* GPL HEADER END
*/
/*
- * Copyright 2008 Sun Microsystems, Inc. All rights reserved
+ * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
* Use is subject to license terms.
*/
/*
/* 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;
- 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;
};
{
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);
+ }
}
}
* 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;
struct bio_vec *bvl;
/* 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 (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 <http://bugzilla.lustre.org/> , along "
+ "report this to <http://bugs.whamcloud.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 "
"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);
}
/* the check is outside of the cycle for performance reason -bzzz */
- if (!test_bit(BIO_RW, &bio->bi_rw)) {
+ if (!cfs_test_bit(BIO_RW, &bio->bi_rw)) {
bio_for_each_segment(bvl, bio, i) {
if (likely(error == 0))
SetPageUptodate(bvl->bv_page);
* 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)
goto failed_2;
iobuf->dr_filter = filter;
- init_waitqueue_head(&iobuf->dr_wait);
- atomic_set(&iobuf->dr_numreqs, 0);
+ 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;
{
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)
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);
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);
}
/* allocate new bio, limited by max BIO size, b=9945 */
- bio = bio_alloc(GFP_NOIO, max(BIO_MAX_PAGES,
+ bio = bio_alloc(GFP_NOIO, min(BIO_MAX_PAGES,
(npages - page_idx) *
blocks_per_page));
if (bio == NULL) {
}
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],
+ 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],
+ 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],
+ 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],
+ lprocfs_oh_tally_log2(&exp->exp_nid_stats->
+ nid_brw_stats->hist[BRW_W_IO_TIME],
jiffies - start_time);
}
}
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",
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) {
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)
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;
struct filter_obd *fo = &obd->u.filter;
- void *wait_handle;
- int total_size = 0, rc2;
- unsigned int qcids[MAXQUOTAS] = {0, 0};
+ 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);
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));
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 < 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);
LASSERT(PageLocked(lnb->page));
LASSERT(!PageWriteback(lnb->page));
- /* truncate might leave tail dirty */
- clear_page_dirty_for_io(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);
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, "i_mutex");
oti->oti_handle = fsfilt_brw_start(obd, objcount, &fso, niocount, res,
/* 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, "direct_io");
- err = fsfilt_commit_wait(obd, inode, wait_handle);
+ if (wait_handle)
+ err = fsfilt_commit_wait(obd, inode, wait_handle);
+ else
+ err = 0;
+
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, "commitrw commit");
cleanup:
+ lquota_pending_commit(filter_quota_interface_ref, obd, qcids,
+ rec_pending, 1);
+
filter_grant_commit(exp, niocount, res);
switch (cleanup_phase) {
* 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 && (fo->fo_writethrough_cache == 0 ||
- i_size_read(inode) > fo->fo_readcache_max_filesize))
- filter_invalidate_cache(obd, obj, nb, inode);
+ 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);
}
git://git.whamcloud.com / fs / lustre-release.git / blobdiff