#include "lov_cl_internal.h"
-/** \addtogroup lov lov @{ */
+/** \addtogroup lov
+ * @{
+ */
static struct cl_lock_closure *lov_closure_get(const struct lu_env *env,
struct cl_lock *parent);
+static int lov_lock_unuse(const struct lu_env *env,
+ const struct cl_lock_slice *slice);
/*****************************************************************************
*
* Lov lock operations.
* they are not initialized at all. As a temp fix, in this case,
* we still borrow the parent's env to call sublock operations.
*/
- if (!cl_object_same(io->ci_obj, parent->cll_descr.cld_obj)) {
+ if (!io || !cl_object_same(io->ci_obj, parent->cll_descr.cld_obj)) {
subenv->lse_env = env;
subenv->lse_io = io;
subenv->lse_sub = NULL;
} else {
- LASSERT(io != NULL);
sub = lov_sub_get(env, lio, lls->sub_stripe);
if (!IS_ERR(sub)) {
subenv->lse_env = sub->sub_env;
lck->lls_sub[idx].sub_lock = lsl;
lck->lls_nr_filled++;
LASSERT(lck->lls_nr_filled <= lck->lls_nr);
- list_add_tail(&link->lll_list, &lsl->lss_parents);
+ cfs_list_add_tail(&link->lll_list, &lsl->lss_parents);
link->lll_idx = idx;
link->lll_super = lck;
cl_lock_get(parent);
}
static int lov_sublock_lock(const struct lu_env *env,
+ struct lov_lock *lck,
struct lov_lock_sub *lls,
struct cl_lock_closure *closure,
struct lov_sublock_env **lsep)
{
- struct cl_lock *child;
- int result = 0;
+ struct lovsub_lock *sublock;
+ struct cl_lock *child;
+ int result = 0;
ENTRY;
- LASSERT(list_empty(&closure->clc_list));
+ LASSERT(cfs_list_empty(&closure->clc_list));
- child = lls->sub_lock->lss_cl.cls_lock;
+ sublock = lls->sub_lock;
+ child = sublock->lss_cl.cls_lock;
result = cl_lock_closure_build(env, child, closure);
if (result == 0) {
struct cl_lock *parent = closure->clc_origin;
LASSERT(cl_lock_is_mutexed(child));
- lls->sub_lock->lss_active = parent;
+ sublock->lss_active = parent;
- if (lsep) {
+ if (unlikely((child->cll_state == CLS_FREEING) ||
+ (child->cll_flags & CLF_CANCELLED))) {
+ struct lov_lock_link *link;
+ /*
+ * we could race with lock deletion which temporarily
+ * put the lock in freeing state, bug 19080.
+ */
+ LASSERT(!(lls->sub_flags & LSF_HELD));
+
+ link = lov_lock_link_find(env, lck, sublock);
+ LASSERT(link != NULL);
+ lov_lock_unlink(env, link, sublock);
+ lov_sublock_unlock(env, sublock, closure, NULL);
+ lck->lls_cancel_race = 1;
+ result = CLO_REPEAT;
+ } else if (lsep) {
struct lov_sublock_env *subenv;
subenv = lov_sublock_env_get(env, parent, lls);
if (IS_ERR(subenv)) {
- lov_sublock_unlock(env, lls->sub_lock,
+ lov_sublock_unlock(env, sublock,
closure, NULL);
result = PTR_ERR(subenv);
} else {
{
int result = 0;
int i;
- int j;
int nr;
- int stripe;
- int start_stripe;
obd_off start;
obd_off end;
obd_off file_start;
file_start = cl_offset(lov2cl(loo), parent->cll_descr.cld_start);
file_end = cl_offset(lov2cl(loo), parent->cll_descr.cld_end + 1) - 1;
- start_stripe = lov_stripe_number(r0->lo_lsm, file_start);
for (i = 0, nr = 0; i < r0->lo_nr; i++) {
/*
* XXX for wide striping smarter algorithm is desirable,
* breaking out of the loop, early.
*/
- stripe = (start_stripe + i) % r0->lo_nr;
- if (lov_stripe_intersects(r0->lo_lsm, stripe,
+ if (lov_stripe_intersects(r0->lo_lsm, i,
file_start, file_end, &start, &end))
nr++;
}
* create sub-locks. At this moment, no other thread can access
* top-lock.
