They are implemented as a workqueue, which means that there are no guarantees of timing nor ordering.

Signed-off-by: Benjamin Tissoires bentiss@kernel.org

--- changes in v4: - dropped __bpf_timer_compute_key() - use a spin_lock instead of a semaphore - ensure bpf_timer_cancel_and_free is not complaining about non sleepable context and use cancel_work() instead of cancel_work_sync() - return -EINVAL if a delay is given to bpf_timer_start() with BPF_F_TIMER_SLEEPABLE

changes in v3: - extracted the implementation in bpf_timer only, without bpf_timer_set_sleepable_cb() - rely on schedule_work() only, from bpf_timer_start() - add semaphore to ensure bpf_timer_work_cb() is accessing consistent data

changes in v2 (compared to the one attaches to v1 0/9): - make use of a kfunc - add a (non-used) BPF_F_TIMER_SLEEPABLE - the callback is *not* called, it makes the kernel crashes --- include/uapi/linux/bpf.h | 4 +++ kernel/bpf/helpers.c | 86 ++++++++++++++++++++++++++++++++++++++++++++++-- 2 files changed, 88 insertions(+), 2 deletions(-)

diff --git a/include/uapi/linux/bpf.h b/include/uapi/linux/bpf.h index 3c42b9f1bada..b90def29d796 100644 --- a/include/uapi/linux/bpf.h +++ b/include/uapi/linux/bpf.h @@ -7461,10 +7461,14 @@ struct bpf_core_relo { * - BPF_F_TIMER_ABS: Timeout passed is absolute time, by default it is * relative to current time. * - BPF_F_TIMER_CPU_PIN: Timer will be pinned to the CPU of the caller. + * - BPF_F_TIMER_SLEEPABLE: Timer will run in a sleepable context, with + * no guarantees of ordering nor timing (consider this as being just + * offloaded immediately). */ enum { BPF_F_TIMER_ABS = (1ULL << 0), BPF_F_TIMER_CPU_PIN = (1ULL << 1), + BPF_F_TIMER_SLEEPABLE = (1ULL << 2), };

/* BPF numbers iterator state */ diff --git a/kernel/bpf/helpers.c b/kernel/bpf/helpers.c index a89587859571..38de73a9df83 100644 --- a/kernel/bpf/helpers.c +++ b/kernel/bpf/helpers.c @@ -1094,14 +1094,20 @@ const struct bpf_func_proto bpf_snprintf_proto = { * bpf_timer_cancel() cancels the timer and decrements prog's refcnt. * Inner maps can contain bpf timers as well. ops->map_release_uref is * freeing the timers when inner map is replaced or deleted by user space. + * + * sleepable_lock protects only the setup of the workqueue, not the callback + * itself. This is done to ensure we don't run concurrently a free of the + * callback or the associated program. */ struct bpf_hrtimer { struct hrtimer timer; + struct work_struct work; struct bpf_map *map; struct bpf_prog *prog; void __rcu *callback_fn; void *value; struct rcu_head rcu; + spinlock_t sleepable_lock; };

/* the actual struct hidden inside uapi struct bpf_timer */ @@ -1114,6 +1120,49 @@ struct bpf_timer_kern { struct bpf_spin_lock lock; } __attribute__((aligned(8)));

+static void bpf_timer_work_cb(struct work_struct *work) +{ + struct bpf_hrtimer *t = container_of(work, struct bpf_hrtimer, work); + struct bpf_map *map = t->map; + bpf_callback_t callback_fn; + void *value = t->value; + unsigned long flags; + void *key; + u32 idx; + + BTF_TYPE_EMIT(struct bpf_timer); + + spin_lock_irqsave(&t->sleepable_lock, flags); + + callback_fn = READ_ONCE(t->callback_fn); + if (!callback_fn) { + spin_unlock_irqrestore(&t->sleepable_lock, flags); + return; + } + + if (map->map_type == BPF_MAP_TYPE_ARRAY) { + struct bpf_array *array = container_of(map, struct bpf_array, map); + + /* compute the key */ + idx = ((char *)value - array->value) / array->elem_size; + key = &idx; + } else { /* hash or lru */ + key = value - round_up(map->key_size, 8); + } + + /* prevent the callback to be freed by bpf_timer_cancel() while running + * so we can release the sleepable lock + */ + bpf_prog_inc(t->prog); + + spin_unlock_irqrestore(&t->sleepable_lock, flags); + + callback_fn((u64)(long)map, (u64)(long)key, (u64)(long)value, 0, 0); + /* The verifier checked that return value is zero. */ + + bpf_prog_put(t->prog); +} + static DEFINE_PER_CPU(struct bpf_hrtimer *, hrtimer_running);

