Whamcloud - gitweb
git://git.whamcloud.com
/
fs
/
lustre-release.git
/ blobdiff
commit
grep
author
committer
pickaxe
?
search:
re
summary
|
shortlog
|
log
|
commit
|
commitdiff
|
tree
raw
| inline |
side by side
LU-12593 osd: up i_append_sem during errors
[fs/lustre-release.git]
/
lustre
/
osd-ldiskfs
/
osd_io.c
diff --git
a/lustre/osd-ldiskfs/osd_io.c
b/lustre/osd-ldiskfs/osd_io.c
index
59c87fc
..
64b995d
100644
(file)
--- a/
lustre/osd-ldiskfs/osd_io.c
+++ b/
lustre/osd-ldiskfs/osd_io.c
@@
-164,7
+164,6
@@
static void dio_complete_routine(struct bio *bio, int error)
{
#endif
struct osd_iobuf *iobuf = bio->bi_private;
- int iter;
struct bio_vec *bvl;
/* CAVEAT EMPTOR: possibly in IRQ context
@@
-191,7
+190,9
@@
static void dio_complete_routine(struct bio *bio, int error)
/* the check is outside of the cycle for performance reason -bzzz */
if (!bio_data_dir(bio)) {
- bio_for_each_segment_all(bvl, bio, iter) {
+ DECLARE_BVEC_ITER_ALL(iter_all);
+
+ bio_for_each_segment_all(bvl, bio, iter_all) {
if (likely(error == 0))
SetPageUptodate(bvl_to_page(bvl));
LASSERT(PageLocked(bvl_to_page(bvl)));
@@
-279,11
+280,11
@@
static int can_be_merged(struct bio *bio, sector_t sector)
static void bio_integrity_fault_inject(struct bio *bio)
{
struct bio_vec *bvec;
-
int i
;
+
DECLARE_BVEC_ITER_ALL(iter_all)
;
void *kaddr;
char *addr;
- bio_for_each_segment_all(bvec, bio, i) {
+ bio_for_each_segment_all(bvec, bio, i
ter_all
) {
struct page *page = bvec->bv_page;
kaddr = kmap(page);
@@
-329,12
+330,13
@@
static int osd_bio_integrity_compare(struct bio *bio, struct block_device *bdev,
bip->bip_vec->bv_offset;
struct bio_vec *bv;
sector_t sector = bio_start_sector(bio);
- unsigned int i, sectors, total;
+ unsigned int sectors, total;
+ DECLARE_BVEC_ITER_ALL(iter_all);
__u16 *expected_guard;
int rc;
total = 0;
- bio_for_each_segment_all(bv, bio, i) {
+ bio_for_each_segment_all(bv, bio, i
ter_all
) {
lnb = iobuf->dr_lnbs[index];
expected_guard = lnb->lnb_guards;
sectors = bv->bv_len / sector_size;
@@
-468,7
+470,7
@@
static int osd_do_bio(struct osd_device *osd, struct inode *inode,
int rc = 0;
bool fault_inject;
bool integrity_enabled;
-
DECLARE_PLUG(plug)
;
+
struct blk_plug plug
;
ENTRY;
fault_inject = OBD_FAIL_CHECK(OBD_FAIL_OST_INTEGRITY_FAULT);
@@
-522,14
+524,13
@@
static int osd_do_bio(struct osd_device *osd, struct inode *inode,
unsigned int bi_size = bio_sectors(bio) << 9;
/* Dang! I have to fragment this I/O */
- CDEBUG(D_INODE,
"bio++ sz %d vcnt %d(%d) "
- "
sectors %d(%d) psg %d(%d) h
sg %d(%d)\n",
+ CDEBUG(D_INODE,
+ "
bio++ sz %d vcnt %d(%d) sectors %d(%d) p
sg %d(%d)\n",
bi_size, bio->bi_vcnt, bio->bi_max_vecs,
bio_sectors(bio),
queue_max_sectors(q),
- bio_phys_segments(q, bio),
- queue_max_phys_segments(q),
- 0, queue_max_hw_segments(q));
+ osd_bio_nr_segs(bio),
+ queue_max_segments(q));
rc = osd_bio_integrity_handle(osd, bio,
iobuf, bio_start_page_idx,
fault_inject, integrity_enabled);
@@
-610,8
+611,9
@@
out:
}
static int osd_map_remote_to_local(loff_t offset, ssize_t len, int *nrpages,
-
struct niobuf_local *
lnb)
+
struct niobuf_local *lnb, int max
lnb)
{
+ int rc = 0;
ENTRY;
*nrpages = 0;
@@
-620,6
+622,11
@@
static int osd_map_remote_to_local(loff_t offset, ssize_t len, int *nrpages,
int poff = offset & (PAGE_SIZE - 1);
int plen = PAGE_SIZE - poff;
+ if (*nrpages >= maxlnb) {
+ rc = -EOVERFLOW;
+ break;
+ }
+
if (plen > len)
plen = len;
lnb->lnb_file_offset = offset;
@@
-631,6
+638,7
@@
static int osd_map_remote_to_local(loff_t offset, ssize_t len, int *nrpages,
lnb->lnb_rc = 0;
lnb->lnb_guard_rpc = 0;
lnb->lnb_guard_disk = 0;
+ lnb->lnb_locked = 0;
LASSERTF(plen <= len, "plen %u, len %lld\n", plen,
(long long) len);
@@
-640,55
+648,64
@@
static int osd_map_remote_to_local(loff_t offset, ssize_t len, int *nrpages,
(*nrpages)++;
}
-
RETURN(0
);
+
RETURN(rc
