*/
/*
* This file is part of Lustre, http://www.lustre.org/
- * Lustre is a trademark of Sun Microsystems, Inc.
*
* lustre/osd/osd_io.c
*
#ifdef HAVE_BIO_ENDIO_USES_ONE_ARG
static void dio_complete_routine(struct bio *bio)
{
- int error = bio->bi_status;
+ int error = blk_status_to_errno(bio->bi_status);
#else
static void dio_complete_routine(struct bio *bio, int error)
{
RETURN(0);
}
+static void osd_mark_page_io_done(struct osd_iobuf *iobuf,
+ struct inode *inode,
+ sector_t start_blocks,
+ sector_t count)
+{
+ struct niobuf_local *lnb;
+ int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
+ pgoff_t pg_start, pg_end;
+
+ pg_start = start_blocks / blocks_per_page;
+ if (start_blocks % blocks_per_page)
+ pg_start++;
+ if (count >= blocks_per_page)
+ pg_end = (start_blocks + count -
+ blocks_per_page) / blocks_per_page;
+ else
+ return; /* nothing to mark */
+ for ( ; pg_start <= pg_end; pg_start++) {
+ lnb = iobuf->dr_lnbs[pg_start];
+ lnb->lnb_flags |= OBD_BRW_DONE;
+ }
+}
+
static int osd_do_bio(struct osd_device *osd, struct inode *inode,
struct osd_iobuf *iobuf, sector_t start_blocks,
sector_t count)
OBD_FREE_PTR(bio_private);
}
+ /* Write only now */
+ if (rc == 0 && iobuf->dr_rw)
+ osd_mark_page_io_done(iobuf, inode,
+ start_blocks, count);
+
RETURN(rc);
}
}
#ifdef HAVE_LDISKFS_JOURNAL_ENSURE_CREDITS
-static int osd_extend_trans(handle_t *handle, int needed,
- struct inode *inode)
-{
- return __ldiskfs_journal_ensure_credits(handle, needed, needed,
- ldiskfs_trans_default_revoke_credits(inode->i_sb));
-}
-
static int osd_extend_restart_trans(handle_t *handle, int needed,
struct inode *inode)
{
- return ldiskfs_journal_ensure_credits(handle, needed,
+ int rc;
+
+ rc = ldiskfs_journal_ensure_credits(handle, needed,
ldiskfs_trans_default_revoke_credits(inode->i_sb));
-}
-#else
-static int osd_extend_trans(handle_t *handle, int needed,
- struct inode *inode)
-{
- if (ldiskfs_handle_has_enough_credits(handle, needed))
- return 0;
+ /* this means journal has been restarted */
+ if (rc > 0)
+ rc = 0;
- return ldiskfs_journal_extend(handle,
- needed - handle->h_buffer_credits);
+ return rc;
}
-
+#else
static int osd_extend_restart_trans(handle_t *handle, int needed,
struct inode *inode)
{
+ int rc;
- int rc = osd_extend_trans(handle, needed, inode);
-
+ if (ldiskfs_handle_has_enough_credits(handle, needed))
+ return 0;
+ rc = ldiskfs_journal_extend(handle,
+ needed - handle->h_buffer_credits);
if (rc <= 0)
return rc;
return osd_do_bio(osd, inode, iobuf, start_blocks, count);
}
+static unsigned int osd_extent_bytes(const struct osd_device *o)
+{
+ unsigned int *extent_bytes_ptr =
+ raw_cpu_ptr(o->od_extent_bytes_percpu);
+
+ if (likely(*extent_bytes_ptr))
+ return *extent_bytes_ptr;
+
+ /* initialize on first access or CPU hotplug */
+ if (!ldiskfs_has_feature_extents(osd_sb(o)))
+ *extent_bytes_ptr = 1 << osd_sb(o)->s_blocksize_bits;
+ else
+ *extent_bytes_ptr = OSD_DEFAULT_EXTENT_BYTES;
+
+ return *extent_bytes_ptr;
+}
+
+#define EXTENT_BYTES_DECAY 64
+static void osd_decay_extent_bytes(struct osd_device *osd,
+ unsigned int new_bytes)
+{
+ unsigned int old_bytes;
+
+ if (!ldiskfs_has_feature_extents(osd_sb(osd)))
+ return;
+
+ old_bytes = osd_extent_bytes(osd);
+ *raw_cpu_ptr(osd->od_extent_bytes_percpu) =
+ (old_bytes * (EXTENT_BYTES_DECAY - 1) +
+ min(new_bytes, OSD_DEFAULT_EXTENT_BYTES) +
+ EXTENT_BYTES_DECAY - 1) / EXTENT_BYTES_DECAY;
+}
static int osd_ldiskfs_map_inode_pages(struct inode *inode,
struct osd_iobuf *iobuf,
struct osd_device *osd,
int create, __u64 user_size,
- int check_credits)
+ int check_credits,
+ struct thandle *thandle)
{
int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
int rc = 0, i = 0, mapped_index = 0;
int clen = 0;
pgoff_t max_page_index;
handle_t *handle = NULL;
- int credits;
sector_t start_blocks = 0, count = 0;
loff_t disk_size = 0;
struct page **page = iobuf->dr_pages;
* transaction to make sure consistency.
