| Message ID | 20190516085810.31077-1-rpenyaev@suse.de (mailing list archive) |
|---|---|
| Headers | show |
| Series | epoll: support pollable epoll from userspace | expand |
On Thu, May 16, 2019 at 10:57:57AM +0200, Roman Penyaev wrote: > When new event comes for some epoll item kernel does the following: > > struct epoll_uitem *uitem; > > /* Each item has a bit (index in user items array), discussed later */ > uitem = user_header->items[epi->bit]; > > if (!atomic_fetch_or(uitem->ready_events, pollflags)) { > i = atomic_add(&ep->user_header->tail, 1); > > item_idx = &user_index[i & index_mask]; > > /* Signal with a bit, user spins on index expecting value > 0 */ > *item_idx = idx + 1; > > /* > * Want index update be flushed from CPU write buffer and > * immediately visible on userspace side to avoid long busy > * loops. > */ That is garbage; smp_wmb() does no such thing. > smp_wmb(); > }
On 5/16/19 2:57 AM, Roman Penyaev wrote: > Hi all, > > This is v3 which introduces pollable epoll from userspace. > > v3: > - Measurements made, represented below. > > - Fix alignment for epoll_uitem structure on all 64-bit archs except > x86-64. epoll_uitem should be always 16 bit, proper BUILD_BUG_ON > is added. (Linus) > > - Check pollflags explicitly on 0 inside work callback, and do nothing > if 0. > > v2: > - No reallocations, the max number of items (thus size of the user ring) > is specified by the caller. > > - Interface is simplified: -ENOSPC is returned on attempt to add a new > epoll item if number is reached the max, nothing more. > > - Alloced pages are accounted using user->locked_vm and limited to > RLIMIT_MEMLOCK value. > > - EPOLLONESHOT is handled. > > This series introduces pollable epoll from userspace, i.e. user creates > epfd with a new EPOLL_USERPOLL flag, mmaps epoll descriptor, gets header > and ring pointers and then consumes ready events from a ring, avoiding > epoll_wait() call. When ring is empty, user has to call epoll_wait() > in order to wait for new events. epoll_wait() returns -ESTALE if user > ring has events in the ring (kind of indication, that user has to consume > events from the user ring first, I could not invent anything better than > returning -ESTALE). > > For user header and user ring allocation I used vmalloc_user(). I found > that it is much easy to reuse remap_vmalloc_range_partial() instead of > dealing with page cache (like aio.c does). What is also nice is that > virtual address is properly aligned on SHMLBA, thus there should not be > any d-cache aliasing problems on archs with vivt or vipt caches. Why aren't we just adding support to io_uring for this instead? Then we don't need yet another entirely new ring, that's is just a little different from what we have. I haven't looked into the details of your implementation, just curious if there's anything that makes using io_uring a non-starter for this purpose?
On 2019-05-31 16:48, Jens Axboe wrote: > On 5/16/19 2:57 AM, Roman Penyaev wrote: >> Hi all, >> >> This is v3 which introduces pollable epoll from userspace. >> >> v3: >> - Measurements made, represented below. >> >> - Fix alignment for epoll_uitem structure on all 64-bit archs except >> x86-64. epoll_uitem should be always 16 bit, proper BUILD_BUG_ON >> is added. (Linus) >> >> - Check pollflags explicitly on 0 inside work callback, and do >> nothing >> if 0. >> >> v2: >> - No reallocations, the max number of items (thus size of the user >> ring) >> is specified by the caller. >> >> - Interface is simplified: -ENOSPC is returned on attempt to add a >> new >> epoll item if number is reached the max, nothing more. >> >> - Alloced pages are accounted using user->locked_vm and limited to >> RLIMIT_MEMLOCK value. >> >> - EPOLLONESHOT is handled. >> >> This series introduces pollable epoll from userspace, i.e. user >> creates >> epfd with a new EPOLL_USERPOLL flag, mmaps epoll descriptor, gets >> header >> and ring pointers and then consumes ready events from a ring, avoiding >> epoll_wait() call. When ring is empty, user has to call epoll_wait() >> in order to wait for new events. epoll_wait() returns -ESTALE if user >> ring has events in the ring (kind of indication, that user has to >> consume >> events from the user ring first, I could not invent anything better >> than >> returning -ESTALE). >> >> For user header and user ring allocation I used vmalloc_user(). I >> found >> that it is much easy to reuse remap_vmalloc_range_partial() instead