> On Jan 26, 2016, at 19:40, Ilya Dryomov <[email protected]> wrote: > > On Tue, Jan 26, 2016 at 12:16 PM, Yan, Zheng <[email protected]> wrote: >> >>> On Jan 26, 2016, at 18:30, Ilya Dryomov <[email protected]> wrote: >>> >>> On Tue, Jan 26, 2016 at 10:24 AM, Dan Carpenter >>> <[email protected]> wrote: >>>> ceph_osdc_alloc_request() returns NULL on error, it never returns error >>>> pointers. >>>> >>>> Fixes: 5be0389dac66 ('ceph: re-send AIO write request when getting >>>> -EOLDSNAP error') >>>> Signed-off-by: Dan Carpenter <[email protected]> >>>> >>>> diff --git a/fs/ceph/file.c b/fs/ceph/file.c >>>> index d37efdd..a52cf9b 100644 >>>> --- a/fs/ceph/file.c >>>> +++ b/fs/ceph/file.c >>>> @@ -698,8 +698,8 @@ static void ceph_aio_retry_work(struct work_struct >>>> *work) >>>> >>>> req = ceph_osdc_alloc_request(orig_req->r_osdc, snapc, 2, >>>> false, GFP_NOFS); >>>> - if (IS_ERR(req)) { >>>> - ret = PTR_ERR(req); >>>> + if (!req) { >>>> + ret = -ENOMEM; >>>> req = orig_req; >>>> goto out; >>>> } >>> >>> Applied, thanks Dan. >>> >>> Zheng, I have an related concern: where do you put snapc (refcount is >>> bumped a few lines above) if ceph_osdc_alloc_request() fails? It looks >>> like it's leaked to me. >>> >>> The BUG_ON(ret == -EOLDSNAPC) also seems a bit bogus, given that ret is >>> either -ENOMEM or ceph_osdc_start_request() retval. >> >> ceph_aio_complete_req treats -EOLDSNAP distinguishingly. Purpose of this >> BUG_ON is detect potential infinite loop. > > Did you miss the part about the snap context? > > I get the purpose of -EOLDSNAPC assert in ceph_direct_read_write(), > where you can actually get it from ceph_osdc_wait_request() - it's > a server-side error code. Asserting it in ceph_aio_retry_work(), in > which only client helpers are called and the only two possible error > codes are -ENOMEM and -EIO doesn't make much sense to me. >
Yeah, removing that BUG_ON is completely OK. Regards, Yan, Zheng > Thanks, > > Ilya