static enum hrtimer_restart bpf_timer_cb(struct hrtimer *hrtimer) @@ -1192,6 +1241,8 @@ BPF_CALL_3(bpf_timer_init, struct bpf_timer_kern *, timer, struct bpf_map *, map t->prog = NULL; rcu_assign_pointer(t->callback_fn, NULL); hrtimer_init(&t->timer, clockid, HRTIMER_MODE_REL_SOFT); + INIT_WORK(&t->work, bpf_timer_work_cb); + spin_lock_init(&t->sleepable_lock); t->timer.function = bpf_timer_cb; WRITE_ONCE(timer->timer, t); /* Guarantee the order between timer->timer and map->usercnt. So @@ -1237,6 +1288,7 @@ BPF_CALL_3(bpf_timer_set_callback, struct bpf_timer_kern *, timer, void *, callb ret = -EINVAL; goto out; } + spin_lock(&t->sleepable_lock); if (!atomic64_read(&t->map->usercnt)) { /* maps with timers must be either held by user space * or pinned in bpffs. Otherwise timer might still be @@ -1263,6 +1315,8 @@ BPF_CALL_3(bpf_timer_set_callback, struct bpf_timer_kern *, timer, void *, callb } rcu_assign_pointer(t->callback_fn, callback_fn); out: + if (t) + spin_unlock(&t->sleepable_lock); __bpf_spin_unlock_irqrestore(&timer->lock); return ret; } @@ -1283,8 +1337,12 @@ BPF_CALL_3(bpf_timer_start, struct bpf_timer_kern *, timer, u64, nsecs, u64, fla

if (in_nmi()) return -EOPNOTSUPP; - if (flags & ~(BPF_F_TIMER_ABS | BPF_F_TIMER_CPU_PIN)) + if (flags & ~(BPF_F_TIMER_ABS | BPF_F_TIMER_CPU_PIN | BPF_F_TIMER_SLEEPABLE)) return -EINVAL; + + if ((flags & BPF_F_TIMER_SLEEPABLE) && nsecs) + return -EINVAL; + __bpf_spin_lock_irqsave(&timer->lock); t = timer->timer; if (!t || !t->prog) { @@ -1300,7 +1358,10 @@ BPF_CALL_3(bpf_timer_start, struct bpf_timer_kern *, timer, u64, nsecs, u64, fla if (flags & BPF_F_TIMER_CPU_PIN) mode |= HRTIMER_MODE_PINNED;

- hrtimer_start(&t->timer, ns_to_ktime(nsecs), mode); + if (flags & BPF_F_TIMER_SLEEPABLE) + schedule_work(&t->work); + else + hrtimer_start(&t->timer, ns_to_ktime(nsecs), mode); out: __bpf_spin_unlock_irqrestore(&timer->lock); return ret; @@ -1348,13 +1409,22 @@ BPF_CALL_1(bpf_timer_cancel, struct bpf_timer_kern *, timer) ret = -EDEADLK; goto out; } + spin_lock(&t->sleepable_lock); drop_prog_refcnt(t); + spin_unlock(&t->sleepable_lock); out: __bpf_spin_unlock_irqrestore(&timer->lock); /* Cancel the timer and wait for associated callback to finish * if it was running. */ ret = ret ?: hrtimer_cancel(&t->timer); + + /* also cancel the sleepable work, but *do not* wait for + * it to finish if it was running as we might not be in a + * sleepable context + */ + ret = ret ?: cancel_work(&t->work); + rcu_read_unlock(); return ret; } @@ -1383,11 +1453,13 @@ void bpf_timer_cancel_and_free(void *val) t = timer->timer; if (!t) goto out; + spin_lock(&t->sleepable_lock); drop_prog_refcnt(t); /* The subsequent bpf_timer_start/cancel() helpers won't be able to use * this timer, since it won't be initialized. */ WRITE_ONCE(timer->timer, NULL); + spin_unlock(&t->sleepable_lock); out: __bpf_spin_unlock_irqrestore(&timer->lock); if (!t) @@ -1410,6 +1482,16 @@ void bpf_timer_cancel_and_free(void *val) */ if (this_cpu_read(hrtimer_running) != t) hrtimer_cancel(&t->timer); + + /* also cancel the sleepable work, but *do not* wait for + * it to finish if it was running as we might not be in a + * sleepable context. Same reason as above, it's fine to + * free 't': the subprog callback will never access it anymore + * and can not reschedule itself since timer->timer = NULL was + * already done. + */ + cancel_work(&t->work); + kfree_rcu(t, rcu); }

