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 <http://jira.whamcloud.com/> , along "
+ "report this to <https://jira.hpdd.intel.com/> , 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 "
if (iobuf->dr_rw == 0) {
wait_event(iobuf->dr_wait,
atomic_read(&iobuf->dr_numreqs) == 0);
+ osd_fini_iobuf(osd, iobuf);
}
if (rc == 0)
plen = len;
lnb->lnb_file_offset = offset;
lnb->lnb_page_offset = poff;
- lnb->len = plen;
- /* lb->flags = rnb->flags; */
- lnb->flags = 0;
- lnb->page = NULL;
- lnb->rc = 0;
+ lnb->lnb_len = plen;
+ /* lnb->lnb_flags = rnb->flags; */
+ lnb->lnb_flags = 0;
+ lnb->lnb_page = NULL;
+ lnb->lnb_rc = 0;
LASSERTF(plen <= len, "plen %u, len %lld\n", plen,
(long long) len);
osd_map_remote_to_local(pos, len, &npages, lnb);
for (i = 0; i < npages; i++, lnb++) {
-
- /* We still set up for ungranted pages so that granted pages
- * can be written to disk as they were promised, and portals
- * needs to keep the pages all aligned properly. */
- lnb->dentry = (void *) obj;
-
- lnb->page = osd_get_page(d, lnb->lnb_file_offset, rw);
- if (lnb->page == NULL)
- GOTO(cleanup, rc = -ENOMEM);
-
- /* DLM locking protects us from write and truncate competing
- * for same region, but truncate can leave dirty page in the
- * cache. it's possible the writeout on a such a page is in
- * progress when we access it. it's also possible that during
- * this writeout we put new (partial) data, but then won't
- * be able to proceed in filter_commitrw_write(). thus let's
- * just wait for writeout completion, should be rare enough.
- * -bzzz */
- wait_on_page_writeback(lnb->page);
- BUG_ON(PageWriteback(lnb->page));
+ lnb->lnb_page = osd_get_page(d, lnb->lnb_file_offset, rw);
+ if (lnb->lnb_page == NULL)
+ GOTO(cleanup, rc = -ENOMEM);
+
+ /* DLM locking protects us from write and truncate competing
+ * for same region, but truncate can leave dirty page in the
+ * cache. it's possible the writeout on a such a page is in
+ * progress when we access it. it's also possible that during
+ * this writeout we put new (partial) data, but then won't
+ * be able to proceed in filter_commitrw_write(). thus let's
+ * just wait for writeout completion, should be rare enough.
+ * -bzzz */
+ wait_on_page_writeback(lnb->lnb_page);
+ BUG_ON(PageWriteback(lnb->lnb_page));
lu_object_get(&d->do_lu);
}
}
static int osd_bufs_put(const struct lu_env *env, struct dt_object *dt,
- struct niobuf_local *lnb, int npages)
+ struct niobuf_local *lnb, int npages)
{
- struct osd_thread_info *oti = osd_oti_get(env);
- struct osd_iobuf *iobuf = &oti->oti_iobuf;
- struct osd_device *d = osd_obj2dev(osd_dt_obj(dt));
- int i;
+ int i;
- /* to do IO stats, notice we do this here because
- * osd_do_bio() doesn't wait for write to complete */
- osd_fini_iobuf(d, iobuf);
+ for (i = 0; i < npages; i++) {
+ if (lnb[i].lnb_page == NULL)
+ continue;
+ LASSERT(PageLocked(lnb[i].lnb_page));
+ unlock_page(lnb[i].lnb_page);