);
}
static struct page *osd_get_page(const struct lu_env *env, struct dt_object *dt,
- loff_t offset, gfp_t gfp_mask)
+ loff_t offset, gfp_t gfp_mask
, bool cache
)
{
struct osd_thread_info *oti = osd_oti_get(env);
struct inode *inode = osd_dt_obj(dt)->oo_inode;
struct osd_device *d = osd_obj2dev(osd_dt_obj(dt));
struct page *page;
- int cur
= oti->oti_dio_pages_used
;
+ int cur;
LASSERT(inode);
- if (
osd_use_page_cache(d)
) {
+ if (
cache
) {
page = find_or_create_page(inode->i_mapping,
- offset >> PAGE_SHIFT,
- gfp_mask);
+ offset >> PAGE_SHIFT, gfp_mask);
if (likely(page))
LASSERT(!test_bit(PG_private_2, &page->flags));
else
lprocfs_counter_add(d->od_stats, LPROC_OSD_NO_PAGE, 1);
- } else {
- LASSERT(oti->oti_dio_pages);
+ return page;
+ }
-
if (unlikely(!oti->oti_dio_pages[cur])
) {
- LASSERT(cur < PTLRPC_MAX_BRW_PAGES);
- page = alloc_page(gfp_mask);
- if (!page)
- return NULL;
-
oti->oti_dio_pages[cur] =
page;
-
}
+
if (inode->i_mapping->nrpages
) {
+ /* consult with pagecache, but do not create new pages */
+ /* this is normally used once */
+ page = find_lock_page(inode->i_mapping, offset >> PAGE_SHIFT);
+ if (page)
+
return
page;
+ }
- page = oti->oti_dio_pages[cur];
- LASSERT(!test_bit(PG_private_2, &page->flags));
- set_bit(PG_private_2, &page->flags);
- oti->oti_dio_pages_used++;
+ LASSERT(oti->oti_dio_pages);
+ cur = oti->oti_dio_pages_used;
- LASSERT(!PageLocked(page));
- lock_page(page);
+ if (unlikely(!oti->oti_dio_pages[cur])) {
+ LASSERT(cur < PTLRPC_MAX_BRW_PAGES);
+ page = alloc_page(gfp_mask);
+ if (!page)
+ return NULL;
+ oti->oti_dio_pages[cur] = page;
+ }
- LASSERT(!page->mapping);
- LASSERT(!PageWriteback(page));
- ClearPageUptodate(page);
+ page = oti->oti_dio_pages[cur];
+ LASSERT(!test_bit(PG_private_2, &page->flags));
+ set_bit(PG_private_2, &page->flags);
+ oti->oti_dio_pages_used++;
- page->index = offset >> PAGE_SHIFT;
- }
+ LASSERT(!PageLocked(page));
+ lock_page(page);
+
+ LASSERT(!page->mapping);
+ LASSERT(!PageWriteback(page));
+ ClearPageUptodate(page);
+
+ page->index = offset >> PAGE_SHIFT;
return page;
}
@@
-738,17
+755,18
@@
static int osd_bufs_put(const struct lu_env *env, struct dt_object *dt,
if (page == NULL)
continue;
- LASSERT(PageLocked(page));
/* if the page isn't cached, then reset uptodate
* to prevent reuse */
if (test_bit(PG_private_2, &page->flags)) {
clear_bit(PG_private_2, &page->flags);
ClearPageUptodate(page);
- unlock_page(page);
+ if (lnb[i].lnb_locked)
+ unlock_page(page);
oti->oti_dio_pages_used--;
} else {
- unlock_page(page);
+ if (lnb[i].lnb_locked)
+ unlock_page(page);
if (pagevec_add(&pvec, page) == 0)
pagevec_release(&pvec);
}
@@
-791,372
+809,100
@@
static int osd_bufs_put(const struct lu_env *env, struct dt_object *dt,
*/
static int osd_bufs_get(const struct lu_env *env, struct dt_object *dt,
loff_t pos, ssize_t len, struct niobuf_local *lnb,
- enum dt_bufs_type rw)
+
int maxlnb,
enum dt_bufs_type rw)
{
struct osd_thread_info *oti = osd_oti_get(env);
struct osd_object *obj = osd_dt_obj(dt);
- int npages, i, rc = 0;
+ struct osd_device *osd = osd_obj2dev(obj);
+ int npages, i, iosize, rc = 0;
+ bool cache, write;
+ loff_t fsize;
gfp_t gfp_mask;
LASSERT(obj->oo_inode);
- if (!osd_use_page_cache(osd_obj2dev(obj))) {
- if (unlikely(!oti->oti_dio_pages)) {
- OBD_ALLOC(oti->oti_dio_pages,
- sizeof(struct page *) * PTLRPC_MAX_BRW_PAGES);
- if (!oti->oti_dio_pages)
- return -ENOMEM;
+ rc = osd_map_remote_to_local(pos, len, &npages, lnb, maxlnb);
+ if (rc)
+ RETURN(rc);
+
+ write = rw & DT_BUFS_TYPE_WRITE;
+
+ fsize = lnb[npages - 1].lnb_file_offset + lnb[npages - 1].lnb_len;
+ iosize = fsize - lnb[0].lnb_file_offset;
+ fsize = max(fsize, i_size_read(obj->oo_inode));
+
+ cache = rw & DT_BUFS_TYPE_READAHEAD;
+ if (cache)
+ goto bypass_checks;
+
+ cache = osd_use_page_cache(osd);
