Re: [Qemu-devel] [RFC 7/8] util/qht: atomically set b->hashes
On 19/09/2016 21:06, Emilio G. Cota wrote: > Let me then just point out that this comes at a small perf loss. > > Running 'taskset -c 0 tests/qht-bench -n 1 -d 10' (i.e. all lookups) 10 times, > we get: > > before the patch: > $ ./mean.pl 34.04 34.24 34.38 34.25 34.18 34.51 34.46 34.44 34.29 34.08 > 34.287 +- 0.160072900059109 > after: > $ ./mean.pl 33.94 34.00 33.52 33.46 33.55 33.71 34.27 34.06 34.28 34.58 > 33.937 +- 0.374731014640279 > > But hey we can live with that. Hmm it shouldn't. I'll take a look at the generated assembly... Paolo
Re: [Qemu-devel] [RFC 7/8] util/qht: atomically set b->hashes
On Mon, Sep 19, 2016 at 20:37:06 +0200, Paolo Bonzini wrote: > On 19/09/2016 20:06, Emilio G. Cota wrote: > > On Mon, Sep 19, 2016 at 16:51:38 +0100, Alex Bennée wrote: > >> > ThreadSanitizer detects a possible race between reading/writing the > >> > hashes. As ordering semantics are already documented for qht we just > >> > need to ensure a race can't tear the hash value so we can use the > >> > relaxed atomic_set/read functions. > > This was discussed here: > > > > https://lists.gnu.org/archive/html/qemu-devel/2016-05/msg03658.html > > > > To reiterate: reading torn hash values is fine, since the retry will > > happen regardless (and all pointers[] remain valid through the RCU > > read-critical section). > > True, but C11 says data races are undefined, not merely unspecified. > seqlock-protected data requires a relaxed read and write, because they > are read concurrently in the read and write sides. Ah I see. Let me then just point out that this comes at a small perf loss. Running 'taskset -c 0 tests/qht-bench -n 1 -d 10' (i.e. all lookups) 10 times, we get: before the patch: $ ./mean.pl 34.04 34.24 34.38 34.25 34.18 34.51 34.46 34.44 34.29 34.08 34.287 +- 0.160072900059109 after: $ ./mean.pl 33.94 34.00 33.52 33.46 33.55 33.71 34.27 34.06 34.28 34.58 33.937 +- 0.374731014640279 But hey we can live with that. Cheers, E.
Re: [Qemu-devel] [RFC 7/8] util/qht: atomically set b->hashes
On 19/09/2016 20:06, Emilio G. Cota wrote: > On Mon, Sep 19, 2016 at 16:51:38 +0100, Alex Bennée wrote: >> > ThreadSanitizer detects a possible race between reading/writing the >> > hashes. As ordering semantics are already documented for qht we just >> > need to ensure a race can't tear the hash value so we can use the >> > relaxed atomic_set/read functions. > This was discussed here: > > https://lists.gnu.org/archive/html/qemu-devel/2016-05/msg03658.html > > To reiterate: reading torn hash values is fine, since the retry will > happen regardless (and all pointers[] remain valid through the RCU > read-critical section). True, but C11 says data races are undefined, not merely unspecified. seqlock-protected data requires a relaxed read and write, because they are read concurrently in the read and write sides. Paolo
Re: [Qemu-devel] [RFC 7/8] util/qht: atomically set b->hashes
On Mon, Sep 19, 2016 at 16:51:38 +0100, Alex Bennée wrote: > ThreadSanitizer detects a possible race between reading/writing the > hashes. As ordering semantics are already documented for qht we just > need to ensure a race can't tear the hash value so we can use the > relaxed atomic_set/read functions. This was discussed here: https://lists.gnu.org/archive/html/qemu-devel/2016-05/msg03658.html To reiterate: reading torn hash values is fine, since the retry will happen regardless (and all pointers[] remain valid through the RCU read-critical section). Couldn't we just tell tsan to ignore it? Thanks, Emilio
[Qemu-devel] [RFC 7/8] util/qht: atomically set b->hashes
ThreadSanitizer detects a possible race between reading/writing the hashes. As ordering semantics are already documented for qht we just need to ensure a race can't tear the hash value so we can use the relaxed atomic_set/read functions. Signed-off-by: Alex Bennée --- util/qht.c | 10 +- 1 file changed, 5 insertions(+), 5 deletions(-) diff --git a/util/qht.c b/util/qht.c index 16a8d79..571639d 100644 --- a/util/qht.c +++ b/util/qht.c @@ -379,7 +379,7 @@ static void qht_bucket_reset__locked(struct qht_bucket *head) if (b->pointers[i] == NULL) { goto done; } -b->hashes[i] = 0; +atomic_set(&b->hashes[i], 0); atomic_set(&b->pointers[i], NULL); } b = b->next; @@ -444,7 +444,7 @@ void *qht_do_lookup(struct qht_bucket *head, qht_lookup_func_t func, do { for (i = 0; i < QHT_BUCKET_ENTRIES; i++) { -if (b->hashes[i] == hash) { +if (atomic_read(&b->hashes[i]) == hash) { /* The pointer is dereferenced before seqlock_read_retry, * so (unlike qht_insert__locked) we need to use * atomic_rcu_read here. @@ -538,8 +538,8 @@ static bool qht_insert__locked(struct qht *ht, struct qht_map *map, if (new) { atomic_rcu_set(&prev->next, b); } -b->hashes[i] = hash; /* smp_wmb() implicit in seqlock_write_begin. */ +atomic_set(&b->hashes[i], hash); atomic_set(&b->pointers[i], p); seqlock_write_end(&head->sequence); return true; @@ -607,10 +607,10 @@ qht_entry_move(struct qht_bucket *to, int i, struct qht_bucket *from, int j) qht_debug_assert(to->pointers[i]); qht_debug_assert(from->pointers[j]); -to->hashes[i] = from->hashes[j]; +atomic_set(&to->hashes[i], from->hashes[j]); atomic_set(&to->pointers[i], from->pointers[j]); -from->hashes[j] = 0; +atomic_set(&from->hashes[j], 0); atomic_set(&from->pointers[j], NULL); } -- 2.9.3