+ page_cache_release(lnb[i].lnb_page);
+ lu_object_put(env, &dt->do_lu);
+ lnb[i].lnb_page = NULL;
+ }
- for (i = 0; i < npages; i++) {
- if (lnb[i].page == NULL)
- continue;
- LASSERT(PageLocked(lnb[i].page));
- unlock_page(lnb[i].page);
- page_cache_release(lnb[i].page);
- lu_object_put(env, &dt->do_lu);
- lnb[i].page = NULL;
- }
- RETURN(0);
+ RETURN(0);
}
#ifdef HAVE_EXT_PBLOCK /* Name changed to ext4_ext_pblock for kernel 2.6.35 */
* 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),
- cpu_to_le16(nex.ee_len), 0);
+ 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),
- cpu_to_le16(nex.ee_len), 0);
+ le16_to_cpu(nex.ee_len), 0);
#endif
goto out;
}
static int osd_ldiskfs_map_inode_pages(struct inode *inode, struct page **page,
int pages, unsigned long *blocks,
- int create, struct mutex *optional_mutex)
+ int create)
{
int rc;
- if (LDISKFS_I(inode)->i_flags & LDISKFS_EXTENTS_FL) {
+ if (ldiskfs_test_inode_flag(inode, LDISKFS_INODE_EXTENTS)) {
rc = osd_ldiskfs_map_ext_inode_pages(inode, page, pages,
blocks, create);
return rc;
}
- if (optional_mutex != NULL)
- mutex_lock(optional_mutex);
rc = osd_ldiskfs_map_bm_inode_pages(inode, page, pages, blocks, create);
- if (optional_mutex != NULL)
- mutex_unlock(optional_mutex);
return rc;
}
do_gettimeofday(&start);
for (i = 0; i < npages; i++) {
- if (cache == 0)
- generic_error_remove_page(inode->i_mapping,
- lnb[i].page);
+ 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
- * subsequent reads can access non-stable data
- */
- ClearPageUptodate(lnb[i].page);
+ /*
+ * till commit the content of the page is undefined
+ * we'll set it uptodate once bulk is done. otherwise
+ * subsequent reads can access non-stable data
+ */
+ ClearPageUptodate(lnb[i].lnb_page);
- if (lnb[i].len == PAGE_CACHE_SIZE)
- continue;
+ if (lnb[i].lnb_len == PAGE_CACHE_SIZE)
+ continue;
- if (maxidx >= lnb[i].page->index) {
- osd_iobuf_add_page(iobuf, lnb[i].page);
- } else {
- long off;
- char *p = kmap(lnb[i].page);
+ if (maxidx >= lnb[i].lnb_page->index) {
+ osd_iobuf_add_page(iobuf, lnb[i].lnb_page);
+ } else {
+ long off;
+ char *p = kmap(lnb[i].lnb_page);
off = lnb[i].lnb_page_offset;
if (off)
memset(p, 0, off);
- off = (lnb[i].lnb_page_offset + lnb[i].len) &
+ off = (lnb[i].lnb_page_offset + lnb[i].lnb_len) &
~CFS_PAGE_MASK;
- if (off)
+ if (off)
memset(p + off, 0, PAGE_CACHE_SIZE - off);
- kunmap(lnb[i].page);
+ kunmap(lnb[i].lnb_page);
}
}
do_gettimeofday(&end);
if (iobuf->dr_npages) {
rc = osd_ldiskfs_map_inode_pages(inode, iobuf->dr_pages,
iobuf->dr_npages,
- iobuf->dr_blocks,
- 0, NULL);
+ iobuf->dr_blocks, 0);
if (likely(rc == 0)) {
rc = osd_do_bio(osd, inode, iobuf);
/* do IO stats for preparation reads */
/* calculate number of extents (probably better to pass nb) */
for (i = 0; i < npages; i++) {
if (i && lnb[i].lnb_file_offset !=
- lnb[i - 1].lnb_file_offset + lnb[i - 1].len)
+ lnb[i - 1].lnb_file_offset + lnb[i - 1].lnb_len)
