Commits
Kirill Tkhai authored and Konstantin Khorenko committed e95c5e36d1d
fuse kio: Fix deadlock at pcs_fuse_submit() error path request_end() takes fc->lock, so we in case of error we bump into deadlock: Call Trace: [<ffffffffb3bb63f5>] _raw_spin_lock+0x75/0xc0 [<ffffffffc170871b>] spin_lock+0x18/0x1b [fuse] [<ffffffffc170ba63>] request_end+0x265/0x72b [fuse] [<ffffffffc18a1b8d>] pcs_fuse_submit+0x9fb/0xaa3 [fuse_kio_pcs] [<ffffffffc18a35c4>] kpcs_req_send+0x793/0xa60 [fuse_kio_pcs] [<ffffffffc170b6ca>] flush_bg_queue+0x14f/0x283 [fuse] [<ffffffffc170d4d4>] fuse_request_send_background_locked+0x50b/0x512 [fuse] [<ffffffffc170d844>] fuse_request_send_background+0x369/0x43f [fuse] [<ffffffffc173028b>] fuse_send_readpages+0x372/0x3b5 [fuse] [<ffffffffc1730c3c>] fuse_readpages+0x28c/0x2f0 [fuse] [<ffffffffb296ba58>] __do_page_cache_readahead+0x518/0x6d0 Fix this by unlocking fc->lock before request_end() call. Note, that it may look strange to have two same lk parameters in pcs_fuse_submit(pfc, req, lk, lk), but the current design interprets requests submitted with locked lk as async and we keep this logic. Generally, I feel we need to improve design in a thing of queueing requests and locking, but we need more inverstigation and thinking here, so let's delay this to next VZ update. https://pmc.acronis.com/browse/VSTOR-16246 Signed-off-by: Kirill Tkhai <ktkhai@virtuozzo.com> Reviewed-by: Pavel Butsykin <pbutsykin@virtuozzo.com>