From aa6bf01d391935a8929333bc2e243084ea0c58db Mon Sep 17 00:00:00 2001 From: Christoph Hellwig Date: Wed, 29 Feb 2012 09:53:48 +0000 Subject: xfs: use per-filesystem I/O completion workqueues The new concurrency managed workqueues are cheap enough that we can create per-filesystem instead of global workqueues. This allows us to remove the trylock or defer scheme on the ilock, which is not helpful once we have outstanding log reservations until finishing a size update. Also allow the default concurrency on this workqueues so that I/O completions blocking on the ilock for one inode do not block process for another inode. Reviewed-by: Dave Chinner Reviewed-by: Mark Tinguely Signed-off-by: Christoph Hellwig Signed-off-by: Ben Myers --- fs/xfs/xfs_aops.c | 39 ++++++++++----------------------------- 1 file changed, 10 insertions(+), 29 deletions(-) (limited to 'fs/xfs/xfs_aops.c') diff --git a/fs/xfs/xfs_aops.c b/fs/xfs/xfs_aops.c index 74b9baf36ac..540a01742c6 100644 --- a/fs/xfs/xfs_aops.c +++ b/fs/xfs/xfs_aops.c @@ -126,21 +126,15 @@ static inline bool xfs_ioend_is_append(struct xfs_ioend *ioend) /* * Update on-disk file size now that data has been written to disk. - * - * This function does not block as blocking on the inode lock in IO completion - * can lead to IO completion order dependency deadlocks.. If it can't get the - * inode ilock it will return EAGAIN. Callers must handle this. */ -STATIC int +STATIC void xfs_setfilesize( - xfs_ioend_t *ioend) + struct xfs_ioend *ioend) { - xfs_inode_t *ip = XFS_I(ioend->io_inode); + struct xfs_inode *ip = XFS_I(ioend->io_inode); xfs_fsize_t isize; - if (!xfs_ilock_nowait(ip, XFS_ILOCK_EXCL)) - return EAGAIN; - + xfs_ilock(ip, XFS_ILOCK_EXCL); isize = xfs_ioend_new_eof(ioend); if (isize) { trace_xfs_setfilesize(ip, ioend->io_offset, ioend->io_size); @@ -149,7 +143,6 @@ xfs_setfilesize( } xfs_iunlock(ip, XFS_ILOCK_EXCL); - return 0; } /* @@ -163,10 +156,12 @@ xfs_finish_ioend( struct xfs_ioend *ioend) { if (atomic_dec_and_test(&ioend->io_remaining)) { + struct xfs_mount *mp = XFS_I(ioend->io_inode)->i_mount; + if (ioend->io_type == IO_UNWRITTEN) - queue_work(xfsconvertd_workqueue, &ioend->io_work); + queue_work(mp->m_unwritten_workqueue, &ioend->io_work); else if (xfs_ioend_is_append(ioend)) - queue_work(xfsdatad_workqueue, &ioend->io_work); + queue_work(mp->m_data_workqueue, &ioend->io_work); else xfs_destroy_ioend(ioend); } @@ -207,23 +202,9 @@ xfs_end_io( * We might have to update the on-disk file size after extending * writes. */ - error = xfs_setfilesize(ioend); - ASSERT(!error || error == EAGAIN); - + xfs_setfilesize(ioend); done: - /* - * If we didn't complete processing of the ioend, requeue it to the - * tail of the workqueue for another attempt later. Otherwise destroy - * it. - */ - if (error == EAGAIN) { - atomic_inc(&ioend->io_remaining); - xfs_finish_ioend(ioend); - /* ensure we don't spin on blocked ioends */ - delay(1); - } else { - xfs_destroy_ioend(ioend); - } + xfs_destroy_ioend(ioend); } /* -- cgit v1.2.3-70-g09d2 From 6923e686f19cb7017fc9777a10e06c2e2b2a2936 Mon Sep 17 00:00:00 2001 From: Christoph Hellwig Date: Wed, 29 Feb 2012 09:53:49 +0000 Subject: xfs: do not require an ioend for new EOF