*/
if (handle && check_credits) {
- /*
- * credits to insert 1 extent into extent tree.
- */
- credits = osd_chunk_trans_blocks(inode, blen);
- rc = osd_extend_trans(handle, credits, inode);
- if (rc < 0)
- GOTO(cleanup, rc);
+ struct osd_thandle *oh;
+
+ LASSERT(thandle != NULL);
+ oh = container_of(thandle, struct osd_thandle,
+ ot_super);
/*
* only issue IO if restart transaction needed,
* as update disk size need hold inode lock, we
* want to avoid that as much as possible.
*/
- if (rc > 0) {
- WARN_ON_ONCE(start_blocks == 0);
+ if (oh->oh_declared_ext <= 0) {
rc = osd_ldiskfs_map_write(inode,
iobuf, osd, start_blocks,
count, &disk_size, user_size);
if (rc)
GOTO(cleanup, rc);
-#ifdef HAVE_LDISKFS_JOURNAL_ENSURE_CREDITS
- rc = ldiskfs_journal_restart(handle, credits,
- ldiskfs_trans_default_revoke_credits(inode->i_sb));
-#else
- rc = ldiskfs_journal_restart(handle, credits);
-#endif
- if (rc)
- GOTO(cleanup, rc);
- start_blocks += count;
- /* reset IO block count */
- count = 0;
+ thandle->th_restart_tran = 1;
+ GOTO(cleanup, rc = -EAGAIN);
}
+
+ if (OBD_FAIL_CHECK(OBD_FAIL_OST_RESTART_IO))
+ oh->oh_declared_ext = 0;
+ else
+ oh->oh_declared_ext--;
}
rc = ldiskfs_map_blocks(handle, inode, &map, create);
if (rc >= 0) {
total++;
break;
}
+ if ((map.m_flags & LDISKFS_MAP_UNWRITTEN) &&
+ !create) {
+ /* don't try to read allocated, but
+ * unwritten blocks, instead fill the
+ * patches with zeros in osd_do_bio() */
+ *(blocks + total) = 0;
+ continue;
+ }
*(blocks + total) = map.m_pblk + c;
/* unmap any possible underlying
* metadata from the block device
}
if (rc == 0 && total < blen) {
+ /*
+ * decay extent blocks if we could not
+ * allocate extent once.
+ */
+ osd_decay_extent_bytes(osd,
+ (total - previous_total) << inode->i_blkbits);
map.m_lblk = fp->index * blocks_per_page + total;
map.m_len = blen - total;
previous_total = total;
}
if (rc != 0)
GOTO(cleanup, rc);
-
+ /*
+ * decay extent blocks if we could allocate
+ * good large extent.
+ */
+ if (total - previous_total >=
+ osd_extent_bytes(osd) >> inode->i_blkbits)
+ osd_decay_extent_bytes(osd,
+ (total - previous_total) << inode->i_blkbits);
/* look for next extent */
fp = NULL;
blocks += blocks_per_page * clen;
if (iobuf->dr_npages) {
rc = osd_ldiskfs_map_inode_pages(inode, iobuf, osd, 0,
- 0, 0);
+ 0, 0, NULL);
if (likely(rc == 0)) {
rc = osd_do_bio(osd, inode, iobuf, 0, 0);
/* do IO stats for preparation reads */
sector_t start;
struct fiemap_extent_info fei = { 0 };
struct fiemap_extent fe = { 0 };
- mm_segment_t saved_fs;
int rc;
if (block >= cached_extent->start && block < cached_extent->end)
fei.fi_extents_max = 1;
fei.fi_extents_start = &fe;
- saved_fs = get_fs();
- set_fs(KERNEL_DS);
rc = inode->i_op->fiemap(inode, &fei, offset, FIEMAP_MAX_OFFSET-offset);
- set_fs(saved_fs);
if (rc != 0)
return 0;
start = fe.fe_logical >> inode->i_blkbits;
+ if (fei.fi_extents_mapped == 0) {
+ /* a special case - no extent found at this offset and forward.
