From 39e6c8208d7b6fb9d2047850fb3327db567b564b Mon Sep 17 00:00:00 2001 From: Eric Dumazet Date: Tue, 28 Feb 2017 10:34:50 -0800 Subject: net: solve a NAPI race While playing with mlx4 hardware timestamping of RX packets, I found that some packets were received by TCP stack with a ~200 ms delay... Since the timestamp was provided by the NIC, and my probe was added in tcp_v4_rcv() while in BH handler, I was confident it was not a sender issue, or a drop in the network. This would happen with a very low probability, but hurting RPC workloads. A NAPI driver normally arms the IRQ after the napi_complete_done(), after NAPI_STATE_SCHED is cleared, so that the hard irq handler can grab it. Problem is that if another point in the stack grabs NAPI_STATE_SCHED bit while IRQ are not disabled, we might have later an IRQ firing and finding this bit set, right before napi_complete_done() clears it. This can happen with busy polling users, or if gro_flush_timeout is used. But some other uses of napi_schedule() in drivers can cause this as well. thread 1 thread 2 (could be on same cpu, or not) // busy polling or napi_watchdog() napi_schedule(); ... napi->poll() device polling: read 2 packets from ring buffer Additional 3rd packet is available. device hard irq // does nothing because NAPI_STATE_SCHED bit is owned by thread 1 napi_schedule(); napi_complete_done(napi, 2); rearm_irq(); Note that rearm_irq() will not force the device to send an additional IRQ for the packet it already signaled (3rd packet in my example) This patch adds a new NAPI_STATE_MISSED bit, that napi_schedule_prep() can set if it could not grab NAPI_STATE_SCHED Then napi_complete_done() properly reschedules the napi to make sure we do not miss something. Since we manipulate multiple bits at once, use cmpxchg() like in sk_busy_loop() to provide proper transactions. In v2, I changed napi_watchdog() to use a relaxed variant of napi_schedule_prep() : No need to set NAPI_STATE_MISSED from this point. In v3, I added more details in the changelog and clears NAPI_STATE_MISSED in busy_poll_stop() In v4, I added the ideas given by Alexander Duyck in v3 review Signed-off-by: Eric Dumazet Cc: Alexander Duyck Signed-off-by: David S. Miller --- net/core/dev.c | 76 ++++++++++++++++++++++++++++++++++++++++++++++++++++++---- 1 file changed, 72 insertions(+), 4 deletions(-) (limited to 'net/core/dev.c') diff --git a/net/core/dev.c b/net/core/dev.c index 304f2deae5f9..e63bf61b19be 100644 --- a/net/core/dev.c +++ b/net/core/dev.c @@ -4883,6 +4883,39 @@ void __napi_schedule(struct napi_struct *n) } EXPORT_SYMBOL(__napi_schedule); +/** + * napi_schedule_prep - check if napi can be scheduled + * @n: napi context + * + * Test if NAPI routine is already running, and if not mark + * it as running. This is used as a condition variable + * insure only one NAPI poll instance runs. We also make + * sure there is no pending NAPI disable. + */ +bool napi_schedule_prep(struct napi_struct *n) +{ + unsigned long val, new; + + do { + val = READ_ONCE(n->state); + if (unlikely(val & NAPIF_STATE_DISABLE)) + return false; + new = val | NAPIF_STATE_SCHED; + + /* Sets STATE_MISSED bit if STATE_SCHED was already set + * This was suggested by Alexander Duyck, as compiler + * emits better code than : + * if (val & NAPIF_STATE_SCHED) + * new |= NAPIF_STATE_MISSED; + */ + new |= (val & NAPIF_STATE_SCHED) / NAPIF_STATE_SCHED * + NAPIF_STATE_MISSED; + } while (cmpxchg(&n->state, val, new) != val); + + return !(val & NAPIF_STATE_SCHED); +} +EXPORT_SYMBOL(napi_schedule_prep); + /** * __napi_schedule_irqoff - schedule for receive * @n: entry to schedule @@ -4897,7 +4930,7 @@ EXPORT_SYMBOL(__napi_schedule_irqoff); bool napi_complete_done(struct napi_struct *n, int work_done) { - unsigned long flags; + unsigned long flags, val, new; /* * 1) Don't let napi dequeue from the cpu poll list @@ -4927,7 +4960,27 @@ bool napi_complete_done(struct napi_struct *n, int work_done) list_del_init(&n->poll_list); local_irq_restore(flags); } - WARN_ON_ONCE(!test_and_clear_bit(NAPI_STATE_SCHED, &n->state)); + + do { + val = READ_ONCE(n->state); + + WARN_ON_ONCE(!