calculation Replace xfs_ioend_new_eof with a new inline xfs_new_eof helper that doesn't require and ioend, and is available also outside of xfs_aops.c. Also make the code a bit more clear by using a normal if statement instead of a slightly misleading MIN(). Reviewed-by: Dave Chinner Reviewed-by: Mark Tinguely Signed-off-by: Christoph Hellwig Signed-off-by: Ben Myers --- fs/xfs/xfs_aops.c | 24 ++++-------------------- fs/xfs/xfs_inode.h | 14 ++++++++++++++ 2 files changed, 18 insertions(+), 20 deletions(-) (limited to 'fs/xfs/xfs_aops.c') diff --git a/fs/xfs/xfs_aops.c b/fs/xfs/xfs_aops.c index 540a01742c6..745492b6c66 100644 --- a/fs/xfs/xfs_aops.c +++ b/fs/xfs/xfs_aops.c @@ -98,23 +98,6 @@ xfs_destroy_ioend( mempool_free(ioend, xfs_ioend_pool); } -/* - * If the end of the current ioend is beyond the current EOF, - * return the new EOF value, otherwise zero. - */ -STATIC xfs_fsize_t -xfs_ioend_new_eof( - xfs_ioend_t *ioend) -{ - xfs_inode_t *ip = XFS_I(ioend->io_inode); - xfs_fsize_t isize; - xfs_fsize_t bsize; - - bsize = ioend->io_offset + ioend->io_size; - isize = MIN(i_size_read(VFS_I(ip)), bsize); - return isize > ip->i_d.di_size ? isize : 0; -} - /* * Fast and loose check if this write could update the on-disk inode size. */ @@ -135,7 +118,7 @@ xfs_setfilesize( xfs_fsize_t isize; xfs_ilock(ip, XFS_ILOCK_EXCL); - isize = xfs_ioend_new_eof(ioend); + isize = xfs_new_eof(ip, ioend->io_offset + ioend->io_size); if (isize) { trace_xfs_setfilesize(ip, ioend->io_offset, ioend->io_size); ip->i_d.di_size = isize; @@ -357,6 +340,7 @@ xfs_submit_ioend_bio( xfs_ioend_t *ioend, struct bio *bio) { + struct xfs_inode *ip = XFS_I(ioend->io_inode); atomic_inc(&ioend->io_remaining); bio->bi_private = ioend; bio->bi_end_io = xfs_end_bio; @@ -365,8 +349,8 @@ xfs_submit_ioend_bio( * If the I/O is beyond EOF we mark the inode dirty immediately * but don't update the inode size until I/O completion. */ - if (xfs_ioend_new_eof(ioend)) - xfs_mark_inode_dirty(XFS_I(ioend->io_inode)); + if (xfs_new_eof(ip, ioend->io_offset + ioend->io_size)) + xfs_mark_inode_dirty(ip); submit_bio(wbc->sync_mode == WB_SYNC_ALL ? WRITE_SYNC : WRITE, bio); } diff --git a/fs/xfs/xfs_inode.h b/fs/xfs/xfs_inode.h index eda49378039..7f90469141d 100644 --- a/fs/xfs/xfs_inode.h +++ b/fs/xfs/xfs_inode.h @@ -274,6 +274,20 @@ static inline xfs_fsize_t XFS_ISIZE(struct xfs_inode *ip) return ip->i_d.di_size; } +/* + * If this I/O goes past the on-disk inode size update it unless it would + * be past the current in-core inode size. + */ +static inline xfs_fsize_t +xfs_new_eof(struct xfs_inode *ip, xfs_fsize_t new_size) +{ + xfs_fsize_t i_size = i_size_read(VFS_I(ip)); + + if (new_size > i_size) + new_size = i_size; + return new_size > ip->i_d.di_size ? new_size : 0; +} + /* * i_flags helper functions */ -- cgit v1.2.3-70-g09d2 From 84803fb78237014cbbc86c0f012b273a199f4691 Mon Sep 17 00:00:00 2001 From: Christoph Hellwig Date: Wed, 29 Feb 2012 09:53:50 +0000 Subject: xfs: log file size updates as part of unwritten extent conversion If we convert and unwritten extent past the current i_size log the size update as part of the extent manipulation transactions instead of doing an unlogged metadata update