extents++;
if (!osd_is_mapped(inode, lnb[i].lnb_file_offset))
*
* XXX we could handle this on per-lnb basis as done by
* grant. */
- if ((lnb[i].flags & OBD_BRW_NOQUOTA) ||
- (lnb[i].flags & (OBD_BRW_FROM_GRANT | OBD_BRW_SYNC)) ==
+ if ((lnb[i].lnb_flags & OBD_BRW_NOQUOTA) ||
+ (lnb[i].lnb_flags & (OBD_BRW_FROM_GRANT | OBD_BRW_SYNC)) ==
OBD_BRW_FROM_GRANT)
ignore_quota = true;
}
* 5 is max tree depth: inode + 4 index blocks
* with blockmaps, depth is 3 at most
*/
- if (LDISKFS_I(inode)->i_flags & LDISKFS_EXTENTS_FL) {
+ if (ldiskfs_test_inode_flag(inode, LDISKFS_INODE_EXTENTS)) {
/*
* many concurrent threads may grow tree by the time
* our transaction starts. so, consider 2 is a min depth
oh->ot_credits += newblocks;
/* make sure the over quota flags were not set */
- lnb[0].flags &= ~(OBD_BRW_OVER_USRQUOTA | OBD_BRW_OVER_GRPQUOTA);
+ lnb[0].lnb_flags &= ~(OBD_BRW_OVER_USRQUOTA | OBD_BRW_OVER_GRPQUOTA);
- rc = osd_declare_inode_qid(env, inode->i_uid, inode->i_gid,
+ rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
quota_space, oh, true, true, &flags,
ignore_quota);
* now, once we support multiple objects BRW, this code needs be
* revised. */
if (flags & QUOTA_FL_OVER_USRQUOTA)
- lnb[0].flags |= OBD_BRW_OVER_USRQUOTA;
+ lnb[0].lnb_flags |= OBD_BRW_OVER_USRQUOTA;
if (flags & QUOTA_FL_OVER_GRPQUOTA)
- lnb[0].flags |= OBD_BRW_OVER_GRPQUOTA;
+ lnb[0].lnb_flags |= OBD_BRW_OVER_GRPQUOTA;
RETURN(rc);
}
ll_vfs_dq_init(inode);
for (i = 0; i < npages; i++) {
- if (lnb[i].rc == -ENOSPC &&
+ if (lnb[i].lnb_rc == -ENOSPC &&
osd_is_mapped(inode, lnb[i].lnb_file_offset)) {
- /* Allow the write to proceed if overwriting an
- * existing block */
- lnb[i].rc = 0;
- }
+ /* Allow the write to proceed if overwriting an
+ * existing block */
+ lnb[i].lnb_rc = 0;
+ }
- if (lnb[i].rc) { /* ENOSPC, network RPC error, etc. */
- CDEBUG(D_INODE, "Skipping [%d] == %d\n", i,
- lnb[i].rc);
- LASSERT(lnb[i].page);
- generic_error_remove_page(inode->i_mapping,lnb[i].page);
- continue;
- }
+ if (lnb[i].lnb_rc) { /* ENOSPC, network RPC error, etc. */
+ CDEBUG(D_INODE, "Skipping [%d] == %d\n", i,
+ lnb[i].lnb_rc);
+ LASSERT(lnb[i].lnb_page);
+ generic_error_remove_page(inode->i_mapping,
+ lnb[i].lnb_page);
+ continue;
+ }
- LASSERT(PageLocked(lnb[i].page));
- LASSERT(!PageWriteback(lnb[i].page));
+ LASSERT(PageLocked(lnb[i].lnb_page));
+ LASSERT(!PageWriteback(lnb[i].lnb_page));
- if (lnb[i].lnb_file_offset + lnb[i].len > isize)
- isize = lnb[i].lnb_file_offset + lnb[i].len;
+ if (lnb[i].lnb_file_offset + lnb[i].lnb_len > isize)
+ isize = lnb[i].lnb_file_offset + lnb[i].lnb_len;
- /*
- * Since write and truncate are serialized by oo_sem, even
- * partial-page truncate should not leave dirty pages in the
- * page cache.
- */
- LASSERT(!PageDirty(lnb[i].page));
+ /*
+ * Since write and truncate are serialized by oo_sem, even
+ * partial-page truncate should not leave dirty pages in the
+ * page cache.