of >> dealing with page cache (like aio.c does). What is also nice is that >> virtual address is properly aligned on SHMLBA, thus there should not >> be >> any d-cache aliasing problems on archs with vivt or vipt caches. > > Why aren't we just adding support to io_uring for this instead? Then we > don't need yet another entirely new ring, that's is just a little > different from what we have. > > I haven't looked into the details of your implementation, just curious > if there's anything that makes using io_uring a non-starter for this > purpose? Afaict the main difference is that you do not need to recharge an fd (submit new poll request in terms of io_uring): once fd has been added to epoll with epoll_ctl() - we get events. When you have thousands of fds - that should matter. Also interesting question is how difficult to modify existing event loops in event libraries in order to support recharging (EPOLLONESHOT in terms of epoll). Maybe Azat who maintains libevent can shed light on this (currently I see that libevent does not support "EPOLLONESHOT" logic). -- Roman
On Thu, May 16, 2019 at 10:57:57AM +0200, Roman Penyaev wrote: > When new event comes for some epoll item kernel does the following: > > struct epoll_uitem *uitem; > > /* Each item has a bit (index in user items array), discussed later */ > uitem = user_header->items[epi->bit]; > > if (!atomic_fetch_or(uitem->ready_events, pollflags)) { > i = atomic_add(&ep->user_header->tail, 1); So this is where you increment tail > > item_idx = &user_index[i & index_mask]; > > /* Signal with a bit, user spins on index expecting value > 0 */ > *item_idx = idx + 1; IUC, this is where you write the idx into shared memory, which is _after_ tail has already been incremented. > } > > Important thing here is that ring can't infinitely grow and corrupt other > elements, because kernel always checks that item was marked as ready, so > userspace has to clear ready_events field. > > On userside events the following code should be used in order to consume > events: > > tail = READ_ONCE(header->tail); > for (i = 0; header->head != tail; header->head++) { > item_idx_ptr = &index[idx & indeces_mask]; > > /* > * Spin here till we see valid index > */ > while (!(idx = __atomic_load_n(item_idx_ptr, __ATOMIC_ACQUIRE))) > ; Which you then try and fix up by busy waiting for @idx to become !0 ?! Why not write the idx first, then increment the ->tail, such that when we see ->tail, we already know idx must be correct? > > item = &header->items[idx - 1]; > > /* > * Mark index as invalid, that is for userspace only, kernel does not care > * and will refill this pointer only when observes that event is cleared, > * which happens below. > */ > *item_idx_ptr = 0; That avoids this store too. > > /* > * Fetch data first, if event is cleared by the kernel we drop the data > * returning false. > */ > event->data = item->event.data; > event->events = __atomic_exchange_n(&item->ready_events, 0, > __ATOMIC_RELEASE); > > } Aside from that, you have to READ/WRITE_ONCE() on ->head, to avoid load/store tearing. That would give something like: kernel: slot = atomic_fetch_inc(&ep->slot); item_idx = &user_index[slot & idx_mask]; WRITE_ONCE(*item_idx, idx); smp_store_release(&ep->user_header->tail, slot); userspace: tail = smp_load_acquire(&header->tail); for (head = READ_ONCE(header->head); head != tail; head++) { idx = READ_ONCE(index[head & idx_mask]); itemp = &header->items[idx]; ... } smp_store_release(&header->head, head);
On 5/31/19 10:02 AM, Roman Penyaev wrote: > On 2019-05-31 16:48, Jens Axboe wrote: >> On 5/16/19 2:57 AM, Roman Penyaev wrote: >>> Hi all, >>> >>> This is v3 which introduces pollable epoll from userspace. >>> >>> v3: >>> - Measurements made, represented below. >>> >>> - Fix alignment for epoll_uitem structure on all 64-bit archs except >>> x86-64. epoll_uitem should be always 16 bit, proper BUILD_BUG_ON >>> is added. (Linus) >>> >>> - Check pollflags explicitly on 0 inside work callback, and do >>> nothing >>> if 0. >>> >>> v2: >>> - No reallocations, the max number of items (thus size of the user >>> ring) >>> is specified by the caller. >>> >>> - Interface is simplified: -ENOSPC is returned on attempt to add a >>> new >>> epoll item if number is reached the max, nothing more. >>> >>> - Alloced pages are accounted using user->locked_vm and limited to >>> RLIMIT_MEMLOCK value. >>> >>> - EPOLLONESHOT is handled. >>> >>> This series introduces