later. Reviewed-by: Dave Chinner Signed-off-by: Christoph Hellwig Reviewed-by: Mark Tinguely Signed-off-by: Ben Myers --- fs/xfs/xfs_aops.c | 11 ++++++----- fs/xfs/xfs_iomap.c | 19 ++++++++++++++++++- 2 files changed, 24 insertions(+), 6 deletions(-) (limited to 'fs/xfs/xfs_aops.c') diff --git a/fs/xfs/xfs_aops.c b/fs/xfs/xfs_aops.c index 745492b6c66..8e11b07bb28 100644 --- a/fs/xfs/xfs_aops.c +++ b/fs/xfs/xfs_aops.c @@ -179,13 +179,14 @@ xfs_end_io( ioend->io_error = -error; goto done; } + } else { + /* + * We might have to update the on-disk file size after + * extending writes. + */ + xfs_setfilesize(ioend); } - /* - * We might have to update the on-disk file size after extending - * writes. - */ - xfs_setfilesize(ioend); done: xfs_destroy_ioend(ioend); } diff --git a/fs/xfs/xfs_iomap.c b/fs/xfs/xfs_iomap.c index 246c7d57c6f..71a464503c4 100644 --- a/fs/xfs/xfs_iomap.c +++ b/fs/xfs/xfs_iomap.c @@ -31,6 +31,7 @@ #include "xfs_ialloc_btree.h" #include "xfs_dinode.h" #include "xfs_inode.h" +#include "xfs_inode_item.h" #include "xfs_btree.h" #include "xfs_bmap.h" #include "xfs_rtalloc.h" @@ -645,6 +646,7 @@ xfs_iomap_write_unwritten( xfs_trans_t *tp; xfs_bmbt_irec_t imap; xfs_bmap_free_t free_list; + xfs_fsize_t i_size; uint resblks; int committed; int error; @@ -705,7 +707,22 @@ xfs_iomap_write_unwritten( if (error) goto error_on_bmapi_transaction; - error = xfs_bmap_finish(&(tp), &(free_list), &committed); + /* + * Log the updated inode size as we go. We have to be careful + * to only log it up to the actual write offset if it is + * halfway into a block. + */ + i_size = XFS_FSB_TO_B(mp, offset_fsb + count_fsb); + if (i_size > offset + count) + i_size = offset + count; + + i_size = xfs_new_eof(ip, i_size); + if (i_size) { + ip->i_d.di_size = i_size; + xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE); + } + + error = xfs_bmap_finish(&tp, &free_list, &committed); if (error) goto error_on_bmapi_transaction; -- cgit v1.2.3-70-g09d2 From 281627df3eb55e1b729b9bb06fff5ff112929646 Mon Sep 17 00:00:00 2001 From: Christoph Hellwig Date: Tue, 13 Mar 2012 08:41:05 +0000 Subject: xfs: log file size updates at I/O completion time Do not use unlogged metadata updates and the VFS dirty bit for updating the file size after writeback. In addition to causing various problems with updates getting delayed for far too long this also drags in the unscalable VFS dirty tracking, and is one of the few remaining unlogged metadata updates. Reviewed-by: Dave Chinner Signed-off-by: Christoph Hellwig Reviewed-by: Mark Tinguely Signed-off-by: Ben Myers --- fs/xfs/xfs_aops.c | 133 ++++++++++++++++++++++++++++++++++++++++++++---------- fs/xfs/xfs_aops.h | 2 + 2 files changed, 111 insertions(+), 24 deletions(-) (limited to 'fs/xfs/xfs_aops.c') diff --git a/fs/xfs/xfs_aops.c b/fs/xfs/xfs_aops.c index 8e11b07bb28..0dbb9e70fe2 100644 --- a/fs/xfs/xfs_aops.c +++ b/fs/xfs/xfs_aops.c @@ -26,6 +26,7 @@ #include "xfs_bmap_btree.h" #include "xfs_dinode.h" #include "xfs_inode.h" +#include "xfs_inode_item.h" #include "xfs_alloc.h" #include "xfs_error.h" #include "xfs_rw.h" @@ -107,25 +108,65 @@ static inline bool xfs_ioend_is_append(struct xfs_ioend *ioend) XFS_I(ioend->io_inode)->i_d.di_size; } +STATIC