*
* You should have received a copy of the GNU General Public License
* version 2 along with this program; If not, see
- * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
- *
- * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
- * CA 95054 USA or visit www.sun.com if you need additional information or
- * have any questions.
+ * http://www.gnu.org/licenses/gpl-2.0.html
*
* GPL HEADER END
*/
* overly wide locks.
*/
static void ldlm_extent_internal_policy_fixup(struct ldlm_lock *req,
- struct ldlm_extent *new_ex,
- int conflicting)
+ struct ldlm_extent *new_ex,
+ int conflicting)
{
- ldlm_mode_t req_mode = req->l_req_mode;
- __u64 req_start = req->l_req_extent.start;
- __u64 req_end = req->l_req_extent.end;
- __u64 req_align, mask;
+ enum ldlm_mode req_mode = req->l_req_mode;
+ __u64 req_start = req->l_req_extent.start;
+ __u64 req_end = req->l_req_extent.end;
+ __u64 req_align, mask;
if (conflicting > 32 && (req_mode == LCK_PW || req_mode == LCK_CW)) {
if (req_end < req_start + LDLM_MAX_GROWN_EXTENT)
* the client requested. Also we need to make sure it's also server
* page size aligned otherwise a server page can be covered by two
* write locks. */
- mask = PAGE_CACHE_SIZE;
+ mask = PAGE_SIZE;
req_align = (req_end + 1) | req_start;
if (req_align != 0 && (req_align & (mask - 1)) == 0) {
while ((req_align & mask) == 0)
new_ex->start = ((new_ex->start - 1) | mask) + 1;
new_ex->end = ((new_ex->end + 1) & ~mask) - 1;
LASSERTF(new_ex->start <= req_start,
- "mask "LPX64" grant start "LPU64" req start "LPU64"\n",
+ "mask %#llx grant start %llu req start %llu\n",
mask, new_ex->start, req_start);
LASSERTF(new_ex->end >= req_end,
- "mask "LPX64" grant end "LPU64" req end "LPU64"\n",
+ "mask %#llx grant end %llu req end %llu\n",
mask, new_ex->end, req_end);
}
static void ldlm_extent_internal_policy_granted(struct ldlm_lock *req,
struct ldlm_extent *new_ex)
{
- struct ldlm_resource *res = req->l_resource;
- ldlm_mode_t req_mode = req->l_req_mode;
- __u64 req_start = req->l_req_extent.start;
- __u64 req_end = req->l_req_extent.end;
- struct ldlm_interval_tree *tree;
- struct interval_node_extent limiter = { new_ex->start, new_ex->end };
- int conflicting = 0;
- int idx;
- ENTRY;
+ struct ldlm_resource *res = req->l_resource;
+ enum ldlm_mode req_mode = req->l_req_mode;
+ __u64 req_start = req->l_req_extent.start;
+ __u64 req_end = req->l_req_extent.end;
+ struct ldlm_interval_tree *tree;
+ struct interval_node_extent limiter = { new_ex->start, new_ex->end };
+ int conflicting = 0;
+ int idx;
+ ENTRY;
- lockmode_verify(req_mode);
+ lockmode_verify(req_mode);
/* Using interval tree to handle the LDLM extent granted locks. */
for (idx = 0; idx < LCK_MODE_NUM; idx++) {
struct ldlm_extent *new_ex)
{
struct ldlm_resource *res = req->l_resource;
- ldlm_mode_t req_mode = req->l_req_mode;
+ enum ldlm_mode req_mode = req->l_req_mode;
__u64 req_start = req->l_req_extent.start;
__u64 req_end = req->l_req_extent.end;
struct ldlm_lock *lock;
static int ldlm_check_contention(struct ldlm_lock *lock, int contended_locks)
{
- struct ldlm_resource *res = lock->l_resource;
- cfs_time_t now = cfs_time_current();
+ struct ldlm_resource *res = lock->l_resource;
+ cfs_time_t now = cfs_time_current();
- if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_SET_CONTENTION))
- return 1;
+ if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_SET_CONTENTION))
+ return 1;