*/
- for (j = 0, nr = 0; j < i; ++j) {
- stripe = (start_stripe + j) % r0->lo_nr;
- if (lov_stripe_intersects(r0->lo_lsm, stripe,
+ for (i = 0, nr = 0; i < r0->lo_nr; ++i) {
+ if (lov_stripe_intersects(r0->lo_lsm, i,
file_start, file_end, &start, &end)) {
struct cl_lock_descr *descr;
descr = &lck->lls_sub[nr].sub_descr;
LASSERT(descr->cld_obj == NULL);
- descr->cld_obj = lovsub2cl(r0->lo_sub[stripe]);
+ descr->cld_obj = lovsub2cl(r0->lo_sub[i]);
descr->cld_start = cl_index(descr->cld_obj, start);
descr->cld_end = cl_index(descr->cld_obj, end);
descr->cld_mode = parent->cll_descr.cld_mode;
descr->cld_gid = parent->cll_descr.cld_gid;
+ descr->cld_enq_flags = parent->cll_descr.cld_enq_flags;
/* XXX has no effect */
lck->lls_sub[nr].sub_got = *descr;
- lck->lls_sub[nr].sub_stripe = stripe;
+ lck->lls_sub[nr].sub_stripe = i;
nr++;
}
}
result = PTR_ERR(sublock);
break;
}
+ cl_lock_get_trust(sublock);
cl_lock_mutex_get(env, sublock);
cl_lock_mutex_get(env, parent);
/*
lov_sublock_adopt(env, lck, sublock, i, link);
cl_lock_mutex_put(env, parent);
} else {
+ OBD_SLAB_FREE_PTR(link, lov_lock_link_kmem);
cl_lock_mutex_put(env, parent);
cl_lock_unhold(env, sublock,
"lov-parent", parent);
}
cl_lock_mutex_put(env, sublock);
+ cl_lock_put(env, sublock);
}
}
/*
}
/**
+ *
+ * \retval 0 if state-transition can proceed
+ * \retval -ve otherwise.
+ */
+static int lov_lock_enqueue_wait(const struct lu_env *env,
+ struct lov_lock *lck,
+ struct cl_lock *sublock)
+{
+ struct cl_lock *lock = lck->lls_cl.cls_lock;
+ struct cl_lock *conflict = sublock->cll_conflict;
+ int result = CLO_REPEAT;
+ ENTRY;
+
+ LASSERT(cl_lock_is_mutexed(lock));
+ LASSERT(cl_lock_is_mutexed(sublock));
+ LASSERT(sublock->cll_state == CLS_QUEUING);
+ LASSERT(conflict != NULL);
+
+ sublock->cll_conflict = NULL;
+ cl_lock_mutex_put(env, lock);
+ cl_lock_mutex_put(env, sublock);
+
+ LASSERT(cl_lock_nr_mutexed(env) == 0);
+
+ cl_lock_mutex_get(env, conflict);
+ cl_lock_cancel(env, conflict);
+ cl_lock_delete(env, conflict);
+ while (conflict->cll_state != CLS_FREEING) {
+ int rc = 0;
+
+ rc = cl_lock_state_wait(env, conflict);
+ if (rc == 0)
+ continue;
+
+ result = lov_subresult(result, rc);
+ break;
+ }
+ cl_lock_mutex_put(env, conflict);
+ lu_ref_del(&conflict->cll_reference, "cancel-wait", sublock);
+ cl_lock_put(env, conflict);
+
+ cl_lock_mutex_get(env, lock);
+ RETURN(result);
+}
+
+/**
* Tries to advance a state machine of a given sub-lock toward enqueuing of
* the top-lock.
*
cl_lock_mutex_get(env, parent);
if (!IS_ERR(sublock)) {
+ cl_lock_get_trust(sublock);
if (parent->cll_state == CLS_QUEUING &&
- lck->lls_sub[idx].sub_lock == NULL)
+ lck->lls_sub[idx].sub_lock == NULL) {
lov_sublock_adopt(env, lck, sublock, idx, link);
- else {
+ } else {
+ OBD_SLAB_FREE_PTR(link, lov_lock_link_kmem);
/* other thread allocated sub-lock, or enqueue is no
* longer going on */
cl_lock_mutex_put(env, parent);
cl_lock_mutex_get(env, parent);
}
cl_lock_mutex_put(env, sublock);
+ cl_lock_put(env, sublock);
result = CLO_REPEAT;
} else
result = PTR_ERR(sublock);
break;
}
sublock = sub->lss_cl.cls_lock;
- rc = lov_sublock_lock(env, lls, closure, &subenv);
+ rc = lov_sublock_lock(env, lck, lls, closure, &subenv);
if (rc == 0) {
lov_sublock_hold(env, lck, i);
rc = lov_lock_enqueue_one(subenv->lse_env, lck, sublock,
subenv->lse_io, enqflags,
i == lck->lls_nr - 1);
minstate = min(minstate, sublock->cll_state);
- /*
- * Don't hold a sub-lock in CLS_CACHED state, see
- * description for lov_lock::lls_sub.
- */
- if (sublock->cll_state > CLS_HELD)
- rc = lov_sublock_release(env, lck, i, 1, rc);
- lov_sublock_unlock(env, sub, closure, subenv);
+ if (rc == CLO_WAIT) {
+ switch (sublock->cll_state) {
+ case CLS_QUEUING:
+ /* take recursive mutex, the lock is
+ * released in lov_lock_enqueue_wait.
