*/
/*
* This file is part of Lustre, http://www.lustre.org/
- * Lustre is a trademark of Sun Microsystems, Inc.
*
* lustre/ofd/ofd_io.c
*
OBD_FREE_PTR(oii);
}
+struct oivm_args {
+ struct ofd_device *od_ofd;
+ struct lu_env od_env;
+ struct lfsck_req_local od_lrl;
+ struct completion *od_started;
+};
+
+#ifndef TASK_IDLE
+#define TASK_IDLE TASK_INTERRUPTIBLE
+#endif
+
/**
* Verification thread to check parent FID consistency.
*
* \retval 0 on successful thread termination
* \retval negative value if thread can't start
*/
-static int ofd_inconsistency_verification_main(void *args)
+static int ofd_inconsistency_verification_main(void *_args)
{
- struct lu_env env;
- struct ofd_device *ofd = args;
- struct ptlrpc_thread *thread = &ofd->ofd_inconsistency_thread;
+ struct oivm_args *args = _args;
+ struct lu_env *env = &args->od_env;
+ struct ofd_device *ofd = args->od_ofd;
struct ofd_inconsistency_item *oii;
- struct lfsck_req_local *lrl = NULL;
- int rc;
+ struct lfsck_req_local *lrl = &args->od_lrl;
ENTRY;
- rc = lu_env_init(&env, LCT_DT_THREAD);
- spin_lock(&ofd->ofd_inconsistency_lock);
- thread_set_flags(thread, rc ? SVC_STOPPED : SVC_RUNNING);
- wake_up_all(&thread->t_ctl_waitq);
- spin_unlock(&ofd->ofd_inconsistency_lock);
- if (rc)
- RETURN(rc);
-
- OBD_ALLOC_PTR(lrl);
- if (unlikely(!lrl))
- GOTO(out_unlocked, rc = -ENOMEM);
-
lrl->lrl_event = LEL_PAIRS_VERIFY_LOCAL;
lrl->lrl_active = LFSCK_TYPE_LAYOUT;
+ complete(args->od_started);
spin_lock(&ofd->ofd_inconsistency_lock);
- while (1) {
- if (unlikely(!thread_is_running(thread)))
- break;
+ while (({set_current_state(TASK_IDLE);
+ !kthread_should_stop(); })) {
while (!list_empty(&ofd->ofd_inconsistency_list)) {
+ __set_current_state(TASK_RUNNING);
oii = list_entry(ofd->ofd_inconsistency_list.next,
struct ofd_inconsistency_item,
oii_list);
list_del_init(&oii->oii_list);
spin_unlock(&ofd->ofd_inconsistency_lock);
- ofd_inconsistency_verify_one(&env, ofd, oii, lrl);
+ ofd_inconsistency_verify_one(env, ofd, oii, lrl);
spin_lock(&ofd->ofd_inconsistency_lock);
}
spin_unlock(&ofd->ofd_inconsistency_lock);
- wait_event_idle(thread->t_ctl_waitq,
- !list_empty(&ofd->ofd_inconsistency_list) ||
- !thread_is_running(thread));
+ schedule();
spin_lock(&ofd->ofd_inconsistency_lock);
}
+ __set_current_state(TASK_RUNNING);
while (!list_empty(&ofd->ofd_inconsistency_list)) {
struct ofd_object *fo;
fo = oii->oii_obj;
spin_unlock(&ofd->ofd_inconsistency_lock);
- ofd_write_lock(&env, fo);
+ ofd_write_lock(env, fo);
fo->ofo_pfid_checking = 0;
- ofd_write_unlock(&env, fo);
+ ofd_write_unlock(env, fo);
- ofd_object_put(&env, fo);
+ ofd_object_put(env, fo);
OBD_FREE_PTR(oii);
spin_lock(&ofd->ofd_inconsistency_lock);
}
- OBD_FREE_PTR(lrl);