int +xfs_setfilesize_trans_alloc( + struct xfs_ioend *ioend) +{ + struct xfs_mount *mp = XFS_I(ioend->io_inode)->i_mount; + struct xfs_trans *tp; + int error; + + tp = xfs_trans_alloc(mp, XFS_TRANS_FSYNC_TS); + + error = xfs_trans_reserve(tp, 0, XFS_FSYNC_TS_LOG_RES(mp), 0, 0, 0); + if (error) { + xfs_trans_cancel(tp, 0); + return error; + } + + ioend->io_append_trans = tp; + + /* + * We hand off the transaction to the completion thread now, so + * clear the flag here. + */ + current_restore_flags_nested(&tp->t_pflags, PF_FSTRANS); + return 0; +} + /* * Update on-disk file size now that data has been written to disk. */ -STATIC void +STATIC int xfs_setfilesize( struct xfs_ioend *ioend) { struct xfs_inode *ip = XFS_I(ioend->io_inode); + struct xfs_trans *tp = ioend->io_append_trans; xfs_fsize_t isize; + /* + * The transaction was allocated in the I/O submission thread, + * thus we need to mark ourselves as beeing in a transaction + * manually. + */ + current_set_flags_nested(&tp->t_pflags, PF_FSTRANS); + xfs_ilock(ip, XFS_ILOCK_EXCL); isize = xfs_new_eof(ip, ioend->io_offset + ioend->io_size); - if (isize) { - trace_xfs_setfilesize(ip, ioend->io_offset, ioend->io_size); - ip->i_d.di_size = isize; - xfs_mark_inode_dirty(ip); + if (!isize) { + xfs_iunlock(ip, XFS_ILOCK_EXCL); + xfs_trans_cancel(tp, 0); + return 0; } - xfs_iunlock(ip, XFS_ILOCK_EXCL); + trace_xfs_setfilesize(ip, ioend->io_offset, ioend->io_size); + + ip->i_d.di_size = isize; + xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL); + xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE); + + return xfs_trans_commit(tp, 0); } /* @@ -143,7 +184,7 @@ xfs_finish_ioend( if (ioend->io_type == IO_UNWRITTEN) queue_work(mp->m_unwritten_workqueue, &ioend->io_work); - else if (xfs_ioend_is_append(ioend)) + else if (ioend->io_append_trans) queue_work(mp->m_data_workqueue, &ioend->io_work); else xfs_destroy_ioend(ioend); @@ -173,18 +214,32 @@ xfs_end_io( * range to normal written extens after the data I/O has finished. */ if (ioend->io_type == IO_UNWRITTEN) { + /* + * For buffered I/O we never preallocate a transaction when + * doing the unwritten extent conversion, but for direct I/O + * we do not know if we are converting an unwritten extent + * or not at the point where we preallocate the transaction. + */ + if (ioend->io_append_trans) { + ASSERT(ioend->io_isdirect); + + current_set_flags_nested( + &ioend->io_append_trans->t_pflags, PF_FSTRANS); + xfs_trans_cancel(ioend->io_append_trans, 0); + } + error = xfs_iomap_write_unwritten(ip, ioend->io_offset, ioend->io_size); if (error) { ioend->io_error = -error; goto done; } + } else if (ioend->io_append_trans) { + error = xfs_setfilesize(ioend); + if (error) + ioend->io_error = -error; } else { - /* - * We might have to update the on-disk file size after - * extending writes. - */ - xfs_setfilesize(ioend); + ASSERT(!xfs_ioend_is_append(ioend)); } done: @@ -224,6 +279,7 @@ xfs_alloc_ioend( */ atomic_set(&ioend->io_remaining, 1); ioend->io_isasync = 0; + ioend->io_isdirect = 0; ioend->io_error = 0; ioend->io_list = NULL; ioend->io_type = type; @@ -234,6 +290,7 @@ xfs_alloc_ioend( ioend->io_size = 0; ioend->io_iocb = NULL; ioend->io_result = 