The branch main has been updated by kib: URL: https://cgit.FreeBSD.org/src/commit/?id=9ef425e560a97cabd1862e803eeb48468f89de18
commit 9ef425e560a97cabd1862e803eeb48468f89de18 Author: Konstantin Belousov <[email protected]> AuthorDate: 2023-08-24 23:16:02 +0000 Commit: Konstantin Belousov <[email protected]> CommitDate: 2024-08-06 04:05:58 +0000 rangelocks: add fast cheating mode Reviewed by: markj, Olivier Certner <[email protected]> Tested by: pho Sponsored by: The FreeBSD Foundation Differential revision: https://reviews.freebsd.org/D41787 --- sys/kern/kern_rangelock.c | 239 ++++++++++++++++++++++++++++++++++++++++++++-- sys/sys/rangelock.h | 2 +- 2 files changed, 234 insertions(+), 7 deletions(-) diff --git a/sys/kern/kern_rangelock.c b/sys/kern/kern_rangelock.c index 2f16569b19aa..5ec443d3199c 100644 --- a/sys/kern/kern_rangelock.c +++ b/sys/kern/kern_rangelock.c @@ -35,9 +35,223 @@ #include <sys/rangelock.h> #include <sys/sleepqueue.h> #include <sys/smr.h> +#include <sys/sysctl.h> #include <vm/uma.h> +/* + * Immediately after initialization (subject to 'rangelock_cheat' + * below), and until a request comes that conflicts with granted ones + * based on type, rangelocks serve requests in the "cheating" mode. + * In this mode, a rangelock behaves like a sxlock, as if each request + * covered the whole range of the protected object. On receiving a + * conflicting request (any request while a write request is + * effective, or any write request while some read ones are + * effective), all requests granted in "cheating" mode are drained, + * and the rangelock then switches to effectively keeping track of the + * precise range of each new request. + * + * Normal sx implementation is not used to not bloat structures (most + * important, vnodes) which embeds rangelocks. + * + * The cheating greatly helps very common pattern where file is first + * written single-threaded, and then read by many processes. + * + * Lock is in cheat mode when RL_CHEAT_CHEATING bit is set in the + * lock->head. Special cookies are returned in this mode, and + * trylocks are same as normal locks but do not drain. + */ + +static int rangelock_cheat = 1; +SYSCTL_INT(_debug, OID_AUTO, rangelock_cheat, CTLFLAG_RWTUN, + &rangelock_cheat, 0, + ""); + +#define RL_CHEAT_MASK 0x7 +#define RL_CHEAT_CHEATING 0x1 +/* #define RL_CHEAT_RLOCKED 0x0 */ +#define RL_CHEAT_WLOCKED 0x2 +#define RL_CHEAT_DRAINING 0x4 + +#define RL_CHEAT_READER 0x8 + +#define RL_RET_CHEAT_RLOCKED 0x1100 +#define RL_RET_CHEAT_WLOCKED 0x2200 + +static bool +rangelock_cheat_lock(struct rangelock *lock, int locktype, bool trylock, + void **cookie) +{ + uintptr_t v, x; + + v = (uintptr_t)atomic_load_ptr(&lock->head); + if ((v & RL_CHEAT_CHEATING) == 0) + return (false); + if ((v & RL_CHEAT_DRAINING) != 0) { +drain: + if (trylock) { + *cookie = NULL; + return (true); + } + sleepq_lock(&lock->head); +drain1: + DROP_GIANT(); + for (;;) { + v = (uintptr_t)atomic_load_ptr(&lock->head); + if ((v & RL_CHEAT_DRAINING) == 0) + break; + sleepq_add(&lock->head, NULL, "ranged1", 0, 0); + sleepq_wait(&lock->head, PRI_USER); + sleepq_lock(&lock->head); + } + sleepq_release(&lock->head); + PICKUP_GIANT(); + return (false); + } + + switch (locktype) { + case RL_LOCK_READ: + for (;;) { + if ((v & RL_CHEAT_WLOCKED) != 0) { + if (trylock) { + *cookie = NULL; + return (true); + } + x = v | RL_CHEAT_DRAINING; + sleepq_lock(&lock->head); + if (atomic_fcmpset_rel_ptr(&lock->head, &v, + x) != 0) + goto drain1; + sleepq_release(&lock->head); + /* Possibly forgive passed conflict */ + continue; + } + x = (v & ~RL_CHEAT_MASK) + RL_CHEAT_READER; + x |= RL_CHEAT_CHEATING; + if (atomic_fcmpset_acq_ptr(&lock->head, &v, x) != 0) + break; + if ((v & RL_CHEAT_CHEATING) == 0) + return (false); + if ((v & RL_CHEAT_DRAINING) != 0) + goto drain; + } + *(uintptr_t *)cookie = RL_RET_CHEAT_RLOCKED; + break; + case RL_LOCK_WRITE: + for (;;) { + if ((v & ~RL_CHEAT_MASK) >= RL_CHEAT_READER || + (v & RL_CHEAT_WLOCKED) != 0) { + if (trylock) { + *cookie = NULL; + return (true); + } + x = v | RL_CHEAT_DRAINING; + sleepq_lock(&lock->head); + if (atomic_fcmpset_rel_ptr(&lock->head, &v, + x) != 0) + goto drain1; + sleepq_release(&lock->head); + /* Possibly forgive passed conflict */ + continue; + } + x = RL_CHEAT_WLOCKED | RL_CHEAT_CHEATING; + if (atomic_fcmpset_acq_ptr(&lock->head, &v, x) != 0) + break; + if ((v & RL_CHEAT_CHEATING) == 0) + return (false); + if ((v & RL_CHEAT_DRAINING) != 0) + goto drain; + } + *(uintptr_t *)cookie = RL_RET_CHEAT_WLOCKED; + break; + default: + __assert_unreachable(); + break; + } + return (true); +} + +static bool +rangelock_cheat_unlock(struct rangelock *lock, void *cookie) +{ + uintptr_t v, x; + + v = (uintptr_t)atomic_load_ptr(&lock->head); + if ((v & RL_CHEAT_CHEATING) == 0) + return (false); + + MPASS((uintptr_t)cookie == RL_RET_CHEAT_WLOCKED || + (uintptr_t)cookie == RL_RET_CHEAT_RLOCKED); + + switch ((uintptr_t)cookie) { + case RL_RET_CHEAT_RLOCKED: + for (;;) { + MPASS((v & ~RL_CHEAT_MASK) >= RL_CHEAT_READER); + MPASS((v & RL_CHEAT_WLOCKED) == 0); + x = (v & ~RL_CHEAT_MASK) - RL_CHEAT_READER; + if ((v & RL_CHEAT_DRAINING) != 0) { + if (x != 0) { + x |= RL_CHEAT_DRAINING | + RL_CHEAT_CHEATING; + if (atomic_fcmpset_rel_ptr(&lock->head, + &v, x) != 0) + break; + } else { + sleepq_lock(&lock->head); + if (atomic_fcmpset_rel_ptr(&lock->head, + &v, x) != 0) { + sleepq_broadcast( + &lock->head, + SLEEPQ_SLEEP, 0, 0); + sleepq_release(&lock->head); + break; + } + sleepq_release(&lock->head); + } + } else { + x |= RL_CHEAT_CHEATING; + if (atomic_fcmpset_rel_ptr(&lock->head, &v, + x) != 0) + break; + } + } + break; + case RL_RET_CHEAT_WLOCKED: + for (;;) { + MPASS((v & RL_CHEAT_WLOCKED) != 0); + if ((v & RL_CHEAT_DRAINING) != 0) { + sleepq_lock(&lock->head); + atomic_store_ptr(&lock->head, 0); + sleepq_broadcast(&lock->head, + SLEEPQ_SLEEP, 0, 0); + sleepq_release(&lock->head); + break; + } else { + if (atomic_fcmpset_ptr(&lock->head, &v, + RL_CHEAT_CHEATING) != 0) + break; + } + } + break; + default: + __assert_unreachable(); + break; + } + return (true); +} + +static bool +rangelock_cheat_destroy(struct rangelock *lock) +{ + uintptr_t v; + + v = (uintptr_t)atomic_load_ptr(&lock->head); + if ((v & RL_CHEAT_CHEATING) == 