+ while (cache) {
+ if (write) {
+ if (!osd->od_writethrough_cache) {
+ cache = false;
+ break;
+ }
+ if (iosize > osd->od_writethrough_max_iosize) {
+ cache = false;
+ break;
+ }
+ } else {
+ if (!osd->od_read_cache) {
+ cache = false;
+ break;
+ }
+ if (iosize > osd->od_readcache_max_iosize) {
+ cache = false;
+ break;
+ }
}
+ /* don't use cache on large files */
+ if (osd->od_readcache_max_filesize &&
+ fsize > osd->od_readcache_max_filesize)
+ cache = false;
+ break;
}
- osd_map_remote_to_local(pos, len, &npages, lnb);
+bypass_checks:
+ if (!cache && unlikely(!oti->oti_dio_pages)) {
+ OBD_ALLOC(oti->oti_dio_pages,
+ sizeof(struct page *) * PTLRPC_MAX_BRW_PAGES);
+ if (!oti->oti_dio_pages)
+ return -ENOMEM;
+ }
/* this could also try less hard for DT_BUFS_TYPE_READAHEAD pages */
gfp_mask = rw & DT_BUFS_TYPE_LOCAL ? (GFP_NOFS | __GFP_HIGHMEM) :
GFP_HIGHUSER;
for (i = 0; i < npages; i++, lnb++) {
lnb->lnb_page = osd_get_page(env, dt, lnb->lnb_file_offset,
- gfp_mask);
+ gfp_mask
, cache
);
if (lnb->lnb_page == NULL)
GOTO(cleanup, rc = -ENOMEM);
+ lnb->lnb_locked = 1;
wait_on_page_writeback(lnb->lnb_page);
BUG_ON(PageWriteback(lnb->lnb_page));
lu_object_get(&dt->do_lu);
}
- RETURN(i);
-
-cleanup:
- if (i > 0)
- osd_bufs_put(env, dt, lnb - i, i);
- return rc;
-}
-
-#ifndef HAVE_LDISKFS_MAP_BLOCKS
-
-#ifdef HAVE_EXT_PBLOCK /* Name changed to ext4_ext_pblock for kernel 2.6.35 */
-#define ldiskfs_ext_pblock(ex) ext_pblock((ex))
-#endif
-
-struct bpointers {
- sector_t *blocks;
- unsigned long start;
- int num;
- int init_num;
- int create;
-};
-
-static long ldiskfs_ext_find_goal(struct inode *inode,
- struct ldiskfs_ext_path *path,
- unsigned long block, int *aflags)
-{
- struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
- unsigned long bg_start;
- unsigned long colour;
- int depth;
-
- if (path) {
- struct ldiskfs_extent *ex;
- depth = path->p_depth;
-
- /* try to predict block placement */
- if ((ex = path[depth].p_ext))
- return ldiskfs_ext_pblock(ex) +
- (block - le32_to_cpu(ex->ee_block));
-
- /* it looks index is empty
- * try to find starting from index itself */
- if (path[depth].p_bh)
- return path[depth].p_bh->b_blocknr;
+#if 0
+ /* XXX: this version doesn't invalidate cached pages, but use them */
+ if (!cache && write && obj->oo_inode->i_mapping->nrpages) {
+ /* do not allow data aliasing, invalidate pagecache */
+ /* XXX: can be quite expensive in mixed case */
+ invalidate_mapping_pages(obj->oo_inode->i_mapping,
+ lnb[0].lnb_file_offset >> PAGE_SHIFT,
+ lnb[npages - 1].lnb_file_offset >> PAGE_SHIFT);
}
-
- /* OK. use inode's group */
- bg_start = (ei->i_block_group * LDISKFS_BLOCKS_PER_GROUP(inode->i_sb)) +
- le32_to_cpu(LDISKFS_SB(inode->i_sb)->s_es->s_first_data_block);
- colour = (current->pid % 16) *
- (LDISKFS_BLOCKS_PER_GROUP(inode->i_sb) / 16);
- return bg_start + colour + block;
-}
-
-static unsigned long new_blocks(handle_t *handle, struct inode *inode,
- struct ldiskfs_ext_path *path,
- unsigned long block, unsigned long *count,
- int *err)
-{
- struct ldiskfs_allocation_request ar;
- unsigned long pblock;
- int aflags;
-
- /* find neighbour allocated blocks */
- ar.lleft = block;
- *err = ldiskfs_ext_search_left(inode, path, &ar.lleft, &ar.pleft);
- if (*err)
- return 0;
- ar.lright = block;
- *err = ldiskfs_ext_search_right(inode, path, &ar.lright, &ar.pright);
- if (*err)
- return 0;
-
- /* allocate new block */
- ar.goal = ldiskfs_ext_find_goal(inode, path, block, &aflags);
- ar.inode = inode;
- ar.logical = block;
- ar.len = *count;
- ar.flags = LDISKFS_MB_HINT_DATA;
- pblock = ldiskfs_mb_new_blocks(handle, &ar, err);
- *count = ar.len;
- return pblock;
-}
-
-static int ldiskfs_ext_new_extent_cb(struct inode *inode,
- struct ldiskfs_ext_path *path,
- struct ldiskfs_ext_cache *cex,
-#ifdef HAVE_EXT_PREPARE_CB_EXTENT
