptr32;
#else /* LITTLE */
uint32_t ptr32;
uint32_t padding; /* Initialized to zero. */
#endif /* ENDIAN */
} ptr;
# ifndef __KERNEL__
struct
{
# if (defined (__BYTE_ORDER) && (__BYTE_ORDER == __BIG_ENDIAN)) || defined (__BIG_ENDIAN)
uint32_t padding; /* Initialized to zero. */
const struct rseq_cs *uptr32;
# else /* LITTLE */
const struct rseq_cs *uptr32;
uint32_t padding; /* Initialized to zero. */
# endif /* ENDIAN */
} uptr;
# endif
#endif
} rseq_cs;I'll leave this out of the patchset for now as we'd need more feedback on its usefulness.
Thanks,
Mathieu
Signature · 3 lines
Mathieu Desnoyers EfficiOS Inc. http://www.efficios.com
Re: [PATCH glibc 5/9] glibc: Perform rseq(2) registration at C startup and thread creation (v17)
In reply to Mathieu Desnoyers* Mathieu Desnoyers:
Quoted 23 lines
Quoted 14 lines
Quoted 14 lines
Quoted 13 lines
sysdeps/unix/sysv/linux/sys/rseq.h line 97
Quoted 10 lines
sysdeps/unix/sysv/linux/sys/rseq.h new file +10
-0+/* struct rseq is aligned on 4 * 8 bytes to ensure it is always +contained within a single cache-line. + +A single struct rseq per thread is allowed. */ +struct rseq + { +/* Restartable sequences cpu_id_start field. Updated by the +kernel. Read by user-space with single-copy atomicity +semantics. This field should only be read by the thread which +registered this data structure. Aligned on 32-bit. Always What does “Aligned on 32-bit” mean in this context? Do you mean to reference 32-*byte* alignment here?
Quoted 8 lines
Quoted 5 lines
No. I really mean 32-bit (4-byte). Being aligned on 32-byte guarantees that this field is aligned at least on 4-byte. This is required by single-copy atomicity semantics.
Should I update this comment to state "Aligned on 4-byte" instead ?
I think this is implied by all Linux ABIs. And the explicit alignment specification for struct rseq makes the alignment 32 bytes.
Quoted 7 lines
Unless a structure ends up being packed, which is of course not the case here.
I would prefer to keep the comment about 32-bit alignment requirement on the specific fields, because the motivation for alignment requirement is much more strict for fields (correctness) than the motivation for alignment of the structure (performance).
But the correctness is already enforced by the compiler, so I fail to see point of mentioning this in the comment.
Anyway, I don't want to make a big deal of it. Please leave it in if you think it is ehlpful.
Quoted 2 lines
x32 should not be an issue as explained above, so I'm very open to add this "uptr" for user-space only.
Okay, then please use anonymous unions and structs as necessary, to ensure that the uptr field can be reached on all platforms in the same way.
Re: [PATCH glibc 5/9] glibc: Perform rseq(2) registration at C startup and thread creation (v17)
Quoted 33 lines · ----- On Apr 28, 2020, at 8:02 AM, Florian Weimer [email protected] wrote:
* Mathieu Desnoyers:
Quoted 23 lines
Quoted 14 lines
Quoted 14 lines
Quoted 13 lines
sysdeps/unix/sysv/linux/sys/rseq.h line 97
Quoted 10 lines
sysdeps/unix/sysv/linux/sys/rseq.h new file +10
-0+/* struct rseq is aligned on 4 * 8 bytes to ensure it is always +contained within a single cache-line. + +A single struct rseq per thread is allowed. */ +struct rseq + { +/* Restartable sequences cpu_id_start field. Updated by the +kernel. Read by user-space with single-copy atomicity +semantics. This field should only be read by the thread which +registered this data structure. Aligned on 32-bit. Always What does “Aligned on 32-bit” mean in this context? Do you mean to reference 32-*byte* alignment here?
Quoted 8 lines
Quoted 5 lines
No. I really mean 32-bit (4-byte). Being aligned on 32-byte guarantees that this field is aligned at least on 4-byte. This is required by single-copy atomicity semantics.
Should I update this comment to state "Aligned on 4-byte" instead ?
I think this is implied by all Linux ABIs. And the explicit alignment specification for struct rseq makes the alignment 32 bytes.
Quoted 7 lines
Unless a structure ends up being packed, which is of course not the case here.