+ */
+ LASSERT(!PageDirty(lnb[i].lnb_page));
- SetPageUptodate(lnb[i].page);
+ SetPageUptodate(lnb[i].lnb_page);
- osd_iobuf_add_page(iobuf, lnb[i].page);
+ osd_iobuf_add_page(iobuf, lnb[i].lnb_page);
}
if (OBD_FAIL_CHECK(OBD_FAIL_OST_MAPBLK_ENOSPC)) {
} else if (iobuf->dr_npages > 0) {
rc = osd_ldiskfs_map_inode_pages(inode, iobuf->dr_pages,
iobuf->dr_npages,
- iobuf->dr_blocks,
- 1, NULL);
+ iobuf->dr_blocks, 1);
} else {
/* no pages to write, no transno is needed */
thandle->th_local = 1;
rc = osd_do_bio(osd, inode, iobuf);
/* we don't do stats here as in read path because
* write is async: we'll do this in osd_put_bufs() */
- }
+ } else {
+ osd_fini_iobuf(osd, iobuf);
+ }
if (unlikely(rc != 0)) {
/* if write fails, we should drop pages from the cache */
for (i = 0; i < npages; i++) {
- if (lnb[i].page == NULL)
- continue;
- LASSERT(PageLocked(lnb[i].page));
- generic_error_remove_page(inode->i_mapping,lnb[i].page);
- }
- }
+ if (lnb[i].lnb_page == NULL)
+ continue;
+ LASSERT(PageLocked(lnb[i].lnb_page));
+ generic_error_remove_page(inode->i_mapping,
+ lnb[i].lnb_page);
+ }
+ }
- RETURN(rc);
+ RETURN(rc);
}
static int osd_read_prep(const struct lu_env *env, struct dt_object *dt,
struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
struct timeval start, end;
unsigned long timediff;
- int rc = 0, i, m = 0, cache = 0;
+ int rc = 0, i, m = 0, cache = 0, cache_hits = 0, cache_misses = 0;
LASSERT(inode);
for (i = 0; i < npages; i++) {
if (i_size_read(inode) <= lnb[i].lnb_file_offset)
- /* If there's no more data, abort early.
- * lnb->rc == 0, so it's easy to detect later. */
- break;
+ /* If there's no more data, abort early.
+ * lnb->lnb_rc == 0, so it's easy to detect later. */
+ break;
+
+ if (i_size_read(inode) <
+ lnb[i].lnb_file_offset + lnb[i].lnb_len - 1)
+ lnb[i].lnb_rc = i_size_read(inode) -
+ lnb[i].lnb_file_offset;
+ else
+ lnb[i].lnb_rc = lnb[i].lnb_len;
+ m += lnb[i].lnb_len;
+
+ if (PageUptodate(lnb[i].lnb_page)) {
+ cache_hits++;
+ } else {
+ cache_misses++;
+ osd_iobuf_add_page(iobuf, lnb[i].lnb_page);
+ }
- if (i_size_read(inode) <
- lnb[i].lnb_file_offset + lnb[i].len - 1)
- lnb[i].rc = i_size_read(inode) - lnb[i].lnb_file_offset;
- else
- lnb[i].rc = lnb[i].len;
- m += lnb[i].len;
-
- lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_ACCESS, 1);
- if (PageUptodate(lnb[i].page)) {
- lprocfs_counter_add(osd->od_stats,
- LPROC_OSD_CACHE_HIT, 1);
- } else {
- lprocfs_counter_add(osd->od_stats,
- LPROC_OSD_CACHE_MISS, 1);
- osd_iobuf_add_page(iobuf, lnb[i].page);
- }
if (cache == 0)
- generic_error_remove_page(inode->i_mapping,lnb[i].page);
+ generic_error_remove_page(inode->i_mapping,
+ lnb[i].lnb_page);
}
do_gettimeofday(&end);
timediff = cfs_timeval_sub(&end, &start, NULL);
lprocfs_counter_add(osd->od_stats, LPROC_OSD_GET_PAGE, timediff);
+ if (cache_hits != 0)
+ lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_HIT,
+ cache_hits);
+ if (cache_misses != 0)
+ lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_MISS,
+ cache_misses);
+ if (cache_hits + cache_misses != 0)
+ lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_ACCESS,
+ cache_hits + cache_misses);
+
if (iobuf->dr_npages) {
rc = osd_ldiskfs_map_inode_pages(inode, iobuf->dr_pages,
iobuf->dr_npages,
- iobuf->dr_blocks,
- 0, NULL);
+ iobuf->dr_blocks, 0);
rc = osd_do_bio(osd, inode, iobuf);
/* IO stats will be done in osd_bufs_put() */
return rc;
}
+static inline int osd_extents_enabled(struct super_block *sb,
+ struct inode *inode)
+{
+ if (inode != NULL) {
+ if (ldiskfs_test_inode_flag(inode, LDISKFS_INODE_EXTENTS))
+ return 1;
+ } else if (LDISKFS_HAS_INCOMPAT_FEATURE(sb,
+ LDISKFS_FEATURE_INCOMPAT_EXTENTS))
+ return 1;
+ return 0;
+}
+
+static inline int osd_calc_bkmap_credits(struct super_block *sb,
+ struct inode *inode,
+ const loff_t size,
+ const loff_t pos,
+ const int blocks)
+{
+ int credits, bits, bs, i;
+
+ bits = sb->s_blocksize_bits;
+ bs = 1 << bits;
+
+ /* legacy blockmap: 3 levels * 3 (bitmap,gd,itself)
+ * we do not expect blockmaps on the large files,
+ * so let's shrink it to 2 levels (4GB files) */
+
+ /* this is default reservation: 2 levels */
+ credits = (blocks + 2) * 3;
+
+ /* actual offset is unknown, hard to optimize */
+ if (pos == -1)
+ return credits;
+
+ /* now check for few specific cases to optimize */
+ if (pos + size <= LDISKFS_NDIR_BLOCKS * bs) {
+ /* no indirects */
+ credits = blocks;
+ /* allocate if not allocated */
+ if (inode == NULL) {
+ credits += blocks * 2;
+ return credits;
+ }
+ for (i = (pos >> bits); i < (pos >> bits) + blocks; i++) {
+ LASSERT(i < LDISKFS_NDIR_BLOCKS);
+ if (LDISKFS_I(inode)->i_data[i] == 0)
+ credits += 2;
+ }
+ } else if (pos + size <= (LDISKFS_NDIR_BLOCKS + 1024) * bs) {
+ /* single indirect */
+ credits = blocks * 3;
+ /* probably indirect block has been allocated already */
+ if (!inode || LDISKFS_I(inode)->i_data[LDISKFS_IND_BLOCK])
+ credits += 3;
+ }
+
+ return credits;
+}
+
static ssize_t osd_declare_write(const struct lu_env *env, struct dt_object *dt,
- const struct lu_buf *buf, loff_t pos,
+ const struct lu_buf *buf, loff_t _pos,
struct thandle *handle)
{
- struct osd_thandle *oh;
- int credits;
- struct inode *inode;
- int rc;
+ struct osd_object *obj = osd_dt_obj(dt);
+ struct inode *inode = obj->oo_inode;
+ struct super_block *sb = osd_sb(osd_obj2dev(obj));
+ struct osd_thandle *oh;
+ int rc = 0, est = 0, credits, blocks, allocated = 0;
+ int bits, bs;
+ int depth, size;
+ loff_t pos;
ENTRY;
+ LASSERT(buf != NULL);
LASSERT(handle != NULL);
oh = container_of0(handle, struct osd_thandle, ot_super);
LASSERT(oh->ot_handle == NULL);
- credits = osd_dto_credits_noquota[DTO_WRITE_BLOCK];
+ size = buf->lb_len;
+ bits = sb->s_blocksize_bits;
+ bs = 1 << bits;
- osd_trans_declare_op(env, oh, OSD_OT_WRITE, credits);
+ if (_pos == -1) {
+ /* if this is an append, then we
+ * should expect cross-block record */
+ pos = 0;
+ } else {