- CDEBUG(D_DLMTRACE, "contended locks = %d\n", contended_locks);
- if (contended_locks > ldlm_res_to_ns(res)->ns_contended_locks)
- res->lr_contention_time = now;
- return cfs_time_before(now, cfs_time_add(res->lr_contention_time,
- cfs_time_seconds(ldlm_res_to_ns(res)->ns_contention_time)));
+ CDEBUG(D_DLMTRACE, "contended locks = %d\n", contended_locks);
+ if (contended_locks > ldlm_res_to_ns(res)->ns_contended_locks)
+ res->lr_contention_time = now;
+ return cfs_time_before(now, cfs_time_add(res->lr_contention_time,
+ cfs_time_seconds(ldlm_res_to_ns(res)->ns_contention_time)));
}
struct ldlm_extent_compat_args {
struct list_head *work_list;
- struct ldlm_lock *lock;
- ldlm_mode_t mode;
- int *locks;
- int *compat;
+ struct ldlm_lock *lock;
+ enum ldlm_mode mode;
+ int *locks;
+ int *compat;
};
static enum interval_iter ldlm_extent_compat_cb(struct interval_node *n,
- void *data)
+ void *data)
{
- struct ldlm_extent_compat_args *priv = data;
- struct ldlm_interval *node = to_ldlm_interval(n);
- struct ldlm_extent *extent;
+ struct ldlm_extent_compat_args *priv = data;
+ struct ldlm_interval *node = to_ldlm_interval(n);
+ struct ldlm_extent *extent;
struct list_head *work_list = priv->work_list;
- struct ldlm_lock *lock, *enq = priv->lock;
- ldlm_mode_t mode = priv->mode;
- int count = 0;
- ENTRY;
+ struct ldlm_lock *lock, *enq = priv->lock;
+ enum ldlm_mode mode = priv->mode;
+ int count = 0;
+ ENTRY;
LASSERT(!list_empty(&node->li_group));
*/
static int
ldlm_extent_compat_queue(struct list_head *queue, struct ldlm_lock *req,
- __u64 *flags, ldlm_error_t *err,
+ __u64 *flags, enum ldlm_error *err,
struct list_head *work_list, int *contended_locks)
{
struct ldlm_resource *res = req->l_resource;
- ldlm_mode_t req_mode = req->l_req_mode;
+ enum ldlm_mode req_mode = req->l_req_mode;
__u64 req_start = req->l_req_extent.start;
__u64 req_end = req->l_req_extent.end;
struct ldlm_lock *lock;
}
/**
+ * This function refresh eviction timer for cancelled lock.
+ * \param[in] lock ldlm lock for refresh
+ * \param[in] arg ldlm prolong arguments, timeout, export, extent
+ * and counter are used
+ */
+void ldlm_lock_prolong_one(struct ldlm_lock *lock,
+ struct ldlm_prolong_args *arg)
+{
+ int timeout;
+
+ if (arg->lpa_export != lock->l_export ||
+ lock->l_flags & LDLM_FL_DESTROYED)
+ /* ignore unrelated locks */
+ return;
+
+ arg->lpa_locks_cnt++;
+
+ if (!(lock->l_flags & LDLM_FL_AST_SENT))
+ /* ignore locks not being cancelled */
+ return;
+
+ /* We are in the middle of the process - BL AST is sent, CANCEL
+ * is ahead. Take half of BL AT + IO AT process time.
+ */
+ timeout = arg->lpa_timeout + (ldlm_bl_timeout(lock) >> 1);
+
+ LDLM_DEBUG(lock, "refreshed to %ds.\n", timeout);
+
+ arg->lpa_blocks_cnt++;
+
+ /* OK. this is a possible lock the user holds doing I/O
+ * let's refresh eviction timer for it.
+ */
+ ldlm_refresh_waiting_lock(lock, timeout);
+}
+EXPORT_SYMBOL(ldlm_lock_prolong_one);
+
+static enum interval_iter ldlm_resource_prolong_cb(struct interval_node *n,
+ void *data)
+{
+ struct ldlm_prolong_args *arg = data;
+ struct ldlm_interval *node = to_ldlm_interval(n);
+ struct ldlm_lock *lock;
+
+ ENTRY;
+
+ LASSERT(!list_empty(&node->li_group));
+
+ list_for_each_entry(lock, &node->li_group, l_sl_policy) {
+ ldlm_lock_prolong_one(lock, arg);
+ }
+
+ RETURN(INTERVAL_ITER_CONT);
+}
+
+/**
+ * Walk through granted tree and prolong locks if they overlaps extent.