- struct ldiskfs_extent *ex,
-#endif
- void *cbdata)
-{
- struct bpointers *bp = cbdata;
- struct ldiskfs_extent nex;
- unsigned long pblock = 0;
- unsigned long tgen;
- int err, i;
- unsigned long count;
- handle_t *handle;
-
-#ifdef LDISKFS_EXT_CACHE_EXTENT /* until kernel 2.6.37 */
- if (cex->ec_type == LDISKFS_EXT_CACHE_EXTENT) {
-#else
- if ((cex->ec_len != 0) && (cex->ec_start != 0)) {
-#endif
- err = EXT_CONTINUE;
- goto map;
- }
-
- if (bp->create == 0) {
- i = 0;
- if (cex->ec_block < bp->start)
- i = bp->start - cex->ec_block;
- if (i >= cex->ec_len)
- CERROR("nothing to do?! i = %d, e_num = %u\n",
- i, cex->ec_len);
- for (; i < cex->ec_len && bp->num; i++) {
- *(bp->blocks) = 0;
- bp->blocks++;
- bp->num--;
- bp->start++;
- }
-
- return EXT_CONTINUE;
- }
-
- tgen = LDISKFS_I(inode)->i_ext_generation;
- count = ldiskfs_ext_calc_credits_for_insert(inode, path);
-
- handle = osd_journal_start(inode, LDISKFS_HT_MISC,
- count + LDISKFS_ALLOC_NEEDED + 1);
- if (IS_ERR(handle)) {
- return PTR_ERR(handle);
- }
-
- if (tgen != LDISKFS_I(inode)->i_ext_generation) {
- /* the tree has changed. so path can be invalid at moment */
- ldiskfs_journal_stop(handle);
- return EXT_REPEAT;
- }
-
- /* In 2.6.32 kernel, ldiskfs_ext_walk_space()'s callback func is not
- * protected by i_data_sem as whole. so we patch it to store
- * generation to path and now verify the tree hasn't changed */
- down_write((&LDISKFS_I(inode)->i_data_sem));
-
- /* validate extent, make sure the extent tree does not changed */
- if (LDISKFS_I(inode)->i_ext_generation != path[0].p_generation) {
- /* cex is invalid, try again */
- up_write(&LDISKFS_I(inode)->i_data_sem);
- ldiskfs_journal_stop(handle);
- return EXT_REPEAT;
- }
-
- count = cex->ec_len;
- pblock = new_blocks(handle, inode, path, cex->ec_block, &count, &err);
- if (!pblock)
- goto out;
- BUG_ON(count > cex->ec_len);
-
- /* insert new extent */
- nex.ee_block = cpu_to_le32(cex->ec_block);
- ldiskfs_ext_store_pblock(&nex, pblock);
- nex.ee_len = cpu_to_le16(count);
- err = ldiskfs_ext_insert_extent(handle, inode, path, &nex, 0);
- if (err) {
- /* free data blocks we just allocated */
- /* not a good idea to call discard here directly,
- * but otherwise we'd need to call it every free() */
- ldiskfs_discard_preallocations(inode);
-#ifdef HAVE_EXT_FREE_BLOCK_WITH_BUFFER_HEAD /* Introduced in 2.6.32-rc7 */
- ldiskfs_free_blocks(handle, inode, NULL,
- ldiskfs_ext_pblock(&nex),
- le16_to_cpu(nex.ee_len), 0);
-#else
- ldiskfs_free_blocks(handle, inode, ldiskfs_ext_pblock(&nex),
- le16_to_cpu(nex.ee_len), 0);
-#endif
- goto out;
- }
-
- /*
- * Putting len of the actual extent we just inserted,
- * we are asking ldiskfs_ext_walk_space() to continue
- * scaning after that block
- */
- cex->ec_len = le16_to_cpu(nex.ee_len);
- cex->ec_start = ldiskfs_ext_pblock(&nex);
- BUG_ON(le16_to_cpu(nex.ee_len) == 0);
- BUG_ON(le32_to_cpu(nex.ee_block) != cex->ec_block);
-
-out:
- up_write((&LDISKFS_I(inode)->i_data_sem));
- ldiskfs_journal_stop(handle);
-map:
- if (err >= 0) {
- /* map blocks */
- if (bp->num == 0) {
- CERROR("hmm. why do we find this extent?\n");
- CERROR("initial space: %lu:%u\n",
- bp->start, bp->init_num);
-#ifdef LDISKFS_EXT_CACHE_EXTENT /* until kernel 2.6.37 */
- CERROR("current extent: %u/%u/%llu %d\n",
- cex->ec_block, cex->ec_len,
- (unsigned long long)cex->ec_start,
- cex->ec_type);
-#else
- CERROR("current extent: %u/%u/%llu\n",
- cex->ec_block, cex->ec_len,
- (unsigned long long)cex->ec_start);
#endif
- }
- i = 0;
- if (cex->ec_block < bp->start)
- i = bp->start - cex->ec_block;
- if (i >= cex->ec_len)
- CERROR("nothing to do?! i = %d, e_num = %u\n",
- i, cex->ec_len);
- for (; i < cex->ec_len && bp->num; i++) {
- *(bp->blocks) = cex->ec_start + i;
- /* unmap any underlying metadata from
- * the block device mapping. b=6998.
