Re: [PATCH 15/19] qht: Fix tsan warnings.

[Emilio G. Cota]

On Fri, May 22, 2020 at 12:07:51 -0400, Robert Foley wrote:
> For example:
> WARNING: ThreadSanitizer: data race (pid=23406)
>   Atomic read of size 4 at 0x7b100003e3c8 by thread T7:
>     #0 __tsan_atomic32_load <null> (qemu-system-aarch64+0x39a36c)
>     #1 qht_do_lookup util/qht.c:495:17 (qemu-system-aarch64+0xd82f7a)
>     #2 qht_lookup_custom util/qht.c:539:11 (qemu-system-aarch64+0xd82f7a)
>   Previous write of size 8 at 0x7b100003e3c8 by thread T6 (mutexes: write 
> M166769147697783108, write M995435858420506688):
>     #0 posix_memalign <null> (qemu-system-aarch64+0x350dd1)
>     #1 qemu_try_memalign util/oslib-posix.c:189:11 
> (qemu-system-aarch64+0xd59317)
>     #2 qemu_memalign util/oslib-posix.c:205:27 (qemu-system-aarch64+0xd5943e)
>     #3 qht_insert__locked util/qht.c:583:9 (qemu-system-aarch64+0xd837c5)
> 
> Signed-off-by: Robert Foley <address@hidden>
> ---
>  util/qht.c | 3 +++
>  1 file changed, 3 insertions(+)
> 
> diff --git a/util/qht.c b/util/qht.c
> index 67e5d5b916..739a53ced0 100644
> --- a/util/qht.c
> +++ b/util/qht.c
> @@ -69,6 +69,7 @@
>  #include "qemu/qht.h"
>  #include "qemu/atomic.h"
>  #include "qemu/rcu.h"
> +#include "qemu/tsan.h"
>  
>  //#define QHT_DEBUG
>  
> @@ -580,10 +581,12 @@ static void *qht_insert__locked(const struct qht *ht, 
> struct qht_map *map,
>          b = b->next;
>      } while (b);
>  
> +    TSAN_ANNOTATE_IGNORE_WRITES_BEGIN();
>      b = qemu_memalign(QHT_BUCKET_ALIGN, sizeof(*b));
>      memset(b, 0, sizeof(*b));
>      new = b;
>      i = 0;
> +    TSAN_ANNOTATE_IGNORE_WRITES_END();

I cannot reproduce this warning post-series with detect_deadlocks=0
but my hypothesis is that this is a side effect of tsan not understanding
the seqlock: tsan sees that below we "publish" this piece of memory with
an atomic write (in atomic_rcu_set), and does not see that with
seqlock_write_begin we have a write memory barrier. I wonder if
what we need instead is to annotate the seqlock functions, not the
callers.

Thanks,

                E.