pnfs_layout_need_return() runs whenever a reference to a layout header is put while a return is pending. It marks the segments covered by the pending return and, if none of them is still in use, lets the caller send the LAYOUTRETURN; otherwise the return stays deferred until they drain. Since commit b739a5bd9d9f ("NFSv4/flexfiles: Cancel I/O if the layout is recalled or revoked"), this re-check also requests cancellation of I/O on matching segments still in use. Changed-layout recalls, error-driven returns, and file bulk returns already request cancellation when initiating the return. As segments drain, the pending return's sequence cutoff can advance, so a later re-check can also cover newly acquired segments. Cancelling their I/O can perpetuate a cancellation and retry cycle. Disable cancellation in the deferred re-check, while retaining segment invalidation and waiting for outstanding references to drain. Explicit cancellation by the initiating paths is unchanged. This was observed on 7.4-rc3 on a one-client one-DS system under sustained direct-read I/O over pNFS/RDMA after restarting the nfs service on the DS, with "NFS/flexfiles: Don't treat a locally aborted read as a DS failure" applied. Once LAYOUTRETURNs were being sent again, the re-check kept cancelling the reads of freshly obtained segments, at about ten transport reconnects per second. With this change, once the DS was back no reads were cancelled and no transports were reconnected. Fixes: b739a5bd9d9f ("NFSv4/flexfiles: Cancel I/O if the layout is recalled or revoked") Cc: stable@vger.kernel.org Signed-off-by: Tim Menninger --- fs/nfs/pnfs.c | 12 +++++++----- 1 file changed, 7 insertions(+), 5 deletions(-) diff --git a/fs/nfs/pnfs.c b/fs/nfs/pnfs.c index 14f4a4b12ffd3..a60df3bdab921 100644 --- a/fs/nfs/pnfs.c +++ b/fs/nfs/pnfs.c @@ -480,15 +480,16 @@ pnfs_mark_layout_stateid_invalid(struct pnfs_layout_hdr *lo, static int pnfs_mark_layout_stateid_return(struct pnfs_layout_hdr *lo, struct list_head *lseg_list, - enum pnfs_iomode iomode, u32 seq) + enum pnfs_iomode iomode, u32 seq, + bool cancel_io) { struct pnfs_layout_range range = { .iomode = iomode, .length = NFS4_MAX_UINT64, }; - return pnfs_mark_matching_lsegs_return(lo, lseg_list, &range, seq, true, - NULL); + return pnfs_mark_matching_lsegs_return(lo, lseg_list, &range, seq, + cancel_io, NULL); } static int @@ -903,7 +904,7 @@ pnfs_layout_free_bulk_destroy_list(struct list_head *layout_list, list_del_init(&lo->plh_bulk_destroy); if (mode == PNFS_LAYOUT_FILE_BULK_RETURN) { pnfs_mark_layout_stateid_return(lo, &lseg_list, - IOMODE_ANY, 0); + IOMODE_ANY, 0, true); } else if (pnfs_mark_layout_stateid_invalid(lo, &lseg_list)) { if (mode == PNFS_LAYOUT_BULK_RETURN) set_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags); @@ -1405,7 +1406,8 @@ pnfs_layout_need_return(struct pnfs_layout_hdr *lo) return false; return pnfs_mark_layout_stateid_return(lo, &lo->plh_return_segs, lo->plh_return_iomode, - lo->plh_return_seq) != -EBUSY; + lo->plh_return_seq, + false) != -EBUSY; } static void pnfs_layoutreturn_before_put_layout_hdr(struct pnfs_layout_hdr *lo) base-commit: aebb7274476403040d582c61478df8f4a7bf9cd3 -- 2.53.0