X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lustre%2Fobdfilter%2Ffilter_io_26.c;h=1c9ad6b6c3d250c0adae96ff76efe8381f6e012c;hp=463ae55a6b1eca943fc2fcc6303d81d690292693;hb=346a17e4d8b5c291d776387ace81a5b74bc24141;hpb=c3e5086897d4ca179520548f13666276991c2965;ds=sidebyside diff --git a/lustre/obdfilter/filter_io_26.c b/lustre/obdfilter/filter_io_26.c index 463ae55..1c9ad6b 100644 --- a/lustre/obdfilter/filter_io_26.c +++ b/lustre/obdfilter/filter_io_26.c @@ -1,30 +1,43 @@ /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*- * vim:expandtab:shiftwidth=8:tabstop=8: * - * linux/fs/obdfilter/filter_io.c + * GPL HEADER START * - * Copyright (c) 2001-2003 Cluster File Systems, Inc. - * Author: Peter Braam - * Author: Andreas Dilger - * Author: Phil Schwan + * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. * - * This file is part of the Lustre file system, http://www.lustre.org - * Lustre is a trademark of Cluster File Systems, Inc. + * This program is free software; you can redistribute it and/or modify + * it under the terms of the GNU General Public License version 2 only, + * as published by the Free Software Foundation. * - * You may have signed or agreed to another license before downloading - * this software. If so, you are bound by the terms and conditions - * of that agreement, and the following does not apply to you. See the - * LICENSE file included with this distribution for more information. + * This program is distributed in the hope that it will be useful, but + * WITHOUT ANY WARRANTY; without even the implied warranty of + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU + * General Public License version 2 for more details (a copy is included + * in the LICENSE file that accompanied this code). * - * If you did not agree to a different license, then this copy of Lustre - * is open source software; you can redistribute it and/or modify it - * under the terms of version 2 of the GNU General Public License as - * published by the Free Software Foundation. + * You should have received a copy of the GNU General Public License + * version 2 along with this program; If not, see + * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf * - * In either case, Lustre is distributed in the hope that it will be - * useful, but WITHOUT ANY WARRANTY; without even the implied warranty - * of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the - * license text for more details. + * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara, + * CA 95054 USA or visit www.sun.com if you need additional information or + * have any questions. + * + * GPL HEADER END + */ +/* + * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved. + * Use is subject to license terms. + */ +/* + * This file is part of Lustre, http://www.lustre.org/ + * Lustre is a trademark of Sun Microsystems, Inc. + * + * lustre/obdfilter/filter_io_26.c + * + * Author: Peter Braam + * Author: Andreas Dilger + * Author: Phil Schwan */ #ifndef AUTOCONF_INCLUDED @@ -45,15 +58,14 @@ /* 512byte block min */ #define MAX_BLOCKS_PER_PAGE (CFS_PAGE_SIZE / 512) struct filter_iobuf { - atomic_t dr_numreqs; /* number of reqs being processed */ - wait_queue_head_t dr_wait; - int dr_max_pages; - int dr_npages; - int dr_error; - struct page **dr_pages; - unsigned long *dr_blocks; - spinlock_t dr_lock; /* IRQ lock */ - unsigned int dr_ignore_quota:1; + cfs_atomic_t dr_numreqs; /* number of reqs being processed */ + cfs_waitq_t dr_wait; + int dr_max_pages; + int dr_npages; + int dr_error; + struct page **dr_pages; + unsigned long *dr_blocks; + unsigned int dr_ignore_quota:1; struct filter_obd *dr_filter; }; @@ -62,26 +74,40 @@ static void