- */
- if (pblock != 0)
- clean_bdev_aliases(inode->i_sb->s_bdev,
- *(bp->blocks), 1);
- bp->blocks++;
- bp->num--;
- bp->start++;
- }
- }
- return err;
-}
-
-static int osd_ldiskfs_map_nblocks(struct inode *inode, unsigned long index,
- int clen, sector_t *blocks, int create)
-{
- int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
- struct bpointers bp;
- int err;
-
- if (index + clen >= inode->i_sb->s_maxbytes >> PAGE_SHIFT)
- return -EFBIG;
-
- bp.blocks = blocks;
- bp.start = index * blocks_per_page;
- bp.init_num = bp.num = clen * blocks_per_page;
- bp.create = create;
-
- CDEBUG(D_OTHER, "blocks %lu-%lu requested for inode %u\n",
- bp.start, bp.start + bp.num - 1, (unsigned)inode->i_ino);
-
- err = ldiskfs_ext_walk_space(inode, bp.start, bp.num,
- ldiskfs_ext_new_extent_cb, &bp);
- ldiskfs_ext_invalidate_cache(inode);
-
- return err;
-}
-static int osd_ldiskfs_map_bm_inode_pages(struct inode *inode,
- struct page **page, int pages,
- sector_t *blocks, int create)
-{
- int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
- pgoff_t bitmap_max_page_index;
- sector_t *b;
- int rc = 0, i;
-
- bitmap_max_page_index = LDISKFS_SB(inode->i_sb)->s_bitmap_maxbytes >>
- PAGE_SHIFT;
- for (i = 0, b = blocks; i < pages; i++, page++) {
- if ((*page)->index + 1 >= bitmap_max_page_index) {
- rc = -EFBIG;
- break;
- }
- rc = ldiskfs_map_inode_page(inode, *page, b, create);
- if (rc) {
- CERROR("ino %lu, blk %llu create %d: rc %d\n",
- inode->i_ino,
- (unsigned long long)*b, create, rc);
- break;
- }
- b += blocks_per_page;
- }
- return rc;
-}
-
-static int osd_ldiskfs_map_ext_inode_pages(struct inode *inode,
- struct page **page,
- int pages, sector_t *blocks,
- int create)
-{
- int rc = 0, i = 0, clen = 0;
- struct page *fp = NULL;
-
- CDEBUG(D_OTHER, "inode %lu: map %d pages from %lu\n",
- inode->i_ino, pages, (*page)->index);
-
- /* pages are sorted already. so, we just have to find
- * contig. space and process them properly */
- while (i < pages) {
- if (fp == NULL) {
- /* start new extent */
- fp = *page++;
- clen = 1;
- i++;
- continue;
- } else if (fp->index + clen == (*page)->index) {
- /* continue the extent */
- page++;
- clen++;
- i++;
- continue;
- }
-
- /* process found extent */
- rc = osd_ldiskfs_map_nblocks(inode, fp->index, clen,
- blocks, create);
- if (rc)
- GOTO(cleanup, rc);
-
- /* look for next extent */
- fp = NULL;
- blocks += clen * (PAGE_SIZE >> inode->i_blkbits);
- }
-
- if (fp)
- rc = osd_ldiskfs_map_nblocks(inode, fp->index, clen,
- blocks, create);
+ RETURN(i);
cleanup:
+ if (i > 0)
+ osd_bufs_put(env, dt, lnb - i, i);
return rc;
}
@@
-1164,22
+910,6
@@
static int osd_ldiskfs_map_inode_pages(struct inode *inode, struct page **page,
int pages, sector_t *blocks,
int create)
{
- int rc;
-
- if (LDISKFS_I(inode)->i_flags & LDISKFS_EXTENTS_FL) {
- rc = osd_ldiskfs_map_ext_inode_pages(inode, page, pages,
- blocks, create);
- return rc;
- }
- rc = osd_ldiskfs_map_bm_inode_pages(inode, page, pages, blocks, create);
-
- return rc;
-}
-#else
-static int osd_ldiskfs_map_inode_pages(struct inode *inode, struct page **page,
- int pages, sector_t *blocks,
- int create)
-{
int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
int rc = 0, i = 0;
struct page *fp = NULL;
@@
-1261,7
+991,6
@@
cont_map:
cleanup:
return rc;
}
-#endif /* HAVE_LDISKFS_MAP_BLOCKS */
static int osd_write_prep(const struct lu_env *env, struct dt_object *dt,
struct niobuf_local *lnb, int npages)
@@
-1270,14
+999,11
@@
static int osd_write_prep(const struct lu_env *env, struct dt_object *dt,
struct osd_iobuf *iobuf = &oti->oti_iobuf;
struct inode *inode = osd_dt_obj(dt)->oo_inode;
struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
- ktime_t start;
- ktime_t end;
+ ktime_t start, end;
s64 timediff;
- ssize_t isize;
- __s64 maxidx;
- int rc = 0;
- int i;
- int cache = 0;
+ ssize_t isize;
+ __s64 maxidx;
