From 146fc63fce9b46ab77ba7959cf69429d9bd88d2a Mon Sep 17 00:00:00 2001 From: Konstantin Belousov Date: Sun, 9 Feb 2020 11:53:12 +0000 Subject: [PATCH] Add a way to manage thread signal mask using shared word, instead of syscall. A new syscall sigfastblock(2) is added which registers a uint32_t variable as containing the count of blocks for signal delivery. Its content is read by kernel on each syscall entry and on AST processing, non-zero count of blocks is interpreted same as the signal mask blocking all signals. The biggest downside of the feature that I see is that memory corruption that affects the registered fast sigblock location, would cause quite strange application misbehavior. For instance, the process would be immune to ^C (but killable by SIGKILL). With consumers (rtld and libthr added), benchmarks do not show a slow-down of the syscalls in micro-measurements, and macro benchmarks like buildworld do not demonstrate a difference. Part of the reason is that buildworld time is dominated by compiler, and clang already links to libthr. On the other hand, small utilities typically used by shell scripts have the total number of syscalls cut by half. The syscall is not exported from the stable libc version namespace on purpose. It is intended to be used only by our C runtime implementation internals. Tested by: pho Disscussed with: cem, emaste, jilles Sponsored by: The FreeBSD Foundation Differential revision: https://reviews.freebsd.org/D12773 --- lib/libc/sys/Makefile.inc | 1 + lib/libc/sys/Symbol.map | 1 + lib/libc/sys/sigfastblock.2 | 166 ++++++++++++++++++++++ sys/compat/freebsd32/syscalls.master | 1 + sys/kern/capabilities.conf | 1 + sys/kern/kern_exec.c | 5 + sys/kern/kern_fork.c | 3 +- sys/kern/kern_sig.c | 198 +++++++++++++++++++++++---- sys/kern/kern_thread.c | 8 +- sys/kern/subr_syscall.c | 7 + sys/kern/subr_trap.c | 81 ++++++++--- sys/kern/syscalls.master | 6 + sys/sys/proc.h | 6 +- sys/sys/signalvar.h | 20 +++ 14 files changed, 458 insertions(+), 46 deletions(-) create mode 100644 lib/libc/sys/sigfastblock.2 diff --git a/lib/libc/sys/Makefile.inc b/lib/libc/sys/Makefile.inc index 4e048b5f3927..d13201c1be7d 100644 --- a/lib/libc/sys/Makefile.inc +++ b/lib/libc/sys/Makefile.inc @@ -317,6 +317,7 @@ MAN+= sctp_generic_recvmsg.2 \ shutdown.2 \ sigaction.2 \ sigaltstack.2 \ + sigfastblock.2 \ sigpending.2 \ sigprocmask.2 \ sigqueue.2 \ diff --git a/lib/libc/sys/Symbol.map b/lib/libc/sys/Symbol.map index a31cf1616ddc..aa60da2a6789 100644 --- a/lib/libc/sys/Symbol.map +++ b/lib/libc/sys/Symbol.map @@ -567,6 +567,7 @@ FBSDprivate_1.0 { __sys_extattr_set_link; _extattrctl; __sys_extattrctl; + __sys_sigfastblock; _fchdir; __sys_fchdir; _fchflags; diff --git a/lib/libc/sys/sigfastblock.2 b/lib/libc/sys/sigfastblock.2 new file mode 100644 index 000000000000..da835abdf9aa --- /dev/null +++ b/lib/libc/sys/sigfastblock.2 @@ -0,0 +1,166 @@ +.\" Copyright (c) 2016 The FreeBSD Foundation, Inc. +.\" +.\" This documentation was written by +.\" Konstantin Belousov under sponsorship +.\" from the FreeBSD Foundation. +.\" +.\" Redistribution and use in source and binary forms, with or without +.\" modification, are permitted provided that the following conditions +.\" are met: +.\" 1. Redistributions of source code must retain the above copyright +.\" notice, this list of conditions and the following disclaimer. +.\" 2. Redistributions in binary form must reproduce the above copyright +.\" notice, this list of conditions and the following disclaimer in the +.