Whamcloud - gitweb
git://git.whamcloud.com
/
fs
/
lustre-release.git
/ blobdiff
commit
grep
author
committer
pickaxe
?
search:
re
summary
|
shortlog
|
log
|
commit
|
commitdiff
|
tree
raw
| inline |
side by side
b=23127 osc lock granted fix
[fs/lustre-release.git]
/
lustre
/
osc
/
osc_request.c
diff --git
a/lustre/osc/osc_request.c
b/lustre/osc/osc_request.c
index
22bf7d9
..
2ba487b
100644
(file)
--- a/
lustre/osc/osc_request.c
+++ b/
lustre/osc/osc_request.c
@@
-1020,6
+1020,11
@@
static int osc_should_shrink_grant(struct client_obd *client)
{
cfs_time_t time = cfs_time_current();
cfs_time_t next_shrink = client->cl_next_shrink_grant;
+
+ if ((client->cl_import->imp_connect_data.ocd_connect_flags &
+ OBD_CONNECT_GRANT_SHRINK) == 0)
+ return 0;
+
if (cfs_time_aftereq(time, next_shrink - 5 * CFS_TICK)) {
if (client->cl_import->imp_state == LUSTRE_IMP_FULL &&
client->cl_avail_grant > GRANT_SHRINK_LIMIT)
@@
-1082,11
+1087,21
@@
static void osc_init_grant(struct client_obd *cli, struct obd_connect_data *ocd)
cli->cl_avail_grant = ocd->ocd_grant;
else
cli->cl_avail_grant = ocd->ocd_grant - cli->cl_dirty;
+
+ if (cli->cl_avail_grant < 0) {
+ CWARN("%s: available grant < 0, the OSS is probably not running"
+ " with patch from bug20278 (%ld) \n",
+ cli->cl_import->imp_obd->obd_name, cli->cl_avail_grant);
+ /* workaround for 1.6 servers which do not have
+ * the patch from bug20278 */
+ cli->cl_avail_grant = ocd->ocd_grant;
+ }
+
client_obd_list_unlock(&cli->cl_loi_list_lock);
- CDEBUG(D_CACHE, "setting cl_avail_grant: %ld cl_lost_grant: %ld \n",
+ CDEBUG(D_CACHE, "%s, setting cl_avail_grant: %ld cl_lost_grant: %ld \n",
+ cli->cl_import->imp_obd->obd_name,
cli->cl_avail_grant, cli->cl_lost_grant);
- LASSERT(cli->cl_avail_grant >= 0);
if (ocd->ocd_connect_flags & OBD_CONNECT_GRANT_SHRINK &&
cfs_list_empty(&cli->cl_grant_shrink_list))
@@
-1172,7
+1187,7
@@
static inline int can_merge_pages(struct brw_page *p1, struct brw_page *p2)
{
if (p1->flag != p2->flag) {
unsigned mask = ~(OBD_BRW_FROM_GRANT|
- OBD_BRW_NOCACHE|OBD_BRW_SYNC);
+ OBD_BRW_NOCACHE|OBD_BRW_SYNC
|OBD_BRW_ASYNC
);
/* warn if we try to combine flags that we don't know to be
* safe to combine */
@@
-1422,6
+1437,10
@@
static int check_write_checksum(struct obdo *oa, const lnet_process_id_t *peer,
return 0;
}
+ /* If this is mmaped file - it can be changed at any time */
+ if (oa->o_valid & OBD_MD_FLFLAGS && oa->o_flags & OBD_FL_MMAP)
+ return 1;
+
if (oa->o_valid & OBD_MD_FLFLAGS)
cksum_type = cksum_type_unpack(oa->o_flags);
else
@@
-1495,14
+1514,14
@@
static int osc_brw_fini_request(struct ptlrpc_request *req, int rc)
}
#endif
+ osc_update_grant(cli, body);
+
if (rc < 0)
RETURN(rc);
if (aa->aa_oa->o_valid & OBD_MD_FLCKSUM)
client_cksum = aa->aa_oa->o_cksum; /* save for later */
- osc_update_grant(cli, body);
-
if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE) {
if (rc > 0) {
CERROR("Unexpected +ve rc %d\n", rc);
@@
-1681,7
+1700,7
@@
int osc_brw_redo_request(struct ptlrpc_request *request,
ENTRY;
if (!osc_should_resend(aa->aa_resends, aa->aa_cli)) {
- CERROR("too many resen