-
- GOTO(out, rc = 0);
-
-out_unlocked:
- spin_lock(&ofd->ofd_inconsistency_lock);
-out:
- thread_set_flags(thread, SVC_STOPPED);
- wake_up_all(&thread->t_ctl_waitq);
spin_unlock(&ofd->ofd_inconsistency_lock);
- lu_env_fini(&env);
- return rc;
+ lu_env_fini(&args->od_env);
+ OBD_FREE_PTR(args);
+ return 0;
}
/**
*/
int ofd_start_inconsistency_verification_thread(struct ofd_device *ofd)
{
- struct ptlrpc_thread *thread = &ofd->ofd_inconsistency_thread;
struct task_struct *task;
+ struct oivm_args *args;
+ DECLARE_COMPLETION_ONSTACK(started);
int rc;
- spin_lock(&ofd->ofd_inconsistency_lock);
- if (unlikely(thread_is_running(thread))) {
- spin_unlock(&ofd->ofd_inconsistency_lock);
-
+ if (ofd->ofd_inconsistency_task)
return -EALREADY;
+
+ OBD_ALLOC_PTR(args);
+ if (!args)
+ return -ENOMEM;
+ rc = lu_env_init(&args->od_env, LCT_DT_THREAD);
+ if (rc) {
+ OBD_FREE_PTR(args);
+ return rc;
}
- thread_set_flags(thread, 0);
- spin_unlock(&ofd->ofd_inconsistency_lock);
- task = kthread_run(ofd_inconsistency_verification_main, ofd,
- "inconsistency_verification");
+ args->od_ofd = ofd;
+ args->od_started = &started;
+ task = kthread_create(ofd_inconsistency_verification_main, args,
+ "inconsistency_verification");
if (IS_ERR(task)) {
rc = PTR_ERR(task);
CERROR("%s: cannot start self_repair thread: rc = %d\n",
ofd_name(ofd), rc);
} else {
rc = 0;
- wait_event_idle(thread->t_ctl_waitq,
- thread_is_running(thread) ||
- thread_is_stopped(thread));
+ spin_lock(&ofd->ofd_inconsistency_lock);
+ if (ofd->ofd_inconsistency_task)
+ rc = -EALREADY;
+ else
+ ofd->ofd_inconsistency_task = task;
+ spin_unlock(&ofd->ofd_inconsistency_lock);
+
+ if (rc)
+ kthread_stop(task);
+ else {
+ wake_up_process(task);
+ wait_for_completion(&started);
+ }
+ }
+ if (rc) {
+ lu_env_fini(&args->od_env);
+ OBD_FREE_PTR(args);
}
return rc;
*/
int ofd_stop_inconsistency_verification_thread(struct ofd_device *ofd)
{
- struct ptlrpc_thread *thread = &ofd->ofd_inconsistency_thread;
+ struct task_struct *task;
spin_lock(&ofd->ofd_inconsistency_lock);
- if (thread_is_init(thread) || thread_is_stopped(thread)) {
- spin_unlock(&ofd->ofd_inconsistency_lock);
+ task = ofd->ofd_inconsistency_task;
+ ofd->ofd_inconsistency_task = NULL;
+ spin_unlock(&ofd->ofd_inconsistency_lock);
+ if (!task)
return -EALREADY;
- }
-
- thread_set_flags(thread, SVC_STOPPING);
- spin_unlock(&ofd->ofd_inconsistency_lock);
- wake_up_all(&thread->t_ctl_waitq);
- wait_event_idle(thread->t_ctl_waitq,
- thread_is_stopped(thread));
+ kthread_stop(task);
return 0;
}
if (list_empty(&ofd->ofd_inconsistency_list))
wakeup = true;
list_add_tail(&oii->oii_list, &ofd->ofd_inconsistency_list);
+ if (wakeup && ofd->ofd_inconsistency_task)