\" documentation and/or other materials provided with the distribution. +.\" +.\" THIS SOFTWARE IS PROVIDED BY THE AUTHORS AND CONTRIBUTORS ``AS IS'' AND +.\" ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE +.\" IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE +.\" ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHORS OR CONTRIBUTORS BE LIABLE +.\" FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL +.\" DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS +.\" OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) +.\" HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT +.\" LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY +.\" OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF +.\" SUCH DAMAGE. +.\" +.\" $FreeBSD$ +.\" +.Dd December 13, 2019 +.Dt SIGFASTBLOCK 2 +.Os +.Sh NAME +.Nm sigfastblock +.Nd controls signals blocking with a simple memory write +.Sh LIBRARY +.Lb libc +.Sh SYNOPSIS +.In sys/signalvar.h +.Ft int +.Fn sigfastblock "int cmd" "void *ptr" +.Sh DESCRIPTION +.Bf -symbolic +This function is not intended for a direct usage by applications. +The functionality is provided for implementing some optimizations in +.Xr ld-elf.so.1 8 +and +.Lb libthr . +.Ef +.Pp +The function configures the kernel facility that allows a thread to +block asynchronous signals delivery with a single write to userspace +memory, avoiding overhead of system calls like +.Xr sigprocmask 2 +for establishing critical sections. +The C runtime uses it to optimize implementation of async-signal-safe +functionality. +.Pp +A thread might register a +.Dv sigblock +variable of type +.Vt int +as a location which is consulted by kernel when calculating the +blocked signal mask for delivery of asynchronous signals. +If the variable indicates that blocking is requested, then the kernel +effectively operates as if the mask containing all blockable signals was +supplied to +.Xr sigprocmask 2 . +.Pp +The variable is supposed to be modified only from the owning thread, +there is no way to guarantee visibility of update from other thread +to kernel when signals are delivered. +.Pp +Lower bits of the sigblock variable are reserved as flags, +which might be set or cleared by kernel at arbitrary moments. +Userspace code should use +.Xr atomic 9 +operations of incrementing and decrementing by +.Dv SIGFASTBLOCK_INC +quantity to recursively block or unblock signals delivery. +.Pp +If a signal would be delivered when unmasked, kernel might set the +.Dv SIGFASTBLOCK_PEND +.Dq pending signal +flag in the sigblock variable. +Userspace should perform +.Dv SIGFASTBLOCK_UNBLOCK +operation when clearing the variable if it notes the pending signal +bit is set, which would deliver the pending signals immediately. +Otherwise, signals delivery might be postponed. +.Pp +The +.Fa cmd +argument specifies one of the following operations: +.Bl -tag -width SIGFASTBLOCK_UNSETPTR +.It Dv SIGFASTBLOCK_SETPTR +Register the variable of type +.Vt int +at location pointed to by the +.Fa ptr +argument as sigblock variable for the calling thread. +.It Dv SIGFASTBLOCK_UNSETPTR +Unregister the currently registered sigblock location. +Kernel stops inferring the blocked mask from non-zero value of its +blocked count. +New location can be registered after previous one is deregistered. +.It Dv SIGFASTBLOCK_UNBLOCK +If there are pending signals which should be delivered to the calling +thread, they are delivered before returning from the call. +The