+ int i, rc = 0;
LASSERT(inode);
@@
-1288,18
+1014,9
@@
static int osd_write_prep(const struct lu_env *env, struct dt_object *dt,
isize = i_size_read(inode);
maxidx = ((isize + PAGE_SIZE - 1) >> PAGE_SHIFT) - 1;
- if (osd->od_writethrough_cache)
- cache = 1;
- if (isize > osd->od_readcache_max_filesize)
- cache = 0;
-
start = ktime_get();
for (i = 0; i < npages; i++) {
- if (cache == 0)
- generic_error_remove_page(inode->i_mapping,
- lnb[i].lnb_page);
-
/*
* till commit the content of the page is undefined
* we'll set it uptodate once bulk is done. otherwise
@@
-1374,7
+1091,7
@@
static int osd_is_mapped(struct dt_object *dt, __u64 offset,
fei.fi_extents_start = &fe;
saved_fs = get_fs();
- set_fs(
get_ds()
);
+ set_fs(
KERNEL_DS
);
rc = inode->i_op->fiemap(inode, &fei, offset, FIEMAP_MAX_OFFSET-offset);
set_fs(saved_fs);
if (rc != 0)
@@
-1409,10
+1126,10
@@
static int osd_declare_write_commit(const struct lu_env *env,
int i;
int newblocks;
int rc = 0;
- int flags = 0;
int credits = 0;
long long quota_space = 0;
struct osd_fextent extent = { 0 };
+ enum osd_quota_local_flags local_flags = 0;
enum osd_qid_declare_flags declare_flags = OSD_QID_BLK;
ENTRY;
@@
-1497,16
+1214,16
@@
static int osd_declare_write_commit(const struct lu_env *env,
rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
i_projid_read(inode), quota_space, oh,
- osd_dt_obj(dt), &flags, declare_flags);
+ osd_dt_obj(dt), &
local_
flags, declare_flags);
/* we need only to store the overquota flags in the first lnb for
* now, once we support multiple objects BRW, this code needs be
* revised. */
- if (flags & QUOTA_FL_OVER_USRQUOTA)
+ if (
local_
flags & QUOTA_FL_OVER_USRQUOTA)
lnb[0].lnb_flags |= OBD_BRW_OVER_USRQUOTA;
- if (flags & QUOTA_FL_OVER_GRPQUOTA)
+ if (
local_
flags & QUOTA_FL_OVER_GRPQUOTA)
lnb[0].lnb_flags |= OBD_BRW_OVER_GRPQUOTA;
- if (flags & QUOTA_FL_OVER_PRJQUOTA)
+ if (
local_
flags & QUOTA_FL_OVER_PRJQUOTA)
lnb[0].lnb_flags |= OBD_BRW_OVER_PRJQUOTA;
if (rc == 0)
@@
-1534,7
+1251,7
@@
static int osd_write_commit(const struct lu_env *env, struct dt_object *dt,
RETURN(rc);
isize = i_size_read(inode);
-
ll_vfs_dq_init
(inode);
+
dquot_initialize
(inode);
for (i = 0; i < npages; i++) {
if (lnb[i].lnb_rc == -ENOSPC &&
@@
-1625,7
+1342,7
@@
static int osd_read_prep(const struct lu_env *env, struct dt_object *dt,
struct osd_iobuf *iobuf = &oti->oti_iobuf;
struct inode *inode = osd_dt_obj(dt)->oo_inode;
struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
- int rc = 0, i, cache
= 0, cache
_hits = 0, cache_misses = 0;
+ int rc = 0, i, cache_hits = 0, cache_misses = 0;
ktime_t start, end;
s64 timediff;
loff_t isize;
@@
-1638,11
+1355,6
@@
static int osd_read_prep(const struct lu_env *env, struct dt_object *dt,
isize = i_size_read(inode);
- if (osd->od_read_cache)
- cache = 1;
- if (isize > osd->od_readcache_max_filesize)
- cache = 0;
-
start = ktime_get();
for (i = 0; i < npages; i++) {
@@
-1651,10
+1363,10
@@
static int osd_read_prep(const struct lu_env *env, struct dt_object *dt,
* lnb->lnb_rc == 0, so it's easy to detect later. */
break;
- if (isize < lnb[i].lnb_file_offset + lnb[i].lnb_len)
- lnb[i].lnb_rc = isize - lnb[i].lnb_file_offset;
- else
-
lnb[i].lnb_rc = lnb[i].lnb_len;
+ /* instead of looking if we go beyong isize, send complete
+ * pages all the time
+ */
+ lnb[i].lnb_rc = lnb[i].lnb_len;
/* Bypass disk read if fail_loc is set properly */
if (OBD_FAIL_CHECK(OBD_FAIL_OST_FAKE_RW))
@@
-1662,14
+1374,17
@@
static int osd_read_prep(const struct lu_env *env, struct dt_object *dt,
if (PageUptodate(lnb[i].lnb_page)) {
cache_hits++;
+ unlock_page(lnb[i].lnb_page);
} else {
cache_misses++;
osd_iobuf_add_page(iobuf, &lnb[i]);
}
-
- if (cache == 0)
- generic_error_remove_page(inode->i_mapping,
- lnb[i].lnb_page);
+ /* no need to unlock in osd_bufs_put(), the sooner page is
+ * unlocked, the earlier another client can access it.
