* Re: [dpdk-stable] [dpdk-dev] [PATCH v3 1/5] event/octeontx2: fix TIM HW race condition
[not found] <20191122154431.17416-1-pbhagavatula@marvell.com>
@ 2019-11-22 16:07 ` Pavan Nikhilesh Bhagavatula
2019-11-23 9:00 ` Jerin Jacob
0 siblings, 1 reply; 2+ messages in thread
From: Pavan Nikhilesh Bhagavatula @ 2019-11-22 16:07 UTC (permalink / raw)
To: Pavan Nikhilesh Bhagavatula, Jerin Jacob Kollanukkaran; +Cc: dev, dpdk stable
+Cc: stable@dpdk.org
>-----Original Message-----
>From: pbhagavatula@marvell.com <pbhagavatula@marvell.com>
>Sent: Friday, November 22, 2019 9:14 PM
>To: Jerin Jacob Kollanukkaran <jerinj@marvell.com>; Pavan Nikhilesh
>Bhagavatula <pbhagavatula@marvell.com>
>Cc: dev@dpdk.org
>Subject: [dpdk-dev] [PATCH v3 1/5] event/octeontx2: fix TIM HW race
>condition
>
>From: Pavan Nikhilesh <pbhagavatula@marvell.com>
>
>Fix HW race condition observed when timeout resolution is low (<5us).
>When HW traverses a given TIM bucket it will clear chunk_remainder,
>but since SW always decreases the chunk_remainder at the start of the
>arm routine it might cause a race where SW updates chunk_remainder
>after HW has cleared it that lead to nasty side effects.
>
>Fixes: 95e4e4ec7469 ("event/octeontx2: add timer arm timeout burst")
>
>Signed-off-by: Pavan Nikhilesh <pbhagavatula@marvell.com>
>---
> drivers/event/octeontx2/otx2_tim_worker.h | 141
>+++++++++++++++++++---
> 1 file changed, 124 insertions(+), 17 deletions(-)
>
>diff --git a/drivers/event/octeontx2/otx2_tim_worker.h
>b/drivers/event/octeontx2/otx2_tim_worker.h
>index 50db6543c..c896b5433 100644
>--- a/drivers/event/octeontx2/otx2_tim_worker.h
>+++ b/drivers/event/octeontx2/otx2_tim_worker.h
>@@ -7,6 +7,13 @@
>
> #include "otx2_tim_evdev.h"
>
>+static inline uint8_t
>+tim_bkt_fetch_lock(uint64_t w1)
>+{
>+ return (w1 >> TIM_BUCKET_W1_S_LOCK) &
>+ TIM_BUCKET_W1_M_LOCK;
>+}
>+
> static inline int16_t
> tim_bkt_fetch_rem(uint64_t w1)
> {
>@@ -188,7 +195,6 @@ tim_insert_chunk(struct otx2_tim_bkt * const
>bkt,
> } else {
> bkt->first_chunk = (uintptr_t)chunk;
> }
>-
> return chunk;
> }
>
>@@ -208,11 +214,38 @@ tim_add_entry_sp(struct otx2_tim_ring *
>const tim_ring,
>
> __retry:
> /* Get Bucket sema*/
>- lock_sema = tim_bkt_fetch_sema(bkt);
>+ lock_sema = tim_bkt_fetch_sema_lock(bkt);
>
> /* Bucket related checks. */
>- if (unlikely(tim_bkt_get_hbt(lock_sema)))
>- goto __retry;
>+ if (unlikely(tim_bkt_get_hbt(lock_sema))) {
>+ if (tim_bkt_get_nent(lock_sema) != 0) {
>+ uint64_t hbt_state;
>+#ifdef RTE_ARCH_ARM64
>+ asm volatile(
>+ " ldaxr %[hbt], [%[w1]]
> \n"
>+ " tbz %[hbt], 33, dne%=
> \n"
>+ " sevl
> \n"
>+ "rty%=: wfe
> \n"
>+ " ldaxr %[hbt], [%[w1]]
> \n"
>+ " tbnz %[hbt], 33, rty%=
> \n"
>+ "dne%=:
> \n"
>+ : [hbt] "=&r" (hbt_state)
>+ : [w1] "r" ((&bkt->w1))
>+ : "memory"
>+ );
>+#else
>+ do {
>+ hbt_state = __atomic_load_n(&bkt-
>>w1,
>+ __ATOMIC_ACQUIRE);
>+ } while (hbt_state & BIT_ULL(33));
>+#endif
>+
>+ if (!(hbt_state & BIT_ULL(34))) {
>+ tim_bkt_dec_lock(bkt);
>+ goto __retry;
>+ }
>+ }
>+ }
>
> /* Insert the work. */
> rem = tim_bkt_fetch_rem(lock_sema);