I would prefer to keep the comment about 32-bit alignment requirement on the specific fields, because the motivation for alignment requirement is much more strict for fields (correctness) than the motivation for alignment of the structure (performance).
Quoted 5 lines
But the correctness is already enforced by the compiler, so I fail to see point of mentioning this in the comment.
Anyway, I don't want to make a big deal of it. Please leave it in if you think it is ehlpful.
I would prefer to leave it in, just to make the requirements plain clear in case those structures are allocated on the heap (for instance).
Quoted 7 lines
Quoted 3 lines
x32 should not be an issue as explained above, so I'm very open to add this "uptr" for user-space only.
Okay, then please use anonymous unions and structs as necessary, to ensure that the uptr field can be reached on all platforms in the same way.
OK, will do!
One issue I'm currently facing when running "make check": because nptl/tst-rseq-nptl.c uses pthread_cancel(), I run into an Abort with:
libgcc_s.so.1 must be installed for pthread_cancel to work Didn't expect signal from child: got `Aborted'
So far I've tested the rest of that file with a patch on top which disables the use of pthread_cancel (), but I'd really like to give it a full coverage before sending this out. In https://sourceware.org/glibc/wiki/Testing/Builds there is a section about "Building glibc with intent to install" which describes that libgcc must be copied manually. My use-case is that I just want to run "make check" in the build directory and make sure it finds the libgcc it needs to succeed using pthread_cancel (). How can I achieve this ?
Thanks,
Mathieu
Signature · 3 lines
Mathieu Desnoyers EfficiOS Inc. http://www.efficios.com
Re: [PATCH glibc 5/9] glibc: Perform rseq(2) registration at C startup and thread creation (v17)
* Mathieu Desnoyers:
Quoted 6 lines
One issue I'm currently facing when running "make check": because
nptl/tst-rseq-nptl.cusespthread_cancel(), I run into an Abort with:libgcc_s.so.1 must be installed for
pthread_cancelto work Didn't expect signal from child: got `Aborted'
This is really unusual. Is the affected test statically linked?
Re: [PATCH glibc 5/9] glibc: Perform rseq(2) registration at C startup and thread creation (v17)
Quoted 10 lines · ----- On Apr 28, 2020, at 8:35 AM, Florian Weimer [email protected] wrote:
* Mathieu Desnoyers:
Quoted 6 lines
One issue I'm currently facing when running "make check": because
nptl/tst-rseq-nptl.cusespthread_cancel(), I run into an Abort with:libgcc_s.so.1 must be installed for
pthread_cancelto work Didn't expect signal from child: got `Aborted'This is really unusual. Is the affected test statically linked?
I built glibc without specifying anything particular, and ran "make check". It indeed seems to be dynamically linked to libc:
ldd tst-rseq-nptl
./tst-rseq-nptl: /lib/x86_64-linux-gnu/libc.so.6: version `GLIBC_2.32' not found (required by ./tst-rseq-nptl) linux-vdso.so.1 (0x00007ffd3a2f3000)
libpthread.so.0 => /lib/x86_64-linux-gnu/libpthread.so.0 (0x00007f0527560000)
libc.so.6 => /lib/x86_64-linux-gnu/libc.so.6 (0x00007f052716f000) /home/efficios/glibc-test5/lib/ld-linux-x86-64.so.2 => /lib64/ld-linux-x86-64.so.2 (0x00007f0527986000)
After make check I have:
cat tst-rseq-nptl.test-result
FAIL: nptl/tst-rseq-nptl
original exit status 134
And if I run
./tst-rseq-nptl
Then I get
libgcc_s.so.1 must be installed for pthread_cancel to work Didn't expect signal from child: got `Aborted' libgcc_s.so.1 must be installed for pthread_cancel to work Aborted (core dumped)
Same result if I do ./testrun.sh nptl/tst-rseq-nptl
Thanks,
Mathieu
Signature · 3 lines
Mathieu Desnoyers EfficiOS Inc. http://www.efficios.com
Re: [PATCH glibc 5/9] glibc: Perform rseq(2) registration at C startup and thread creation (v17)
* Mathieu Desnoyers:
Quoted 16 lines
Quoted 10 lines · ----- On Apr 28, 2020, at 8:35 AM, Florian Weimer [email protected] wrote:
* Mathieu Desnoyers:
Quoted 6 lines
One issue I'm currently facing when running "make check": because
nptl/tst-rseq-nptl.cusespthread_cancel(), I run into an Abort with:libgcc_s.so.1 must be installed for
pthread_cancelto work Didn't expect signal from child: got `Aborted'This is really unusual. Is the affected test statically linked?