+ * notice real unlock_page() can be called few lines
+ * below after osd_do_bio(). lnb is a per-thread, so it's
+ * fine to have PG_locked and lnb_locked inconsistent here */
+ lnb[i].lnb_locked = 0;
}
end = ktime_get();
timediff = ktime_us_delta(end, start);
@@
-1685,16
+1400,22
@@
static int osd_read_prep(const struct lu_env *env, struct dt_object *dt,
lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_ACCESS,
cache_hits + cache_misses);
-
if (iobuf->dr_npages) {
+ if (iobuf->dr_npages) {
rc = osd_ldiskfs_map_inode_pages(inode, iobuf->dr_pages,
iobuf->dr_npages,
iobuf->dr_blocks, 0);
-
rc = osd_do_bio(osd, inode, iobuf);
+ rc = osd_do_bio(osd, inode, iobuf);
- /* IO stats will be done in osd_bufs_put() */
- }
+ /* IO stats will be done in osd_bufs_put() */
- RETURN(rc);
+ /* early release to let others read data during the bulk */
+ for (i = 0; i < iobuf->dr_npages; i++) {
+ LASSERT(PageLocked(iobuf->dr_pages[i]));
+ unlock_page(iobuf->dr_pages[i]);
+ }
+ }
+
+ RETURN(rc);
}
/*
@@
-1732,7
+1453,8
@@
int osd_ldiskfs_read(struct inode *inode, void *buf, int size, loff_t *offs)
loff_t diff = i_size_read(inode) - *offs;
spin_unlock(&inode->i_lock);
if (diff < 0) {
- CDEBUG(D_EXT2, "size %llu is too short to read @%llu\n",
+ CDEBUG(D_OTHER,
+ "size %llu is too short to read @%llu\n",
i_size_read(inode), *offs);
return -EBADR;
} else if (diff == 0) {
@@
-1959,9
+1681,11
@@
static int osd_ldiskfs_writelink(struct inode *inode, char *buffer, int buflen)
return 0;
}
-int osd_ldiskfs_write_record(struct inode *inode, void *buf, int bufsize,
- int write_NUL, loff_t *offs, handle_t *handle)
+static int osd_ldiskfs_write_record(struct dt_object *dt, void *buf,
+ int bufsize, int write_NUL, loff_t *offs,
+ handle_t *handle)
{
+ struct inode *inode = osd_dt_obj(dt)->oo_inode;
struct buffer_head *bh = NULL;
loff_t offset = *offs;
loff_t new_size = i_size_read(inode);
@@
-1971,6
+1695,8
@@
int osd_ldiskfs_write_record(struct inode *inode, void *buf, int bufsize,
int size;
int boffs;
int dirty_inode = 0;
+ struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
+ bool create, sparse, sync = false;
if (write_NUL) {
/*
@@
-1982,8
+1708,14
@@
int osd_ldiskfs_write_record(struct inode *inode, void *buf, int bufsize,
++bufsize;
}
+ /* sparse checking is racy, but sparse is very rare case, leave as is */
+ sparse = (new_size > 0 && (inode->i_blocks >> (inode->i_blkbits - 9)) <
+ ((new_size - 1) >> inode->i_blkbits) + 1);
+
while (bufsize > 0) {
int credits = handle->h_buffer_credits;
+ unsigned long last_block = (new_size == 0) ? 0 :
+ (new_size - 1) >> inode->i_blkbits;
if (bh)
brelse(bh);
@@
-1991,7
+1723,39
@@
int osd_ldiskfs_write_record(struct inode *inode, void *buf, int bufsize,
block = offset >> inode->i_blkbits;
boffs = offset & (blocksize - 1);
size = min(blocksize - boffs, bufsize);
- bh = __ldiskfs_bread(handle, inode, block, 1);
+ sync = (block > last_block || new_size == 0 || sparse);
+
+ if (sync)
+ down(&ei->i_append_sem);
+
+ bh = __ldiskfs_bread(handle, inode, block, 0);
+
+ if (unlikely(IS_ERR_OR_NULL(bh) && !sync))
+ CWARN("%s: adding bh without locking off %llu (block %lu, "
+ "size %d, offs %llu)\n", inode->i_sb->s_id,
+ offset, block, bufsize, *offs);
+
+ if (IS_ERR_OR_NULL(bh)) {
+ struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
+ int flags = LDISKFS_GET_BLOCKS_CREATE;
+
+ /* while the file system is being mounted, avoid
+ * preallocation otherwise mount can take a long
+ * time as mballoc cache is cold.