>@@ -224,14 +257,15 @@ tim_add_entry_sp(struct otx2_tim_ring *
>const tim_ring,
> chunk = tim_insert_chunk(bkt, tim_ring);
>
> if (unlikely(chunk == NULL)) {
>- tim_bkt_set_rem(bkt, 0);
>+ bkt->chunk_remainder = 0;
>+ tim_bkt_dec_lock(bkt);
> tim->impl_opaque[0] = 0;
> tim->impl_opaque[1] = 0;
> tim->state = RTE_EVENT_TIMER_ERROR;
> return -ENOMEM;
> }
> bkt->current_chunk = (uintptr_t)chunk;
>- tim_bkt_set_rem(bkt, tim_ring->nb_chunk_slots - 1);
>+ bkt->chunk_remainder = tim_ring->nb_chunk_slots - 1;
> } else {
> chunk = (struct otx2_tim_ent *)(uintptr_t)bkt-
>>current_chunk;
> chunk += tim_ring->nb_chunk_slots - rem;
>@@ -241,6 +275,7 @@ tim_add_entry_sp(struct otx2_tim_ring * const
>tim_ring,
> *chunk = *pent;
>
> tim_bkt_inc_nent(bkt);
>+ tim_bkt_dec_lock(bkt);
>
> tim->impl_opaque[0] = (uintptr_t)chunk;
> tim->impl_opaque[1] = (uintptr_t)bkt;
>@@ -263,19 +298,60 @@ tim_add_entry_mp(struct otx2_tim_ring *
>const tim_ring,
>
> __retry:
> bkt = tim_get_target_bucket(tim_ring, rel_bkt, flags);
>-
> /* Get Bucket sema*/
> lock_sema = tim_bkt_fetch_sema_lock(bkt);
>
> /* Bucket related checks. */
> if (unlikely(tim_bkt_get_hbt(lock_sema))) {
>- tim_bkt_dec_lock(bkt);
>- goto __retry;
>+ if (tim_bkt_get_nent(lock_sema) != 0) {
>+ uint64_t hbt_state;
>+#ifdef RTE_ARCH_ARM64
>+ asm volatile(
>+ " ldaxr %[hbt], [%[w1]]
> \n"
>+ " tbz %[hbt], 33, dne%=
> \n"
>+ " sevl
> \n"
>+ "rty%=: wfe
> \n"
>+ " ldaxr %[hbt], [%[w1]]
> \n"
>+ " tbnz %[hbt], 33, rty%=
> \n"
>+ "dne%=:
> \n"
>+ : [hbt] "=&r" (hbt_state)
>+ : [w1] "r" ((&bkt->w1))
>+ : "memory"
>+ );
>+#else
>+ do {
>+ hbt_state = __atomic_load_n(&bkt-
>>w1,
>+ __ATOMIC_ACQUIRE);
>+ } while (hbt_state & BIT_ULL(33));
>+#endif
>+
>+ if (!(hbt_state & BIT_ULL(34))) {
>+ tim_bkt_dec_lock(bkt);
>+ goto __retry;
>+ }
>+ }
> }
>
> rem = tim_bkt_fetch_rem(lock_sema);
>-
> if (rem < 0) {
>+#ifdef RTE_ARCH_ARM64
>+ asm volatile(
>+ " ldaxrh %w[rem], [%[crem]]
> \n"
>+ " tbz %w[rem], 15, dne%=
> \n"
>+ " sevl
> \n"
>+ "rty%=: wfe
> \n"
>+ " ldaxrh %w[rem], [%[crem]]
> \n"
>+ " tbnz %w[rem], 15, rty%=
> \n"
>+ "dne%=:
> \n"
>+ : [rem] "=&r" (rem)
>+ : [crem] "r" (&bkt->chunk_remainder)
>+ : "memory"
>+ );
>+#else
>+ while (__atomic_load_n(&bkt->chunk_remainder,
>+ __ATOMIC_ACQUIRE) < 0)
>+ ;
>+#endif
> /* Goto diff bucket. */
> tim_bkt_dec_lock(bkt);
> goto __retry;
>@@ -294,17 +370,23 @@ tim_add_entry_mp(struct otx2_tim_ring *
>const tim_ring,
> tim->state = RTE_EVENT_TIMER_ERROR;
> return -ENOMEM;
> }
>- bkt->current_chunk = (uintptr_t)chunk;
>- tim_bkt_set_rem(bkt, tim_ring->nb_chunk_slots - 1);
>+ *chunk = *pent;
>+ while (tim_bkt_fetch_lock(lock_sema) !=
>+ (-tim_bkt_fetch_rem(lock_sema)))
>+ lock_sema = __atomic_load_n(&bkt->w1,
>__ATOMIC_ACQUIRE);
>+
>+ bkt->current_chunk = (uintptr_t)chunk;
>+ __atomic_store_n(&bkt->chunk_remainder,
>+ tim_ring->nb_chunk_slots - 1,
>__ATOMIC_RELEASE);
> } else {
>- chunk = (struct otx2_tim_ent *)(uintptr_t)bkt-
>>current_chunk;
>+ chunk = (struct otx2_tim_ent *)bkt->current_chunk;
> chunk += tim_ring->nb_chunk_slots - rem;
>+ *chunk = *pent;
> }
>
> /* Copy work entry. */
>- *chunk = *pent;