+ * we can consider this as a hole to EOF. it's safe to cache
+ * as other threads can not allocate/punch blocks this thread
+ * is working on (LDLM). */
+ cached_extent->start = block;
+ cached_extent->end = i_size_read(inode) >> inode->i_blkbits;
+ cached_extent->mapped = 0;
+ return 0;
+ }
if (start > block) {
cached_extent->start = block;
return cached_extent->mapped;
}
+#define MAX_EXTENTS_PER_WRITE 100
static int osd_declare_write_commit(const struct lu_env *env,
struct dt_object *dt,
struct niobuf_local *lnb, int npages,
struct osd_fextent mapped = { 0 }, extent = { 0 };
enum osd_quota_local_flags local_flags = 0;
enum osd_qid_declare_flags declare_flags = OSD_QID_BLK;
+ unsigned int extent_bytes;
ENTRY;
LASSERT(handle != NULL);
oh = container_of(handle, struct osd_thandle, ot_super);
LASSERT(oh->ot_handle == NULL);
+ /*
+ * We track a decaying average extent blocks per filesystem,
+ * for most of time, it will be 1M, with filesystem becoming
+ * heavily-fragmented, it will be reduced to 4K at the worst.
+ */
+ extent_bytes = osd_extent_bytes(osd);
+ LASSERT(extent_bytes >= (1 << osd_sb(osd)->s_blocksize));
+
/* calculate number of extents (probably better to pass nb) */
for (i = 0; i < npages; i++) {
/* ignore quota for the whole request if any page is from
continue;
}
+ if (lnb[i].lnb_flags & OBD_BRW_DONE) {
+ lnb[i].lnb_flags |= OBD_BRW_MAPPED;
+ continue;
+ }
+
/* count only unmapped changes */
newblocks++;
if (lnb[i].lnb_file_offset != extent.end || extent.end == 0) {
- extents++;
+ if (extent.end != 0)
+ extents += (extent.end - extent.start +
+ extent_bytes - 1) / extent_bytes;
+ extent.start = lnb[i].lnb_file_offset;
extent.end = lnb[i].lnb_file_offset + lnb[i].lnb_len;
} else {
extent.end += lnb[i].lnb_len;
*/
if (!newblocks)
goto out_declare;
+
+ extents += (extent.end - extent.start +
+ extent_bytes - 1) / extent_bytes;
+ /**
+ * with system space usage growing up, mballoc codes won't
+ * try best to scan block group to align best free extent as
+ * we can. So extent bytes per extent could be decayed to a
+ * very small value, this could make us reserve too many credits.
+ * We could be more optimistic in the credit reservations, even
+ * in a case where the filesystem is nearly full, it is extremely
+ * unlikely that the worst case would ever be hit.
+ */
+ if (extents > MAX_EXTENTS_PER_WRITE)
+ extents = MAX_EXTENTS_PER_WRITE;
+
/*
* each extent can go into new leaf causing a split
* 5 is max tree depth: inode + 4 index blocks
credits += depth * extents;
}
- /*
- * try a bit more extents to avoid restart
- * as much as possible in normal case.
- */
- if (npages > 1 && extents)
- extents <<= 1;
+ oh->oh_declared_ext = extents;
/* quota space for metadata blocks */
quota_space += depth * extents * LDISKFS_BLOCK_SIZE(osd_sb(osd));
else
credits += extents;
+ CDEBUG(D_INODE,
+ "%s: inode #%lu extent_bytes %u extents %d credits %d\n",
+ osd_ino2name(inode), inode->i_ino, extent_bytes, extents,
+ credits);
+
out_declare:
osd_trans_declare_op(env, oh, OSD_OT_WRITE, credits);
struct inode *inode = osd_dt_obj(dt)->oo_inode;
struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
int rc = 0, i, check_credits = 0;
- struct osd_thandle *oh = container_of(thandle,
- struct osd_thandle, ot_super);
- unsigned int save_credits = oh->ot_credits;
LASSERT(inode);
continue;
}
+ if (lnb[i].lnb_flags & OBD_BRW_DONE)
+ continue;
+
if (!(lnb[i].lnb_flags & OBD_BRW_MAPPED))
check_credits = 1;
} else if (iobuf->dr_npages > 0) {
rc = osd_ldiskfs_map_inode_pages(inode, iobuf, osd,
1, user_size,
- check_credits);
- /*
- * Write might restart transaction, extend credits
- * if needed for operations such as attribute set.