+ *
+ * \param[in] arg prolong args
+ */
+void ldlm_resource_prolong(struct ldlm_prolong_args *arg)
+{
+ struct ldlm_interval_tree *tree;
+ struct ldlm_resource *res;
+ struct interval_node_extent ex = { .start = arg->lpa_extent.start,
+ .end = arg->lpa_extent.end };
+ int idx;
+
+ ENTRY;
+
+ res = ldlm_resource_get(arg->lpa_export->exp_obd->obd_namespace, NULL,
+ &arg->lpa_resid, LDLM_EXTENT, 0);
+ if (IS_ERR(res)) {
+ CDEBUG(D_DLMTRACE, "Failed to get resource for resid %llu/%llu\n",
+ arg->lpa_resid.name[0], arg->lpa_resid.name[1]);
+ RETURN_EXIT;
+ }
+
+ lock_res(res);
+ for (idx = 0; idx < LCK_MODE_NUM; idx++) {
+ tree = &res->lr_itree[idx];
+ if (tree->lit_root == NULL) /* empty tree, skipped */
+ continue;
+
+ /* There is no possibility to check for the groupID
+ * so all the group locks are considered as valid
+ * here, especially because the client is supposed
+ * to check it has such a lock before sending an RPC.
+ */
+ if (!(tree->lit_mode & arg->lpa_mode))
+ continue;
+
+ interval_search(tree->lit_root, &ex,
+ ldlm_resource_prolong_cb, arg);
+ }
+
+ unlock_res(res);
+ ldlm_resource_putref(res);
+
+ EXIT;
+}
+EXPORT_SYMBOL(ldlm_resource_prolong);
+
+
+/**
* Discard all AST work items from list.
*
* If for whatever reason we do not want to send ASTs to conflicting locks
* would be collected and ASTs sent.
*/
int ldlm_process_extent_lock(struct ldlm_lock *lock, __u64 *flags,
- int first_enq, ldlm_error_t *err,
+ int first_enq, enum ldlm_error *err,
struct list_head *work_list)
{
struct ldlm_resource *res = lock->l_resource;
}
#endif /* HAVE_SERVER_SUPPORT */
+struct ldlm_kms_shift_args {
+ __u64 old_kms;
+ __u64 kms;
+ bool complete;
+};
+
+/* Callback for interval_iterate functions, used by ldlm_extent_shift_Kms */
+static enum interval_iter ldlm_kms_shift_cb(struct interval_node *n,
+ void *args)
+{
+ struct ldlm_kms_shift_args *arg = args;
+ struct ldlm_interval *node = to_ldlm_interval(n);
+ struct ldlm_lock *tmplock;
+ struct ldlm_lock *lock = NULL;
+
+ ENTRY;
+
+ /* Since all locks in an interval have the same extent, we can just
+ * use the first lock without kms_ignore set. */
+ list_for_each_entry(tmplock, &node->li_group, l_sl_policy) {
+ if (ldlm_is_kms_ignore(tmplock))
+ continue;
+
+ lock = tmplock;
+
+ break;
+ }
+
+ /* No locks in this interval without kms_ignore set */
+ if (!lock)
+ RETURN(INTERVAL_ITER_CONT);
+
+ /* If we find a lock with a greater or equal kms, we are not the
+ * highest lock (or we share that distinction with another lock), and
+ * don't need to update KMS. Return old_kms and stop looking. */
+ if (lock->l_policy_data.l_extent.end >= arg->old_kms) {
+ arg->kms = arg->old_kms;
+ arg->complete = true;
+ RETURN(INTERVAL_ITER_STOP);
+ }
+
+ if (lock->l_policy_data.l_extent.end + 1 > arg->kms)
+ arg->kms = lock->l_policy_data.l_extent.end + 1;
+
+ /* Since interval_iterate_reverse starts with the highest lock and
+ * works down, for PW locks, we only need to check if we should update
+ * the kms, then stop walking the tree. PR locks are not exclusive, so
+ * the highest start does not imply the highest end and we must
+ * continue. (Only one group lock is allowed per resource, so this is
+ * irrelevant for group locks.)*/
+ if (lock->l_granted_mode == LCK_PW)
+ RETURN(INTERVAL_ITER_STOP);
+ else
+ RETURN(INTERVAL_ITER_CONT);
+}
+
/* When a lock is cancelled by a client, the KMS may undergo change if this
- * is the "highest lock". This function returns the new KMS value.
+ * is the "highest lock". This function returns the new KMS value, updating
+ * it only if we were the highest lock.
+ *
* Caller must hold lr_lock already.