(val & NAPIF_STATE_SCHED)); + + new = val & ~(NAPIF_STATE_MISSED | NAPIF_STATE_SCHED); + + /* If STATE_MISSED was set, leave STATE_SCHED set, + * because we will call napi->poll() one more time. + * This C code was suggested by Alexander Duyck to help gcc. + */ + new |= (val & NAPIF_STATE_MISSED) / NAPIF_STATE_MISSED * + NAPIF_STATE_SCHED; + } while (cmpxchg(&n->state, val, new) != val); + + if (unlikely(val & NAPIF_STATE_MISSED)) { + __napi_schedule(n); + return false; + } + return true; } EXPORT_SYMBOL(napi_complete_done); @@ -4953,6 +5006,16 @@ static void busy_poll_stop(struct napi_struct *napi, void *have_poll_lock) { int rc; + /* Busy polling means there is a high chance device driver hard irq + * could not grab NAPI_STATE_SCHED, and that NAPI_STATE_MISSED was + * set in napi_schedule_prep(). + * Since we are about to call napi->poll() once more, we can safely + * clear NAPI_STATE_MISSED. + * + * Note: x86 could use a single "lock and ..." instruction + * to perform these two clear_bit() + */ + clear_bit(NAPI_STATE_MISSED, &napi->state); clear_bit(NAPI_STATE_IN_BUSY_POLL, &napi->state); local_bh_disable(); @@ -5088,8 +5151,13 @@ static enum hrtimer_restart napi_watchdog(struct hrtimer *timer) struct napi_struct *napi; napi = container_of(timer, struct napi_struct, timer); - if (napi->gro_list) - napi_schedule_irqoff(napi); + + /* Note : we use a relaxed variant of napi_schedule_prep() not setting + * NAPI_STATE_MISSED, since we do not react to a device IRQ. + */ + if (napi->gro_list && !napi_disable_pending(napi) && + !test_and_set_bit(NAPI_STATE_SCHED, &napi->state)) + __napi_schedule_irqoff(napi); return HRTIMER_NORESTART; } -- cgit v1.2.3-70-g09d2 From 13baa00ad01bb3a9f893e3a08cbc2d072fc0c15d Mon Sep 17 00:00:00 2001 From: Eric Dumazet Date: Wed, 1 Mar 2017 14:28:39 -0800 Subject: net: net_enable_timestamp() can be called from irq contexts It is now very clear that silly TCP listeners might play with enabling/disabling timestamping while new children are added to their accept queue. Meaning net_enable_timestamp() can be called from BH context while current state of the static key is not enabled. Lets play safe and allow all contexts. The work queue is scheduled only under the problematic cases, which are the static key enable/disable transition, to not slow down critical paths. This extends and improves what we did in commit 5fa8bbda38c6 ("net: use a work queue to defer net_disable_timestamp() work") Fixes: b90e5794c5bd ("net: dont call jump_label_dec from irq context") Signed-off-by: Eric Dumazet Reported-by: Dmitry Vyukov Signed-off-by: David S. Miller --- net/core/dev.c | 35 +++++++++++++++++++++++++++++++---- 1 file changed, 31 insertions(+), 4 deletions(-) (limited to 'net/core/dev.c') diff --git a/net/core/dev.c b/net/core/dev.c index e63bf61b19be..8637b2b71f3d 100644 --- a/net/core/dev.c +++ b/net/core/dev.c @@ -1698,27 +1698,54 @@ EXPORT_SYMBOL_GPL(net_dec_egress_queue); static struct static_key netstamp_needed __read_mostly; #ifdef HAVE_JUMP_LABEL static atomic_t netstamp_needed_deferred; +static atomic_t netstamp_wanted; static void netstamp_clear(struct work_struct *work) { int deferred = atomic_xchg(&netstamp_needed_deferred, 0); + int wanted; - while (deferred--) - static_key_slow_dec(&netstamp_needed); + wanted = atomic_add_return(deferred, &netstamp_wanted); + if (wanted > 0) + static_key_enable(&netstamp_needed); + else + static_key_disable(&netstamp_needed); } static DECLARE_WORK(netstamp_work, netstamp_clear); #endif void net_enable_timestamp(void) { +#ifdef HAVE_JUMP_LABEL + int wanted; + + while (1) { + wanted = atomic_read(&netstamp_wanted); + if (wanted <= 0) + break; + if (atomic_cmpxchg(&netstamp_wanted, wanted, wanted + 1) == wanted) + return; + } + atomic_inc(&netstamp_needed_deferred); + schedule_work(&netstamp_work); +#else static_key_slow_inc(&netstamp_needed); +#endif } EXPORT_SYMBOL(net_enable_timestamp); void net_disable_timestamp(void) { #ifdef HAVE_JUMP_LABEL - /* net_disable_timestamp() can be called from non process context */ - atomic_inc(&netstamp_needed_deferred); + int wanted; + + while (1) { + wanted = atomic_read(&netstamp_wanted); + if (wanted <= 1) + break; + if (atomic_cmpxchg(&netstamp_wanted, wanted, wanted - 1) == wanted) + return; + } + atomic_dec(&netstamp_needed_deferred); schedule_work(&netstamp_work); #else static_key_slow_dec(&netstamp_needed); -- cgit v1.2.3-70-g09d2