On 2/11/20 1:01 PM, Pavel Begunkov wrote: > On 10/02/2020 23:56, Jens Axboe wrote: >> As mentioned in the previous email, here are the three patches that add >> support for multiple waitqueues for polling with io_uring. >> >> Patches 1-2 are just basic prep patches, and should not have any >> functional changes in them. Patch 3 adds support for allocating a new >> io_poll_iocb unit if we get multiple additions through our queue proc >> for the wait queues. This new 'poll' addition is queued up as well, and >> it grabs a reference to the original poll request. >> >> Please do review, would love to get this (long standing) issue fixed as >> it's a real problem for various folks. >> > > I need to dig a bit deeper into poll to understand some moments, but there is a > question: don't we won't to support arbitrary number of waitqueues then? In theory, probably. But in practice, don't think anything exists with > 2 waitqueues. As mentioned, even the 2 waitqueue case is limited to less than a handful of users. But the patch should for sure check this, and -EINVAL if we get a third entry attempted. -- Jens Axboe