d
retries, returning error\n");
+ CERROR("too many resen
t
retries, returning error\n");
RETURN(-EIO);
}
@@
-1938,7
+1957,7
@@
static int lop_makes_rpc(struct client_obd *cli, struct loi_oap_pages *lop,
if (cmd & OBD_BRW_WRITE) {
/* trigger a write rpc stream as long as there are dirtiers
* waiting for space. as they're waiting, they're not going to
- * create more pages to coal
l
esce with what's waiting.. */
+ * create more pages to coalesce with what's waiting.. */
if (!cfs_list_empty(&cli->cl_cache_waiters)) {
CDEBUG(D_CACHE, "cache waiters forcing RPC\n");
RETURN(1);
@@
-2162,9
+2181,20
@@
static int brw_interpret(const struct lu_env *env,
rc = osc_brw_fini_request(req, rc);
CDEBUG(D_INODE, "request %p aa %p rc %d\n", req, aa, rc);
if (osc_recoverable_error(rc)) {
- rc = osc_brw_redo_request(req, aa);
- if (rc == 0)
- RETURN(0);
+ /* Only retry once for mmaped files since the mmaped page
+ * might be modified at anytime. We have to retry at least
+ * once in case there WAS really a corruption of the page
+ * on the network, that was not caused by mmap() modifying
+ * the page. Bug11742 */
+ if ((rc == -EAGAIN) && (aa->aa_resends > 0) &&
+ aa->aa_oa->o_valid & OBD_MD_FLFLAGS &&
+ aa->aa_oa->o_flags & OBD_FL_MMAP) {
+ rc = 0;
+ } else {
+ rc = osc_brw_redo_request(req, aa);
+ if (rc == 0)
+ RETURN(0);
+ }
}
if (aa->aa_ocapa) {
@@
-2196,7
+2226,7
@@
static int brw_interpret(const struct lu_env *env,
}
OBDO_FREE(aa->aa_oa);
} else { /* from async_internal() */
-
i
nt i;
+
obd_cou
nt i;
for (i = 0; i < aa->aa_page_count; i++)
osc_release_write_grant(aa->aa_cli, aa->aa_ppga[i], 1);
@@
-2209,6
+2239,7
@@
static int brw_interpret(const struct lu_env *env,
if (!async)
cl_req_completion(env, aa->aa_clerq, rc);
osc_release_ppga(aa->aa_ppga, aa->aa_page_count);
+
RETURN(rc);
}
@@
-2230,11
+2261,14
@@
static struct ptlrpc_request *osc_build_req(const struct lu_env *env,
enum cl_req_type crt = (cmd & OBD_BRW_WRITE) ? CRT_WRITE : CRT_READ;
struct ldlm_lock *lock = NULL;
struct cl_req_attr crattr;
- int i, rc;
+ int i, rc
, mpflag = 0
;
ENTRY;
LASSERT(!cfs_list_empty(rpc_list));
+ if (cmd & OBD_BRW_MEMALLOC)
+ mpflag = cfs_memory_pressure_get_and_set();
+
memset(&crattr, 0, sizeof crattr);
OBD_ALLOC(pga, sizeof(*pga) * page_count);
if (pga == NULL)
@@
-2290,6
+2324,9
@@
static struct ptlrpc_request *osc_build_req(const struct lu_env *env,
GOTO(out, req = ERR_PTR(rc));
}
+ if (cmd & OBD_BRW_MEMALLOC)
+ req->rq_memalloc = 1;
+
/* Need to update the timestamps after the request is built in case
* we race with setattr (locally or in queue at OST). If OST gets
* later setattr before earlier BRW (as determined by the request xid),
@@
-2306,6
+2343,9
@@
static struct ptlrpc_request *osc_build_req(const struct lu_env *env,
CFS_INIT_LIST_HEAD(rpc_list);
aa->aa_clerq = clerq;
out:
+ if (cmd & OBD_BRW_MEMALLOC)
+ cfs_memory_pressure_restore(mpflag);
+
capa_put(crattr.cra_capa);
if (IS_ERR(req)) {
if (oa)
@@
-2340,8
+2380,9
@@
out:
* \param cmd OBD_BRW_* macroses
* \param lop pending pages
*
- * \return zero if pages successfully add to send queue.