>- tim_bkt_dec_lock(bkt);
> tim_bkt_inc_nent(bkt);
>+ tim_bkt_dec_lock(bkt);
> tim->impl_opaque[0] = (uintptr_t)chunk;
> tim->impl_opaque[1] = (uintptr_t)bkt;
> tim->state = RTE_EVENT_TIMER_ARMED;
>@@ -360,8 +442,33 @@ tim_add_entry_brst(struct otx2_tim_ring *
>const tim_ring,
>
> /* Bucket related checks. */
> if (unlikely(tim_bkt_get_hbt(lock_sema))) {
>- tim_bkt_dec_lock(bkt);
>- goto __retry;
>+ if (tim_bkt_get_nent(lock_sema) != 0) {
>+ uint64_t hbt_state;
>+#ifdef RTE_ARCH_ARM64
>+ asm volatile(
>+ " ldaxr %[hbt], [%[w1]]
> \n"
>+ " tbz %[hbt], 33, dne%=
> \n"
>+ " sevl
> \n"
>+ "rty%=: wfe
> \n"
>+ " ldaxr %[hbt], [%[w1]]
> \n"
>+ " tbnz %[hbt], 33, rty%=
> \n"
>+ "dne%=:
> \n"
>+ : [hbt] "=&r" (hbt_state)
>+ : [w1] "r" ((&bkt->w1))
>+ : "memory"
>+ );
>+#else
>+ do {
>+ hbt_state = __atomic_load_n(&bkt-
>>w1,
>+ __ATOMIC_ACQUIRE);
>+ } while (hbt_state & BIT_ULL(33));
>+#endif
>+
>+ if (!(hbt_state & BIT_ULL(34))) {
>+ tim_bkt_dec_lock(bkt);
>+ goto __retry;
>+ }
>+ }
> }
>
> chunk_remainder = tim_bkt_fetch_rem(lock_sema);
>--
>2.17.1
^ permalink raw reply [flat|nested] 2+ messages in thread
* Re: [dpdk-stable] [dpdk-dev] [PATCH v3 1/5] event/octeontx2: fix TIM HW race condition
2019-11-22 16:07 ` [dpdk-stable] [dpdk-dev] [PATCH v3 1/5] event/octeontx2: fix TIM HW race condition Pavan Nikhilesh Bhagavatula
@ 2019-11-23 9:00 ` Jerin Jacob
0 siblings, 0 replies; 2+ messages in thread
From: Jerin Jacob @ 2019-11-23 9:00 UTC (permalink / raw)
To: Pavan Nikhilesh Bhagavatula; +Cc: Jerin Jacob Kollanukkaran, dev, dpdk stable
On Sat, Nov 23, 2019 at 1:07 AM Pavan Nikhilesh Bhagavatula
<pbhagavatula@marvell.com> wrote:
>
> +Cc: stable@dpdk.org
>
> >-----Original Message-----
> >From: pbhagavatula@marvell.com <pbhagavatula@marvell.com>
> >Sent: Friday, November 22, 2019 9:14 PM
> >To: Jerin Jacob Kollanukkaran <jerinj@marvell.com>; Pavan Nikhilesh
> >Bhagavatula <pbhagavatula@marvell.com>
> >Cc: dev@dpdk.org
> >Subject: [dpdk-dev] [PATCH v3 1/5] event/octeontx2: fix TIM HW race
> >condition
> >
> >From: Pavan Nikhilesh <pbhagavatula@marvell.com>
> >
> >Fix HW race condition observed when timeout resolution is low (<5us).
> >When HW traverses a given TIM bucket it will clear chunk_remainder,
> >but since SW always decreases the chunk_remainder at the start of the
> >arm routine it might cause a race where SW updates chunk_remainder
> >after HW has cleared it that lead to nasty side effects.
> >
> >Fixes: 95e4e4ec7469 ("event/octeontx2: add timer arm timeout burst")
> >
> >Signed-off-by: Pavan Nikhilesh <pbhagavatula@marvell.com>
Series applied to dpdk-next-eventdev/master. Thanks.
^ permalink raw reply [flat|nested] 2+ messages in thread
end of thread, other threads:[~2019-11-23 9:00 UTC | newest]
Thread overview: 2+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
[not found] <20191122154431.17416-1-pbhagavatula@marvell.com>
2019-11-22 16:07 ` [dpdk-stable] [dpdk-dev] [PATCH v3 1/5] event/octeontx2: fix TIM HW race condition Pavan Nikhilesh Bhagavatula
2019-11-23 9:00 ` Jerin Jacob
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).