pollable epoll from userspace, i.e. user >>> creates >>> epfd with a new EPOLL_USERPOLL flag, mmaps epoll descriptor, gets >>> header >>> and ring pointers and then consumes ready events from a ring, avoiding >>> epoll_wait() call. When ring is empty, user has to call epoll_wait() >>> in order to wait for new events. epoll_wait() returns -ESTALE if user >>> ring has events in the ring (kind of indication, that user has to >>> consume >>> events from the user ring first, I could not invent anything better >>> than >>> returning -ESTALE). >>> >>> For user header and user ring allocation I used vmalloc_user(). I >>> found >>> that it is much easy to reuse remap_vmalloc_range_partial() instead of >>> dealing with page cache (like aio.c does). What is also nice is that >>> virtual address is properly aligned on SHMLBA, thus there should not >>> be >>> any d-cache aliasing problems on archs with vivt or vipt caches. >> >> Why aren't we just adding support to io_uring for this instead? Then we >> don't need yet another entirely new ring, that's is just a little >> different from what we have. >> >> I haven't looked into the details of your implementation, just curious >> if there's anything that makes using io_uring a non-starter for this >> purpose? > > Afaict the main difference is that you do not need to recharge an fd > (submit new poll request in terms of io_uring): once fd has been added > to > epoll with epoll_ctl() - we get events. When you have thousands of fds > - > that should matter. > > Also interesting question is how difficult to modify existing event > loops > in event libraries in order to support recharging (EPOLLONESHOT in terms > of epoll). > > Maybe Azat who maintains libevent can shed light on this (currently I > see > that libevent does not support "EPOLLONESHOT" logic). In terms of existing io_uring poll support, which is what I'm guessing you're referring to, it is indeed just one-shot. But there's no reason why we can't have it persist until explicitly canceled with POLL_REMOVE.
On 2019-05-31 18:33, Peter Zijlstra wrote: > On Thu, May 16, 2019 at 10:57:57AM +0200, Roman Penyaev wrote: >> When new event comes for some epoll item kernel does the following: >> >> struct epoll_uitem *uitem; >> >> /* Each item has a bit (index in user items array), discussed later >> */ >> uitem = user_header->items[epi->bit]; >> >> if (!atomic_fetch_or(uitem->ready_events, pollflags)) { >> i = atomic_add(&ep->user_header->tail, 1); > > So this is where you increment tail > >> >> item_idx = &user_index[i & index_mask]; >> >> /* Signal with a bit, user spins on index expecting value > 0 */ >> *item_idx = idx + 1; > > IUC, this is where you write the idx into shared memory, which is > _after_ tail has already been incremented. > >> } >> >> Important thing here is that ring can't infinitely grow and corrupt >> other >> elements, because kernel always checks that item was marked as ready, >> so >> userspace has to clear ready_events field. >> >> On userside events the following code should be used in order to >> consume >> events: >> >> tail = READ_ONCE(header->tail); >> for (i = 0; header->head != tail; header->head++) { >> item_idx_ptr = &index[idx & indeces_mask]; >> >> /* >> * Spin here till we see valid index >> */ >> while (!(idx = __atomic_load_n(item_idx_ptr, __ATOMIC_ACQUIRE))) >> ; > > Which you then try and fix up by busy waiting for @idx to become !0 ?! > > Why not write the idx first, then increment the ->tail, such that when > we see ->tail, we already know idx must be correct? > >> >> item = &header->items[idx - 1]; >> >> /* >> * Mark index as invalid, that is for userspace only, kernel does >> not care >> * and will refill this pointer only when observes that event is >> cleared, >> * which happens below. >> */ >> *item_idx_ptr = 0; > > That avoids this store too. > >> >> /* >> * Fetch data first, if event is cleared by the kernel we drop >> the data >> * returning false. >> */ >> event->data = item->event.data; >> event->events = __atomic_exchange_n(&item->ready_events, 0, >> __ATOMIC_RELEASE); >> >> } > > Aside from that, you have to READ/WRITE_ONCE() on ->head, to avoid > load/store tearing. Yes, clear. Thanks. > > > That would give something like: > > kernel: > > slot = atomic_fetch_inc(&ep->slot); > item_idx = &user_index[slot & idx_mask]; > WRITE_ONCE(*item_idx, idx); > smp_store_release(&ep->user_header->tail, slot); This can't be called from many cpus, tail can be overwritten with "old" value. That what I try to solve. -- Roman