0) + return (false); + MPASS(v == RL_CHEAT_CHEATING); + return (true); +} + /* * Implementation of range locks based on the paper * https://doi.org/10.1145/3342195.3387533 @@ -112,16 +326,17 @@ void rangelock_init(struct rangelock *lock) { lock->sleepers = false; - atomic_store_ptr(&lock->head, NULL); + atomic_store_ptr(&lock->head, rangelock_cheat ? RL_CHEAT_CHEATING : 0); } void rangelock_destroy(struct rangelock *lock) { struct rl_q_entry *e, *ep; - struct thread *td; MPASS(!lock->sleepers); + if (rangelock_cheat_destroy(lock)) + return; for (e = (struct rl_q_entry *)atomic_load_ptr(&lock->head); e != NULL; e = rl_e_unmark(ep)) { ep = atomic_load_ptr(&e->rl_q_next); @@ -190,6 +405,9 @@ rangelock_unlock_int(struct rangelock *lock, struct rl_q_entry *e) void rangelock_unlock(struct rangelock *lock, void *cookie) { + if (rangelock_cheat_unlock(lock, cookie)) + return; + sleepq_lock(&lock->sleepers); rangelock_unlock_int(lock, cookie); sleepq_release(&lock->sleepers); @@ -297,7 +515,7 @@ rl_w_validate(struct rangelock *lock, struct rl_q_entry *e, { struct rl_q_entry *cur, *next, **prev; - prev = &lock->head; + prev = (struct rl_q_entry **)&lock->head; cur = rl_q_load(prev); MPASS(!rl_e_is_marked(cur)); /* head is not marked */ for (;;) { @@ -337,7 +555,7 @@ rl_insert(struct rangelock *lock, struct rl_q_entry *e, bool trylock, int r; again: - prev = &lock->head; + prev = (struct rl_q_entry **)&lock->head; cur = rl_q_load(prev); if (cur == NULL && rl_q_cas(prev, NULL, e)) return (RL_LOCK_SUCCESS); @@ -401,8 +619,11 @@ rangelock_lock_int(struct rangelock *lock, bool trylock, vm_ooffset_t start, { struct rl_q_entry *e, *free, *x, *xp; struct thread *td; + void *cookie; enum RL_INSERT_RES res; + if (rangelock_cheat_lock(lock, locktype, trylock, &cookie)) + return (cookie); td = curthread; for (res = RL_LOCK_RETRY; res == RL_LOCK_RETRY;) { free = NULL; @@ -446,7 +667,7 @@ rangelock_tryrlock(struct rangelock *lock, vm_ooffset_t start, vm_ooffset_t end) void * rangelock_wlock(struct rangelock *lock, vm_ooffset_t start, vm_ooffset_t end) { - return (rangelock_lock_int(lock, true, start, end, RL_LOCK_WRITE)); + return (rangelock_lock_int(lock, false, start, end, RL_LOCK_WRITE)); } void * @@ -470,6 +691,7 @@ DB_SHOW_COMMAND(rangelock, db_show_rangelock) { struct rangelock *lock; struct rl_q_entry *e, *x; + uintptr_t v; if (!have_addr) { db_printf("show rangelock addr\n"); @@ -478,7 +700,12 @@ DB_SHOW_COMMAND(rangelock, db_show_rangelock) lock = (struct rangelock *)addr; db_printf("rangelock %p sleepers %d\n", lock, lock->sleepers); - for (e = lock->head;;) { + v = lock->head; + if ((v & RL_CHEAT_CHEATING) != 0) { + db_printf(" cheating head %#jx\n", (uintmax_t)v); + return; + } + for (e = (struct rl_q_entry *)(lock->head);;) { x = rl_e_is_marked(e) ? rl_e_unmark(e) : e; if (x == NULL) break; diff --git a/sys/sys/rangelock.h b/sys/sys/rangelock.h index 60f394b67677..127f101ddc2e 100644 --- a/sys/sys/rangelock.h +++ b/sys/sys/rangelock.h @@ -48,7 +48,7 @@ struct rl_q_entry; * owners are for read. */ struct rangelock { - struct rl_q_entry *head; + uintptr_t head; bool sleepers; };
