X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lustre%2Fobdfilter%2Ffilter_io_26.c;h=ad109e4b00e0b888f663334bf1f13edc53923184;hp=5ce65c7b31130be4c60fa5c1c3d3b82393b0a99d;hb=904ce4f8c9a4eb2b0dd1dff02e2159b0a73a1a99;hpb=4721137e38a657ab5fdccb2b75c7a7d0e3957a4d diff --git a/lustre/obdfilter/filter_io_26.c b/lustre/obdfilter/filter_io_26.c index 5ce65c7..ad109e4 100644 --- a/lustre/obdfilter/filter_io_26.c +++ b/lustre/obdfilter/filter_io_26.c @@ -28,6 +28,7 @@ #include #include // XXX kill me soon #include +#include #define DEBUG_SUBSYSTEM S_FILTER @@ -35,17 +36,18 @@ #include #include "filter_internal.h" -#warning "implement writeback mode -bzzz" - /* 512byte block min */ #define MAX_BLOCKS_PER_PAGE (PAGE_SIZE / 512) struct dio_request { - atomic_t numreqs; /* number of reqs being processed */ - struct bio *bio_list; /* list of completed bios */ - wait_queue_head_t wait; - int created[MAX_BLOCKS_PER_PAGE]; - unsigned long blocks[MAX_BLOCKS_PER_PAGE]; - spinlock_t lock; + atomic_t dr_numreqs; /* number of reqs being processed */ + struct bio *dr_bios; /* list of completed bios */ + 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; }; static int dio_complete_routine(struct bio *bio, unsigned int done, int error) @@ -53,20 +55,46 @@ static int dio_complete_routine(struct bio *bio, unsigned int done, int error) struct dio_request *dreq = bio->bi_private; unsigned long flags; - spin_lock_irqsave(&dreq->lock, flags); - bio->bi_private = dreq->bio_list; - dreq->bio_list = bio; - spin_unlock_irqrestore(&dreq->lock, flags); - if (atomic_dec_and_test(&dreq->numreqs)) - wake_up(&dreq->wait); + if (bio->bi_size) { + CWARN("gets called against non-complete bio 0x%p: %d/%d/%d\n", + bio, bio->bi_size, done, error); + return 1; + } + + if (dreq == 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."); + 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_private); + return 0; + } + + spin_lock_irqsave(&dreq->dr_lock, flags); + bio->bi_private = dreq->dr_bios; + dreq->dr_bios = bio; + if (dreq->dr_error == 0) + dreq->dr_error = error; + spin_unlock_irqrestore(&dreq->dr_lock, flags); + + if (atomic_dec_and_test(&dreq->dr_numreqs)) + wake_up(&dreq->dr_wait); return 0; } static int can_be_merged(struct bio *bio, sector_t sector) { - int size; - + unsigned int size; if (!bio) return 0; @@ -74,6 +102,315 @@ static int can_be_merged(struct bio *bio, sector_t sector) return bio->bi_sector + size == sector ? 1 : 0; } + +int filter_alloc_iobuf(int rw, int num_pages, void **ret) +{ + struct dio_request *dreq; + + LASSERTF(rw == OBD_BRW_WRITE || rw == OBD_BRW_READ, "%x\n", rw); + + OBD_ALLOC(dreq, sizeof(*dreq)); + if (dreq == NULL) + goto failed_0; + + OBD_ALLOC(dreq->dr_pages, num_pages * sizeof(*dreq->dr_pages)); + if (dreq->dr_pages == NULL) + goto failed_1; + + OBD_ALLOC(dreq->dr_blocks, + MAX_BLOCKS_PER_PAGE * num_pages * sizeof(*dreq->dr_blocks)); + if (dreq->dr_blocks == NULL) + goto failed_2; + + dreq->dr_bios = NULL; + init_waitqueue_head(&dreq->dr_wait); + atomic_set(&dreq->dr_numreqs, 0); + spin_lock_init(&dreq->dr_lock); + dreq->dr_max_pages = num_pages; + dreq->dr_npages = 0; + + *ret = dreq; + RETURN(0); + + failed_2: + OBD_FREE(dreq->dr_pages, + num_pages * sizeof(*dreq->dr_pages)); + failed_1: + OBD_FREE(dreq, sizeof(*dreq)); + failed_0: + RETURN(-ENOMEM); +} + +void filter_free_iobuf(void *iobuf) +{ + struct dio_request *dreq = iobuf; + int num_pages = dreq->dr_max_pages; + + /* free all bios */ + while (dreq->dr_bios) { + struct bio *bio = dreq->dr_bios; + dreq->dr_bios = bio->bi_private; + bio_put(bio); + } + + OBD_FREE(dreq->dr_blocks, + MAX_BLOCKS_PER_PAGE * num_pages * sizeof(*dreq->dr_blocks)); + OBD_FREE(dreq->dr_pages, + num_pages * sizeof(*dreq->dr_pages)); + OBD_FREE(dreq, sizeof(*dreq)); +} + +int filter_iobuf_add_page(struct obd_device *obd, void *iobuf, + struct inode *inode, struct page *page) +{ + struct dio_request *dreq = iobuf; + + LASSERT (dreq->dr_npages < dreq->dr_max_pages); + dreq->dr_pages[dreq->dr_npages++] = page; + + return 0; +} + +int filter_do_bio(struct obd_device *obd, struct inode *inode, + struct dio_request *dreq, int rw) +{ + int blocks_per_page = PAGE_SIZE >> inode->i_blkbits; + struct page **pages = dreq->dr_pages; + int npages = dreq->dr_npages; + unsigned long *blocks = dreq->dr_blocks; + int total_blocks = npages * blocks_per_page; + int sector_bits = inode->i_sb->s_blocksize_bits - 9; + unsigned int blocksize = inode->i_sb->s_blocksize; + struct bio *bio = NULL; + struct page *page; + unsigned int page_offset; + sector_t sector; + int nblocks; + int block_idx; + int page_idx; + int i; + int rc = 0; + ENTRY; + + LASSERT(dreq->dr_npages == npages); + LASSERT(total_blocks <= OBDFILTER_CREATED_SCRATCHPAD_ENTRIES); + + for (page_idx = 0, block_idx = 0; + page_idx < npages; + page_idx++, block_idx += blocks_per_page) { + + page = pages[page_idx]; + LASSERT (block_idx + blocks_per_page <= total_blocks); + + for (i = 0, page_offset = 0; + i < blocks_per_page; + i += nblocks, page_offset += blocksize * nblocks) { + + nblocks = 1; + + if (blocks[block_idx + i] == 0) { /* hole */ + LASSERT(rw == OBD_BRW_READ); + memset(kmap(page) + page_offset, 0, blocksize); + kunmap(page); + continue; + } + + sector = blocks[block_idx + i] << sector_bits; + + /* Additional contiguous file blocks? */ + while (i + nblocks < blocks_per_page && + (sector + nblocks*(blocksize>>9)) == + (blocks[block_idx + i + nblocks] << sector_bits)) + nblocks++; + + if (bio != NULL && + can_be_merged(bio, sector) && + bio_add_page(bio, page, + blocksize * nblocks, page_offset) != 0) + continue; /* added this frag OK */ + + if (bio != NULL) { + request_queue_t *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", + bio->bi_size, + bio->bi_vcnt, bio->bi_max_vecs, + bio->bi_size >> 9, q->max_sectors, + bio_phys_segments(q, bio), + q->max_phys_segments, + bio_hw_segments(q, bio), + q->max_hw_segments); + + atomic_inc(&dreq->dr_numreqs); + rc = fsfilt_send_bio(rw, obd, inode, bio); + if (rc < 0) { + CERROR("Can't send bio: %d\n", rc); + /* OK do dec; we do the waiting */ + atomic_dec(&dreq->dr_numreqs); + goto out; + } + rc = 0; + + bio = NULL; + } + + /* allocate new bio */ + bio = bio_alloc(GFP_NOIO, + (npages - page_idx) * blocks_per_page); + if (bio == NULL) { + CERROR ("Can't allocate bio\n"); + rc = -ENOMEM; + goto out; + } + + bio->bi_bdev = inode->i_sb->s_bdev; + bio->bi_sector = sector; + bio->bi_end_io = dio_complete_routine; + bio->bi_private = dreq; + + rc = bio_add_page(bio, page, + blocksize * nblocks, page_offset); + LASSERT (rc != 0); + } + } + + if (bio != NULL) { + atomic_inc(&dreq->dr_numreqs); + rc = fsfilt_send_bio(rw, obd, inode, bio); + if (rc >= 0) { + rc = 0; + } else { + CERROR("Can't send bio: %d\n", rc); + /* OK do dec; we do the waiting */ + atomic_dec(&dreq->dr_numreqs); + } + } + + out: + wait_event(dreq->dr_wait, atomic_read(&dreq->dr_numreqs) == 0); + + if (rc == 0) + rc = dreq->dr_error; + 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_clear_page_cache(struct inode *inode, + struct dio_request *iobuf) +{ + struct page *page; + int i, rc, rc2; + + /* This is nearly generic_osync_inode, without the waiting on the inode + rc = generic_osync_inode(inode, inode->i_mapping, + OSYNC_DATA|OSYNC_METADATA); + */ + rc = filemap_fdatawrite(inode->i_mapping); + rc2 = sync_mapping_buffers(inode->i_mapping); + if (rc == 0) + rc = rc2; + rc2 = filemap_fdatawait(inode->i_mapping); + if (rc == 0) + rc = rc2; + 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) { + wait_on_page_writeback(page); + ll_truncate_complete_page(page); + } + + unlock_page(page); + page_cache_release(page); + } + return 0; +} +/* Must be called with i_sem taken for writes; this will drop it */ +int filter_direct_io(int rw, struct dentry *dchild, void *iobuf, + struct obd_export *exp, struct iattr *attr, + struct obd_trans_info *oti, void **wait_handle) +{ + struct obd_device *obd = exp->exp_obd; + struct inode *inode = dchild->d_inode; + struct dio_request *dreq = iobuf; + struct semaphore *sem = NULL; + int rc, rc2, create = 0; + ENTRY; + + LASSERTF(rw == OBD_BRW_WRITE || rw == OBD_BRW_READ, "%x\n", rw); + LASSERTF(dreq->dr_npages <= dreq->dr_max_pages, "%d,%d\n", + dreq->dr_npages, dreq->dr_max_pages); + + if (dreq->dr_npages == 0) + RETURN(0); + + if (dreq->dr_npages > OBDFILTER_CREATED_SCRATCHPAD_ENTRIES) + RETURN(-EINVAL); + + if (rw == OBD_BRW_WRITE) { + create = 1; + //sem = &obd->u.filter.fo_alloc_lock; + } + + rc = fsfilt_map_inode_pages(obd, inode, + dreq->dr_pages, dreq->dr_npages, + dreq->dr_blocks, + obdfilter_created_scratchpad, + create, sem); + + if (rw == OBD_BRW_WRITE) { + if (rc == 0) { + int blocks_per_page = PAGE_SIZE >> inode->i_blkbits; + filter_tally_write(&obd->u.filter, dreq->dr_pages, + dreq->dr_npages, dreq->dr_blocks, + blocks_per_page); + if (attr->ia_size > inode->i_size) + attr->ia_valid |= ATTR_SIZE; + rc = fsfilt_setattr(obd, dchild, + oti->oti_handle, attr, 0); + } + + up(&inode->i_sem); + + rc2 = filter_finish_transno(exp, oti, 0); + if (rc2 != 0) + CERROR("can't close transaction: %d\n", rc); + rc = (rc == 0) ? rc2 : rc; + + rc2 = fsfilt_commit_async(obd,inode,oti->oti_handle,wait_handle); + rc = (rc == 0) ? rc2 : rc; + + if (rc != 0) + RETURN(rc); + } + + rc = filter_clear_page_cache(inode, dreq); + if (rc != 0) + RETURN(rc); + + RETURN(filter_do_bio(obd, inode, dreq, rw)); +} + /* See if there are unallocated parts in given file region */ static int filter_range_is_mapped(struct inode *inode, obd_size offset, int len) { @@ -95,167 +432,125 @@ static int filter_range_is_mapped(struct inode *inode, obd_size offset, int len) return 1; } -int filter_commitrw_write(struct obd_export *exp, struct obdo *oa, int objcount, - struct obd_ioobj *obj, int niocount, +int