On 2019-05-31 18:54, Jens Axboe wrote: > On 5/31/19 10:02 AM, Roman Penyaev wrote: >> On 2019-05-31 16:48, Jens Axboe wrote: >>> On 5/16/19 2:57 AM, Roman Penyaev wrote: >>>> Hi all, >>>> >>>> This is v3 which introduces pollable epoll from userspace. >>>> >>>> v3: >>>> - Measurements made, represented below. >>>> >>>> - Fix alignment for epoll_uitem structure on all 64-bit archs >>>> except >>>> x86-64. epoll_uitem should be always 16 bit, proper >>>> BUILD_BUG_ON >>>> is added. (Linus) >>>> >>>> - Check pollflags explicitly on 0 inside work callback, and do >>>> nothing >>>> if 0. >>>> >>>> v2: >>>> - No reallocations, the max number of items (thus size of the >>>> user >>>> ring) >>>> is specified by the caller. >>>> >>>> - Interface is simplified: -ENOSPC is returned on attempt to add >>>> a >>>> new >>>> epoll item if number is reached the max, nothing more. >>>> >>>> - Alloced pages are accounted using user->locked_vm and limited >>>> to >>>> RLIMIT_MEMLOCK value. >>>> >>>> - EPOLLONESHOT is handled. >>>> >>>> This series introduces pollable epoll from userspace, i.e. user >>>> creates >>>> epfd with a new EPOLL_USERPOLL flag, mmaps epoll descriptor, gets >>>> header >>>> and ring pointers and then consumes ready events from a ring, >>>> avoiding >>>> epoll_wait() call. When ring is empty, user has to call >>>> epoll_wait() >>>> in order to wait for new events. epoll_wait() returns -ESTALE if >>>> user >>>> ring has events in the ring (kind of indication, that user has to >>>> consume >>>> events from the user ring first, I could not invent anything better >>>> than >>>> returning -ESTALE). >>>> >>>> For user header and user ring allocation I used vmalloc_user(). I >>>> found >>>> that it is much easy to reuse remap_vmalloc_range_partial() instead >>>> of >>>> dealing with page cache (like aio.c does). What is also nice is >>>> that >>>> virtual address is properly aligned on SHMLBA, thus there should not >>>> be >>>> any d-cache aliasing problems on archs with vivt or vipt caches. >>> >>> Why aren't we just adding support to io_uring for this instead? Then >>> we >>> don't need yet another entirely new ring, that's is just a little >>> different from what we have. >>> >>> I haven't looked into the details of your implementation, just >>> curious >>> if there's anything that makes using io_uring a non-starter for this >>> purpose? >> >> Afaict the main difference is that you do not need to recharge an fd >> (submit new poll request in terms of io_uring): once fd has been added >> to >> epoll with epoll_ctl() - we get events. When you have thousands of >> fds >> - >> that should matter. >> >> Also interesting question is how difficult to modify existing event >> loops >> in event libraries in order to support recharging (EPOLLONESHOT in >> terms >> of epoll). >> >> Maybe Azat who maintains libevent can shed light on this (currently I >> see >> that libevent does not support "EPOLLONESHOT" logic). > > In terms of existing io_uring poll support, which is what I'm guessing > you're referring to, it is indeed just one-shot. Yes, yes. > But there's no reason why we can't have it persist until explicitly > canceled with POLL_REMOVE. It seems not so easy. The main problem is that with only a ring it is impossible to figure out on kernel side what event bits have been already seen by the userspace and what bits are new. So every new cqe has to be added to a completion ring on each wake_up_interruptible() call. (I mean when fd wants to report that something is ready). IMO that can lead to many duplicate events (tens? hundreds? honestly no idea), which userspace has to handle with subsequent read/write calls. It can kill all performance benefits of a uring. In uepoll this is solved with another piece of shared memory, where userspace atomically clears bits and kernel side sets bits. If kernel observes that bits were set (i.e. userspace has not seen this event) - new index is added to a ring. Can we extend the io_uring API to support this behavior? Also would be great if we can make event path lockless. On a big number of fds and frequent events - this matters, please take a look, recently I did some measurements: https://lkml.org/lkml/2018/12/12/305 -- Roman