I built glibc without specifying anything particular, and ran "make check". It indeed seems to be dynamically linked to libc:
Quoted 6 lines
ldd tst-rseq-nptl
./tst-rseq-nptl:/lib/x86_64-linux-gnu/libc.so.6: version `GLIBC_2.32' not found (required by ./tst-rseq-nptl) linux-vdso.so.1 (0x00007ffd3a2f3000)
libpthread.so.0 =>/lib/x86_64-linux-gnu/libpthread.so.0(0x00007f0527560000)
libc.so.6 =>/lib/x86_64-linux-gnu/libc.so.6(0x00007f052716f000)/home/efficios/glibc-test5/lib/ld-linux-x86-64.so.2=>/lib64/ld-linux-x86-64.so.2(0x00007f0527986000)
That's expected if the installed glibc is older than the built glibc.
Quoted 5 lines
After make check I have:
cat tst-rseq-nptl.test-result
FAIL: nptl/tst-rseq-nptl
original exit status 134
What's in the tst-rseq-nptl.out file?
Quoted 10 lines
And if I run
./tst-rseq-nptl
Then I get
libgcc_s.so.1 must be installed for
pthread_cancelto work Didn't expect signal from child: got `Aborted' libgcc_s.so.1 must be installed forpthread_cancelto work Aborted (core dumped)
I'm puzzled why you don't get a GLIBC_2.32 version error in this case. Do you build with --enable-hardcoded-path-in-tests?
Quoted 1 line
Same result if I do ./testrun.sh nptl/tst-rseq-nptl
That one definitely should work.
I expect you might see this if libgcc_s.so.1 is installed into a multiarch subdirectory that upstream glibc does not search. (The Debian patches are unfortunately not upstream.)
I think on my system, the built glibc can find the system libgcc_s via /etc/ld.so.cache, so I haven't seen this issue yet.
Re: [PATCH glibc 5/9] glibc: Perform rseq(2) registration at C startup and thread creation (v17)
Quoted 29 lines · ----- On Apr 28, 2020, at 8:54 AM, Florian Weimer [email protected] wrote:
* Mathieu Desnoyers:
Quoted 16 lines
Quoted 10 lines · ----- On Apr 28, 2020, at 8:35 AM, Florian Weimer [email protected] wrote:
* Mathieu Desnoyers:
Quoted 6 lines
One issue I'm currently facing when running "make check": because
nptl/tst-rseq-nptl.cusespthread_cancel(), I run into an Abort with:libgcc_s.so.1 must be installed for
pthread_cancelto work Didn't expect signal from child: got `Aborted'This is really unusual. Is the affected test statically linked?
I built glibc without specifying anything particular, and ran "make check". It indeed seems to be dynamically linked to libc:
Quoted 8 lines
ldd tst-rseq-nptl
./tst-rseq-nptl:/lib/x86_64-linux-gnu/libc.so.6: version `GLIBC_2.32' not found (required by ./tst-rseq-nptl)
linux-vdso.so.1 (0x00007ffd3a2f3000)
libpthread.so.0 =>/lib/x86_64-linux-gnu/libpthread.so.0(0x00007f0527560000) libc.so.6 =>/lib/x86_64-linux-gnu/libc.so.6(0x00007f052716f000)/home/efficios/glibc-test5/lib/ld-linux-x86-64.so.2=>/lib64/ld-linux-x86-64.so.2(0x00007f0527986000)That's expected if the installed glibc is older than the built glibc.
Quoted 7 lines
Quoted 5 lines
After make check I have:
cat tst-rseq-nptl.test-result
FAIL: nptl/tst-rseq-nptl
original exit status 134What's in the tst-rseq-nptl.out file?
It contains:
Didn't expect signal from child: got `Aborted'
Quoted 6 lines
Quoted 6 lines
And if I run
./tst-rseq-nptl
Then I get
Quoted 7 lines
Quoted 4 lines
libgcc_s.so.1 must be installed for
pthread_cancelto work Didn't expect signal from child: got `Aborted' libgcc_s.so.1 must be installed forpthread_cancelto work Aborted (core dumped)I'm puzzled why you don't get a GLIBC_2.32 version error in this case. Do you build with --enable-hardcoded-path-in-tests?