*
* NB: A lock on [x,y] protects a KMS of up to y + 1 bytes! */
__u64 ldlm_extent_shift_kms(struct ldlm_lock *lock, __u64 old_kms)
{
- struct ldlm_resource *res = lock->l_resource;
- struct list_head *tmp;
- struct ldlm_lock *lck;
- __u64 kms = 0;
- ENTRY;
+ struct ldlm_resource *res = lock->l_resource;
+ struct ldlm_interval_tree *tree;
+ struct ldlm_kms_shift_args args;
+ int idx = 0;
- /* don't let another thread in ldlm_extent_shift_kms race in
- * just after we finish and take our lock into account in its
- * calculation of the kms */
- ldlm_set_kms_ignore(lock);
+ ENTRY;
- list_for_each(tmp, &res->lr_granted) {
- lck = list_entry(tmp, struct ldlm_lock, l_res_link);
+ args.old_kms = old_kms;
+ args.kms = 0;
+ args.complete = false;
- if (ldlm_is_kms_ignore(lck))
- continue;
+ /* don't let another thread in ldlm_extent_shift_kms race in
+ * just after we finish and take our lock into account in its
+ * calculation of the kms */
+ ldlm_set_kms_ignore(lock);
- if (lck->l_policy_data.l_extent.end >= old_kms)
- RETURN(old_kms);
+ /* We iterate over the lock trees, looking for the largest kms smaller
+ * than the current one. */
+ for (idx = 0; idx < LCK_MODE_NUM; idx++) {
+ tree = &res->lr_itree[idx];
+
+ /* If our already known kms is >= than the highest 'end' in
+ * this tree, we don't need to check this tree, because
+ * the kms from a tree can be lower than in_max_high (due to
+ * kms_ignore), but it can never be higher. */
+ if (!tree->lit_root || args.kms >= tree->lit_root->in_max_high)
+ continue;
+
+ interval_iterate_reverse(tree->lit_root, ldlm_kms_shift_cb,
+ &args);
+
+ /* this tells us we're not the highest lock, so we don't need
+ * to check the remaining trees */
+ if (args.complete)
+ break;
+ }
- /* This extent _has_ to be smaller than old_kms (checked above)
- * so kms can only ever be smaller or the same as old_kms. */
- if (lck->l_policy_data.l_extent.end + 1 > kms)
- kms = lck->l_policy_data.l_extent.end + 1;
- }
- LASSERTF(kms <= old_kms, "kms "LPU64" old_kms "LPU64"\n", kms, old_kms);
+ LASSERTF(args.kms <= args.old_kms, "kms %llu old_kms %llu\n", args.kms,
+ args.old_kms);
- RETURN(kms);
+ RETURN(args.kms);
}
EXPORT_SYMBOL(ldlm_extent_shift_kms);
return list_empty(&n->li_group) ? n : NULL;
}
-static inline int lock_mode_to_index(ldlm_mode_t mode)
+static inline int ldlm_mode_to_index(enum ldlm_mode mode)
{
- int index;
-
- LASSERT(mode != 0);
- LASSERT(IS_PO2(mode));
- for (index = -1; mode; index++, mode >>= 1) ;
- LASSERT(index < LCK_MODE_NUM);
- return index;
+ int index;
+
+ LASSERT(mode != 0);
+ LASSERT(IS_PO2(mode));
+ for (index = -1; mode != 0; index++, mode >>= 1)
+ /* do nothing */;
+ LASSERT(index < LCK_MODE_NUM);
+ return index;
}
/** Add newly granted lock into interval tree for the resource. */
struct interval_node *found, **root;
struct ldlm_interval *node;
struct ldlm_extent *extent;
- int idx;
+ int idx, rc;
LASSERT(lock->l_granted_mode == lock->l_req_mode);
LASSERT(node != NULL);
LASSERT(!interval_is_intree(&node->li_node));
- idx = lock_mode_to_index(lock->l_granted_mode);
- LASSERT(lock->l_granted_mode == 1 << idx);
- LASSERT(lock->l_granted_mode == res->lr_itree[idx].lit_mode);
+ idx = ldlm_mode_to_index(lock->l_granted_mode);
+ LASSERT(lock->l_granted_mode == 1 << idx);
+ LASSERT(lock->l_granted_mode == res->lr_itree[idx].lit_mode);
/* node extent initialize */