On 5/31/19 1:45 PM, Roman Penyaev wrote: > On 2019-05-31 18:54, Jens Axboe wrote: >> On 5/31/19 10:02 AM, Roman Penyaev wrote: >>> On 2019-05-31 16:48, Jens Axboe wrote: >>>> On 5/16/19 2:57 AM, Roman Penyaev wrote: >>>>> Hi all, >>>>> >>>>> This is v3 which introduces pollable epoll from userspace. >>>>> >>>>> v3: >>>>> - Measurements made, represented below. >>>>> >>>>> - Fix alignment for epoll_uitem structure on all 64-bit archs >>>>> except >>>>> x86-64. epoll_uitem should be always 16 bit, proper >>>>> BUILD_BUG_ON >>>>> is added. (Linus) >>>>> >>>>> - Check pollflags explicitly on 0 inside work callback, and do >>>>> nothing >>>>> if 0. >>>>> >>>>> v2: >>>>> - No reallocations, the max number of items (thus size of the >>>>> user >>>>> ring) >>>>> is specified by the caller. >>>>> >>>>> - Interface is simplified: -ENOSPC is returned on attempt to add >>>>> a >>>>> new >>>>> epoll item if number is reached the max, nothing more. >>>>> >>>>> - Alloced pages are accounted using user->locked_vm and limited >>>>> to >>>>> RLIMIT_MEMLOCK value. >>>>> >>>>> - EPOLLONESHOT is handled. >>>>> >>>>> This series introduces pollable epoll from userspace, i.e. user >>>>> creates >>>>> epfd with a new EPOLL_USERPOLL flag, mmaps epoll descriptor, gets >>>>> header >>>>> and ring pointers and then consumes ready events from a ring, >>>>> avoiding >>>>> epoll_wait() call. When ring is empty, user has to call >>>>> epoll_wait() >>>>> in order to wait for new events. epoll_wait() returns -ESTALE if >>>>> user >>>>> ring has events in the ring (kind of indication, that user has to >>>>> consume >>>>> events from the user ring first, I could not invent anything better >>>>> than >>>>> returning -ESTALE). >>>>> >>>>> For user header and user ring allocation I used vmalloc_user(). I >>>>> found >>>>> that it is much easy to reuse remap_vmalloc_range_partial() instead >>>>> of >>>>> dealing with page cache (like aio.c does). What is also nice is >>>>> that >>>>> virtual address is properly aligned on SHMLBA, thus there should not >>>>> be >>>>> any d-cache aliasing problems on archs with vivt or vipt caches. >>>> >>>> Why aren't we just adding support to io_uring for this instead? Then >>>> we >>>> don't need yet another entirely new ring, that's is just a little >>>> different from what we have. >>>> >>>> I haven't looked into the details of your implementation, just >>>> curious >>>> if there's anything that makes using io_uring a non-starter for this >>>> purpose? >>> >>> Afaict the main difference is that you do not need to recharge an fd >>> (submit new poll request in terms of io_uring): once fd has been added >>> to >>> epoll with epoll_ctl() - we get events. When you have thousands of >>> fds >>> - >>> that should matter. >>> >>> Also interesting question is how difficult to modify existing event >>> loops >>> in event libraries in order to support recharging (EPOLLONESHOT in >>> terms >>> of epoll). >>> >>> Maybe Azat who maintains libevent can shed light on this (currently I >>> see >>> that libevent does not support "EPOLLONESHOT" logic). >> >> In terms of existing io_uring poll support, which is what I'm guessing >> you're referring to, it is indeed just one-shot. > > Yes, yes. > >> But there's no reason why we can't have it persist until explicitly >> canceled with POLL_REMOVE. > > It seems not so easy. The main problem is that with only a ring it is > impossible to figure out on kernel side what event bits have been > already > seen by the userspace and what bits are new. So every new cqe has to > be added to a completion ring on each wake_up_interruptible() call. > (I mean when fd wants to report that something is ready). > > IMO that can lead to many duplicate events (tens? hundreds? honestly no > idea), which userspace has to handle with subsequent read/write calls. > It can kill all performance benefits of a uring. > > In uepoll this is solved with another piece of shared memory, where > userspace atomically clears bits and kernel side sets bits. If kernel > observes that bits were set (i.e. userspace has not seen this event) > - new index is added to a ring. Those are good points. > Can we extend the io_uring API to support this behavior? Also would > be great if we can make event path lockless. On a big number of fds > and frequent events - this matters, please take a look, recently I > did some measurements: https://lkml.org/lkml/2018/12/12/305 Yeah, I'd be happy to entertain that idea, and lockless completions as well. We already do that for polled IO, but consider any "normal" completion to be IRQ driven and hence need locking.