record_start_io(struct filter_iobuf *iobuf, int rw, int size, { struct filter_obd *filter = iobuf->dr_filter; - atomic_inc(&iobuf->dr_numreqs); + cfs_atomic_inc(&iobuf->dr_numreqs); if (rw == OBD_BRW_READ) { - atomic_inc(&filter->fo_r_in_flight); + cfs_atomic_inc(&filter->fo_r_in_flight); lprocfs_oh_tally(&filter->fo_filter_stats.hist[BRW_R_RPC_HIST], - atomic_read(&filter->fo_r_in_flight)); - lprocfs_oh_tally_log2(&filter->fo_filter_stats.hist[BRW_R_DISK_IOSIZE], + cfs_atomic_read(&filter->fo_r_in_flight)); + lprocfs_oh_tally_log2(&filter-> + fo_filter_stats.hist[BRW_R_DISK_IOSIZE], size); - lprocfs_oh_tally(&exp->exp_filter_data.fed_brw_stats.hist[BRW_R_RPC_HIST], - atomic_read(&filter->fo_r_in_flight)); - lprocfs_oh_tally_log2(&exp->exp_filter_data.fed_brw_stats.hist[BRW_R_DISK_IOSIZE], size); + if (exp->exp_nid_stats && exp->exp_nid_stats->nid_brw_stats) { + lprocfs_oh_tally(&exp->exp_nid_stats->nid_brw_stats-> + hist[BRW_R_RPC_HIST], + cfs_atomic_read(&filter-> \ + fo_r_in_flight)); + lprocfs_oh_tally_log2(&exp->exp_nid_stats-> + nid_brw_stats->hist[BRW_R_DISK_IOSIZE], + size); + } } else { - atomic_inc(&filter->fo_w_in_flight); + cfs_atomic_inc(&filter->fo_w_in_flight); lprocfs_oh_tally(&filter->fo_filter_stats.hist[BRW_W_RPC_HIST], - atomic_read(&filter->fo_w_in_flight)); - lprocfs_oh_tally_log2(&filter->fo_filter_stats.hist[BRW_W_DISK_IOSIZE], + cfs_atomic_read(&filter->fo_w_in_flight)); + lprocfs_oh_tally_log2(&filter-> + fo_filter_stats.hist[BRW_W_DISK_IOSIZE], size); - lprocfs_oh_tally(&exp->exp_filter_data.fed_brw_stats.hist[BRW_W_RPC_HIST], - atomic_read(&filter->fo_w_in_flight)); - lprocfs_oh_tally_log2(&exp->exp_filter_data.fed_brw_stats.hist[BRW_W_DISK_IOSIZE], size); + if (exp->exp_nid_stats && exp->exp_nid_stats->nid_brw_stats) { + lprocfs_oh_tally(&exp->exp_nid_stats->nid_brw_stats-> + hist[BRW_W_RPC_HIST], + cfs_atomic_read(&filter-> \ + fo_w_in_flight)); + lprocfs_oh_tally_log2(&exp->exp_nid_stats-> + nid_brw_stats->hist[BRW_W_DISK_IOSIZE], + size); + } } } @@ -89,54 +115,83 @@ static void record_finish_io(struct filter_iobuf *iobuf, int rw, int rc) { struct filter_obd *filter = iobuf->dr_filter; - /* CAVEAT EMPTOR: possibly in IRQ context + /* CAVEAT EMPTOR: possibly in IRQ context * DO NOT record procfs stats here!!! */ if (rw == OBD_BRW_READ) - atomic_dec(&filter->fo_r_in_flight); + cfs_atomic_dec(&filter->fo_r_in_flight); else - atomic_dec(&filter->fo_w_in_flight); + cfs_atomic_dec(&filter->fo_w_in_flight); - if (atomic_dec_and_test(&iobuf->dr_numreqs)) - wake_up(&iobuf->dr_wait); + if (cfs_atomic_dec_and_test(&iobuf->dr_numreqs)) + cfs_waitq_signal(&iobuf->dr_wait); } +#ifdef HAVE_BIO_ENDIO_2ARG +#define DIO_RETURN(a) +static void dio_complete_routine(struct bio *bio, int error) +#else +#define DIO_RETURN(a) return(a) static int dio_complete_routine(struct bio *bio, unsigned int done, int error) +#endif { struct filter_iobuf *iobuf = bio->bi_private; - unsigned long flags; + struct bio_vec *bvl; + int i; - /* CAVEAT EMPTOR: possibly in IRQ context + /* CAVEAT EMPTOR: possibly in IRQ context * DO NOT record procfs stats here!!! */ - if (bio->bi_size) /* Not complete */ - return 1; +#ifdef HAVE_BIO_ENDIO_2ARG + /* The "bi_size" check was needed for kernels < 2.6.24 in order to + * handle the case where a SCSI request error caused this callback + * to be called before all of the biovecs had been processed. + * Without this check the server thread will hang. In newer kernels + * the bio_end_io routine is never called for partial completions, + * so this check is no longer needed. */ + if (bio->bi_size) /* Not complete */ + DIO_RETURN(1); +#endif - if (iobuf == NULL) { + if (unlikely(iobuf == NULL)) { CERROR("***** bio->bi_private is NULL! This should never " "happen. Normally, I would crash here, but instead I " "will dump the bio contents to the console. Please " - "report this to CFS, along with any interesting " - "messages leading up to this point (like SCSI errors, " - "perhaps). Because bi_private is NULL, I can't wake up " - "the thread that initiated this I/O -- so you will " - "probably have to reboot this node.