+ * XXX: this is a workaround until we have a proper
+ * fix in mballoc
+ * XXX: works with extent-based files only */
+ if (!osd->od_cl_seq)
+ flags |= LDISKFS_GET_BLOCKS_NO_NORMALIZE;
+ bh = __ldiskfs_bread(handle, inode, block, flags);
+ create = true;
+ } else {
+ if (sync) {
+ up(&ei->i_append_sem);
+ sync = false;
+ }
+ create = false;
+ }
if (IS_ERR_OR_NULL(bh)) {
if (bh == NULL) {
err = -EIO;
@@
-2016,7
+1780,14
@@
int osd_ldiskfs_write_record(struct inode *inode, void *buf, int bufsize,
LASSERTF(boffs + size <= bh->b_size,
"boffs %d size %d bh->b_size %lu\n",
boffs, size, (unsigned long)bh->b_size);
- memcpy(bh->b_data + boffs, buf, size);
+ if (create) {
+ memset(bh->b_data, 0, bh->b_size);
+ if (sync) {
+ up(&ei->i_append_sem);
+ sync = false;
+ }
+ }
+ memcpy(bh->b_data + boffs, buf, size);
err = ldiskfs_handle_dirty_metadata(handle, NULL, bh);
if (err)
break;
@@
-2027,8
+1798,11
@@
int osd_ldiskfs_write_record(struct inode *inode, void *buf, int bufsize,
bufsize -= size;
buf += size;
}
- if (bh)
- brelse(bh);
+ if (sync)
+ up(&ei->i_append_sem);
+
+ if (bh)
+ brelse(bh);
if (write_NUL)
--new_size;
@@
-2064,7
+1838,7
@@
static ssize_t osd_write(const struct lu_env *env, struct dt_object *dt,
LASSERT(handle != NULL);
LASSERT(inode != NULL);
-
ll_vfs_dq_init
(inode);
+
dquot_initialize
(inode);
/* XXX: don't check: one declared chunk can be used many times */
/* osd_trans_exec_op(env, handle, OSD_OT_WRITE); */
@@
-2082,9
+1856,8
@@
static ssize_t osd_write(const struct lu_env *env, struct dt_object *dt,
if (is_link && (buf->lb_len < sizeof(LDISKFS_I(inode)->i_data)))
result = osd_ldiskfs_writelink(inode, buf->lb_buf, buf->lb_len);
else
- result = osd_ldiskfs_write_record(inode, buf->lb_buf,
- buf->lb_len, is_link, pos,
- oh->ot_handle);
+ result = osd_ldiskfs_write_record(dt, buf->lb_buf, buf->lb_len,
+ is_link, pos, oh->ot_handle);
if (result == 0)
result = buf->lb_len;
@@
-2144,7
+1917,7
@@
static int osd_punch(const struct lu_env *env, struct dt_object *dt,
LASSERT(dt_object_exists(dt));
LASSERT(osd_invariant(obj));
LASSERT(inode != NULL);
-
ll_vfs_dq_init
(inode);
+
dquot_initialize
(inode);
LASSERT(th);
oh = container_of(th, struct osd_thandle, ot_super);
@@
-2234,7
+2007,7
@@
static int osd_fiemap_get(const struct lu_env *env, struct dt_object *dt,
struct inode *inode = osd_dt_obj(dt)->oo_inode;
u64 len;
int rc;
-
+ mm_segment_t cur_fs;
LASSERT(inode);
if (inode->i_op->fiemap == NULL)
@@
-2254,27
+2027,34
@@
static int osd_fiemap_get(const struct lu_env *env, struct dt_object *dt,
if (fieinfo.fi_flags & FIEMAP_FLAG_SYNC)
filemap_write_and_wait(inode->i_mapping);
+ /* Save previous value address limit */
+ cur_fs = get_fs();
+ /* Set the address limit of the kernel */
+ set_fs(KERNEL_DS);
+
rc = inode->i_op->fiemap(inode, &fieinfo, fm->fm_start, len);
fm->fm_flags = fieinfo.fi_flags;
fm->fm_mapped_extents = fieinfo.fi_extents_mapped;
+ /* Restore the previous address limt */
+ set_fs(cur_fs);
+
return rc;
}
static int osd_ladvise(const struct lu_env *env, struct dt_object *dt,
__u64 start, __u64 end, enum lu_ladvise_type advice)
{
-
int rc = 0
;
-
struct inode *inode = osd_dt_obj(dt)->oo_inode
;
+
struct osd_object *obj = osd_dt_obj(dt)
;
+
int rc = 0
;
ENTRY;
switch (advice) {
case LU_LADVISE_DONTNEED:
- if (end == 0)
- break;
- invalidate_mapping_pages(inode->i_mapping,
- start >> PAGE_SHIFT,
- (end - 1) >> PAGE_SHIFT);
+ if (end)
+ invalidate_mapping_pages(obj->oo_inode->i_mapping,
+ start >> PAGE_SHIFT,
+ (end - 1) >> PAGE_SHIFT);
break;
default:
rc = -ENOTSUPP;
@@
-2386,7
+2166,9
@@
void osd_execute_truncate(struct osd_object *obj)
return;
}
+ inode_lock(inode);
ldiskfs_truncate(inode);
+ inode_unlock(inode);
/*
* For a partial-page truncate, flush the page to disk immediately to