Just tried with this, and it fails the same way. However, the output of ldd nptl/tst-rseq-nptl changes:
linux-vdso.so.1 (0x00007ffc235c9000)
libpthread.so.0 => /home/efficios/git/glibc-build/nptl/libpthread.so.0 (0x00007fd308278000) libc.so.6 => /home/efficios/git/glibc-build/libc.so.6 (0x00007fd307ebc000) /home/efficios/git/glibc-build/elf/ld.so => /lib64/ld-linux-x86-64.so.2 (0x00007fd30869e000)
Quoted 7 lines
Quoted 1 line
Same result if I do ./testrun.sh nptl/tst-rseq-nptl
That one definitely should work.
I expect you might see this if libgcc_s.so.1 is installed into a multiarch subdirectory that upstream glibc does not search. (The Debian patches are unfortunately not upstream.)
My test environment is a Ubuntu 18.04.1 LTS.
Quoted 3 lines
I think on my system, the built glibc can find the system
libgcc_svia/etc/ld.so.cache, so I haven't seen this issue yet.
On my system, libgcc_s is provided here:
/lib/x86_64-linux-gnu/libgcc_s.so.1
by this package:
Package: libgcc1
Architecture: amd64
Version: 1:8.4.0-1ubuntu1~18.04
Thanks,
Mathieu
Signature · 3 lines
Mathieu Desnoyers EfficiOS Inc. http://www.efficios.com
Re: [PATCH glibc 5/9] glibc: Perform rseq(2) registration at C startup and thread creation (v17)
Quoted 15 lines · Mathieu Desnoyers
Quoted 5 lines · ----- On Apr 28, 2020, at 8:54 AM, Florian Weimer [email protected] wrote:
That one definitely should work.
I expect you might see this if libgcc_s.so.1 is installed into a multiarch subdirectory that upstream glibc does not search. (The Debian patches are unfortunately not upstream.)
My test environment is a Ubuntu 18.04.1 LTS.
Quoted 3 lines
I think on my system, the built glibc can find the system
libgcc_svia/etc/ld.so.cache, so I haven't seen this issue yet.On my system,
libgcc_sis provided here:
Quoted 7 lines
/lib/x86_64-linux-gnu/libgcc_s.so.1by this package:
Package: libgcc1
Architecture: amd64
Version: 1:8.4.0-1ubuntu1~18.04
before running the tests
cp $CC --print-file-name libgcc_s.so.1 glibc/build/dir cp $CC --print-file-name libstdc++.so.6 glibc/build/dir
so those toolchain libs are in the search path
of the newly built libc when running tests.
Re: [PATCH glibc 5/9] glibc: Perform rseq(2) registration at C startup and thread creation (v17)
* Szabolcs Nagy:
Quoted 24 lines
Quoted 15 lines · Mathieu Desnoyers
Quoted 5 lines · ----- On Apr 28, 2020, at 8:54 AM, Florian Weimer [email protected] wrote:
That one definitely should work.
I expect you might see this if libgcc_s.so.1 is installed into a multiarch subdirectory that upstream glibc does not search. (The Debian patches are unfortunately not upstream.)
My test environment is a Ubuntu 18.04.1 LTS.
Quoted 3 lines
I think on my system, the built glibc can find the system
libgcc_svia/etc/ld.so.cache, so I haven't seen this issue yet.On my system,
libgcc_sis provided here:Quoted 7 lines
/lib/x86_64-linux-gnu/libgcc_s.so.1by this package:
Package: libgcc1
Architecture: amd64
Version: 1:8.4.0-1ubuntu1~18.04
Quoted 7 lines
before running the tests
cp
$CC --print-file-name libgcc_s.so.1glibc/build/dir cp$CC --print-file-name libstdc++.so.6glibc/build/dirso those toolchain libs are in the search path
of the newly built libc when running tests.
Do you actually see the need for these steps yourself?
I guess the correct fix would be to upstream the Debian multiarch changes and activate them automatically with a configure check on systems that use multiarch paths.
Re: [PATCH glibc 5/9] glibc: Perform rseq(2) registration at C startup and thread creation (v17)
Quoted 31 lines · Florian Weimer
* Szabolcs Nagy:
Quoted 29 lines
Quoted 15 lines · Mathieu Desnoyers
Quoted 5 lines · ----- On Apr 28, 2020, at 8:54 AM, Florian Weimer [email protected] wrote:
That one definitely should work.
