Hi Christoph On 4/9/19 5:37 PM, Christoph Hellwig wrote: > On Tue, Apr 09, 2019 at 01:21:54PM +0800, Jianchao Wang wrote: >> For the IORING_SETUP_IOPOLL case, all of the submission and >> completion are handled under ctx->uring_lock or in SQ poll thread >> context, so io_get_req and io_put_req has been serialized well. >> The only exception is the asynchronous workqueue context where could >> free the io_kiocb for error. To overcome this, we allocate a new >> io_kiocb and free the previous inlined one. >> >> Based on this, we introduce the preallocated reqs list per ctx and >> needn't to provide any lock to serialize the updating of list. The >> performacne benefits from this. The test result of following fio >> command > > I really don't like the idea of exposing this to userspace. Is > there any good reason to not simply always allocate inline request > up to a certain ring size? > Sorry, I cannot get your point. There is nothing exposed to userspace. We will try to allocated fixed 128 per-ctx preallocated reqs if IORING_SETUP_IOPOLL. When this inlined reqs are used up, it will allocated reqs in old fashion. Thanks Jianchao