On 2019-05-31 23:09, Jens Axboe wrote: > On 5/31/19 1:45 PM, Roman Penyaev wrote: >> On 2019-05-31 18:54, Jens Axboe wrote: >>> On 5/31/19 10:02 AM, Roman Penyaev wrote: >>>> On 2019-05-31 16:48, Jens Axboe wrote: >>>>> On 5/16/19 2:57 AM, Roman Penyaev wrote: >>>>>> Hi all, >>>>>> >>>>>> This is v3 which introduces pollable epoll from userspace. >>>>>> >>>>>> v3: >>>>>> - Measurements made, represented below. >>>>>> >>>>>> - Fix alignment for epoll_uitem structure on all 64-bit archs >>>>>> except >>>>>> x86-64. epoll_uitem should be always 16 bit, proper >>>>>> BUILD_BUG_ON >>>>>> is added. (Linus) >>>>>> >>>>>> - Check pollflags explicitly on 0 inside work callback, and do >>>>>> nothing >>>>>> if 0. >>>>>> >>>>>> v2: >>>>>> - No reallocations, the max number of items (thus size of the >>>>>> user >>>>>> ring) >>>>>> is specified by the caller. >>>>>> >>>>>> - Interface is simplified: -ENOSPC is returned on attempt to >>>>>> add >>>>>> a >>>>>> new >>>>>> epoll item if number is reached the max, nothing more. >>>>>> >>>>>> - Alloced pages are accounted using user->locked_vm and >>>>>> limited >>>>>> to >>>>>> RLIMIT_MEMLOCK value. >>>>>> >>>>>> - EPOLLONESHOT is handled. >>>>>> >>>>>> This series introduces pollable epoll from userspace, i.e. user >>>>>> creates >>>>>> epfd with a new EPOLL_USERPOLL flag, mmaps epoll descriptor, gets >>>>>> header >>>>>> and ring pointers and then consumes ready events from a ring, >>>>>> avoiding >>>>>> epoll_wait() call. When ring is empty, user has to call >>>>>> epoll_wait() >>>>>> in order to wait for new events. epoll_wait() returns -ESTALE if >>>>>> user >>>>>> ring has events in the ring (kind of indication, that user has to >>>>>> consume >>>>>> events from the user ring first, I could not invent anything >>>>>> better >>>>>> than >>>>>> returning -ESTALE). >>>>>> >>>>>> For user header and user ring allocation I used vmalloc_user(). I >>>>>> found >>>>>> that it is much easy to reuse remap_vmalloc_range_partial() >>>>>> instead >>>>>> of >>>>>> dealing with page cache (like aio.c does). What is also nice is >>>>>> that >>>>>> virtual address is properly aligned on SHMLBA, thus there should >>>>>> not >>>>>> be >>>>>> any d-cache aliasing problems on archs with vivt or vipt caches. >>>>> >>>>> Why aren't we just adding support to io_uring for this instead? >>>>> Then >>>>> we >>>>> don't need yet another entirely new ring, that's is just a little >>>>> different from what we have. >>>>> >>>>> I haven't looked into the details of your implementation, just >>>>> curious >>>>> if there's anything that makes using io_uring a non-starter for >>>>> this >>>>> purpose? >>>> >>>> Afaict the main difference is that you do not need to recharge an fd >>>> (submit new poll request in terms of io_uring): once fd has been >>>> added >>>> to >>>> epoll with epoll_ctl() - we get events. When you have thousands of >>>> fds >>>> - >>>> that should matter. >>>> >>>> Also interesting question is how difficult to modify existing event >>>> loops >>>> in event libraries in order to support recharging (EPOLLONESHOT in >>>> terms >>>> of epoll). >>>> >>>> Maybe Azat who maintains libevent can shed light on this (currently >>>> I >>>> see >>>> that libevent does not support "EPOLLONESHOT" logic). >>> >>> In terms of existing io_uring poll support, which is what I'm >>> guessing >>> you're referring to, it is indeed just one-shot. >> >> Yes, yes. >> >>> But there's no reason why we can't have it persist until explicitly >>> canceled with POLL_REMOVE. >> >> It seems not so easy. The main problem is that with only a ring it is >> impossible to figure out on kernel side what event bits have been >> already >> seen by the userspace and what bits are new. So every new cqe has to >> be added to a completion ring on each wake_up_interruptible() call. >> (I mean when fd wants to report that something is ready). >> >> IMO that can lead to many duplicate events (tens? hundreds? honestly >> no >> idea), which userspace has to handle with subsequent read/write calls. >> It can kill all performance benefits of a uring. >> >> In uepoll this is solved with another piece of shared memory, where >> userspace atomically clears bits and kernel side sets bits. If kernel >> observes that bits were set (i.e. userspace has not seen this event) >> - new index is added to a ring. > > Those are good points. > >> Can we extend the io_uring API to support this behavior? Also would >> be great if we can make event path lockless. On a big number of fds >> and frequent events - this matters, please take a look, recently I >> did some measurements: https://lkml.org/lkml/2018/12/12/305 > > Yeah, I'd