Index: kern/exec_elf.c =================================================================== RCS file: /cvs/src/sys/kern/exec_elf.c,v diff -u -p -r1.191 exec_elf.c --- kern/exec_elf.c 15 Sep 2024 23:13:19 -0000 1.191 +++ kern/exec_elf.c 27 Sep 2024 04:46:07 -0000 @@ -1023,6 +1023,12 @@ exec_elf_fixup(struct proc *p, struct ex a->au_v = p->p_p->ps_timekeep; a++; +#ifdef MULTIPROCESSOR + a->au_id = AUX_openbsd_ncpus; + a->au_v = ncpus; + a++; +#endif + a->au_id = AUX_null; a->au_v = 0; a++; Index: kern/sys_futex.c =================================================================== RCS file: /cvs/src/sys/kern/sys_futex.c,v diff -u -p -r1.22 sys_futex.c --- kern/sys_futex.c 14 Aug 2023 07:42:34 -0000 1.22 +++ kern/sys_futex.c 27 Sep 2024 04:46:07 -0000 @@ -36,41 +36,56 @@ * Kernel representation of a futex. */ struct futex { - LIST_ENTRY(futex) ft_list; /* list of all futexes */ - TAILQ_HEAD(, proc) ft_threads; /* sleeping queue */ + TAILQ_ENTRY(futex) ft_entry; /* list of all futexes */ + struct process *ft_ps; struct uvm_object *ft_obj; /* UVM object */ struct vm_amap *ft_amap; /* UVM amap */ voff_t ft_off; /* UVM offset */ - unsigned int ft_refcnt; /* # of references */ + + volatile unsigned int ft_wait; }; -/* Syscall helpers. */ -int futex_wait(uint32_t *, uint32_t, const struct timespec *, int); -int futex_wake(uint32_t *, uint32_t, int); -int futex_requeue(uint32_t *, uint32_t, uint32_t *, uint32_t, int); - -/* Flags for futex_get(). */ -#define FT_CREATE 0x1 /* Create a futex if it doesn't exist. */ -#define FT_PRIVATE 0x2 /* Futex is process-private. */ +TAILQ_HEAD(futexen, futex); -struct futex *futex_get(uint32_t *, int); -void futex_put(struct futex *); +struct futex_bucket { + struct futexen fb_list; + struct rwlock fb_lock; + uint32_t fb_id; /* for lock ordering */ +} __aligned(64); -/* - * The global futex lock serializes futex(2) calls so that no wakeup - * event is lost, and protects all futex lists and futex states. - */ -struct rwlock ftlock = RWLOCK_INITIALIZER("futex"); -static struct futex_list ftlist_shared = - LIST_HEAD_INITIALIZER(ftlist_shared); -struct pool ftpool; +/* Syscall helpers. */ +static int futex_wait(struct proc *, uint32_t *, uint32_t, + const struct timespec *, int); +static int futex_wake(struct proc *, uint32_t *, uint32_t, int, + register_t *); +static int futex_requeue(struct proc *, uint32_t *, uint32_t, + uint32_t *, uint32_t, int, register_t *); + +/* Flags for futex_get(). kernel private flags sit in FUTEX_OP_MASK space */ +#define FT_PRIVATE FUTEX_PRIVATE_FLAG /* Futex is process-private. */ + +#define FUTEX_BUCKET_BITS 6 +#define FUTEX_BUCKET_SIZE (1U << FUTEX_BUCKET_BITS) +#define FUTEX_BUCKET_MASK (FUTEX_BUCKET_SIZE - 1) +static struct futex_bucket futex_hash[FUTEX_BUCKET_SIZE]; void futex_init(void) { - pool_init(&ftpool, sizeof(struct futex), 0, IPL_NONE, - PR_WAITOK | PR_RWLOCK, "futexpl", NULL); + struct futex_bucket *fb; + unsigned int i; + + for (i = 0; i < nitems(futex_hash); i++) { + fb = &futex_hash[i]; + + TAILQ_INIT(&fb->fb_list); + rw_init(&fb->fb_lock, "futexlk"); + + fb->fb_id = arc4random(); + fb->fb_id &= ~FUTEX_BUCKET_MASK; + fb->fb_id |= i; + } } int @@ -88,65 +103,51 @@ sys_futex(struct proc *p, void *v, regis uint32_t val = SCARG(uap, val); const struct timespec *timeout = SCARG(uap, timeout); void *g = SCARG(uap, g); - int flags = 0; + int flags = op & FUTEX_FLAG_MASK; int error = 0; - if (op & FUTEX_PRIVATE_FLAG) - flags |= FT_PRIVATE; - - rw_enter_write(&ftlock); - switch (op) { + switch (op & FUTEX_OP_MASK) { case FUTEX_WAIT: - case FUTEX_WAIT_PRIVATE: - error = futex_wait(uaddr, val, timeout, flags); + error = futex_wait(p, uaddr, val, timeout, flags); break; case FUTEX_WAKE: - case FUTEX_WAKE_PRIVATE: - *retval = futex_wake(uaddr, val, flags); + error = futex_wake(p, uaddr, val, flags, retval); break; case FUTEX_REQUEUE: - case FUTEX_REQUEUE_PRIVATE: - *retval = futex_requeue(uaddr, val, g, (u_long)timeout, flags); + error = futex_requeue(p, uaddr, val, g, + (u_long)timeout, flags, retval); break; default: error = ENOSYS; break; } - rw_exit_write(&ftlock); return error; } -/* - * Return an existing futex matching userspace address ``uaddr''. - * - * If such futex does not exist and FT_CREATE is given, create it. - */ -struct futex * -futex_get(uint32_t *uaddr, int flags) +static void +futex_addrs(struct proc *p, struct futex *f, uint32_t *uaddr, int flags) { - struct proc *p = curproc; vm_map_t map = &p->p_vmspace->vm_map; vm_map_entry_t entry; struct uvm_object *obj = NULL; struct vm_amap *amap = NULL; voff_t off = (vaddr_t)uaddr; - struct futex *f; - struct futex_list *ftlist = &p->p_p->ps_ftlist; + struct process *ps; - rw_assert_wrlock(&ftlock); + if (ISSET(flags, FT_PRIVATE)) + ps = p->p_p; + else { + ps = NULL; - if (!(flags & FT_PRIVATE)) { vm_map_lock_read(map); if (uvm_map_lookup_entry(map, (vaddr_t)uaddr, &entry) && entry->inheritance == MAP_INHERIT_SHARE) { if (UVM_ET_ISOBJ(entry)) { - ftlist = &ftlist_shared; obj = entry->object.uvm_obj; off = entry->offset + ((vaddr_t)uaddr - entry->start); } else if (entry->aref.ar_amap) { - ftlist = &ftlist_shared; amap = entry->aref.ar_amap; off = ptoa(entry->aref.ar_pageoff) + ((vaddr_t)uaddr - entry->start); @@ -155,47 +156,18 @@ futex_get(uint32_t *uaddr, int flags) vm_map_unlock_read(map); } - LIST_FOREACH(f, ftlist, ft_list) { - if (f->ft_obj == obj && f->ft_amap == amap && - f->ft_off == off) { - f->ft_refcnt++; - break; - } - } - - if ((f == NULL) && (flags & FT_CREATE)) { - /* - * We rely on the rwlock to ensure that no other thread - * create the same futex. - */ - f = pool_get(&ftpool, PR_WAITOK); - TAILQ_INIT(&f->ft_threads); - f->ft_obj = obj; - f->ft_amap = amap; - f->ft_off = off; - f->ft_refcnt = 1; - LIST_INSERT_HEAD(ftlist, f, ft_list); - } - - return f; + f->ft_ps = ps; + f->ft_obj = obj; + f->ft_amap = amap; + f->ft_off = off; } -/* - * Release a given futex. - */ -void -futex_put(struct futex *f) +static inline struct futex_bucket * +futex_get_bucket(struct futex *f) { - rw_assert_wrlock(&ftlock); + uint32_t key = f->ft_off >> 3; /* watevs */ - KASSERT(f->ft_refcnt > 0); - - --f->ft_refcnt; - if (f->ft_refcnt == 0) { - KASSERT(TAILQ_EMPTY(&f->ft_threads)); - LIST_REMOVE(f, ft_list); - pool_put(&ftpool, f); - } + return (&futex_hash[key & FUTEX_BUCKET_MASK]); } /* @@ -203,69 +175,79 @@ futex_put(struct futex *f) * ``uaddr''. Let it sleep for the specified ``timeout'' time, or * indefinitely if the argument is NULL. */ -int -futex_wait(uint32_t *uaddr, uint32_t val, const struct timespec *timeout, - int flags) +static int +futex_wait(struct proc *p, uint32_t *uaddr, uint32_t val, + const struct timespec *timeout, int flags) { - struct proc *p = curproc; - struct futex *f; + struct futex f; + struct futex_bucket *fb; uint64_t nsecs = INFSLP; uint32_t cval; int error; - /* - * After reading the value a race is still possible but - * we deal with it by serializing all futex syscalls. - */ - rw_assert_wrlock(&ftlock); - - /* - * Read user space futex