extent = &lock->l_policy_data.l_extent;
- interval_set(&node->li_node, extent->start, extent->end);
+
+ rc = interval_set(&node->li_node, extent->start, extent->end);
+ LASSERT(!rc);
root = &res->lr_itree[idx].lit_root;
found = interval_insert(&node->li_node, root);
/* even though we use interval tree to manage the extent lock, we also
* add the locks into grant list, for debug purpose, .. */
ldlm_resource_add_lock(res, &res->lr_granted, lock);
+
+ if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_GRANT_CHECK)) {
+ struct ldlm_lock *lck;
+
+ list_for_each_entry_reverse(lck, &res->lr_granted,
+ l_res_link) {
+ if (lck == lock)
+ continue;
+ if (lockmode_compat(lck->l_granted_mode,
+ lock->l_granted_mode))
+ continue;
+ if (ldlm_extent_overlap(&lck->l_req_extent,
+ &lock->l_req_extent)) {
+ CDEBUG(D_ERROR, "granting conflicting lock %p "
+ "%p\n", lck, lock);
+ ldlm_resource_dump(D_ERROR, res);
+ LBUG();
+ }
+ }
+ }
}
/** Remove cancelled lock from resource interval tree. */
void ldlm_extent_unlink_lock(struct ldlm_lock *lock)
{
- struct ldlm_resource *res = lock->l_resource;
- struct ldlm_interval *node = lock->l_tree_node;
- struct ldlm_interval_tree *tree;
- int idx;
+ struct ldlm_resource *res = lock->l_resource;
+ struct ldlm_interval *node = lock->l_tree_node;
+ struct ldlm_interval_tree *tree;
+ int idx;
- if (!node || !interval_is_intree(&node->li_node)) /* duplicate unlink */
- return;
+ if (!node || !interval_is_intree(&node->li_node)) /* duplicate unlink */
+ return;
- idx = lock_mode_to_index(lock->l_granted_mode);
- LASSERT(lock->l_granted_mode == 1 << idx);
- tree = &res->lr_itree[idx];
+ idx = ldlm_mode_to_index(lock->l_granted_mode);
+ LASSERT(lock->l_granted_mode == 1 << idx);
+ tree = &res->lr_itree[idx];
- LASSERT(tree->lit_root != NULL); /* assure the tree is not null */
+ LASSERT(tree->lit_root != NULL); /* assure the tree is not null */
- tree->lit_size--;
- node = ldlm_interval_detach(lock);
- if (node) {
- interval_erase(&node->li_node, &tree->lit_root);
- ldlm_interval_free(node);
- }
+ tree->lit_size--;
+ node = ldlm_interval_detach(lock);
+ if (node) {
+ interval_erase(&node->li_node, &tree->lit_root);
+ ldlm_interval_free(node);
+ }
}
-void ldlm_extent_policy_wire_to_local(const ldlm_wire_policy_data_t *wpolicy,
- ldlm_policy_data_t *lpolicy)
+void ldlm_extent_policy_wire_to_local(const union ldlm_wire_policy_data *wpolicy,
+ union ldlm_policy_data *lpolicy)
{
- memset(lpolicy, 0, sizeof(*lpolicy));
- lpolicy->l_extent.start = wpolicy->l_extent.start;
- lpolicy->l_extent.end = wpolicy->l_extent.end;
- lpolicy->l_extent.gid = wpolicy->l_extent.gid;
+ lpolicy->l_extent.start = wpolicy->l_extent.start;
+ lpolicy->l_extent.end = wpolicy->l_extent.end;
+ lpolicy->l_extent.gid = wpolicy->l_extent.gid;
}
-void ldlm_extent_policy_local_to_wire(const ldlm_policy_data_t *lpolicy,
- ldlm_wire_policy_data_t *wpolicy)
+void ldlm_extent_policy_local_to_wire(const union ldlm_policy_data *lpolicy,
+ union ldlm_wire_policy_data *wpolicy)
{
- memset(wpolicy, 0, sizeof(*wpolicy));
- wpolicy->l_extent.start = lpolicy->l_extent.start;
- wpolicy->l_extent.end = lpolicy->l_extent.end;
- wpolicy->l_extent.gid = lpolicy->l_extent.gid;
+ memset(wpolicy, 0, sizeof(*wpolicy));
+ wpolicy->l_extent.start = lpolicy->l_extent.start;
+ wpolicy->l_extent.end = lpolicy->l_extent.end;
+ wpolicy->l_extent.gid = lpolicy->l_extent.gid;
}