+ wake_up_process(ofd->ofd_inconsistency_task);
spin_unlock(&ofd->ofd_inconsistency_lock);
- if (wakeup)
- wake_up_all(&ofd->ofd_inconsistency_thread.t_ctl_waitq);
/* XXX: When the found inconsistency exceeds some threshold,
* we can trigger the LFSCK to scan part of the system
}
+/*
+ * Lazy ATIME update to refresh atime every ofd_atime_diff
+ * seconds so that external scanning tool can see it actual
+ * within that period and be able to identify accessed files
+ */
+static void ofd_handle_atime(const struct lu_env *env, struct ofd_device *ofd,
+ struct ofd_object *fo, time64_t atime)
+{
+ struct lu_attr *la;
+ struct dt_object *o;
+ struct thandle *th;
+ int rc;
+
+ if (ofd->ofd_atime_diff == 0)
+ return;
+
+ la = &ofd_info(env)->fti_attr2;
+ o = ofd_object_child(fo);
+
+ if (unlikely(fo->ofo_atime_ondisk == 0)) {
+ rc = dt_attr_get(env, o, la);
+ if (unlikely(rc))
+ return;
+ LASSERT(la->la_valid & LA_ATIME);
+ if (la->la_atime == 0)
+ la->la_atime = la->la_mtime;
+ fo->ofo_atime_ondisk = la->la_atime;
+ }
+ if (atime - fo->ofo_atime_ondisk < ofd->ofd_atime_diff)
+ return;
+
+ /* atime hasn't been updated too long, update it */
+ fo->ofo_atime_ondisk = atime;
+
+ th = ofd_trans_create(env, ofd);
+ if (IS_ERR(th)) {
+ CERROR("%s: cannot create transaction: rc = %d\n",
+ ofd_name(ofd), (int)PTR_ERR(th));
+ return;
+ }
+
+ la->la_valid = LA_ATIME;
+ rc = dt_declare_attr_set(env, o, la, th);
+ if (rc)
+ GOTO(out_tx, rc);
+
+ rc = dt_trans_start_local(env, ofd->ofd_osd , th);
+ if (rc) {
+ CERROR("%s: cannot start transaction: rc = %d\n",
+ ofd_name(ofd), rc);
+ GOTO(out_tx, rc);
+ }
+
+ ofd_read_lock(env, fo);
+ if (ofd_object_exists(fo)) {
+ la->la_atime = fo->ofo_atime_ondisk;
+ rc = dt_attr_set(env, o, la, th);
+ }
+
+ ofd_read_unlock(env, fo);
+
+out_tx:
+ ofd_trans_stop(env, ofd, th, rc);
+}
+
/**
* Prepare buffers for read request processing.
*
int i, j, rc, tot_bytes = 0;
enum dt_bufs_type dbt = DT_BUFS_TYPE_READ;
int maxlnb = *nr_local;
+ __u64 begin, end;
+ ktime_t kstart = ktime_get();
ENTRY;
LASSERT(env != NULL);
RETURN(PTR_ERR(fo));
LASSERT(fo != NULL);
+ ofd_info(env)->fti_obj = fo;
+
+ if (oa->o_valid & OBD_MD_FLATIME)
+ ofd_handle_atime(env, ofd, fo, oa->o_atime);
+
ofd_read_lock(env, fo);
if (!ofd_object_exists(fo))
GOTO(unlock, rc = -ENOENT);
if (ptlrpc_connection_is_local(exp->exp_connection))
dbt |= DT_BUFS_TYPE_LOCAL;
+ begin = -1;
+ end = 0;
+
for (*nr_local = 0, i = 0, j = 0; i < niocount; i++) {
+ begin = min_t(__u64, begin, rnb[i].rnb_offset);
+ end = max_t(__u64, end, rnb[i].rnb_offset + rnb[i].rnb_len);
if (OBD_FAIL_CHECK(OBD_FAIL_OST_2BIG_NIOBUF))
rnb[i].rnb_len = 100 * 1024 * 1024;
}
LASSERT(*nr_local > 0 && *nr_local <= PTLRPC_MAX_BRW_PAGES);
