X-Git-Url: https://git.whamcloud.com/?a=blobdiff_plain;f=lustre%2Fobdfilter%2Ffilter_io_26.c;h=cafb080abd6e98a1db502e76e4dd46eae2cf19ae;hb=51206e8cd42134400fa0b6259a92d7138f3dc984;hp=d6a356e5255dce3536bd27547af20855d84572c2;hpb=f95393b0d0a59cf3dc2f29cffc35dcc4cc9d7728;p=fs%2Flustre-release.git
diff --git a/lustre/obdfilter/filter_io_26.c b/lustre/obdfilter/filter_io_26.c
index d6a356e..cafb080 100644
--- a/lustre/obdfilter/filter_io_26.c
+++ b/lustre/obdfilter/filter_io_26.c
@@ -28,6 +28,8 @@
/*
* Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
* Use is subject to license terms.
+ *
+ * Copyright (c) 2011, 2012, Whamcloud, Inc.
*/
/*
* This file is part of Lustre, http://www.lustre.org/
@@ -103,7 +105,7 @@ static void record_start_io(struct filter_iobuf *iobuf, int rw, int size,
lprocfs_oh_tally(&exp->exp_nid_stats->nid_brw_stats->
hist[BRW_W_RPC_HIST],
cfs_atomic_read(&filter-> \
- fo_r_in_flight));
+ fo_w_in_flight));
lprocfs_oh_tally_log2(&exp->exp_nid_stats->
nid_brw_stats->hist[BRW_W_DISK_IOSIZE],
size);
@@ -127,7 +129,13 @@ static void record_finish_io(struct filter_iobuf *iobuf, int rw, int rc)
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;
@@ -136,14 +144,22 @@ static int dio_complete_routine(struct bio *bio, unsigned int done, int error)
/* 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 , along "
+ "report this to , 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 "
@@ -154,7 +170,7 @@ static int dio_complete_routine(struct bio *bio, unsigned int done, int error)
bio->bi_rw, bio->bi_vcnt, bio->bi_idx, bio->bi_size,
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 */
@@ -185,7 +201,7 @@ static int dio_complete_routine(struct bio *bio, unsigned int done, int error)
* 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)
@@ -356,19 +372,22 @@ int filter_do_bio(struct obd_export *exp, struct inode *inode,
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);
@@ -381,7 +400,7 @@ int filter_do_bio(struct obd_export *exp, struct inode *inode,
}
/* 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) {
@@ -464,7 +483,7 @@ int filter_direct_io(int rw, struct dentry *dchild, struct filter_iobuf *iobuf,
struct inode *inode = dchild->d_inode;
int blocks_per_page = CFS_PAGE_SIZE >> inode->i_blkbits;
int rc, rc2, create;
- cfs_semaphore_t *sem;
+ cfs_mutex_t *mutex;
ENTRY;
LASSERTF(iobuf->dr_npages <= iobuf->dr_max_pages, "%d,%d\n",
@@ -475,20 +494,25 @@ int filter_direct_io(int rw, struct dentry *dchild, struct filter_iobuf *iobuf,
if (iobuf->dr_npages == 0)
RETURN(0);
create = 0;
- sem = NULL;
+ mutex = NULL;
} else {
LASSERTF(rw == OBD_BRW_WRITE, "%x\n", rw);
LASSERT(iobuf->dr_npages > 0);
create = 1;
- sem = &obd->u.filter.fo_alloc_lock;
+ mutex = &obd->u.filter.fo_alloc_lock;
lquota_enforce(filter_quota_interface_ref, obd,
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);
+ obdfilter_created_scratchpad, create, mutex);
+ }
if (rw == OBD_BRW_WRITE) {
if (rc == 0) {
@@ -503,6 +527,13 @@ int filter_direct_io(int rw, struct dentry *dchild, struct filter_iobuf *iobuf,
UNLOCK_INODE_MUTEX(inode);
+ /* 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);
@@ -510,8 +541,11 @@ int filter_direct_io(int rw, struct dentry *dchild, struct filter_iobuf *iobuf,
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)
@@ -572,10 +606,12 @@ int filter_commitrw_write(struct obd_export *exp, struct obdo *oa,
int i, err, cleanup_phase = 0;
struct obd_device *obd = exp->exp_obd;
struct filter_obd *fo = &obd->u.filter;
- void *wait_handle;
+ 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);
@@ -643,6 +679,10 @@ int filter_commitrw_write(struct obd_export *exp, struct obdo *oa,
(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
@@ -653,11 +693,15 @@ int filter_commitrw_write(struct obd_export *exp, struct obdo *oa,
if (rc == -ENOTCONN)
GOTO(cleanup, rc);
+ if (OBD_FAIL_CHECK(OBD_FAIL_OST_DQACQ_NET))
+ GOTO(cleanup, rc = -EINPROGRESS);
+
push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
cleanup_phase = 2;
- DQUOT_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,
@@ -674,6 +718,9 @@ int filter_commitrw_write(struct obd_export *exp, struct obdo *oa,
fsfilt_check_slow(obd, now, "brw_start");
+ /* Locking order: i_mutex -> journal_lock -> dqptr_sem. LU-952 */
+ ll_vfs_dq_init(inode);
+
i = OBD_MD_FLATIME | OBD_MD_FLMTIME | OBD_MD_FLCTIME;
/* If the inode still has SUID+SGID bits set (see filter_precreate())
@@ -715,25 +762,55 @@ int filter_commitrw_write(struct obd_export *exp, struct obdo *oa,
/* filter_direct_io drops i_mutex */
rc = filter_direct_io(OBD_BRW_WRITE, res->dentry, iobuf, exp, &iattr,
- oti, &wait_handle);
+ 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);
+ obdo_from_inode(oa, inode, (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)
- LASSERTF(oti->oti_transno <= obd->obd_last_committed,
- "oti_transno "LPU64" last_committed "LPU64"\n",
- oti->oti_transno, obd->obd_last_committed);
+ /* In rare cases fsfilt_commit_wait() will wake up and return after
+ * the transaction has finished its work and updated j_commit_sequence
+ * but the commit callbacks have not been run yet. Wait here until
+ * that is finished so that clients requesting sync IO don't see the
+ * reply transno < last_committed. LU-753 */
+ if (unlikely(obd->obd_replayable && !rc && wait_handle &&
+ oti->oti_transno > obd->obd_last_committed)) {
+ cfs_waitq_t wq;
+ struct l_wait_info lwi =
+ LWI_TIMEOUT_INTERVAL(cfs_time_seconds(5),
+ (cfs_duration_t)((HZ + 4)/5),
+ NULL, NULL);
+ cfs_waitq_init(&wq);
+ l_wait_event(wq,
+ oti->oti_transno <= obd->obd_last_committed,
+ &lwi);
+
+ /* commit callback isn't done after waiting for 5 secs ? */
+ if (unlikely(oti->oti_transno > obd->obd_last_committed))
+ CERROR("transno:"LPU64" > last_committed:"LPU64"\n",
+ oti->oti_transno, obd->obd_last_committed);
+ }
fsfilt_check_slow(obd, now, "commitrw commit");