I expect you might see this if libgcc_s.so.1 is installed into a multiarch subdirectory that upstream glibc does not search. (The Debian patches are unfortunately not upstream.)
My test environment is a Ubuntu 18.04.1 LTS.
Quoted 3 lines
I think on my system, the built glibc can find the system
libgcc_svia/etc/ld.so.cache, so I haven't seen this issue yet.On my system,
libgcc_sis provided here:Quoted 7 lines
/lib/x86_64-linux-gnu/libgcc_s.so.1by this package:
Package: libgcc1
Architecture: amd64
Version: 1:8.4.0-1ubuntu1~18.04before running the tests
cp
$CC --print-file-name libgcc_s.so.1glibc/build/dir cp$CC --print-file-name libstdc++.so.6glibc/build/dir
Quoted 8 lines
Quoted 2 lines
so those toolchain libs are in the search path
of the newly built libc when running tests.Do you actually see the need for these steps yourself?
I guess the correct fix would be to upstream the Debian multiarch changes and activate them automatically with a configure check on systems that use multiarch paths.
cancel tests work for me on an ubuntu system because of /etc/ld.so.cache, but that may not be present or the system may not be glibc based at all.
i always do the cp because i build gcc myself (usually close to current master) and don't install it to the system path which means at compile time and runtime different libraries are used if i dont copy
Re: [PATCH glibc 5/9] glibc: Perform rseq(2) registration at C startup and thread creation (v17)
In reply to Mathieu DesnoyersQuoted 2 lines · ----- On Apr 28, 2020, at 8:33 AM, Mathieu Desnoyers…
----- On Apr 28, 2020, at 8:02 AM, Florian Weimer [email protected] wrote:
[...]
Quoted 9 lines
Quoted 7 lines
Quoted 3 lines
x32 should not be an issue as explained above, so I'm very open to add this "uptr" for user-space only.
Okay, then please use anonymous unions and structs as necessary, to ensure that the uptr field can be reached on all platforms in the same way.
OK, will do!
What I came up with looks like this. User-space can use rseq_cs.uptr.ptr both on 32-bit and 64-bit to update the pointer:
/* Restartable sequences rseq_cs field.
Contains NULL when no critical section is active for the current thread, or holds a pointer to the currently active struct rseq_cs.
Updated by user-space, which sets the address of the currently active rseq_cs at the beginning of assembly instruction sequence block, and set to NULL by the kernel when it restarts an assembly instruction sequence block, as well as when the kernel detects that it is preempting or delivering a signal outside of the range targeted by the rseq_cs. Also needs to be set to NULL by user-space before reclaiming memory that contains the targeted struct rseq_cs.
Read and set by the kernel. Set by user-space with single-copy atomicity semantics. This field should only be updated by the thread which registered this data structure. Aligned on 64-bit.
User-space may perform the update through the rseq_cs.uptr.ptr field. The padding needs to be initialized to zero on 32-bit. */
union
{
uint64_t ptr64;
#ifdef __LP64__
uint64_t ptr;
#else
struct
{
# if (defined (__BYTE_ORDER) && (__BYTE_ORDER == __BIG_ENDIAN)) || defined (__BIG_ENDIAN)
uint32_t padding; /* Initialized to zero. */
uint32_t ptr32;
# else /* LITTLE */
uint32_t ptr32;
uint32_t padding; /* Initialized to zero. */
# endif /* ENDIAN */
} ptr;
#endif#ifndef __KERNEL__
struct
{
# ifdef __LP64__
const struct rseq_cs *ptr;
# else
# if (defined (__BYTE_ORDER) && (__BYTE_ORDER == __BIG_ENDIAN)) || defined (__BIG_ENDIAN)
uint32_t padding; /* Initialized to zero. */
const struct rseq_cs *ptr;
# else /* LITTLE */
const struct rseq_cs *ptr;
uint32_t padding; /* Initialized to zero. */
# endif /* ENDIAN */
# endif
} uptr;
#endif
} rseq_cs;Thanks,
Mathieu
Signature · 3 lines
Mathieu Desnoyers EfficiOS Inc. http://www.efficios.com
Re: [PATCH glibc 5/9] glibc: Perform rseq(2) registration at C startup and thread creation (v17)
* Mathieu Desnoyers:
Quoted 9 lines
Quoted 2 lines · ----- On Apr 28, 2020, at 8:33 AM, Mathieu Desnoyers…
----- On Apr 28, 2020, at 8:02 AM, Florian Weimer [email protected] wrote:
[...]