0; + ioend->io_append_trans = NULL; INIT_WORK(&ioend->io_work, xfs_end_io); return ioend; @@ -341,18 +398,9 @@ xfs_submit_ioend_bio( xfs_ioend_t *ioend, struct bio *bio) { - struct xfs_inode *ip = XFS_I(ioend->io_inode); atomic_inc(&ioend->io_remaining); bio->bi_private = ioend; bio->bi_end_io = xfs_end_bio; - - /* - * If the I/O is beyond EOF we mark the inode dirty immediately - * but don't update the inode size until I/O completion. - */ - if (xfs_new_eof(ip, ioend->io_offset + ioend->io_size)) - xfs_mark_inode_dirty(ip); - submit_bio(wbc->sync_mode == WB_SYNC_ALL ? WRITE_SYNC : WRITE, bio); } @@ -999,8 +1047,20 @@ xfs_vm_writepage( wbc, end_index); } - if (iohead) + if (iohead) { + /* + * Reserve log space if we might write beyond the on-disk + * inode size. + */ + if (ioend->io_type != IO_UNWRITTEN && + xfs_ioend_is_append(ioend)) { + err = xfs_setfilesize_trans_alloc(ioend); + if (err) + goto error; + } + xfs_submit_ioend(wbc, iohead); + } return 0; @@ -1280,17 +1340,32 @@ xfs_vm_direct_IO( { struct inode *inode = iocb->ki_filp->f_mapping->host; struct block_device *bdev = xfs_find_bdev_for_inode(inode); + struct xfs_ioend *ioend = NULL; ssize_t ret; if (rw & WRITE) { - iocb->private = xfs_alloc_ioend(inode, IO_DIRECT); + size_t size = iov_length(iov, nr_segs); + + /* + * We need to preallocate a transaction for a size update + * here. In the case that this write both updates the size + * and converts at least on unwritten extent we will cancel + * the still clean transaction after the I/O has finished. + */ + iocb->private = ioend = xfs_alloc_ioend(inode, IO_DIRECT); + if (offset + size > XFS_I(inode)->i_d.di_size) { + ret = xfs_setfilesize_trans_alloc(ioend); + if (ret) + goto out_destroy_ioend; + ioend->io_isdirect = 1; + } ret = __blockdev_direct_IO(rw, iocb, inode, bdev, iov, offset, nr_segs, xfs_get_blocks_direct, xfs_end_io_direct_write, NULL, 0); if (ret != -EIOCBQUEUED && iocb->private) - xfs_destroy_ioend(iocb->private); + goto out_trans_cancel; } else { ret = __blockdev_direct_IO(rw, iocb, inode, bdev, iov, offset, nr_segs, @@ -1299,6 +1374,16 @@ xfs_vm_direct_IO( } return ret; + +out_trans_cancel: + if (ioend->io_append_trans) { + current_set_flags_nested(&ioend->io_append_trans->t_pflags, + PF_FSTRANS); + xfs_trans_cancel(ioend->io_append_trans, 0); + } +out_destroy_ioend: + xfs_destroy_ioend(ioend); + return ret; } STATIC void diff --git a/fs/xfs/xfs_aops.h b/fs/xfs/xfs_aops.h index 06e4caf3820..84eafbcb0d9 100644 --- a/fs/xfs/xfs_aops.h +++ b/fs/xfs/xfs_aops.h @@ -46,12 +46,14 @@ typedef struct xfs_ioend { int io_error; /* I/O error code */ atomic_t io_remaining; /* hold count */ unsigned int io_isasync : 1; /* needs aio_complete */ + unsigned int io_isdirect : 1;/* direct I/O */ struct inode *io_inode; /* file being written to */ struct buffer_head *io_buffer_head;/* buffer linked list head */ struct buffer_head *io_buffer_tail;/* buffer linked list tail */ size_t io_size; /* size of the extent */ xfs_off_t io_offset; /* offset in the file */ struct work_struct io_work; /* xfsdatad work queue */ + struct xfs_trans *io_append_trans;/* xact. for size update */ struct kiocb *io_iocb; int io_result; } xfs_ioend_t; -- cgit v1.2.3-70-g09d2