filter_commitrw_write(struct obd_export *exp, struct obdo *oa, + int objcount, struct obd_ioobj *obj, int niocount, struct niobuf_local *res, struct obd_trans_info *oti, int rc) { - struct obd_device *obd = exp->exp_obd; - struct obd_run_ctxt saved; struct niobuf_local *lnb; + struct dio_request *dreq = NULL; + struct lvfs_run_ctxt saved; struct fsfilt_objinfo fso; - struct iattr iattr = { .ia_valid = ATTR_SIZE, .ia_size = 0, }; + struct iattr iattr = { 0 }; struct inode *inode = NULL; - int i, k, cleanup_phase = 0, err; - unsigned long now = jiffies; /* DEBUGGING OST TIMEOUTS */ - int blocks_per_page; - struct dio_request *dreq; - struct bio *bio = NULL; + unsigned long now = jiffies; + int i, err, cleanup_phase = 0; + struct obd_device *obd = exp->exp_obd; + void *wait_handle = NULL; + int total_size = 0; + loff_t old_size; ENTRY; + LASSERT(oti != NULL); LASSERT(objcount == 1); LASSERT(current->journal_info == NULL); if (rc != 0) GOTO(cleanup, rc); - - inode = res->dentry->d_inode; - blocks_per_page = PAGE_SIZE >> inode->i_blkbits; - LASSERT(blocks_per_page <= MAX_BLOCKS_PER_PAGE); - - OBD_ALLOC(dreq, sizeof(*dreq)); - if (dreq == NULL) - RETURN(-ENOMEM); - dreq->bio_list = NULL; - init_waitqueue_head(&dreq->wait); - atomic_set(&dreq->numreqs, 0); - spin_lock_init(&dreq->lock); - + + rc = filter_alloc_iobuf(OBD_BRW_WRITE, obj->ioo_bufcnt, (void **)&dreq); + if (rc) + GOTO(cleanup, rc); cleanup_phase = 1; + fso.fso_dentry = res->dentry; fso.fso_bufcnt = obj->ioo_bufcnt; + inode = res->dentry->d_inode; - push_ctxt(&saved, &obd->obd_ctxt, NULL); - cleanup_phase = 2; - - oti->oti_handle = fsfilt_brw_start(obd, objcount, &fso, niocount, res, oti); - if (IS_ERR(oti->oti_handle)) { - rc = PTR_ERR(oti->oti_handle); - CDEBUG(rc == -ENOSPC ? D_INODE : D_ERROR, - "error starting transaction: rc = %d\n", rc); - oti->oti_handle = NULL; - GOTO(cleanup, rc); - } - - if (time_after(jiffies, now + 15 * HZ)) - CERROR("slow brw_start %lus\n", (jiffies - now) / HZ); - - iattr_from_obdo(&iattr,oa,OBD_MD_FLATIME|OBD_MD_FLMTIME|OBD_MD_FLCTIME); for (i = 0, lnb = res; i < obj->ioo_bufcnt; i++, lnb++) { loff_t this_size; - sector_t sector; - int offs; /* 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 (lnb->rc) /* ENOSPC, network RPC error */ + if (lnb->rc) { /* ENOSPC, network RPC error, etc. */ + CDEBUG(D_INODE, "Skipping [%d] == %d\n", i, lnb->rc); continue; - - /* get block number for next page */ - rc = fsfilt_map_inode_page(obd, inode, lnb->page, dreq->blocks, - dreq->created, 1); - if (rc) - GOTO(cleanup, rc); - - for (k = 0; k < blocks_per_page; k++) { - sector = dreq->blocks[k] *(inode->i_sb->s_blocksize>>9); - offs = k * inode->i_sb->s_blocksize; - - if (!bio || !can_be_merged(bio, sector) || - !bio_add_page(bio, lnb->page, lnb->len, offs)) { - if (bio) { - atomic_inc(&dreq->numreqs); - submit_bio(WRITE, bio); - bio = NULL; - } - /* allocate new bio */ - bio = bio_alloc(GFP_NOIO, obj->ioo_bufcnt); - bio->bi_bdev = inode->i_sb->s_bdev; - bio->bi_sector = sector; - bio->bi_end_io = dio_complete_routine; - bio->bi_private = dreq; - - if (!bio_add_page(bio, lnb->page, lnb->len, 0)) - LBUG(); - } } - /* We expect these pages to be in offset order, but we'll + err = filter_iobuf_add_page(obd, dreq, inode, lnb->page); + LASSERT (err == 0); + + total_size += lnb->len; + + /* we expect these pages to be in offset order, but we'll * be forgiving */ this_size = lnb->offset + lnb->len; if (this_size > iattr.ia_size) iattr.ia_size = this_size; } +#if 0 + /* I use this when I'm checking our lovely 1M I/Os reach the disk -eeb */ + if (total_size != (1<<20)) + CWARN("total size %d (%d pages)\n", + total_size, total_size/PAGE_SIZE); +#endif + push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL); + cleanup_phase = 2; -#warning This probably needs filemap_fdatasync() like filter_io_24 (bug 2366) - if (bio) { - atomic_inc(&dreq->numreqs); - submit_bio(WRITE, bio); + down(&inode->i_sem); + old_size = inode->i_size; + oti->oti_handle = fsfilt_brw_start(obd, objcount, &fso, niocount, res, + oti); + if (IS_ERR(oti->oti_handle)) { + up(&inode->i_sem); + rc = PTR_ERR(oti->oti_handle); + CDEBUG(rc == -ENOSPC ? D_INODE : D_ERROR, + "error starting transaction: rc = %d\n", rc); + oti->oti_handle = NULL; + GOTO(cleanup, rc); } + /* have to call fsfilt_commit() from this point on */ - /* time to wait for I/O completion */ - wait_event(dreq->wait, atomic_read(&dreq->numreqs) == 0); + fsfilt_check_slow(now, obd_timeout, "brw_start"); - /* free all bios */ - while (dreq->bio_list) { - bio = dreq->bio_list; - dreq->bio_list = bio->bi_private; - bio_put(bio); + iattr_from_obdo(&iattr,oa,OBD_MD_FLATIME|OBD_MD_FLMTIME|OBD_MD_FLCTIME); + /* filter_direct_io drops i_sem */ + rc = filter_direct_io(OBD_BRW_WRITE, res->dentry, dreq, exp, &iattr, + oti, &wait_handle); + +#if 0 + if (inode->i_size != old_size) { + struct llog_cookie *cookie = obdo_logcookie(oa); + struct lustre_id *id = obdo_id(oa); + filter_log_sz_change(obd, id, oa->o_easize, cookie, inode); } +#endif - if (rc == 0) { - down(&inode->i_sem); - if (iattr.ia_size > inode->i_size) { - CDEBUG(D_INFO, "setting i_size to "LPU64"\n", - iattr.ia_size); - fsfilt_setattr(obd, res->dentry, oti->oti_handle, - &iattr, 0); - } - up(&inode->i_sem); - } + if (rc == 0) + obdo_from_inode(oa, inode, FILTER_VALID_FLAGS); - if (time_after(jiffies, now + 15 * HZ)) - CERROR("slow direct_io %lus\n", (jiffies - now) / HZ); + fsfilt_check_slow(now, obd_timeout, "direct_io"); - rc = filter_finish_transno(exp, oti, rc); - err = fsfilt_commit(obd, inode, oti->oti_handle, obd_sync_filter); - if (err) + err = fsfilt_commit_wait(obd, inode, wait_handle); + if (rc == 0) rc = err; - if (obd_sync_filter) - LASSERT(oti->oti_transno <= obd->obd_last_committed); - if (time_after(jiffies, now + 15 * HZ)) - CERROR("slow commitrw commit %lus\n", (jiffies - now) / HZ); + + fsfilt_check_slow(now, obd_timeout, "commitrw commit"); cleanup: filter_grant_commit(exp, niocount, res); switch (cleanup_phase) { case 2: - pop_ctxt(&saved, &obd->obd_ctxt, NULL); + pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL); LASSERT(current->journal_info == NULL); case 1: - OBD_FREE(dreq, sizeof(*dreq)); + filter_free_iobuf(dreq); case 0: - for (i = 0, lnb = res; i < obj->ioo_bufcnt; i++, lnb++) { - /* flip_.. gets a ref, while free_page only frees - * when it decrefs to 0 */ - if (rc == 0) - flip_into_page_cache(inode, lnb->page); - __free_page(lnb->page); - } + filter_free_dio_pages(objcount, obj, niocount, res); f_dput(res->dentry); }