- */
- if (rc == 0) {
- handle_t *handle = ldiskfs_journal_current_handle();
-
- LASSERT(handle != NULL);
- rc = osd_extend_restart_trans(handle, save_credits,
- inode);
- }
+ check_credits,
+ thandle);
} else {
/* no pages to write, no transno is needed */
thandle->th_local = 1;
}
- if (rc != 0)
+ if (rc != 0 && !thandle->th_restart_tran)
osd_fini_iobuf(osd, iobuf);
osd_trans_exec_check(env, thandle, OSD_OT_WRITE);
- if (unlikely(rc != 0)) {
+ if (unlikely(rc != 0 && !thandle->th_restart_tran)) {
/* if write fails, we should drop pages from the cache */
for (i = 0; i < npages; i++) {
if (lnb[i].lnb_page == NULL)
if (iobuf->dr_npages) {
rc = osd_ldiskfs_map_inode_pages(inode, iobuf, osd, 0,
- 0, 0);
+ 0, 0, NULL);
if (!rc)
rc = osd_do_bio(osd, inode, iobuf, 0, 0);
++bufsize;
}
- dirty_inode = test_and_set_bit(LDISKFS_INODE_JOURNAL_DATA,
+ /* only the first flag-set matters */
+ dirty_inode = !test_and_set_bit(LDISKFS_INODE_JOURNAL_DATA,
&ei->i_flags);
/* sparse checking is racy, but sparse is very rare case, leave as is */
ENTRY;
/*
- * Only mode == 0 (which is standard prealloc) is supported now.
+ * mode == 0 (which is standard prealloc) and PUNCH is supported
* Rest of mode options is not supported yet.
*/
- if (mode & ~FALLOC_FL_KEEP_SIZE)
+ if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
+ RETURN(-EOPNOTSUPP);
+
+ /* disable fallocate completely */
+ if (osd_dev(dt->do_lu.lo_dev)->od_fallocate_zero_blocks < 0)
RETURN(-EOPNOTSUPP);
LASSERT(th);
LASSERT(inode);
+ if (mode & FALLOC_FL_PUNCH_HOLE) {
+ rc = osd_declare_inode_qid(env, i_uid_read(inode),
+ i_gid_read(inode),
+ i_projid_read(inode), 0, oh,
+ osd_dt_obj(dt), NULL, OSD_QID_BLK);
+ if (rc == 0)
+ rc = osd_trunc_lock(osd_dt_obj(dt), oh, false);
+ RETURN(rc);
+ }
+
/* quota space for metadata blocks
* approximate metadata estimate should be good enough.
*/
RETURN(rc);
}
-static int osd_fallocate(const struct lu_env *env, struct dt_object *dt,
- __u64 start, __u64 end, int mode, struct thandle *th)
+static int osd_fallocate_preallocate(const struct lu_env *env,
+ struct dt_object *dt,
+ __u64 start, __u64 end, int mode,
+ struct thandle *th)
{
struct osd_thandle *oh = container_of(th, struct osd_thandle, ot_super);
handle_t *handle = ldiskfs_journal_current_handle();
boff = start >> inode->i_blkbits;
blen = (ALIGN(end, 1 << inode->i_blkbits) >> inode->i_blkbits) - boff;
- flags = LDISKFS_GET_BLOCKS_CREATE;
+ /* Create and mark new extents as either zero or unwritten */
+ flags = osd_dev(dt->do_lu.lo_dev)->od_fallocate_zero_blocks ?