\n"); + "report this to , along " + "with any interesting messages leading up to this point " + "(like SCSI errors, perhaps). Because bi_private is " + "NULL, I can't wake up the thread that initiated this " + "IO - you will probably have to reboot this node.\n"); CERROR("bi_next: %p, bi_flags: %lx, bi_rw: %lu, bi_vcnt: %d, " "bi_idx: %d, bi->size: %d, bi_end_io: %p, bi_cnt: %d, " "bi_private: %p\n", bio->bi_next, bio->bi_flags, bio->bi_rw, bio->bi_vcnt, bio->bi_idx, bio->bi_size, - bio->bi_end_io, atomic_read(&bio->bi_cnt), + bio->bi_end_io, cfs_atomic_read(&bio->bi_cnt), bio->bi_private); - return 0; + DIO_RETURN(0); } - spin_lock_irqsave(&iobuf->dr_lock, flags); - if (iobuf->dr_error == 0) - iobuf->dr_error = error; - spin_unlock_irqrestore(&iobuf->dr_lock, flags); + /* the check is outside of the cycle for performance reason -bzzz */ + if (!cfs_test_bit(BIO_RW, &bio->bi_rw)) { + bio_for_each_segment(bvl, bio, i) { + if (likely(error == 0)) + SetPageUptodate(bvl->bv_page); + LASSERT(PageLocked(bvl->bv_page)); + ClearPageConstant(bvl->bv_page); + } + record_finish_io(iobuf, OBD_BRW_READ, error); + } else { + if (mapping_cap_page_constant_write(iobuf->dr_pages[0]->mapping)){ + bio_for_each_segment(bvl, bio, i) { + ClearPageConstant(bvl->bv_page); + } + } + record_finish_io(iobuf, OBD_BRW_WRITE, error); + } - record_finish_io(iobuf, test_bit(BIO_RW, &bio->bi_rw) ? - OBD_BRW_WRITE : OBD_BRW_READ, error); + /* any real error is good enough -bzzz */ + if (error != 0 && iobuf->dr_error == 0) + iobuf->dr_error = error; /* Completed bios used to be chained off iobuf->dr_bios and freed in * filter_clear_dreq(). It was then possible to exhaust the biovec-256 @@ -144,7 +199,7 @@ static int dio_complete_routine(struct bio *bio, unsigned int done, int error) * deadlocking the OST. The bios are now released as soon as complete * so the pool cannot be exhausted while IOs are competing. bug 10076 */ bio_put(bio); - return 0; + DIO_RETURN(0); } static int can_be_merged(struct bio *bio, sector_t sector) @@ -179,9 +234,8 @@ struct filter_iobuf *filter_alloc_iobuf(struct filter_obd *filter, goto failed_2; iobuf->dr_filter = filter; - init_waitqueue_head(&iobuf->dr_wait); - atomic_set(&iobuf->dr_numreqs, 0); - spin_lock_init(&iobuf->dr_lock); + cfs_waitq_init(&iobuf->dr_wait); + cfs_atomic_set(&iobuf->dr_numreqs, 0); iobuf->dr_max_pages = num_pages; iobuf->dr_npages = 0; iobuf->dr_error = 0; @@ -201,7 +255,7 @@ static void filter_clear_iobuf(struct filter_iobuf *iobuf) { iobuf->dr_npages = 0; iobuf->dr_error = 0; - atomic_set(&iobuf->dr_numreqs, 0); + cfs_atomic_set(&iobuf->dr_numreqs, 0); } void filter_free_iobuf(struct filter_iobuf *iobuf) @@ -220,7 +274,8 @@ void filter_free_iobuf(struct filter_iobuf *iobuf) void filter_iobuf_put(struct filter_obd *filter, struct filter_iobuf *iobuf, struct obd_trans_info *oti) { - int thread_id = oti ? oti->oti_thread_id : -1; + int thread_id = (oti && oti->oti_thread) ? + oti->oti_thread->t_id : -1; if (unlikely(thread_id < 0)) { filter_free_iobuf(iobuf); @@ -298,6 +353,16 @@ int filter_do_bio(struct obd_export *exp, struct inode *inode, sector_bits)) nblocks++; + /* I only set the page to be constant only if it + * is mapped to a contiguous underlying