value - */ - if ((error = copyin32(uaddr, &cval))) - return error; - - /* If the value changed, stop here. */ - if (cval != val) - return EAGAIN; - if (timeout != NULL) { struct timespec ts; - if ((error = copyin(timeout, &ts, sizeof(ts)))) - return error; + error = copyin(timeout, &ts, sizeof(ts)); + if (error != 0) + return (error); #ifdef KTRACE if (KTRPOINT(p, KTR_STRUCT)) ktrreltimespec(p, &ts); #endif if (ts.tv_sec < 0 || !timespecisvalid(&ts)) - return EINVAL; + return (EINVAL); + nsecs = MAX(1, MIN(TIMESPEC_TO_NSEC(&ts), MAXTSLP)); } - f = futex_get(uaddr, flags | FT_CREATE); - TAILQ_INSERT_TAIL(&f->ft_threads, p, p_fut_link); - p->p_futex = f; - - error = rwsleep_nsec(p, &ftlock, PWAIT|PCATCH, "fsleep", nsecs); - if (error == ERESTART) - error = ECANCELED; - else if (error == EWOULDBLOCK) { - /* A race occurred between a wakeup and a timeout. */ - if (p->p_futex == NULL) - error = 0; - else - error = ETIMEDOUT; - } + futex_addrs(p, &f, uaddr, flags); + fb = futex_get_bucket(&f); + + /* + * After reading the value a race is still possible but + * we deal with it by serializing futex syscalls. + */ + error = rw_enter(&fb->fb_lock, RW_WRITE|RW_INTR); + if (error != 0) + return (error); - /* Remove ourself if we haven't been awaken. */ - if ((f = p->p_futex) != NULL) { - p->p_futex = NULL; - TAILQ_REMOVE(&f->ft_threads, p, p_fut_link); - futex_put(f); + /* + * Read user space futex value + */ + error = copyin32(uaddr, &cval); + if (error != 0) + goto exit; + + /* If the value changed, stop here. */ + if (cval != val) { + error = EAGAIN; + goto exit; } - return error; + TAILQ_INSERT_TAIL(&fb->fb_list, &f, ft_entry); + f.ft_wait = 1; + do { + error = rwsleep_nsec(&f, &fb->fb_lock, + PWAIT|PCATCH, "fsleep", nsecs); + if (error != 0) { + switch (error) { + case ERESTART: + error = ECANCELED; + break; + case EWOULDBLOCK: + error = f.ft_wait ? ETIMEDOUT : 0; + break; + } + + break; + } + } while (f.ft_wait); + TAILQ_REMOVE(&fb->fb_list, &f, ft_entry); + +exit: + rw_exit_write(&fb->fb_lock); + return (error); } /* @@ -273,46 +255,131 @@ futex_wait(uint32_t *uaddr, uint32_t val * ``uaddr'' and requeue at most ``m'' sibling threads on a futex at * address ``uaddr2''. */ -int -futex_requeue(uint32_t *uaddr, uint32_t n, uint32_t *uaddr2, uint32_t m, - int flags) +static int +futex_requeue(struct proc *p, uint32_t *uaddr, uint32_t n, + uint32_t *uaddr2, uint32_t m, int flags, register_t *retval) { - struct futex *f, *g; - struct proc *p; + struct futex okey, nkey; + struct futex *f, *nf = NULL; + struct futex_bucket *ofb, *nfb; uint32_t count = 0; - rw_assert_wrlock(&ftlock); + if (m > INT_MAX) + return (EINVAL); + if (m == 0) + return (futex_wake(p, uaddr, n, flags, retval)); + + futex_addrs(p, &okey, uaddr, flags); + ofb = futex_get_bucket(&okey); + futex_addrs(p, &nkey, uaddr2, flags); + nfb = futex_get_bucket(&nkey); + + if (ofb->fb_id < nfb->fb_id) { + rw_enter_write(&ofb->fb_lock); + rw_enter_write(&nfb->fb_lock); + } else if (ofb->fb_id > nfb->fb_id) { + rw_enter_write(&nfb->fb_lock); + rw_enter_write(&ofb->fb_lock); + } else + rw_enter_write(&ofb->fb_lock); + + TAILQ_FOREACH(f, &ofb->fb_list, ft_entry) { + if (f->ft_off != okey.ft_off || + f->ft_ps != okey.ft_ps || + f->ft_obj != okey.ft_obj || + f->ft_amap != okey.ft_amap) + continue; - f = futex_get(uaddr, flags); - if (f == NULL) - return 0; - - while ((p = TAILQ_FIRST(&f->ft_threads)) != NULL && (count < (n + m))) { - p->p_futex = NULL; - TAILQ_REMOVE(&f->ft_threads, p, p_fut_link); - futex_put(f); - - if (count < n) { - wakeup_one(p); - } else if (uaddr2 != NULL) { - g = futex_get(uaddr2, FT_CREATE); - TAILQ_INSERT_TAIL(&g->ft_threads, p, p_fut_link); - p->p_futex = g; + f->ft_wait = 0; + wakeup_one(f); + + if (++count == n) { + nf = TAILQ_NEXT(f, ft_entry); + break; } - count++; } - futex_put(f); + /* move matching futexes to the new bucket */ + while (nf != NULL) { + f = nf; + nf = TAILQ_NEXT(f, ft_entry); + + if (f->ft_off != okey.ft_off || + f->ft_ps != okey.ft_ps || + f->ft_obj != okey.ft_obj || + f->ft_amap != okey.ft_amap) + continue; + + TAILQ_REMOVE(&ofb->fb_list, f, ft_entry); + /* it should only be ft_off that changes, but eh */ + f->ft_ps = nkey.ft_ps; + f->ft_obj = nkey.ft_obj; + f->ft_amap = nkey.ft_amap; + f->ft_off = nkey.ft_off; + TAILQ_INSERT_TAIL(&nfb->fb_list, f, ft_entry); - return count; + if (--m == 0) + break; + } + + if (ofb->fb_id < nfb->fb_id) { + rw_exit_write(&nfb->fb_lock); + rw_exit_write(&ofb->fb_lock); + } else if (ofb->fb_id > nfb->fb_id) { + rw_exit_write(&ofb->fb_lock); + rw_exit_write(&nfb->fb_lock); + } else + rw_exit_write(&ofb->fb_lock); + + *retval = count; + + return (0); } /* * Wakeup at most ``n'' sibling threads sleeping on a futex at address * ``uaddr''. */ -int -futex_wake(uint32_t *uaddr, uint32_t n, int flags) +static int +futex_wake(struct proc *p, uint32_t *uaddr, uint32_t n, int flags, + register_t *retval) { - return futex_requeue(uaddr, n, NULL, 0, flags); + struct futex key; + struct futex *f; + struct futex_bucket *fb; + int count = 0; + int error; + + if (n > INT_MAX) + return (EINVAL); + if (n == 0) { + *retval = 0; + return (0); + } + + futex_addrs(p, &key, uaddr, flags); + fb = futex_get_bucket(&key); + + error = rw_enter(&fb->fb_lock, RW_READ|RW_INTR); + if (error != 0) + return (error); + + TAILQ_FOREACH(f, &fb->fb_list, ft_entry) { + if (f->ft_off != key.ft_off || + f->ft_ps != key.ft_ps || + f->ft_obj != key.ft_obj || + f->ft_amap != key.ft_amap) + continue; + + f->ft_wait = 0; + wakeup_one(f); + + if (++count == n) + break; + } + + rw_exit_read(&fb->fb_lock); + + *retval = count; + return (0); } Index: sys/exec_elf.h =================================================================== RCS file: /cvs/src/sys/sys/exec_elf.h,v diff -u -p -r1.105 exec_elf.h --- sys/exec_elf.h 14 Jul 2024 09:48:49 -0000 1.105 +++ sys/exec_elf.h 27 Sep 2024 04:46:07 -0000 @@ -734,6 +734,7 @@ enum AuxID { AUX_sun_gid = 2002, /* egid */ AUX_sun_rgid = 2003, /* rgid */ AUX_openbsd_timekeep = 4000, /* userland clock_gettime */ + AUX_openbsd_ncpus = 4001, /* ncpus */ }; struct elf_args { @@ -822,7 +823,7 @@ extern Elf_Dyn _DYNAMIC[]; /* * How many entries are in the AuxInfo array we pass to the process? */ -#define ELF_AUX_ENTRIES 11 +#define ELF_AUX_ENTRIES 12 #define ELF_AUX_WORDS (sizeof(AuxInfo) * ELF_AUX_ENTRIES / sizeof(char *)) struct exec_package; Index: sys/futex.h =================================================================== RCS file: /cvs/src/sys/sys/futex.h,v diff -u -p -r1.2 futex.h --- sys/futex.h 3 Jun 2018 15:09:26 -0000 1.2 +++ sys/futex.h 27 Sep 2024 04:46:07 -0000 @@ -28,11 +28,15 @@ int futex(volatile uint32_t *, int, int, __END_DECLS #endif /* ! _KERNEL */ +#define FUTEX_OP_MASK 0x007f + #define FUTEX_WAIT 1 #define FUTEX_WAKE 2 #define FUTEX_REQUEUE 3 -#define FUTEX_PRIVATE_FLAG 128 +#define FUTEX_FLAG_MASK 0xff80 + +#define FUTEX_PRIVATE_FLAG 0x0080 #define FUTEX_WAIT_PRIVATE (FUTEX_WAIT | FUTEX_PRIVATE_FLAG) #define FUTEX_WAKE_PRIVATE (FUTEX_WAKE | FUTEX_PRIVATE_FLAG)