sigblock variable should have zero blocking count, and indicate +that the pending signal exists. +Effectively this means that the variable should have the value +.Dv SIGFASTBLOCK_PEND . +.El +.Sh RETURN VALUES +.Rv -std +.Sh ERRORS +The operation may fail with the following errors: +.Bl -tag -width Er +.It Bq Er EBUSY +The +.Dv SIGFASTBLOCK_SETPTR +attempted while the sigblock address was already registered. +The +.Dv SIGFASTBLOCK_UNBLOCK +was called while sigblock variable value is not equal to +.Dv SIGFASTBLOCK_PEND . +.It Bq Er EINVAL +The variable address passed to +.Dv SIGFASTBLOCK_SETPTR +is not aligned naturally. +The +.Dv SIGFASTBLOCK_UNSETPTR +operation was attempted without prior successfull call to +.Dv SIGFASTBLOCK_SETPTR . +.It Bq Er EFAULT +Attempt to read or write to the sigblock variable failed. +Note that kernel generates the +.Dv SIGSEGV +signal if an attempt to read from the sigblock variable faulted +during implicit accesses from syscall entry. +.El +.Sh SEE ALSO +.Xr kill 2 , +.Xr signal 2 , +.Xr sigprocmask 2 , +.Xr libthr 3 , +.Xr ld-elf.so.1 8 +.Sh STANDARDS +The +.Nm +function is non-standard, although a similar functionality is a common +optimization provided by several other systems. +.Sh HISTORY +The +.Nm +function was introduced in +.Fx 13.0 . +.Sh BUGS +The +.Nm +symbol is currently not exported by libc, on purpose. +Consumers should either use the +.Dv __sys_fast_sigblock +symbol from the private libc namespace, or utilize +.Xr syscall 2 . diff --git a/sys/compat/freebsd32/syscalls.master b/sys/compat/freebsd32/syscalls.master index b4e58730cc65..a82d75d26e22 100644 --- a/sys/compat/freebsd32/syscalls.master +++ b/sys/compat/freebsd32/syscalls.master @@ -1159,5 +1159,6 @@ int shmflags, const char *name); } 572 AUE_SHMRENAME NOPROTO { int shm_rename(const char *path_from, \ const char *path_to, int flags); } +573 AUE_NULL NOPROTO { int sigfastblock(int cmd, uint32_t *ptr); } ; vim: syntax=off diff --git a/sys/kern/capabilities.conf b/sys/kern/capabilities.conf index 897c64145514..0b5aa2f83fd8 100644 --- a/sys/kern/capabilities.conf +++ b/sys/kern/capabilities.conf @@ -668,6 +668,7 @@ shutdown sigaction sigaltstack sigblock +sigfastblock sigpending sigprocmask sigqueue diff --git a/sys/kern/kern_exec.c b/sys/kern/kern_exec.c index 69b5c72af233..3fa76d659a37 100644 --- a/sys/kern/kern_exec.c +++ b/sys/kern/kern_exec.c @@ -1025,6 +1025,7 @@ exec_new_vmspace(struct image_params *imgp, struct sysentvec *sv) int error; struct proc *p = imgp->proc; struct vmspace *vmspace = p->p_vmspace; + struct thread *td = curthread; vm_object_t obj; struct rlimit rlim_stack; vm_offset_t sv_minuser, stack_addr; @@ -1034,6 +1035,10 @@ exec_new_vmspace(struct image_params *imgp, struct sysentvec *sv) imgp->vmspace_destroyed = 1; imgp->sysent = sv; + td->td_pflags &= ~TDP_SIGFASTBLOCK; + td->td_sigblock_ptr = NULL; + td->td_sigblock_val = 0; + /* May be called with Giant held */ EVENTHANDLER_DIRECT_INVOKE(process_exec, p, imgp); diff --git a/sys/kern/kern_fork.c b/sys/kern/kern_fork.c index e8ac950a5d78..958e119927fb 100644 --- a/sys/kern/kern_fork.c +++ b/sys/kern/kern_fork.c @@ -563,7 +563,8 @@ do_fork(struct thread *td, struct fork_req *fr, struct proc *p2, struct thread * * been preserved. */ p2->p_flag |= p1->p_flag & P_SUGID; - td2->td_pflags |= (td->td_pflags & TDP_ALTSTACK) | TDP_FORKING; + td2->td_pflags |= (td->td_pflags & (TDP_ALTSTACK | + TDP_SIGFASTBLOCK)) | TDP_FORKING; SESS_LOCK(p1->p_session); if (p1->p_session->s_ttyvp != NULL && p1->p_flag & P_CONTROLT) p2->p_flag |= P_CONTROLT; diff --git a/sys/kern/kern_sig.c b/sys/kern/kern_sig.c index 0d14b0665f4c..5d6a677748ea 100644 --- a/sys/kern/kern_sig.c +++ b/sys/kern/kern_sig.c @@ -114,7 +114,7 @@ static int sig_suspend_threads(struct thread *, struct proc *, int); static int filt_sigattach(struct knote *kn); static void filt_sigdetach(struct knote *kn); static int filt_signal(struct knote *kn, long hint); -static struct thread *sigtd(struct proc *p, int sig, int prop); +static struct thread *sigtd(struct proc *p, int sig, bool fast_sigblock); static void sigqueue_start(void); static uma_zone_t ksiginfo_zone = NULL; @@ -238,7 +238,7 @@ static int sigproptbl[NSIG] = { [SIGUSR2] = SIGPROP_KILL, }; -static void reschedule_signals(struct proc *p, sigset_t block, int flags); +sigset_t fastblock_mask; static void sigqueue_start(void) @@ -249,6 +249,8 @@ sigqueue_start(void) p31b_setcfg(CTL_P1003_1B_REALTIME_SIGNALS, _POSIX_REALTIME_SIGNALS); p31b_setcfg(CTL_P1003_1B_RTSIG_MAX, SIGRTMAX - SIGRTMIN + 1); p31b_setcfg(CTL_P1003_1B_SIGQUEUE_MAX, max_pending_per_proc); + SIGFILLSET(fastblock_mask); + SIG_CANTMASK(fastblock_mask); } ksiginfo_t * @@ -1995,8 +1997,8 @@ trapsignal(struct thread *td, ksiginfo_t *ksi) { struct sigacts *ps; struct proc *p; - int sig; - int code; + sigset_t sigmask; + int code, sig; p = td->td_proc; sig = ksi->ksi_signo; @@ -2006,8 +2008,11 @@ trapsignal(struct thread *td, ksiginfo_t *ksi) PROC_LOCK(p); ps = p->p_sigacts; mtx_lock(&ps->ps_mtx); + sigmask = td->td_sigmask; + if (td->td_sigblock_val != 0) + SIGSETOR(sigmask, fastblock_mask); if ((p->p_flag & P_TRACED) == 0 && SIGISMEMBER(ps->ps_sigcatch, sig) && - !SIGISMEMBER(td->td_sigmask, sig)) { + !SIGISMEMBER(sigmask, sig)) { #ifdef KTRACE if (KTRPOINT(curthread, KTR_PSIG)) ktrpsig(sig, ps->ps_sigact[_SIG_IDX(sig)], @@ -2023,13 +2028,14 @@ trapsignal(struct thread *td, ksiginfo_t *ksi) * masking the signal or process is ignoring the * signal. */ - if (kern_forcesigexit && - (SIGISMEMBER(td->td_sigmask, sig) || - ps->ps_sigact[_SIG_IDX(sig)] == SIG_IGN)) { + if (kern_forcesigexit && (SIGISMEMBER(sigmask, sig) || + ps->ps_sigact[_SIG_IDX(sig)] == SIG_IGN)) { SIGDELSET(td->td_sigmask, sig); SIGDELSET(ps->ps_sigcatch, sig); SIGDELSET(ps->ps_sigignore, sig); ps->ps_sigact[_SIG_IDX(sig)] = SIG_DFL; + td->td_pflags &= ~TDP_SIGFASTBLOCK; + td->td_sigblock_val = 0; } mtx_unlock(&ps->ps_mtx); p->p_sig = sig; /* XXX to verify code */ @@ -2039,21 +2045,24 @@ trapsignal(struct thread *td, ksiginfo_t *ksi) } static struct thread * -sigtd(struct proc *p, int sig, int prop) +sigtd(struct proc *p, int sig, bool fast_sigblock) { struct thread *td, *signal_td; PROC_LOCK_ASSERT(p, MA_OWNED); + MPASS(!fast_sigblock || p == curproc); /* * Check if current thread can handle the signal without * switching context to another thread. */ - if (curproc == p && !SIGISMEMBER(curthread->td_sigmask, sig)) + if (curproc == p && !SIGISMEMBER(curthread->td_sigmask, sig) && + (!fast_sigblock || curthread->td_sigblock_val == 0)) return (curthread); signal_td = NULL; FOREACH_THREAD_IN_PROC(p, td) { - if (!SIGISMEMBER(td->td_sigmask, sig)) { + if (!SIGISMEMBER(td->td_sigmask, sig) && (!fast_sigblock || + td != curthread || td->td_sigblock_val == 0)) { signal_td = td; break; } @@ -2167,7 +2176,7 @@ tdsendsignal(struct proc *p, struct thread *td, int sig, ksiginfo_t *ksi) prop = sigprop(sig); if (td == NULL) { - td = sigtd(p, sig, prop); + td = sigtd(p, sig, false); sigqueue = &p->p_sigqueue; } else sigqueue = &td->td_sigqueue; @@ -2562,7 +2571,6 @@ ptracestop(struct thread *td, int sig, ksiginfo_t *si) struct proc *p = td->td_proc; struct thread *td2; ksiginfo_t ksi; - int prop; PROC_LOCK_ASSERT(p, MA_OWNED); KASSERT(!