In this patch, bpf_timer_set_sleepable_cb() is functionally equivalent to bpf_timer_set_callback(), to the exception that it enforces the timer to be started with BPF_F_TIMER_SLEEPABLE.

But given that bpf_timer_set_callback() is a helper when bpf_timer_set_sleepable_cb() is a kfunc, we need to teach the verifier about its attached callback. Marking that callback as sleepable will be done in a separate patch

Signed-off-by: Benjamin Tissoires bentiss@kernel.org

---

changes in v4: - added a new (ignored) argument to the kfunc so that we do not need to wlak the stack

new in v3 (split from v2 02/10) --- kernel/bpf/helpers.c | 46 +++++++++++++++++++++++++++++++++++++-- kernel/bpf/verifier.c | 60 +++++++++++++++++++++++++++++++++++++++++++++++++-- 2 files changed, 102 insertions(+), 4 deletions(-)

diff --git a/kernel/bpf/helpers.c b/kernel/bpf/helpers.c index 38de73a9df83..65c07c0df263 100644 --- a/kernel/bpf/helpers.c +++ b/kernel/bpf/helpers.c @@ -1108,6 +1108,7 @@ struct bpf_hrtimer { void *value; struct rcu_head rcu; spinlock_t sleepable_lock; + bool is_sleepable; };

/* the actual struct hidden inside uapi struct bpf_timer */ @@ -1273,8 +1274,8 @@ static const struct bpf_func_proto bpf_timer_init_proto = { .arg3_type = ARG_ANYTHING, };

-BPF_CALL_3(bpf_timer_set_callback, struct bpf_timer_kern *, timer, void *, callback_fn, - struct bpf_prog_aux *, aux) +static int __bpf_timer_set_callback(struct bpf_timer_kern *timer, void *callback_fn, + struct bpf_prog_aux *aux, bool is_sleepable) { struct bpf_prog *prev, *prog = aux->prog; struct bpf_hrtimer *t; @@ -1314,6 +1315,7 @@ BPF_CALL_3(bpf_timer_set_callback, struct bpf_timer_kern *, timer, void *, callb t->prog = prog; } rcu_assign_pointer(t->callback_fn, callback_fn); + t->is_sleepable = is_sleepable; out: if (t) spin_unlock(&t->sleepable_lock); @@ -1321,6 +1323,12 @@ BPF_CALL_3(bpf_timer_set_callback, struct bpf_timer_kern *, timer, void *, callb return ret; }

+BPF_CALL_3(bpf_timer_set_callback, struct bpf_timer_kern *, timer, void *, callback_fn, + struct bpf_prog_aux *, aux) +{ + return __bpf_timer_set_callback(timer, callback_fn, aux, false); +} + static const struct bpf_func_proto bpf_timer_set_callback_proto = { .func = bpf_timer_set_callback, .gpl_only = true, @@ -1350,6 +1358,11 @@ BPF_CALL_3(bpf_timer_start, struct bpf_timer_kern *, timer, u64, nsecs, u64, fla goto out; }

+ if (t->is_sleepable && !(flags & BPF_F_TIMER_SLEEPABLE)) { + ret = -EINVAL; + goto out; + } + if (flags & BPF_F_TIMER_ABS) mode = HRTIMER_MODE_ABS_SOFT; else @@ -2627,6 +2640,34 @@ __bpf_kfunc void bpf_throw(u64 cookie) WARN(1, "A call to BPF exception callback should never return\n"); }