Quoted 4 lines
Quoted 4 lines
Quoted 3 lines
x32 should not be an issue as explained above, so I'm very open to add this "uptr" for user-space only.
Quoted 8 lines
Quoted 5 lines
Quoted 3 lines
Okay, then please use anonymous unions and structs as necessary, to ensure that the uptr field can be reached on all platforms in the same way.
OK, will do!
What I came up with looks like this. User-space can use rseq_cs.uptr.ptr both on 32-bit and 64-bit to update the pointer:
Agreed, this should work.
Re: [PATCH glibc 5/9] glibc: Perform rseq(2) registration at C startup and thread creation (v17)
In reply to Mathieu Desnoyers* Mathieu Desnoyers via Libc-alpha:
Quoted 8 lines
elf/libc_early_init.c +0
-0@@ -18,10 +18,13 @@#include <ctype.h> #include <libc-early-init.h>
elf/libc_early_init.c line 30
Quoted 10 lines
elf/libc_early_init.c +3
-0+#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 (); }
I think the registration must be restricted to the primary namespace. Otherwise, LD_AUDIT will register the area to the secondary libc (in the audit module), not the primary libc for the entire process.
I think the easiest way to implement this for now is a flag argument for __libc_early_init (as the upstream __libc_multiple_libcs is not entirely accurate). I will submit a patch.
Re: [PATCH glibc 5/9] glibc: Perform rseq(2) registration at C startup and thread creation (v17)
Quoted 25 lines · ----- On Apr 27, 2020, at 7:59 AM, Florian Weimer [email protected] wrote:
* Mathieu Desnoyers via Libc-alpha:
Quoted 8 lines
elf/libc_early_init.c +0
-0@@ -18,10 +18,13 @@#include <ctype.h> #include <libc-early-init.h> elf/libc_early_init.c line 30
Quoted 10 lines
elf/libc_early_init.c +3
-0+#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 (); }I think the registration must be restricted to the primary namespace. Otherwise,
LD_AUDITwill register the area to the secondary libc (in the audit module), not the primary libc for the entire process.
Quoted 3 lines
I think the easiest way to implement this for now is a flag argument for
__libc_early_init(as the upstream__libc_multiple_libcsis not entirely accurate). I will submit a patch.
OK, once I get the patch, I will pick it up in my series.
Thanks,
Mathieu
Signature · 3 lines
Mathieu Desnoyers EfficiOS Inc. http://www.efficios.com
Re: [PATCH glibc 5/9] glibc: Perform rseq(2) registration at C startup and thread creation (v17)
* Mathieu Desnoyers:
Quoted 27 lines
Quoted 25 lines · ----- On Apr 27, 2020, at 7:59 AM, Florian Weimer [email protected] wrote:
* Mathieu Desnoyers via Libc-alpha:
Quoted 8 lines
elf/libc_early_init.c +0
-0@@ -18,10 +18,13 @@#include <ctype.h> #include <libc-early-init.h> elf/libc_early_init.c line 30
Quoted 10 lines
elf/libc_early_init.c +3
-0+#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 (); }I think the registration must be restricted to the primary namespace. Otherwise,
LD_AUDITwill register the area to the secondary libc (in the audit module), not the primary libc for the entire process.
Quoted 5 lines
Quoted 3 lines
I think the easiest way to implement this for now is a flag argument for
__libc_early_init(as the upstream__libc_multiple_libcsis not entirely accurate). I will submit a patch.OK, once I get the patch, I will pick it up in my series.
There should be no need for that, it can be reviewed and committed separately:
<https://sourceware.org/pipermail/libc-alpha/2020-April/113182.html>
PATCH 6/9 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 34 lines
sysdeps/unix/sysv/linux/sched_getcpu.c +15
-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);
sysdeps/unix/sysv/linux/sched_getcpu.c sched_getcpu (void) · line 57
Quoted 10 lines
sysdeps/unix/sysv/linux/sched_getcpu.c +10
-0+ + return cpu_id >= 0 ? cpu_id : vsyscall_sched_getcpu (); +} +#else +int +sched_getcpu (void ) +{ +return vsyscall_sched_getcpu (); +} +#endif
This patch needs to be rebased on current master, but the direction looks right to me.