disk block(s). + * It will then make sure the corresponding device + * cache of raid5 will be overwritten by this page. + * - jay */ + if ((rw == OBD_BRW_WRITE) && + (nblocks == blocks_per_page) && + mapping_cap_page_constant_write(inode->i_mapping)) + SetPageConstant(page); + if (bio != NULL && can_be_merged(bio, sector) && bio_add_page(bio, page, @@ -305,19 +370,22 @@ int filter_do_bio(struct obd_export *exp, struct inode *inode, continue; /* added this frag OK */ if (bio != NULL) { - request_queue_t *q = + struct request_queue *q = bdev_get_queue(bio->bi_bdev); /* Dang! I have to fragment this I/O */ CDEBUG(D_INODE, "bio++ sz %d vcnt %d(%d) " - "sectors %d(%d) psg %d(%d) hsg %d(%d)\n", + "sectors %d(%d) psg %d(%d) hsg %d(%d) " + "sector %llu next %llu\n", bio->bi_size, bio->bi_vcnt, bio->bi_max_vecs, - bio->bi_size >> 9, q->max_sectors, + bio->bi_size >> 9, queue_max_sectors(q), bio_phys_segments(q, bio), - q->max_phys_segments, + queue_max_phys_segments(q), bio_hw_segments(q, bio), - q->max_hw_segments); + queue_max_hw_segments(q), + (unsigned long long)bio->bi_sector, + (unsigned long long)sector); record_start_io(iobuf, rw, bio->bi_size, exp); rc = fsfilt_send_bio(rw, obd, inode, bio); @@ -329,9 +397,10 @@ int filter_do_bio(struct obd_export *exp, struct inode *inode, frags++; } - /* allocate new bio */ - bio = bio_alloc(GFP_NOIO, - (npages - page_idx) * blocks_per_page); + /* allocate new bio, limited by max BIO size, b=9945 */ + bio = bio_alloc(GFP_NOIO, min(BIO_MAX_PAGES, + (npages - page_idx) * + blocks_per_page)); if (bio == NULL) { CERROR("Can't allocate bio %u*%u = %u pages\n", (npages - page_idx), blocks_per_page, @@ -364,24 +433,38 @@ int filter_do_bio(struct obd_export *exp, struct inode *inode, } out: - wait_event(iobuf->dr_wait, atomic_read(&iobuf->dr_numreqs) == 0); + cfs_wait_event(iobuf->dr_wait, + cfs_atomic_read(&iobuf->dr_numreqs) == 0); if (rw == OBD_BRW_READ) { - lprocfs_oh_tally(&obd->u.filter.fo_filter_stats.hist[BRW_R_DIO_FRAGS], + lprocfs_oh_tally(&obd->u.filter.fo_filter_stats. + hist[BRW_R_DIO_FRAGS], frags); - lprocfs_oh_tally(&exp->exp_filter_data.fed_brw_stats.hist[BRW_R_DIO_FRAGS], - frags); - lprocfs_oh_tally_log2(&obd->u.filter.fo_filter_stats.hist[BRW_R_IO_TIME], + lprocfs_oh_tally_log2(&obd->u.filter. + fo_filter_stats.hist[BRW_R_IO_TIME], jiffies - start_time); - lprocfs_oh_tally_log2(&exp->exp_filter_data.fed_brw_stats.hist[BRW_R_IO_TIME], jiffies - start_time); + if (exp->exp_nid_stats && exp->exp_nid_stats->nid_brw_stats) { + lprocfs_oh_tally(&exp->exp_nid_stats->nid_brw_stats-> + hist[BRW_R_DIO_FRAGS], + frags); + lprocfs_oh_tally_log2(&exp->exp_nid_stats-> + nid_brw_stats->hist[BRW_R_IO_TIME], + jiffies - start_time); + } } else { - lprocfs_oh_tally(&obd->u.filter.fo_filter_stats.hist[BRW_W_DIO_FRAGS], - frags); - lprocfs_oh_tally(&exp->exp_filter_data.fed_brw_stats.hist[BRW_W_DIO_FRAGS], - frags); - lprocfs_oh_tally_log2(&obd->u.filter.fo_filter_stats.hist[BRW_W_IO_TIME], + lprocfs_oh_tally(&obd->u.filter.fo_filter_stats. + hist[BRW_W_DIO_FRAGS], frags); + lprocfs_oh_tally_log2(&obd->u.filter.fo_filter_stats. + hist[BRW_W_IO_TIME], jiffies - start_time); - lprocfs_oh_tally_log2(&exp->exp_filter_data.fed_brw_stats.hist[BRW_W_IO_TIME], jiffies - start_time); + if (exp->exp_nid_stats && exp->exp_nid_stats->nid_brw_stats) { + lprocfs_oh_tally(&exp->exp_nid_stats->nid_brw_stats-> + hist[BRW_W_DIO_FRAGS], + frags); + lprocfs_oh_tally_log2(&exp->exp_nid_stats-> + nid_brw_stats->hist[BRW_W_IO_TIME], + jiffies - start_time); + } } if (rc == 0) @@ -389,109 +472,6 @@ int filter_do_bio(struct obd_export *exp, struct inode *inode, RETURN(rc); } -/* These are our hacks to