+ LDISKFS_GET_BLOCKS_CREATE_ZERO :
+ LDISKFS_GET_BLOCKS_CREATE_UNWRIT_EXT;
if (mode & FALLOC_FL_KEEP_SIZE)
flags |= LDISKFS_GET_BLOCKS_KEEP_SIZE;
RETURN(rc);
}
+static int osd_fallocate_punch(const struct lu_env *env, struct dt_object *dt,
+ __u64 start, __u64 end, int mode,
+ struct thandle *th)
+{
+ struct osd_object *obj = osd_dt_obj(dt);
+ struct inode *inode = obj->oo_inode;
+ struct osd_access_lock *al;
+ struct osd_thandle *oh;
+ int rc = 0, found = 0;
+
+ ENTRY;
+
+ LASSERT(dt_object_exists(dt));
+ LASSERT(osd_invariant(obj));
+ LASSERT(inode != NULL);
+
+ dquot_initialize(inode);
+
+ LASSERT(th);
+ oh = container_of(th, struct osd_thandle, ot_super);
+ LASSERT(oh->ot_handle->h_transaction != NULL);
+
+ list_for_each_entry(al, &oh->ot_trunc_locks, tl_list) {
+ if (obj != al->tl_obj)
+ continue;
+ LASSERT(al->tl_shared == 0);
+ found = 1;
+ /* do actual punch in osd_trans_stop() */
+ al->tl_start = start;
+ al->tl_end = end;
+ al->tl_mode = mode;
+ al->tl_punch = true;
+ break;
+ }
+
+ RETURN(rc);
+}
+
+static int osd_fallocate(const struct lu_env *env, struct dt_object *dt,
+ __u64 start, __u64 end, int mode, struct thandle *th)
+{
+ int rc;
+
+ ENTRY;
+
+ if (mode & FALLOC_FL_PUNCH_HOLE) {
+ /* punch */
+ rc = osd_fallocate_punch(env, dt, start, end, mode, th);
+ } else {
+ /* standard preallocate */
+ rc = osd_fallocate_preallocate(env, dt, start, end, mode, th);
+ }
+ RETURN(rc);
+}
+
static int osd_declare_punch(const struct lu_env *env, struct dt_object *dt,
__u64 start, __u64 end, struct thandle *th)
{
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)
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;
}
}
}
+/*
+ * For a partial-page truncate, flush the page to disk immediately to
+ * avoid data corruption during direct disk write. b=17397
+ */
+static void osd_partial_page_flush(struct osd_device *d, struct inode *inode,
+ loff_t offset)
+{
+ if (!(offset & ~PAGE_MASK))
+ return;
+
+ if (osd_use_page_cache(d)) {
+ filemap_fdatawrite_range(inode->i_mapping, offset, offset + 1);
+ } else {
+ /* Notice we use "wait" version to ensure I/O is complete */
+ filemap_write_and_wait_range(inode->i_mapping, offset,
+ offset + 1);
+ invalidate_mapping_pages(inode->i_mapping, offset >> PAGE_SHIFT,
+ offset >> PAGE_SHIFT);
+ }
+}
+
void osd_execute_truncate(struct osd_object *obj)
{
struct osd_device *d = osd_obj2dev(obj);
spin_unlock(&inode->i_lock);
osd_dirty_inode(inode, I_DIRTY_DATASYNC);
}
+ osd_partial_page_flush(d, inode, size);
+}
- /*
- * For a partial-page truncate, flush the page to disk immediately to
- * avoid data corruption during direct disk write. b=17397
- */
- if ((size & ~PAGE_MASK) == 0)
- return;
- if (osd_use_page_cache(d)) {
- filemap_fdatawrite_range(inode->i_mapping, size, size + 1);
- } else {
- /* Notice we use "wait" version to ensure I/O is complete */
- filemap_write_and_wait_range(inode->i_mapping, size, size + 1);
- invalidate_mapping_pages(inode->i_mapping, size >> PAGE_SHIFT,
- size >> PAGE_SHIFT);
- }
+void osd_execute_punch(const struct lu_env *env, struct osd_object *obj,
+ loff_t start, loff_t end, int mode)
+{
+ struct osd_device *d = osd_obj2dev(obj);
+ struct inode *inode = obj->oo_inode;
+ struct file *file = osd_quasi_file(env, inode);
+
+ file->f_op->fallocate(file, mode, start, end - start);
+ osd_partial_page_flush(d, inode, start);
+ osd_partial_page_flush(d, inode, end - 1);
}
-void osd_process_truncates(struct list_head *list)
+void osd_process_truncates(const struct lu_env *env, struct list_head *list)
{
struct osd_access_lock *al;
list_for_each_entry(al, list, tl_list) {
if (al->tl_shared)
continue;
- if (!al->tl_truncate)
- continue;
- osd_execute_truncate(al->tl_obj);
+ if (al->tl_truncate)
+ osd_execute_truncate(al->tl_obj);
+ else if (al->tl_punch)
+ osd_execute_punch(env, al->tl_obj, al->tl_start,
+ al->tl_end, al->tl_mode);
}
}
git://git.whamcloud.com / fs / lustre-release.git / blobdiff