(p->p_flag & P_WEXIT), ("Stopping exiting process")); @@ -2659,8 +2667,7 @@ ptracestop(struct thread *td, int sig, ksiginfo_t *si) ksiginfo_init(&ksi); ksi.ksi_signo = td->td_xsig; ksi.ksi_flags |= KSI_PTRACE; - prop = sigprop(td->td_xsig); - td2 = sigtd(p, td->td_xsig, prop); + td2 = sigtd(p, td->td_xsig, false); tdsendsignal(p, td2, td->td_xsig, &ksi); if (td != td2) return (0); @@ -2669,33 +2676,45 @@ ptracestop(struct thread *td, int sig, ksiginfo_t *si) return (td->td_xsig); } -static void +void reschedule_signals(struct proc *p, sigset_t block, int flags) { struct sigacts *ps; struct thread *td; int sig; + bool fastblk, pslocked; PROC_LOCK_ASSERT(p, MA_OWNED); ps = p->p_sigacts; - mtx_assert(&ps->ps_mtx, (flags & SIGPROCMASK_PS_LOCKED) != 0 ? - MA_OWNED : MA_NOTOWNED); + pslocked = (flags & SIGPROCMASK_PS_LOCKED) != 0; + mtx_assert(&ps->ps_mtx, pslocked ? MA_OWNED : MA_NOTOWNED); if (SIGISEMPTY(p->p_siglist)) return; SIGSETAND(block, p->p_siglist); + fastblk = (flags & SIGPROCMASK_FASTBLK) != 0; while ((sig = sig_ffs(&block)) != 0) { SIGDELSET(block, sig); - td = sigtd(p, sig, 0); + td = sigtd(p, sig, fastblk); + + /* + * If sigtd() selected us despite sigfastblock is + * blocking, do not activate AST or wake us, to avoid + * loop in AST handler. + */ + if (fastblk && td == curthread) + continue; + signotify(td); - if (!(flags & SIGPROCMASK_PS_LOCKED)) + if (!pslocked) mtx_lock(&ps->ps_mtx); if (p->p_flag & P_TRACED || (SIGISMEMBER(ps->ps_sigcatch, sig) && - !SIGISMEMBER(td->td_sigmask, sig))) + !SIGISMEMBER(td->td_sigmask, sig))) { tdsigwakeup(td, sig, SIG_CATCH, (SIGISMEMBER(ps->ps_sigintr, sig) ? EINTR : - ERESTART)); - if (!(flags & SIGPROCMASK_PS_LOCKED)) + ERESTART)); + } + if (!pslocked) mtx_unlock(&ps->ps_mtx); } } @@ -2844,6 +2863,24 @@ issignal(struct thread *td) SIG_STOPSIGMASK(sigpending); if (SIGISEMPTY(sigpending)) /* no signal to send */ return (0); + + /* + * Do fast sigblock if requested by usermode. Since + * we do know that there was a signal pending at this + * point, set the FAST_SIGBLOCK_PEND as indicator for + * usermode to perform a dummy call to + * FAST_SIGBLOCK_UNBLOCK, which causes immediate + * delivery of postponed pending signal. + */ + if ((td->td_pflags & TDP_SIGFASTBLOCK) != 0) { + if (td->td_sigblock_val != 0) + SIGSETNAND(sigpending, fastblock_mask); + if (SIGISEMPTY(sigpending)) { + td->td_pflags |= TDP_SIGFASTPENDING; + return (0); + } + } + if ((p->p_flag & (P_TRACED | P_PPTRACE)) == P_TRACED && (p->p_flag2 & P2_PTRACE_FSTP) != 0 && SIGISMEMBER(sigpending, SIGSTOP)) { @@ -3914,3 +3951,118 @@ sig_drop_caught(struct proc *p) sigqueue_delete_proc(p, sig); } } + +int +sys_sigfastblock(struct thread *td, struct sigfastblock_args *uap) +{ + struct proc *p; + int error, res; + uint32_t oldval; + + error = 0; + switch (uap->cmd) { + case SIGFASTBLOCK_SETPTR: + if ((td->td_pflags & TDP_SIGFASTBLOCK) != 0) { + error = EBUSY; + break; + } + if (((uintptr_t)(uap->ptr) & (sizeof(uint32_t) - 1)) != 0) { + error = EINVAL; + break; + } + td->td_pflags |= TDP_SIGFASTBLOCK; + td->td_sigblock_ptr = uap->ptr; + break; + + case SIGFASTBLOCK_UNBLOCK: + if ((td->td_pflags & TDP_SIGFASTBLOCK) != 0) { + error = EINVAL; + break; + } +again: + res = casueword32(td->td_sigblock_ptr, SIGFASTBLOCK_PEND, + &oldval, 0); + if (res == -1) { + error = EFAULT; + break; + } + if (res == 1) { + if (oldval != SIGFASTBLOCK_PEND) { + error = EBUSY; + break; + } + error = thread_check_susp(td, false); + if (error != 0) + break; + goto again; + } + td->td_sigblock_val = 0; + + /* + * Rely on normal ast mechanism to deliver pending + * signals to current thread. But notify others about + * fake unblock. + */ + p = td->td_proc; + if (error == 0 && p->p_numthreads != 1) { + PROC_LOCK(p); + reschedule_signals(p, td->td_sigmask, 0); + PROC_UNLOCK(p); + } + break; + + case SIGFASTBLOCK_UNSETPTR: + if ((td->td_pflags & TDP_SIGFASTBLOCK) == 0) { + error = EINVAL; + break; + } + res = fueword32(td->td_sigblock_ptr, &oldval); + if (res == -1) { + error = EFAULT; + break; + } + if (oldval != 0 && oldval != SIGFASTBLOCK_PEND) { + error = EBUSY; + break; + } + td->td_pflags &= ~TDP_SIGFASTBLOCK; + td->td_sigblock_val = 0; + break; + + default: + error = EINVAL; + break; + } + return (error); +} + +void +fetch_sigfastblock(struct thread *td) +{ + + if ((td->td_pflags & TDP_SIGFASTBLOCK) == 0) + return; + if (fueword32(td->td_sigblock_ptr, &td->td_sigblock_val) == -1) { + fetch_sigfastblock_failed(td, false); + return; + } + td->td_sigblock_val &= ~SIGFASTBLOCK_FLAGS; +} + +void +fetch_sigfastblock_failed(struct thread *td, bool write) +{ + ksiginfo_t ksi; + + /* + * Prevent further fetches and SIGSEGVs, allowing thread to + * issue syscalls despite corruption. + */ + td->td_pflags &= ~TDP_SIGFASTBLOCK; + + ksiginfo_init_trap(&ksi); + ksi.ksi_signo = SIGSEGV; + ksi.ksi_code = write ? SEGV_ACCERR : SEGV_MAPERR; + ksi.ksi_addr = td->td_sigblock_ptr; + trapsignal(td, &ksi); +} diff --git a/sys/kern/kern_thread.c b/sys/kern/kern_thread.c index 83c22309fd70..b3746553f2ac 100644 --- a/sys/kern/kern_thread.c +++ b/sys/kern/kern_thread.c @@ -82,9 +82,9 @@ _Static_assert(offsetof(struct thread, td_flags) == 0xfc, "struct thread KBI td_flags"); _Static_assert(offsetof(struct thread, td_pflags) == 0x104, "struct thread KBI td_pflags"); -_Static_assert(offsetof(struct thread, td_frame) == 0x480, +_Static_assert(offsetof(struct thread, td_frame) == 0x490, "struct thread KBI td_frame"); -_Static_assert(offsetof(struct thread, td_emuldata) == 0x690, +_Static_assert(offsetof(struct thread, td_emuldata) == 0x6a0, "struct thread KBI td_emuldata"); _Static_assert(offsetof(struct proc, p_flag) == 0xb0, "struct proc KBI p_flag"); @@ -102,9 +102,9 @@ _Static_assert(offsetof(struct thread, td_flags) == 0x98, "struct thread KBI td_flags"); _Static_assert(offsetof(struct thread, td_pflags) == 0xa0, "struct thread KBI td_pflags"); -_Static_assert(offsetof(struct thread, td_frame) == 0x2f0, +_Static_assert(offsetof(struct thread, td_frame) == 0x2f8, "struct thread KBI td_frame"); -_Static_assert(offsetof(struct thread, td_emuldata) == 0x338, +_Static_assert(offsetof(struct thread, td_emuldata) == 0x340, "struct thread KBI td_emuldata"); _Static_assert(offsetof(struct proc, p_flag) == 0x68, "struct proc KBI p_flag"); diff --git a/sys/kern/subr_syscall.c b/sys/kern/subr_syscall.c index 951e7b682623..a6f8e9d6b7d1 100644 --- a/sys/kern/subr_syscall.c +++ b/sys/kern/subr_syscall.c @@ -140,6 +140,13 @@ syscallenter(struct thread *td) /* Let system calls set td_errno directly. */ td->td_pflags &= ~TDP_NERRNO; + /* + * Fetch fast sigblock value at the time of syscall + * entry because sleepqueue primitives might call + * cursig(). + */ + fetch_sigfastblock(td); + AUDIT_SYSCALL_ENTER(sa->code, td); error = (sa->callp->sy_call)(td, sa->args); AUDIT_SYSCALL_EXIT(error, td); diff --git a/sys/kern/subr_trap.c b/sys/kern/subr_trap.c index 3285272b1855..a7f842a1948e 100644 --- a/sys/kern/subr_trap.c +++ b/sys/kern/subr_trap.c @@ -116,12 +116,16 @@ userret(struct thread *td, struct trapframe *frame) if (p->p_numthreads == 1) { PROC_LOCK(p); thread_lock(td); - if ((p->p_flag & P_PPWAIT) == 0) { - KASSERT(!SIGPENDING(td) || (td->td_flags & - (TDF_NEEDSIGCHK | TDF_ASTPENDING)) == - (TDF_NEEDSIGCHK | TDF_ASTPENDING), - ("failed to set signal flags for ast p %p " - "td %p fl %x", p, td, td->td_flags)); + if ((p->p_flag & P_PPWAIT) == 0 && + (td->td_pflags & TDP_SIGFASTBLOCK) == 0) { + if (SIGPENDING(td) && (td->td_flags & + (TDF_NEEDSIGCHK | TDF_ASTPENDING)) != + (TDF_NEEDSIGCHK | TDF_ASTPENDING)) { + thread_unlock(td); + panic( + "failed to set signal flags for ast p %p td %p fl %x", + p, td, td->td_flags); + } } thread_unlock(td); PROC_UNLOCK(p); @@ -218,8 +222,8 @@ ast(struct trapframe *framep) { struct thread *td; struct proc *p; - int flags; - int sig; + uint32_t oldval; + int flags, sig, res; td = curthread; p = td->td_proc; @@ -298,12 +302,16 @@ ast(struct trapframe *framep) * the reason for looping check for AST condition. * See comment in userret() about P_PPWAIT. */ - if ((p->p_flag & P_PPWAIT) == 0) { - KASSERT(!SIGPENDING(td) || (td->td_flags & - (TDF_NEEDSIGCHK | TDF_ASTPENDING)) == - (TDF_NEEDSIGCHK | TDF_ASTPENDING), - ("failed2 to set signal flags for ast p %p td %p " - "fl %x %x", p, td, flags, td->td_flags)); + if ((p->p_flag & P_PPWAIT) == 0 && + (td->td_pflags & TDP_SIGFASTBLOCK) == 0) { + if (SIGPENDING(td) && (td->td_flags & + (TDF_NEEDSIGCHK | TDF_ASTPENDING)) != + (TDF_NEEDSIGCHK | TDF_ASTPENDING)) { + thread_unlock(td); /* fix dumps */ + panic( + "failed2 to set signal flags for ast p %p td %p fl %x %x", + p, td, flags, td->td_flags); + } } thread_unlock(td); PROC_UNLOCK(p); @@ -317,15 +325,54 @@ ast(struct trapframe *framep) */ if (flags & TDF_NEEDSIGCHK || p->p_pendingcnt > 0 || !SIGISEMPTY(p->p_siglist)) { + fetch_sigfastblock(td); PROC_LOCK(p); mtx_lock(&p->p_sigacts->ps_mtx); - while ((sig = cursig(td)) != 0) { - KASSERT(sig >= 0, ("sig %d", sig)); - postsig(sig); + if ((td->td_pflags & TDP_SIGFASTBLOCK) != 0 && + td->td_sigblock_val != 0) { + reschedule_signals(p, fastblock_mask, + SIGPROCMASK_PS_LOCKED | SIGPROCMASK_FASTBLK); + } else { + while ((sig = cursig(td)) != 0) { + KASSERT(sig >= 0, ("sig %d", sig)); + postsig(sig); + } } mtx_unlock(&p->p_sigacts->ps_mtx); PROC_UNLOCK(p); } + + /* + * Handle deferred update of the fast sigblock value, after + * the postsig() loop was performed. + */ + if (td->td_pflags & TDP_SIGFASTPENDING) { + td->td_pflags &= ~TDP_SIGFASTPENDING; + res = fueword32(td->td_sigblock_ptr, &oldval); + if (res == -1) { + fetch_sigfastblock_failed(td, false); + } else { + for (;;) { + oldval |= SIGFASTBLOCK_PEND; + res = casueword32(td->td_sigblock_ptr, oldval, + &oldval, oldval | SIGFASTBLOCK_PEND); + if (res == -1) { + fetch_sigfastblock_failed(td, true); + break; + } + if (res == 0) { + td->td_sigblock_val = oldval & + ~SIGFASTBLOCK_FLAGS; + break; + } + MPASS(res == 1); + res = thread_check_susp(td, false); + if (res != 0) + break; + } + } + } + /* * We need to check to see if we have to exit or wait due to a * single threading requirement or some other STOP condition. diff --git a/sys/kern/syscalls.master b/sys/kern/syscalls.master index 