keep our directio/bh IO coherent with ext3's - * page cache use. Most notably ext3 reads file data into the page - * cache when it is zeroing the tail of partial-block truncates and - * leaves it there, sometimes generating io from it at later truncates. - * This removes the partial page and its buffers from the page cache, - * so it should only ever cause a wait in rare cases, as otherwise we - * always do full-page IO to the OST. - * - * The call to truncate_complete_page() will call journal_invalidatepage() - * to free the buffers and drop the page from cache. The buffers should - * not be dirty, because we already called fdatasync/fdatawait on them. - */ -static int filter_sync_inode_data(struct inode *inode, int locked) -{ - int rc = 0; - - /* This is nearly do_fsync(), without the waiting on the inode */ - /* XXX: in 2.6.16 (at least) we don't need to hold i_mutex over - * filemap_fdatawrite() and filemap_fdatawait(), so we may no longer - * need this lock here at all. */ - if (!locked) - LOCK_INODE_MUTEX(inode); - if (inode->i_mapping->nrpages) { -#ifdef PF_SYNCWRITE - current->flags |= PF_SYNCWRITE; -#endif - rc = filemap_fdatawrite(inode->i_mapping); - if (rc == 0) - rc = filemap_fdatawait(inode->i_mapping); -#ifdef PF_SYNCWRITE - current->flags &= ~PF_SYNCWRITE; -#endif - } - if (!locked) - UNLOCK_INODE_MUTEX(inode); - - return rc; -} -/* Clear pages from the mapping before we do direct IO to that offset. - * Now that the only source of such pages in the truncate path flushes - * these pages to disk and then discards them, this is error condition. - * If add back read cache this will happen again. This could be disabled - * until that time if we never see the below error. */ -static int filter_clear_page_cache(struct inode *inode, - struct filter_iobuf *iobuf) -{ - struct page *page; - int i, rc; - - rc = filter_sync_inode_data(inode, 0); - if (rc != 0) - RETURN(rc); - - /* be careful to call this after fsync_inode_data_buffers has waited - * for IO to complete before we evict it from the cache */ - for (i = 0; i < iobuf->dr_npages; i++) { - page = find_lock_page(inode->i_mapping, - iobuf->dr_pages[i]->index); - if (page == NULL) - continue; - if (page->mapping != NULL) { - CERROR("page %lu (%d/%d) in page cache during write!\n", - page->index, i, iobuf->dr_npages); - wait_on_page_writeback(page); - ll_truncate_complete_page(page); - } - - unlock_page(page); - page_cache_release(page); - } - - return 0; -} - -int filter_clear_truncated_page(struct inode *inode) -{ - struct page *page; - int rc; - - /* Truncate on page boundary, so nothing to flush? */ - if (!(i_size_read(inode) & ~CFS_PAGE_MASK)) - return 0; - - rc = filter_sync_inode_data(inode, 1); - if (rc != 0) - RETURN(rc); - - /* be careful to call this after fsync_inode_data_buffers has waited - * for IO to complete before we evict it from the cache */ - page = find_lock_page(inode->i_mapping, - i_size_read(inode) >> CFS_PAGE_SHIFT); - if (page) { - if (page->mapping != NULL) { - wait_on_page_writeback(page); - ll_truncate_complete_page(page); - } - unlock_page(page); - page_cache_release(page); - } - - return 0; -} - /* Must be called with i_mutex taken for writes; this will drop it */ int filter_direct_io(int rw, struct dentry *dchild, struct filter_iobuf *iobuf, struct obd_export *exp, struct iattr *attr, @@ -501,7 +481,7 @@ int filter_direct_io(int rw, struct dentry *dchild, struct filter_iobuf *iobuf, struct inode *inode = dchild->d_inode; int blocks_per_page = CFS_PAGE_SIZE >> inode->i_blkbits; int rc, rc2, create; - struct semaphore *sem; + cfs_semaphore_t *sem; ENTRY; LASSERTF(iobuf->dr_npages <= iobuf->dr_max_pages, "%d,%d\n", @@ -519,12 +499,18 @@ int filter_direct_io(int rw, struct dentry *dchild, struct filter_iobuf *iobuf, create = 1; sem = &obd->u.filter.fo_alloc_lock; - lquota_enforce(filter_quota_interface_ref, obd, iobuf->dr_ignore_quota); + lquota_enforce(filter_quota_interface_ref, obd, + iobuf->dr_ignore_quota); } - rc = fsfilt_map_inode_pages(obd, inode, iobuf->dr_pages, + if (rw == OBD_BRW_WRITE && + OBD_FAIL_CHECK(OBD_FAIL_OST_MAPBLK_ENOSPC)) { + rc = -ENOSPC; + } else { + rc = fsfilt_map_inode_pages(obd, inode, iobuf->dr_pages, iobuf->dr_npages, iobuf->dr_blocks, obdfilter_created_scratchpad, create, sem); + } if (rw == OBD_BRW_WRITE) { if (rc == 0) { @@ -539,15 +525,25 @@ int filter_direct_io(int rw, struct dentry *dchild, struct filter_iobuf *iobuf, UNLOCK_INODE_MUTEX(inode); - rc2 = filter_finish_transno(exp, oti, 0, 0); + /* Force commit to make the just-deleted blocks + * reusable. LU-456 */ + if (rc == -ENOSPC) { + fsfilt_commit(obd, inode, oti->oti_handle, 1); + RETURN(rc); + } + + rc2 = filter_finish_transno(exp, inode, oti, 0, 0); if (rc2 != 0) { CERROR("can't close transaction: %d\n", rc2); if (rc == 0) rc = rc2; } - rc2 = fsfilt_commit_async(obd,inode,oti->oti_handle, - wait_handle); + if (wait_handle) + rc2 = fsfilt_commit_async(obd,inode,oti->oti_handle, + wait_handle); + else + rc2 = fsfilt_commit(obd, inode, oti->oti_handle, 0); if (rc == 0) rc = rc2; if (rc != 0) @@ -557,10 +553,6 @@ int filter_direct_io(int rw, struct dentry *dchild, struct filter_iobuf *iobuf, iobuf->dr_blocks, blocks_per_page, 0); } - rc = filter_clear_page_cache(inode, iobuf); - if (rc != 0) - RETURN(rc); - RETURN(filter_do_bio(exp, inode, iobuf, rw)); } @@ -585,8 +577,20 @@ static int filter_range_is_mapped(struct inode *inode, obd_size offset, int len) return 1; } +/* + * interesting use cases on how it interacts with VM: + * + * - vm writeout -- shouldn't see our pages as we don't mark them dirty + * though vm can find partial page left dirty by truncate. in this + * usual writeout is used unless our write rewrite that page - then we + * drop PG_dirty with PG_lock held. + * + * - else? + * + */ int filter_commitrw_write(struct obd_export *exp, struct obdo *oa, - int objcount, struct obd_ioobj *obj, int niocount, + int objcount, struct obd_ioobj *obj, + struct niobuf_remote *nb, int niocount, struct niobuf_local *res, struct obd_trans_info *oti, int rc) { @@ -595,13 +599,17 @@ int filter_commitrw_write(struct obd_export *exp, struct obdo *oa, struct lvfs_run_ctxt saved; struct fsfilt_objinfo fso; struct iattr iattr = { 0 }; - struct inode *inode = NULL; + struct inode *inode = res->dentry->d_inode; unsigned long now = jiffies; int i, err, cleanup_phase = 0; struct obd_device *obd = exp->exp_obd; - void *wait_handle; - int total_size = 0, rc2; - unsigned int qcids[MAXQUOTAS] = {0, 0}; + struct filter_obd *fo = &obd->u.filter; + void *wait_handle = NULL; + int total_size = 0; + unsigned int qcids[MAXQUOTAS] = { oa->o_uid, oa->o_gid }; + int rec_pending[MAXQUOTAS] = { 0, 0 }, quota_pages = 0; + int sync_journal_commit = obd->u.filter.fo_syncjournal; + int retries = 0; ENTRY; LASSERT(oti != NULL); @@ -611,22 +619,6 @@ int filter_commitrw_write(struct obd_export *exp, struct obdo *oa, if (rc != 0) GOTO(cleanup, rc); - /* Unfortunately, if quota master is too busy to handle the - * pre-dqacq in time and quota hash on ost is used up, we - * have to wait for the completion of in flight dqacq/dqrel, - * then try again */ - if ((rc2 = lquota_chkquota(filter_quota_interface_ref, obd, oa->o_uid, - oa->o_gid, niocount)) == QUOTA_RET_ACQUOTA) { - OBD_FAIL_TIMEOUT(OBD_FAIL_OST_HOLD_WRITE_RPC, 90); - lquota_acquire(filter_quota_interface_ref, obd, oa->o_uid, - oa->o_gid); - } - - if (rc2 < 0) { - rc = rc2; - GOTO(cleanup, rc); - } - iobuf = filter_iobuf_get(&obd->u.filter, oti); if (IS_ERR(iobuf)) GOTO(cleanup, rc = PTR_ERR(iobuf)); @@ -634,22 +626,36 @@ int filter_commitrw_write(struct obd_export *exp, struct obdo *oa, fso.fso_dentry = res->dentry; fso.fso_bufcnt = obj->ioo_bufcnt; - inode = res->dentry->d_inode; iobuf->dr_ignore_quota = 0; - for (i = 0, lnb = res; i < obj->ioo_bufcnt; i++, lnb++) { + for (i = 0, lnb = res; i < niocount; i++, lnb++) { loff_t this_size; + __u32 flags = lnb->flags; - /* If overwriting an existing block, we don't need a grant */ - if (!(lnb->flags & OBD_BRW_GRANTED) && lnb->rc == -ENOSPC && - filter_range_is_mapped(inode, lnb->offset, lnb->len)) - lnb->rc = 0; + if (filter_range_is_mapped(inode, lnb->offset, lnb->len)) { + /* If overwriting an existing block, + * we don't need a grant */ + if (!(flags & OBD_BRW_GRANTED) && lnb->rc == -ENOSPC) + lnb->rc = 0; + } else { + quota_pages++; + } if (lnb->rc) { /* ENOSPC, network RPC error, etc. */ CDEBUG(D_INODE, "Skipping [%d] == %d\n", i, lnb->rc); continue; } + LASSERT(PageLocked(lnb->page)); + LASSERT(!PageWriteback(lnb->page)); + + /* since write & truncate are serialized by the i_alloc_sem, + * even partial truncate should not leave dirty pages in + * the page cache */ + LASSERT(!PageDirty(lnb->page)); + + SetPageUptodate(lnb->page); + err = filter_iobuf_add_page(obd, iobuf, inode, lnb->page); LASSERT (err == 0); @@ -661,20 +667,39 @@ int filter_commitrw_write(struct obd_export *exp, struct obdo *oa, if (this_size > iattr.ia_size) iattr.ia_size = this_size; - /* if one page is a write-back page from client cache, or it's - * written by root, then mark the whole io request as ignore - * quota request */ - if (lnb->flags & (OBD_BRW_FROM_GRANT | OBD_BRW_NOQUOTA)) + /* if one page is a write-back page from client cache and + * not from direct_io, or it's written by root, then mark + * the whole io request as ignore quota request, remote + * client can not break through quota. */ + if (exp_connect_rmtclient(exp)) + flags &= ~OBD_BRW_NOQUOTA; + if ((flags & OBD_BRW_NOQUOTA) || + (flags & (OBD_BRW_FROM_GRANT | OBD_BRW_SYNC)) == + OBD_BRW_FROM_GRANT) iobuf->dr_ignore_quota = 1; + + if (!(lnb->flags & OBD_BRW_ASYNC)) { + sync_journal_commit = 1; + } } + /* we try to get enough quota to write here, and let ldiskfs + * decide if it is out of quota or not b=14783 */ + rc = lquota_chkquota(filter_quota_interface_ref, obd, exp, qcids, + rec_pending, quota_pages, oti, LQUOTA_FLAGS_BLK, + (void *)inode, obj->ioo_bufcnt); + if (rc == -ENOTCONN) + GOTO(cleanup, rc); + push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL); cleanup_phase = 2; - DQUOT_INIT(inode); + ll_vfs_dq_init(inode); + fsfilt_check_slow(obd, now, "quota init"); +retry: LOCK_INODE_MUTEX(inode); - fsfilt_check_slow(obd, now, obd_timeout, "i_mutex"); + fsfilt_check_slow(obd, now, "i_mutex"); oti->oti_handle = fsfilt_brw_start(obd, objcount, &fso, niocount, res, oti); if (IS_ERR(oti->oti_handle)) { @@ -687,7 +712,7 @@ int filter_commitrw_write(struct obd_export *exp, struct obdo *oa, } /* have to call fsfilt_commit() from this point on */ - fsfilt_check_slow(obd, now, obd_timeout, "brw_start"); + fsfilt_check_slow(obd, now, "brw_start"); i = OBD_MD_FLATIME | OBD_MD_FLMTIME | OBD_MD_FLCTIME; @@ -707,7 +732,7 @@ int filter_commitrw_write(struct obd_export *exp, struct obdo *oa, CDEBUG(D_INODE, "update UID/GID to %lu/%lu\n", (unsigned long)oa->o_uid, (unsigned long)oa->o_gid); - cap_raise(current->cap_effective, CAP_SYS_RESOURCE); + cfs_cap_raise(CFS_CAP_SYS_RESOURCE); iattr.ia_valid |= ATTR_MODE; iattr.ia_mode = inode->i_mode; @@ -722,39 +747,52 @@ int filter_commitrw_write(struct obd_export *exp, struct obdo *oa, * in the inode before filter_direct_io() - see bug 10357. */ save = iattr.ia_valid; iattr.ia_valid &= (ATTR_UID | ATTR_GID); - rc = fsfilt_setattr(obd, res->dentry, oti->oti_handle, &iattr, 0); - CDEBUG(D_QUOTA, "set uid(%u)/gid(%u) to ino(%lu). rc(%d)\n", + rc = fsfilt_setattr(obd, res->dentry, oti->oti_handle,&iattr,0); + CDEBUG(D_QUOTA, "set uid(%u)/gid(%u) to ino(%lu). rc(%d)\n", iattr.ia_uid, iattr.ia_gid, inode->i_ino, rc); iattr.ia_valid = save & ~(ATTR_UID | ATTR_GID); } /* filter_direct_io drops i_mutex */ rc = filter_direct_io(OBD_BRW_WRITE, res->dentry, iobuf, exp, &iattr, - oti, &wait_handle); - if (rc == 0) - obdo_from_inode(oa, inode, - FILTER_VALID_FLAGS |OBD_MD_FLUID |OBD_MD_FLGID); - else - obdo_from_inode(oa, inode, OBD_MD_FLUID | OBD_MD_FLGID); + oti, sync_journal_commit ? &wait_handle : NULL); + if (rc == -ENOSPC && retries++ < 3) { + CDEBUG(D_INODE, "retry after force commit, retries:%d\n", + retries); + oti->oti_handle = NULL; + fsfilt_check_slow(obd, now, "direct_io"); + goto retry; + } + + obdo_from_inode(oa, inode, NULL, rc == 0 ? FILTER_VALID_FLAGS : 0 | + OBD_MD_FLUID |OBD_MD_FLGID); lquota_getflag(filter_quota_interface_ref, obd, oa); - fsfilt_check_slow(obd, now, obd_timeout, "direct_io"); + fsfilt_check_slow(obd, now, "direct_io"); + + if (wait_handle) + err = fsfilt_commit_wait(obd, inode, wait_handle); + else + err = 0; - err = fsfilt_commit_wait(obd, inode, wait_handle); if (err) { CERROR("Failure to commit OST transaction (%d)?\n", err); - rc = err; + if (rc == 0) + rc = err; } - if (obd->obd_replayable && !rc) + if (obd->obd_replayable && !rc && wait_handle) LASSERTF(oti->oti_transno <= obd->obd_last_committed, "oti_transno "LPU64" last_committed "LPU64"\n", oti->oti_transno, obd->obd_last_committed); - fsfilt_check_slow(obd, now, obd_timeout, "commitrw commit"); + fsfilt_check_slow(obd, now, "commitrw commit"); cleanup: + lquota_pending_commit(filter_quota_interface_ref, obd, qcids, + rec_pending, 1); + filter_grant_commit(exp, niocount, res); switch (cleanup_phase) { @@ -768,16 +806,50 @@ cleanup: * lnb->page automatically returns back into per-thread page * pool (bug 5137) */ - f_dput(res->dentry); + break; } /* trigger quota pre-acquire */ - qcids[USRQUOTA] = oa->o_uid; - qcids[GRPQUOTA] = oa->o_gid; err = lquota_adjust(filter_quota_interface_ref, obd, qcids, NULL, rc, FSFILT_OP_CREATE); - CDEBUG(err ? D_ERROR : D_QUOTA, - "filter adjust qunit! (rc:%d)\n", err); + CDEBUG(err ? D_ERROR : D_QUOTA, "filter adjust qunit! " + "(rc:%d, uid:%u, gid:%u)\n", + err, qcids[USRQUOTA], qcids[GRPQUOTA]); + if (qcids[USRQUOTA] != oa->o_uid || qcids[GRPQUOTA] != oa->o_gid) { + qcids[USRQUOTA] = oa->o_uid; + qcids[GRPQUOTA] = oa->o_gid; + err = lquota_adjust(filter_quota_interface_ref, obd, qcids, + NULL, rc, FSFILT_OP_CREATE); + CDEBUG(err ? D_ERROR : D_QUOTA, "filter adjust qunit! " + "(rc:%d, uid:%u, gid:%u)\n", + err, qcids[USRQUOTA], qcids[GRPQUOTA]); + } + + for (i = 0, lnb = res; i < niocount; i++, lnb++) { + if (lnb->page == NULL) + continue; + + if (rc) + /* If the write has failed, the page cache may + * not be consitent with what is on disk, so + * force pages to be reread next time it is + * accessed */ + ClearPageUptodate(lnb->page); + + LASSERT(PageLocked(lnb->page)); + unlock_page(lnb->page); + + page_cache_release(lnb->page); + lnb->page = NULL; + } + f_dput(res->dentry); + + if (inode) { + if (fo->fo_writethrough_cache == 0 || + i_size_read(inode) > fo->fo_readcache_max_filesize) + filter_release_cache(obd, obj, nb, inode); + up_read(&inode->i_alloc_sem); + } RETURN(rc); }