- * \return not zere if error occurring.
+ * \return zero if no page added to send queue.
+ * \return 1 if pages successfully added to send queue.
+ * \return negative on errors.
*/
static int
osc_send_oap_rpc(const struct lu_env *env, struct client_obd *cli,
@@
-2357,7
+2398,7
@@
osc_send_oap_rpc(const struct lu_env *env, struct client_obd *cli,
CFS_LIST_HEAD(tmp_list);
unsigned int ending_offset;
unsigned starting_offset = 0;
- int srvlock = 0;
+ int srvlock = 0
, mem_tight = 0
;
struct cl_object *clob = NULL;
ENTRY;
@@
-2409,7
+2450,7
@@
osc_send_oap_rpc(const struct lu_env *env, struct client_obd *cli,
* until completion unlocks it. commit_write submits a page
* as not ready because its unlock will happen unconditionally
* as the call returns. if we race with commit_write giving
- * us that page we don
t'
want to create a hole in the page
+ * us that page we don
't
want to create a hole in the page
* stream, so we stop and leave the rpc to be fired by
* another dirtier or kupdated interval (the not ready page
* will still be on the dirty list). we could call in
@@
-2501,6
+2542,8
@@
osc_send_oap_rpc(const struct lu_env *env, struct client_obd *cli,
/* now put the page back in our accounting */
cfs_list_add_tail(&oap->oap_rpc_item, &rpc_list);
+ if (oap->oap_brw_flags & OBD_BRW_MEMALLOC)
+ mem_tight = 1;
if (page_count == 0)
srvlock = !!(oap->oap_brw_flags & OBD_BRW_SRVLOCK);
if (++page_count >= cli->cl_max_pages_per_rpc)
@@
-2536,7
+2579,8
@@
osc_send_oap_rpc(const struct lu_env *env, struct client_obd *cli,
RETURN(0);
}
- req = osc_build_req(env, cli, &rpc_list, page_count, cmd);
+ req = osc_build_req(env, cli, &rpc_list, page_count,
+ mem_tight ? (cmd | OBD_BRW_MEMALLOC) : cmd);
if (IS_ERR(req)) {
LASSERT(cfs_list_empty(&rpc_list));
loi_list_maint(cli, loi);
@@
-2720,7
+2764,7
@@
void osc_check_rpcs(const struct lu_env *env, struct client_obd *cli)
race_counter++;
}
- /* attempt some inter-object balancing by issu
e
ing rpcs
+ /* attempt some inter-object balancing by issuing rpcs
* for each object in turn */
if (!cfs_list_empty(&loi->loi_hp_ready_item))
cfs_list_del_init(&loi->loi_hp_ready_item);
@@
-2946,7
+2990,7
@@
int osc_queue_async_io(const struct lu_env *env,
oap->oap_count = count;
oap->oap_brw_flags = brw_flags;
/* Give a hint to OST that requests are coming from kswapd - bug19529 */
- if (
lib
cfs_memory_pressure_get())
+ if (cfs_memory_pressure_get())
oap->oap_brw_flags |= OBD_BRW_MEMALLOC;
cfs_spin_lock(&oap->oap_lock);
oap->oap_async_flags = async_flags;
@@
-3183,6
+3227,9
@@
static int osc_enqueue_interpret(const struct lu_env *env,
* osc_enqueue_fini(). */
ldlm_lock_addref(&handle, mode);
+ /* Let CP AST to grant the lock first. */
+ OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_ENQ_RACE, 1);
+
/* Complete obtaining the lock procedure. */
rc = ldlm_cli_enqueue_fini(aa->oa_exp, req, aa->oa_ei->ei_type, 1,
mode, aa->oa_flags, aa->oa_lvb,