- rc = dt_attr_get(env, ofd_object_child(fo), la);
- if (unlikely(rc))
- GOTO(buf_put, rc);
-
rc = dt_read_prep(env, ofd_object_child(fo), lnb, *nr_local);
if (unlikely(rc))
GOTO(buf_put, rc);
- ofd_counter_incr(exp, LPROC_OFD_STATS_READ, jobid, tot_bytes);
+ ofd_access(env, ofd,
+ &(struct lu_fid) {
+ .f_seq = oa->o_parent_seq,
+ .f_oid = oa->o_parent_oid,
+ .f_ver = oa->o_stripe_idx,
+ },
+ begin, end,
+ tot_bytes,
+ niocount,
+ READ);
+
+ ofd_counter_incr(exp, LPROC_OFD_STATS_READ_BYTES, jobid, tot_bytes);
+ ofd_counter_incr(exp, LPROC_OFD_STATS_READ, jobid,
+ ktime_us_delta(ktime_get(), kstart));
RETURN(0);
buf_put:
int i, j, k, rc = 0, tot_bytes = 0;
enum dt_bufs_type dbt = DT_BUFS_TYPE_WRITE;
int maxlnb = *nr_local;
+ __u64 begin, end;
+ ktime_t kstart = ktime_get();
+ struct range_lock *range = &ofd_info(env)->fti_write_range;
ENTRY;
LASSERT(env != NULL);
LASSERT(objcount == 1);
if (unlikely(exp->exp_obd->obd_recovering)) {
- u64 seq = fid_seq(fid);
- u64 oid = fid_oid(fid);
+ u64 seq = ostid_seq(&oa->o_oi);
+ u64 oid = ostid_id(&oa->o_oi);
struct ofd_seq *oseq;
oseq = ofd_seq_load(env, ofd, seq);
GOTO(out, rc = PTR_ERR(fo));
LASSERT(fo != NULL);
+ ofd_info(env)->fti_obj = fo;
+
ofd_read_lock(env, fo);
if (!ofd_object_exists(fo)) {
CERROR("%s: BRW to missing obj "DOSTID"\n",
if (ptlrpc_connection_is_local(exp->exp_connection))
dbt |= DT_BUFS_TYPE_LOCAL;
+ begin = -1;
+ end = 0;
+
/* parse remote buffers to local buffers and prepare the latter */
for (*nr_local = 0, i = 0, j = 0; i < obj->ioo_bufcnt; i++) {
+ begin = min_t(__u64, begin, rnb[i].rnb_offset);
+ end = max_t(__u64, end, rnb[i].rnb_offset + rnb[i].rnb_len);
+
if (OBD_FAIL_CHECK(OBD_FAIL_OST_2BIG_NIOBUF))
rnb[i].rnb_len += PAGE_SIZE;
rc = dt_bufs_get(env, ofd_object_child(fo),
GOTO(err, rc);
ofd_read_unlock(env, fo);
- ofd_counter_incr(exp, LPROC_OFD_STATS_WRITE, jobid, tot_bytes);
+
+ ofd_access(env, ofd,
+ &(struct lu_fid) {
+ .f_seq = oa->o_parent_seq,
+ .f_oid = oa->o_parent_oid,
+ .f_ver = oa->o_stripe_idx,
+ },
+ begin, end,
+ tot_bytes,
+ obj->ioo_bufcnt,
+ WRITE);
+
+ /*
+ * Reordering precautions: make sure that request processing that
+ * was able to receive its bulk data should not get reordered with
+ * overlapping BRW requests, e.g.
+ * 1) BRW1 sent, bulk data received, but disk I/O delayed
+ * 2) BRW1 resent and fully processed
+ * 3) the page was unlocked on the client and its writeback bit reset
+ * 4) BRW2 sent and fully processed
+ * 5) BRW1 processing wakes up and writes stale data to disk
+ * If on step 1 bulk data was not received, client resend will invalidate
+ * its bulk descriptor and the RPC will be dropped due to failed bulk
+ * transfer, which is just fine.