+/** + * bpf_timer_set_sleepable_cb_impl() - Configure the timer to call %callback_fn + * static function in a sleepable context. + * @timer: The bpf_timer that needs to be configured + * @callback_fn: a static bpf function + * + * @returns %0 on success. %-EINVAL if %timer was not initialized with + * bpf_timer_init() earlier. %-EPERM if %timer is in a map that doesn't + * have any user references. + * The user space should either hold a file descriptor to a map with timers + * or pin such map in bpffs. When map is unpinned or file descriptor is + * closed all timers in the map will be cancelled and freed. + * + * This kfunc is equivalent to %bpf_timer_set_callback except that it tells + * the verifier that the target callback is run in a sleepable context. + */ +__bpf_kfunc int bpf_timer_set_sleepable_cb_impl(struct bpf_timer_kern *timer, + int (callback_fn)(void *map, int *key, struct bpf_timer *timer), + void *aux__ign) +{ + struct bpf_prog_aux *aux = (struct bpf_prog_aux *)aux__ign; + + if (!aux) + return -EINVAL; + + return __bpf_timer_set_callback(timer, (void *)callback_fn, aux, true); +} + __bpf_kfunc_end_defs();

BTF_KFUNCS_START(generic_btf_ids) @@ -2703,6 +2744,7 @@ BTF_ID_FLAGS(func, bpf_dynptr_is_null) BTF_ID_FLAGS(func, bpf_dynptr_is_rdonly) BTF_ID_FLAGS(func, bpf_dynptr_size) BTF_ID_FLAGS(func, bpf_dynptr_clone) +BTF_ID_FLAGS(func, bpf_timer_set_sleepable_cb_impl) BTF_KFUNCS_END(common_btf_ids)

static const struct btf_kfunc_id_set common_kfunc_set = { diff --git a/kernel/bpf/verifier.c b/kernel/bpf/verifier.c index 1483ebc0ee73..53f85e114a33 100644 --- a/kernel/bpf/verifier.c +++ b/kernel/bpf/verifier.c @@ -501,8 +501,12 @@ static bool is_dynptr_ref_function(enum bpf_func_id func_id) }

static bool is_sync_callback_calling_kfunc(u32 btf_id); +static bool is_async_callback_calling_kfunc(u32 btf_id); +static bool is_callback_calling_kfunc(u32 btf_id); static bool is_bpf_throw_kfunc(struct bpf_insn *insn);

+static bool is_bpf_timer_set_sleepable_cb_impl_kfunc(u32 btf_id); + static bool is_sync_callback_calling_function(enum bpf_func_id func_id) { return func_id == BPF_FUNC_for_each_map_elem || @@ -530,7 +534,8 @@ static bool is_sync_callback_calling_insn(struct bpf_insn *insn)

static bool is_async_callback_calling_insn(struct bpf_insn *insn) { - return bpf_helper_call(insn) && is_async_callback_calling_function(insn->imm); + return (bpf_helper_call(insn) && is_async_callback_calling_function(insn->imm)) || + (bpf_pseudo_kfunc_call(insn) && is_async_callback_calling_kfunc(insn->imm)); }

static bool is_may_goto_insn(struct bpf_insn *insn) @@ -9471,7 +9476,7 @@ static int push_callback_call(struct bpf_verifier_env *env, struct bpf_insn *ins */ env->subprog_info[subprog].is_cb = true; if (bpf_pseudo_kfunc_call(insn) && - !is_sync_callback_calling_kfunc(insn->imm)) { + !is_callback_calling_kfunc(insn->imm)) { verbose(env, "verifier bug: kfunc %s#%d not marked as callback-calling\n", func_id_name(insn->imm), insn->imm); return -EFAULT; @@ -10981,6 +10986,7 @@ enum special_kfunc_type { KF_bpf_percpu_obj_drop_impl, KF_bpf_throw, KF_bpf_iter_css_task_new, + KF_bpf_timer_set_sleepable_cb_impl, };

BTF_SET_START(special_kfunc_set) @@ -11007,6 +11013,7 @@ BTF_ID(func, bpf_throw) #ifdef CONFIG_CGROUPS BTF_ID(func, bpf_iter_css_task_new) #endif +BTF_ID(func, bpf_timer_set_sleepable_cb_impl) BTF_SET_END(special_kfunc_set)

BTF_ID_LIST(special_kfunc_list) @@ -11037,6 +11044,7 @@ BTF_ID(func, bpf_iter_css_task_new) #else BTF_ID_UNUSED #endif +BTF_ID(func, bpf_timer_set_sleepable_cb_impl)

static bool is_kfunc_ret_null(struct bpf_kfunc_call_arg_meta *meta) { @@ -11365,12 +11373,28 @@ static bool is_sync_callback_calling_kfunc(u32 btf_id) return btf_id == special_kfunc_list[KF_bpf_rbtree_add_impl]; }