+ pos = _pos;
+ }
- inode = osd_dt_obj(dt)->oo_inode;
+ /* blocks to modify */
+ blocks = ((pos + size + bs - 1) >> bits) - (pos >> bits);
+ LASSERT(blocks > 0);
+
+ if (inode != NULL && _pos != -1) {
+ /* object size in blocks */
+ est = (i_size_read(inode) + bs - 1) >> bits;
+ allocated = inode->i_blocks >> (bits - 9);
+ if (pos + size <= i_size_read(inode) && est <= allocated) {
+ /* looks like an overwrite, no need to modify tree */
+ credits = blocks;
+ /* no need to modify i_size */
+ goto out;
+ }
+ }
- /* we may declare write to non-exist llog */
- if (inode == NULL)
- RETURN(0);
+ if (osd_extents_enabled(sb, inode)) {
+ /*
+ * many concurrent threads may grow tree by the time
+ * our transaction starts. so, consider 2 is a min depth
+ * for every level we may need to allocate a new block
+ * and take some entries from the old one. so, 3 blocks
+ * to allocate (bitmap, gd, itself) + old block - 4 per
+ * level.
+ */
+ depth = inode != NULL ? ext_depth(inode) : 0;
+ depth = max(depth, 1) + 1;
+ credits = depth;
+ /* if not append, then split may need to modify
+ * existing blocks moving entries into the new ones */
+ if (_pos == -1)
+ credits += depth;
+ /* blocks to store data: bitmap,gd,itself */
+ credits += blocks * 3;
+ } else {
+ credits = osd_calc_bkmap_credits(sb, inode, size, _pos, blocks);
+ }
+ /* if inode is created as part of the transaction,
+ * then it's counted already by the creation method */
+ if (inode != NULL)
+ credits++;
+
+out:
+
+ osd_trans_declare_op(env, oh, OSD_OT_WRITE, credits);
/* dt_declare_write() is usually called for system objects, such
* as llog or last_rcvd files. We needn't enforce quota on those
* objects, so always set the lqi_space as 0. */
- rc = osd_declare_inode_qid(env, inode->i_uid, inode->i_gid, 0, oh,
- true, true, NULL, false);
+ if (inode != NULL)
+ rc = osd_declare_inode_qid(env, i_uid_read(inode),
+ i_gid_read(inode), 0, oh, true,
+ true, NULL, false);
RETURN(rc);
}
err);
break;
}
- LASSERTF(boffs + size <= bh->b_size,
- "boffs %d size %d bh->b_size %lu",
- boffs, size, (unsigned long)bh->b_size);
+ 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);
err = ldiskfs_journal_dirty_metadata(handle, bh);
if (err)
inode = osd_dt_obj(dt)->oo_inode;
LASSERT(inode);
- rc = osd_declare_inode_qid(env, inode->i_uid, inode->i_gid, 0, oh,
- true, true, NULL, false);
+ rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
+ 0, oh, true, true, NULL, false);
RETURN(rc);
}
#ifdef HAVE_INODEOPS_TRUNCATE
if (inode->i_op->truncate) {
inode->i_op->truncate(inode);
- } else {
-#endif
- if (!(inode->i_state & (I_NEW|I_FREEING)))
- mutex_lock(&inode->i_mutex);
- ldiskfs_truncate(inode);
- if (!(inode->i_state & (I_NEW|I_FREEING)))
- mutex_unlock(&inode->i_mutex);
-#ifdef HAVE_INODEOPS_TRUNCATE
- }
+ } else
#endif
+ ldiskfs_truncate(inode);
/*
* For a partial-page truncate, flush the page to disk immediately to
git://git.whamcloud.com / fs / lustre-release.git / blobdiff