+ */
+ range_lock_init(range,
+ rnb[0].rnb_offset,
+ rnb[obj->ioo_bufcnt - 1].rnb_offset +
+ rnb[obj->ioo_bufcnt - 1].rnb_len - 1);
+ range_lock(&fo->ofo_write_tree, range);
+ ofd_info(env)->fti_range_locked = 1;
+
+ ofd_counter_incr(exp, LPROC_OFD_STATS_WRITE_BYTES, jobid, tot_bytes);
+ ofd_counter_incr(exp, LPROC_OFD_STATS_WRITE, jobid,
+ ktime_us_delta(ktime_get(), kstart));
RETURN(0);
err:
dt_bufs_put(env, ofd_object_child(fo), lnb, *nr_local);
int rc = 0;
if (*nr_local > PTLRPC_MAX_BRW_PAGES) {
- CERROR("%s: bulk has too many pages %d, which exceeds the"
- "maximum pages per RPC of %d\n",
+ CERROR("%s: bulk has too many pages %d, which exceeds the maximum pages per RPC of %d\n",
exp->exp_obd->obd_name, *nr_local, PTLRPC_MAX_BRW_PAGES);
RETURN(-EPROTO);
}
rc = ofd_preprw_read(env, exp, ofd, fid, &info->fti_attr, oa,
obj->ioo_bufcnt, rnb, nr_local, lnb,
jobid);
- obdo_from_la(oa, &info->fti_attr, LA_ATIME);
} else {
CERROR("%s: wrong cmd %d received!\n",
exp->exp_obd->obd_name, cmd);
LASSERT(niocount > 0);
- fo = ofd_object_find(env, ofd, fid);
- if (IS_ERR(fo))
- RETURN(PTR_ERR(fo));
+ fo = ofd_info(env)->fti_obj;
LASSERT(fo != NULL);
LASSERT(ofd_object_exists(fo));
dt_bufs_put(env, ofd_object_child(fo), lnb, niocount);
ofd_read_unlock(env, fo);
ofd_object_put(env, fo);
- /* second put is pair to object_get in ofd_preprw_read */
- ofd_object_put(env, fo);
RETURN(0);
}
int rc = 0;
int rc2 = 0;
int retries = 0;
- int i;
+ int i, restart = 0;
bool soft_sync = false;
bool cb_registered = false;
bool fake_write = false;
+ struct range_lock *range = &ofd_info(env)->fti_write_range;
ENTRY;
LASSERT(objcount == 1);
- fo = ofd_object_find(env, ofd, fid);
+ fo = ofd_info(env)->fti_obj;
LASSERT(fo != NULL);
o = ofd_object_child(fo);
GOTO(out_stop, rc);
}
+ /* don't update atime on disk if it is older */
+ if (la->la_valid & LA_ATIME && la->la_atime <= fo->ofo_atime_ondisk)
+ la->la_valid &= ~LA_ATIME;
+
if (la->la_valid) {
/* update [mac]time if needed */
rc = dt_declare_attr_set(env, o, la, th);
if (!ofd_object_exists(fo))
GOTO(out_unlock, rc = -ENOENT);
- if (likely(!fake_write)) {
- rc = dt_write_commit(env, o, lnb, niocount, th);
- if (rc)
- GOTO(out_unlock, rc);
- }
-
/* Don't update timestamps if this write is older than a
* setattr which modifies the timestamps. b=10150 */
if (la->la_valid && tgt_fmd_check(exp, fid, info->fti_xid)) {
rc = dt_attr_set(env, o, la, th);
if (rc)
GOTO(out_unlock, rc);
+ if (la->la_valid & LA_ATIME)
+ fo->ofo_atime_ondisk = la->la_atime;
+ }
+
+ if (likely(!fake_write)) {
+ OBD_FAIL_TIMEOUT_ORSET(OBD_FAIL_OST_WR_ATTR_DELAY,
+ OBD_FAIL_ONCE, cfs_fail_val);
+ rc = dt_write_commit(env, o, lnb, niocount, th, oa->o_size);
+ if (rc) {
+ restart = th->th_restart_tran;
+ GOTO(out_unlock, rc);
+ }
}
/* get attr to return */
granted = 0;
}
- rc2 = ofd_trans_stop(env, ofd, th, rc);
+ rc2 = ofd_trans_stop(env, ofd, th, restart ? 0 : rc);
if (!rc)
rc = rc2;
if (rc == -ENOSPC && retries++ < 3) {
goto retry;
}
+ if (restart) {
+ retries++;
+ restart = 0;
+ if (retries % 10000 == 0)
+ CERROR("%s: restart IO write too many times: %d\n",
+ ofd_name(ofd), retries);
+ CDEBUG(D_INODE, "retry transaction, retries:%d\n",