+static bool is_async_callback_calling_kfunc(u32 btf_id) +{ + return btf_id == special_kfunc_list[KF_bpf_timer_set_sleepable_cb_impl]; +} + static bool is_bpf_throw_kfunc(struct bpf_insn *insn) { return bpf_pseudo_kfunc_call(insn) && insn->off == 0 && insn->imm == special_kfunc_list[KF_bpf_throw]; }

+static bool is_bpf_timer_set_sleepable_cb_impl_kfunc(u32 btf_id) +{ + return btf_id == special_kfunc_list[KF_bpf_timer_set_sleepable_cb_impl]; +} + +static bool is_callback_calling_kfunc(u32 btf_id) +{ + return is_sync_callback_calling_kfunc(btf_id) || + is_async_callback_calling_kfunc(btf_id); +} + static bool is_rbtree_lock_required_kfunc(u32 btf_id) { return is_bpf_rbtree_api_kfunc(btf_id); @@ -12151,6 +12175,16 @@ static int check_kfunc_call(struct bpf_verifier_env *env, struct bpf_insn *insn, } }

+ if (is_async_callback_calling_kfunc(meta.func_id)) { + err = push_callback_call(env, insn, insn_idx, meta.subprogno, + set_timer_callback_state); + if (err) { + verbose(env, "kfunc %s#%d failed callback verification\n", + func_name, meta.func_id); + return err; + } + } + rcu_lock = is_kfunc_bpf_rcu_read_lock(&meta); rcu_unlock = is_kfunc_bpf_rcu_read_unlock(&meta);

@@ -19544,6 +19578,28 @@ static int fixup_kfunc_call(struct bpf_verifier_env *env, struct bpf_insn *insn, desc->func_id == special_kfunc_list[KF_bpf_rdonly_cast]) { insn_buf[0] = BPF_MOV64_REG(BPF_REG_0, BPF_REG_1); *cnt = 1; + } else if (is_bpf_timer_set_sleepable_cb_impl_kfunc(desc->func_id)) { + /* The verifier will process callback_fn as many times as necessary + * with different maps and the register states prepared by + * set_timer_callback_state will be accurate. + * + * The following use case is valid: + * map1 is shared by prog1, prog2, prog3. + * prog1 calls bpf_timer_init for some map1 elements + * prog2 calls bpf_timer_set_callback for some map1 elements. + * Those that were not bpf_timer_init-ed will return -EINVAL. + * prog3 calls bpf_timer_start for some map1 elements. + * Those that were not both bpf_timer_init-ed and + * bpf_timer_set_callback-ed will return -EINVAL. + */ + struct bpf_insn ld_addrs[2] = { + BPF_LD_IMM64(BPF_REG_3, (long)env->prog->aux), + }; + + insn_buf[0] = ld_addrs[0]; + insn_buf[1] = ld_addrs[1]; + insn_buf[2] = *insn; + *cnt = 3; } return 0; }

cp include/uapi/linux/bpf.h tools/include/uapi/linux/bpf.h

Signed-off-by: Benjamin Tissoires bentiss@kernel.org

---

new in v4 --- tools/include/uapi/linux/bpf.h | 4 ++++ 1 file changed, 4 insertions(+)

diff --git a/tools/include/uapi/linux/bpf.h b/tools/include/uapi/linux/bpf.h index 3c42b9f1bada..b90def29d796 100644 --- a/tools/include/uapi/linux/bpf.h +++ b/tools/include/uapi/linux/bpf.h @@ -7461,10 +7461,14 @@ struct bpf_core_relo { * - BPF_F_TIMER_ABS: Timeout passed is absolute time, by default it is * relative to current time. * - BPF_F_TIMER_CPU_PIN: Timer will be pinned to the CPU of the caller. + * - BPF_F_TIMER_SLEEPABLE: Timer will run in a sleepable context, with + * no guarantees of ordering nor timing (consider this as being just + * offloaded immediately). */ enum { BPF_F_TIMER_ABS = (1ULL << 0), BPF_F_TIMER_CPU_PIN = (1ULL << 1), + BPF_F_TIMER_SLEEPABLE = (1ULL << 2), };

/* BPF numbers iterator state */