8c7a3bd06a6c..47a42020d2c6 100644 --- a/sys/kern/syscalls.master +++ b/sys/kern/syscalls.master @@ -3212,6 +3212,12 @@ int flags ); } +573 AUE_NULL STD { + int sigfastblock( + int cmd, + _Inout_opt_ uint32_t *ptr + ); + } ; Please copy any additions and changes to the following compatability tables: ; sys/compat/freebsd32/syscalls.master diff --git a/sys/sys/proc.h b/sys/sys/proc.h index 2a85b035a7b4..f657a4cc7af5 100644 --- a/sys/sys/proc.h +++ b/sys/sys/proc.h @@ -322,6 +322,9 @@ struct thread { uintptr_t td_rb_inact; /* (k) Current in-action mutex loc. */ struct syscall_args td_sa; /* (kx) Syscall parameters. Copied on fork for child tracing. */ + void *td_sigblock_ptr; /* (k) uptr for fast sigblock. */ + uint32_t td_sigblock_val; /* (k) fast sigblock value read at + td_sigblock_ptr on kern entry */ #define td_endcopy td_pcb /* @@ -486,7 +489,7 @@ do { \ #define TDP_ALTSTACK 0x00000020 /* Have alternate signal stack. */ #define TDP_DEADLKTREAT 0x00000040 /* Lock acquisition - deadlock treatment. */ #define TDP_NOFAULTING 0x00000080 /* Do not handle page faults. */ -#define TDP_UNUSED9 0x00000100 /* --available-- */ +#define TDP_SIGFASTBLOCK 0x00000100 /* Fast sigblock active */ #define TDP_OWEUPC 0x00000200 /* Call addupc() at next AST. */ #define TDP_ITHREAD 0x00000400 /* Thread is an interrupt thread. */ #define TDP_SYNCIO 0x00000800 /* Local override, disable async i/o. */ @@ -509,6 +512,7 @@ do { \ #define TDP_UIOHELD 0x10000000 /* Current uio has pages held in td_ma */ #define TDP_FORKING 0x20000000 /* Thread is being created through fork() */ #define TDP_EXECVMSPC 0x40000000 /* Execve destroyed old vmspace */ +#define TDP_SIGFASTPENDING 0x80000000 /* Pending signal due to sigfastblock */ /* * Reasons that the current thread can not be run yet. diff --git a/sys/sys/signalvar.h b/sys/sys/signalvar.h index ed82ff123209..66a87c0e4b23 100644 --- a/sys/sys/signalvar.h +++ b/sys/sys/signalvar.h @@ -256,7 +256,23 @@ typedef struct sigqueue { /* Flags for ksi_flags */ #define SQ_INIT 0x01 +/* + * Fast_sigblock + */ +#define SIGFASTBLOCK_SETPTR 1 +#define SIGFASTBLOCK_UNBLOCK 2 +#define SIGFASTBLOCK_UNSETPTR 3 + +#define SIGFASTBLOCK_PEND 0x1 +#define SIGFASTBLOCK_FLAGS 0xf +#define SIGFASTBLOCK_INC 0x10 + +#ifndef _KERNEL +int __sys_sigfastblock(int cmd, void *ptr); +#endif + #ifdef _KERNEL +extern sigset_t fastblock_mask; /* Return nonzero if process p has an unmasked pending signal. */ #define SIGPENDING(td) \ @@ -328,6 +344,7 @@ extern struct mtx sigio_lock; #define SIGPROCMASK_OLD 0x0001 #define SIGPROCMASK_PROC_LOCKED 0x0002 #define SIGPROCMASK_PS_LOCKED 0x0004 +#define SIGPROCMASK_FASTBLK 0x0008 /* * Modes for sigdeferstop(). Manages behaviour of @@ -365,6 +382,8 @@ sigallowstop(int prev) int cursig(struct thread *td); void execsigs(struct proc *p); +void fetch_sigfastblock(struct thread *td); +void fetch_sigfastblock_failed(struct thread *td, bool write); void gsignal(int pgid, int sig, ksiginfo_t *ksi); void killproc(struct proc *p, char *why); ksiginfo_t * ksiginfo_alloc(int wait); @@ -375,6 +394,7 @@ void pgsignal(struct pgrp *pgrp, int sig, int checkctty, ksiginfo_t *ksi); int postsig(int sig); void kern_psignal(struct proc *p, int sig); int ptracestop(struct thread *td, int sig, ksiginfo_t *si); +void reschedule_signals(struct proc *p, sigset_t block, int flags); void sendsig(sig_t catcher, ksiginfo_t *ksi, sigset_t *retmask); struct sigacts *sigacts_alloc(void); void sigacts_copy(struct sigacts *dest, struct sigacts *src);