be happy to entertain that idea, and lockless completions as > well. We already do that for polled IO, but consider any "normal" > completion to be IRQ driven and hence need locking. I would like to contribute as much as I can. "Subscription" on events along with lockless ring seems reasonable to do for io_uring. I still tend to think that uepoll and io_uring poll can coexist, at least because it can be difficult to adopt current event libraries to async nature of "add fd" / "remove add" requests of the io_uring, e.g. when epoll_ctl() is called in order to remove fd, the caller expects no events come after epoll_ctl() returns. Async behavior can break the event loop. What can help is ability to wait on particular request, which seems not possible without ugly tricks, right? (Under ugly tricks I mean something as: wait for any event, traverse the completion ring in order to meet particular completion, repeat if nothing is found). Also epoll_ctl() can be called from another thread in order to add/remove fd, and I suppose that is also successfully used by event loop libraries or users of these libraries (not quite sure though, but can imagine why it can be useful). To fix that will require introducing locks on submission path of io_uring callers (I mean on user side, inside these libraries), which can impact performance for generic cases (only submission though). What I want to say is that polling using io_uring can be used in some new io/event stacks, but adoption of current event libraries can be non trivial, where old plain epoll with a ring can be an easiest way. But of course that's only my speculation. -- Roman
Hi all, This is v3 which introduces pollable epoll from userspace. v3: - Measurements made, represented below. - Fix alignment for epoll_uitem structure on all 64-bit archs except x86-64. epoll_uitem should be always 16 bit, proper BUILD_BUG_ON is added. (Linus) - Check pollflags explicitly on 0 inside work callback, and do nothing if 0. v2: - No reallocations, the max number of items (thus size of the user ring) is specified by the caller. - Interface is simplified: -ENOSPC is returned on attempt to add a new epoll item if number is reached the max, nothing more. - Alloced pages are accounted using user->locked_vm and limited to RLIMIT_MEMLOCK value. - EPOLLONESHOT is handled. This series introduces pollable epoll from userspace, i.e. user creates epfd with a new EPOLL_USERPOLL flag, mmaps epoll descriptor, gets header and ring pointers and then consumes ready events from a ring, avoiding epoll_wait() call. When ring is empty, user has to call epoll_wait() in order to wait for new events. epoll_wait() returns -ESTALE if user ring has events in the ring (kind of indication, that user has to consume events from the user ring first, I could not invent anything better than returning -ESTALE). For user header and user ring allocation I used vmalloc_user(). I found that it is much easy to reuse remap_vmalloc_range_partial() instead of dealing with page cache (like aio.c does). What is also nice is that virtual address is properly aligned on SHMLBA, thus there should not be any d-cache aliasing problems on archs with vivt or vipt caches. ** Measurements In order to measure polling from userspace libevent was modified [1] and bench_http benchmark (client and server) was used: o EPOLLET, original epoll: 20000 requests in 0.551306 sec. (36277.49 throughput) Each took about 5.54 msec latency 1600000bytes read. 0 errors. o EPOLLET + polling from userspace: 20000 requests in 0.475585 sec. (42053.47 throughput) Each took about 4.78 msec latency 1600000bytes read. 0 errors. So harvesting events from userspace gives 15% gain. Though bench_http is not ideal benchmark, but at least it is the part of libevent and was easy to modify. Worth to mention that uepoll is very sensible to CPU, e.g. the gain above is observed on desktop "Intel(R) Core(TM) i7-6820HQ CPU @ 2.70GHz", but on "Intel(R) Xeon(R) Silver 4110 CPU @ 2.10GHz" measurements are almost the same for both runs. ** Limitations 1. Expect always EPOLLET flag for new epoll items (Edge Triggered behavior) obviously we can't call vfs_epoll() from userpace to have level triggered behaviour. 