+ retries);
+ goto retry;
+ }
if (!soft_sync)
/* reset fed_soft_sync_count upon non-SOFT_SYNC RPC */
atomic_set(&fed->fed_soft_sync_count, 0);
dt_commit_async(env, ofd->ofd_osd);
out:
+ if (info->fti_range_locked) {
+ range_unlock(&fo->ofo_write_tree, range);
+ info->fti_range_locked = 0;
+ }
dt_bufs_put(env, o, lnb, niocount);
ofd_object_put(env, fo);
- /* second put is pair to object_get in ofd_preprw_write */
- ofd_object_put(env, fo);
if (granted > 0)
tgt_grant_commit(exp, granted, old_rc);
RETURN(rc);
struct ofd_thread_info *info = ofd_info(env);
struct ofd_device *ofd = ofd_exp(exp);
const struct lu_fid *fid = &oa->o_oi.oi_fid;
+ struct ldlm_namespace *ns = ofd->ofd_namespace;
+ struct ldlm_resource *rs = NULL;
__u64 valid;
int rc = 0;
if (cmd == OBD_BRW_WRITE) {
struct lu_nodemap *nodemap;
+ __u32 mapped_uid, mapped_gid;
+
+ nodemap = nodemap_get_from_exp(exp);
+ mapped_uid = nodemap_map_id(nodemap, NODEMAP_UID,
+ NODEMAP_FS_TO_CLIENT,
+ oa->o_uid);
+ mapped_gid = nodemap_map_id(nodemap, NODEMAP_GID,
+ NODEMAP_FS_TO_CLIENT,
+ oa->o_gid);
+
+ if (!IS_ERR_OR_NULL(nodemap)) {
+ /* do not bypass quota enforcement if squashed uid */
+ if (unlikely(mapped_uid == nodemap->nm_squash_uid)) {
+ int idx;
+
+ for (idx = 0; idx < npages; idx++)
+ lnb[idx].lnb_flags &= ~OBD_BRW_NOQUOTA;
+ }
+ nodemap_putref(nodemap);
+ }
valid = OBD_MD_FLUID | OBD_MD_FLGID | OBD_MD_FLPROJID |
OBD_MD_FLATIME | OBD_MD_FLMTIME | OBD_MD_FLCTIME;
oa->o_valid |= OBD_MD_FLALLQUOTA;
}
- /* Convert back to client IDs. LU-9671.
- * nodemap_get_from_exp() may fail due to nodemap deactivated,
- * server ID will be returned back to client in that case. */
- nodemap = nodemap_get_from_exp(exp);
- if (nodemap != NULL && !IS_ERR(nodemap)) {
- oa->o_uid = nodemap_map_id(nodemap, NODEMAP_UID,
- NODEMAP_FS_TO_CLIENT,
- oa->o_uid);
- oa->o_gid = nodemap_map_id(nodemap, NODEMAP_GID,
- NODEMAP_FS_TO_CLIENT,
- oa->o_gid);
- nodemap_putref(nodemap);
- }
- } else if (cmd == OBD_BRW_READ) {
- struct ldlm_namespace *ns = ofd->ofd_namespace;
-
- /* If oa != NULL then ofd_preprw_read updated the inode
- * atime and we should update the lvb so that other glimpses
- * will also get the updated value. bug 5972 */
- if (oa && ns && ns->ns_lvbo && ns->ns_lvbo->lvbo_update) {
- struct ldlm_resource *rs = NULL;
-
+ /**
+ * Update LVB after writing finish for server lock, see
+ * comments in ldlm_lock_decref_internal(), If this is a
+ * local lock on a server namespace and this was the last
+ * reference, lock will be destroyed directly thus there
+ * is no chance for ldlm_request_cancel() to update lvb.
+ */
+ if (rc == 0 && (rnb[0].rnb_flags & OBD_BRW_SRVLOCK)) {
ost_fid_build_resid(fid, &info->fti_resid);
rs = ldlm_resource_get(ns, NULL, &info->fti_resid,
LDLM_EXTENT, 0);
ldlm_resource_putref(rs);
}
}
+
+ /* Convert back to client IDs. LU-9671.
+ * nodemap_get_from_exp() may fail due to nodemap deactivated,
+ * server ID will be returned back to client in that case. */
+ oa->o_uid = mapped_uid;
+ oa->o_gid = mapped_gid;
+ } else if (cmd == OBD_BRW_READ) {
rc = ofd_commitrw_read(env, ofd, fid, objcount,
npages, lnb);
if (old_rc)
git://git.whamcloud.com / fs / lustre-release.git / blobdiff