Messages in this thread | | | Date | Fri, 03 Oct 2014 12:25:23 -0600 | From | Jens Axboe <> | Subject | Re: [PATCH] aio: Fix return code of io_submit() (RFC) |
| |
On 2014-10-03 12:21, Kent Overstreet wrote: > On Fri, Oct 03, 2014 at 12:13:39PM -0600, Jens Axboe wrote: >> On 2014-10-03 12:08, Kent Overstreet wrote: >>> io_submit() could return -EAGAIN on memory allocation failure when it should >>> really have been returning -ENOMEM. This could confuse applications (i.e. fio) >>> since -EAGAIN means "too many requests outstanding, wait until completions have >>> been reaped" and if the application actually was tracking outstanding >>> completions this wouldn't make a lot of sense. >>> >>> NOTE: >>> >>> the man page seems to imply that the current behaviour (-EAGAIN on allocation >>> failure) has always been the case. I don't think it makes a lot of sense, but >>> this should probably be discussed more widely in case applications have somehow >>> come to rely on the current behaviour... >> >> We can't really feasibly fix this, is my worry. Fio does track the pending >> requests and does not get into a getevents() forever wait if it gets -EAGAIN >> on submission. But before the fix, it would loop forever in submission in >> -EAGAIN. >> >> How are applications supposed to deal with ENOMEM? I think the answer here >> is that they can't, it would be a fatal condition. AIO must provide isn't >> own guarantee of progress, with a mempool or similar. > > Well, even though the AIO code doesn't currently return -ENOMEM we definitely do > have random other driver/filesystem code that will return -ENOMEM if a random > GFP_KERNEL allocation fails (e.g. the dio code, if allocating a struct dio > fails). So I think there's precedent for this, and having it be a fatal error > when the system is under major memory pressure is not a crazy thing to do too. > > But OTOH maybe we should just use a mempool there. > > The argument against making it a mempool would be "we don't want io_submit() to > block; even if that's not the case today, we at least have a chance of fixing it > with the current setup. If we can't allocate memory for our asynchronous state, > we really can't do anything there except block or fail".
It'll block anyway in other places, if we run out of resources there. But good point on the other potential -ENOMEM cases, it's not a new condition (potentially).
> I'm not sure I have strong feelings one way or the other.
Me neither...
-- Jens Axboe
| |