2. No support for EPOLLWAKEUP events are consumed from userspace, thus no way to call __pm_relax() 3. No support for EPOLLEXCLUSIVE If device does not pass pollflags to wake_up() there is no way to call poll() from the context under spinlock, thus special work is scheduled to offload polling. In this specific case we can't support exclusive wakeups, because we do not know actual result of scheduled work and have to wake up every waiter. ** Principle of operation * Basic structures shared with userspace: In order to consume events from userspace all inserted items should be stored in items array, which has original epoll_event field and u32 field for keeping ready events, i.e. each item has the following struct: struct epoll_uitem { __poll_t ready_events; struct epoll_event event; }; BUILD_BUG_ON(sizeof(struct epoll_uitem) != 16); And the following is a header, which is seen by userspace: struct epoll_uheader { u32 magic; /* epoll user header magic */ u32 header_length; /* length of the header + items */ u32 index_length; /* length of the index ring, always pow2 */ u32 max_items_nr; /* max num of items */ u32 head; /* updated by userland */ u32 int tail; /* updated by kernel */ struct epoll_uitem items[] __aligned(128); }; /* Header is 128 bytes, thus items are aligned on CPU cache */ BUILD_BUG_ON(sizeof(struct epoll_uheader) != 128); In order to poll epfd from userspace application has to call: epoll_create2(EPOLL_USERPOLL, max_items_nr); Ready events are kept in a ring buffer, which is simply an index table, where each element points to an item in a header: unsinged int *user_index; * How is new event accounted on kernel side? Hot it is consumed from * userspace? When new event comes for some epoll item kernel does the following: struct epoll_uitem *uitem; /* Each item has a bit (index in user items array), discussed later */ uitem = user_header->items[epi->bit]; if (!atomic_fetch_or(uitem->ready_events, pollflags)) { i = atomic_add(&ep->user_header->tail, 1); item_idx = &user_index[i & index_mask]; /* Signal with a bit, user spins on index expecting value > 0 */ *item_idx = idx + 1; /* * Want index update be flushed from CPU write buffer and * immediately visible on userspace side to avoid long busy * loops. */ smp_wmb(); } Important thing here is that ring can't infinitely grow and corrupt other elements, because kernel always checks that item was marked as ready, so userspace has to clear ready_events field. On userside events the following code should be used in order to consume events: tail = READ_ONCE(header->tail); for (i = 0; header->head != tail; header->head++) { item_idx_ptr = &index[idx & indeces_mask]; /* * Spin here till we see valid index */ while (!(idx = __atomic_load_n(item_idx_ptr, __ATOMIC_ACQUIRE))) ; item = &header->items[idx - 1]; /* * Mark index as invalid, that is for userspace only, kernel does not care * and will refill this pointer only when observes that event is cleared, * which happens below. */ *item_idx_ptr = 0; /* * Fetch data first, if event is cleared by the kernel we drop the data * returning false. */ event->data = item->event.data; event->events = __atomic_exchange_n(&item->ready_events, 0, __ATOMIC_RELEASE); } * How new epoll item gets its index inside user items array? Kernel has a bitmap for that and gets free bit on attempt to insert a new epoll item. When bitmap is full -ENOSPC is returned. * Is there any testing app available? There is a small app [2] which starts many threads with many event fds and produces many events, while single consumer fetches them from userspace and goes to kernel from time to time in order to wait. [1] https://github.com/libevent/libevent/pull/801 [2] https://github.com/rouming/test-tools/blob/master/userpolled-epoll.c Roman Penyaev (13): epoll: move private helpers from a header to the source epoll: introduce user structures for polling from userspace epoll: allocate user header and user events ring for polling from userspace epoll: some sanity flags checks for epoll syscalls for polling from userspace epoll: offload polling to a work in case of epfd polled from userspace epoll: introduce helpers for adding/removing events to uring epoll: call ep_add_event_to_uring() from ep_poll_callback() epoll: support polling from userspace for ep_insert() epoll: support polling from userspace for ep_remove() epoll: support polling from userspace for ep_modify() epoll: support polling from userspace for ep_poll() epoll: support mapping for epfd when polled from userspace epoll: implement epoll_create2() syscall arch/x86/entry/syscalls/syscall_32.tbl | 1 + arch/x86/entry/syscalls/syscall_64.tbl | 1 + fs/eventpoll.c | 714 ++++++++++++++++++++++--- include/linux/syscalls.h | 1 + include/uapi/asm-generic/unistd.h | 4 +- include/uapi/linux/eventpoll.h | 46 +- kernel/sys_ni.c | 1 + 7 files changed, 669 insertions(+), 99 deletions(-) Signed-off-by: Roman Penyaev <rpenyaev@suse.de> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Al Viro <viro@zeniv.linux.org.uk> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: linux-fsdevel@vger.kernel.org Cc: linux-kernel@vger.kernel.org