Re: [Qemu-devel] [QEMU-devel][PATCH v4 0/2] Fix concurrent aio_poll/set_fd_handler.

[Paolo Bonzini <pbonzini@redhat.com>]

[-- Attachment #1: Type: text/plain, Size: 1483 bytes --]

On 20/12/18 17:37, Stefan Hajnoczi wrote:
> On Thu, Dec 20, 2018 at 04:20:28PM +0100, remy.noel@blade-group.com wrote:
>> From: Remy Noel <remy.noel@blade-group.com>
>>
>> It is possible for an io_poll/read/write callback to be concurrently executed along
>> with an aio_set_fd_handlers. This can cause all sorts of problems, like
>> a NULL callback or a bad opaque pointer.
>>
>> V2:
>>     * Do not use RCU anymore as it inccurs a performance loss
>> V3:
>>     * Don't drop revents when a handler is modified [Stefan]
>> V4:
>>     * Unregister fd from ctx epoll when removing fd_handler [Paolo]
>>
>> Remy Noel (2):
>>   aio-posix: Unregister fd from ctx epoll when removing fd_handler.
>>   aio-posix: Fix concurrent aio_poll/set_fd_handler.
>>
>>  util/aio-posix.c | 90 +++++++++++++++++++++++++++++-------------------
>>  util/aio-win32.c | 67 ++++++++++++++++-------------------
>>  2 files changed, 84 insertions(+), 73 deletions(-)
>>
>> -- 
>> 2.19.2
>>
> 
> Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
Reviewed-by: Paolo Bonzini <pbonzini@redhat.com>

FWIW, I had missed the early version that used RCU, but lockcnt is
already very RCU-like, so not using RCU is the right thing to do.  The
difference between lockcnt and RCU is that cleanup is done by the reader
instead of a separate thread.  Because we know that reader/writer
concurrency is very rare for AioContext handlers, the tradeoffs favor
lockcnt over RCU.

Paolo


[-- Attachment #2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 488 bytes --]