+ */
+ cl_lock_mutex_get(env, sublock);
+ lov_sublock_unlock(env, sub, closure,
+ subenv);
+ rc = lov_lock_enqueue_wait(env, lck,
+ sublock);
+ break;
+ case CLS_CACHED:
+ rc = lov_sublock_release(env, lck, i,
+ 1, rc);
+ default:
+ lov_sublock_unlock(env, sub, closure,
+ subenv);
+ break;
+ }
+ } else {
+ LASSERT(sublock->cll_conflict == NULL);
+ lov_sublock_unlock(env, sub, closure, subenv);
+ }
}
result = lov_subresult(result, rc);
- if (result < 0)
+ if (result != 0)
break;
}
cl_lock_closure_fini(closure);
/* top-lock state cannot change concurrently, because single
* thread (one that released the last hold) carries unlocking
* to the completion. */
- LASSERT(slice->cls_lock->cll_state == CLS_UNLOCKING);
+ LASSERT(slice->cls_lock->cll_state == CLS_INTRANSIT);
lls = &lck->lls_sub[i];
sub = lls->sub_lock;
if (sub == NULL)
continue;
sublock = sub->lss_cl.cls_lock;
- rc = lov_sublock_lock(env, lls, closure, &subenv);
+ rc = lov_sublock_lock(env, lck, lls, closure, &subenv);
if (rc == 0) {
- if (lck->lls_sub[i].sub_flags & LSF_HELD) {
+ if (lls->sub_flags & LSF_HELD) {
LASSERT(sublock->cll_state == CLS_HELD);
rc = cl_unuse_try(subenv->lse_env, sublock);
- if (rc != CLO_WAIT)
- rc = lov_sublock_release(env, lck,
- i, 0, rc);
+ rc = lov_sublock_release(env, lck, i, 0, rc);
}
lov_sublock_unlock(env, sub, closure, subenv);
}
result = lov_subresult(result, rc);
- if (result < 0)
- break;
}
- if (result == 0 && lck->lls_unuse_race) {
- lck->lls_unuse_race = 0;
+
+ if (result == 0 && lck->lls_cancel_race) {
+ lck->lls_cancel_race = 0;
result = -ESTALE;
}
cl_lock_closure_fini(closure);
RETURN(result);
}
+
+static void lov_lock_cancel(const struct lu_env *env,
+ const struct cl_lock_slice *slice)
+{
+ struct lov_lock *lck = cl2lov_lock(slice);
+ struct cl_lock_closure *closure = lov_closure_get(env, slice->cls_lock);
+ int i;
+ int result;
+
+ ENTRY;
+
+ for (result = 0, i = 0; i < lck->lls_nr; ++i) {
+ int rc;
+ struct lovsub_lock *sub;
+ struct cl_lock *sublock;
+ struct lov_lock_sub *lls;
+ struct lov_sublock_env *subenv;
+
+ /* top-lock state cannot change concurrently, because single
+ * thread (one that released the last hold) carries unlocking
+ * to the completion. */
+ lls = &lck->lls_sub[i];
+ sub = lls->sub_lock;
+ if (sub == NULL)
+ continue;
+
+ sublock = sub->lss_cl.cls_lock;
+ rc = lov_sublock_lock(env, lck, lls, closure, &subenv);
+ if (rc == 0) {
+ if (!(lls->sub_flags & LSF_HELD)) {
+ lov_sublock_unlock(env, sub, closure, subenv);
+ continue;
+ }
+
+ switch(sublock->cll_state) {
+ case CLS_HELD:
+ rc = cl_unuse_try(subenv->lse_env,
+ sublock);
+ lov_sublock_release(env, lck, i, 0, 0);
+ break;
+ case CLS_ENQUEUED:
+ /* TODO: it's not a good idea to cancel this
+ * lock because it's innocent. But it's
+ * acceptable. The better way would be to
+ * define a new lock method to unhold the
+ * dlm lock. */
+ cl_lock_cancel(env, sublock);
+ default:
+ lov_sublock_release(env, lck, i, 1, 0);
+ break;
+ }
+ lov_sublock_unlock(env, sub, closure, subenv);
+ }
+
+ if (rc == CLO_REPEAT) {
+ --i;
+ continue;
+ }
+
+ result = lov_subresult(result, rc);
+ }
+
+ if (result)
+ CL_LOCK_DEBUG(D_ERROR, env, slice->cls_lock,
+ "lov_lock_cancel fails with %d.\n", result);
+
+ cl_lock_closure_fini(closure);
+}
+
static int lov_lock_wait(const struct lu_env *env,
const struct cl_lock_slice *slice)
{
sub = lls->sub_lock;
LASSERT(sub != NULL);
sublock = sub->lss_cl.cls_lock;
- rc = lov_sublock_lock(env, lls, closure, &subenv);
+ rc = lov_sublock_lock(env, lck, lls, closure, &subenv);
if (rc == 0) {
LASSERT(sublock->cll_state >= CLS_ENQUEUED);
if (sublock->cll_state < CLS_HELD)
lov_sublock_unlock(env, sub, closure, subenv);
}
result = lov_subresult(result, rc);
- if (result < 0)
+ if (result != 0)
break;
}
cl_lock_closure_fini(closure);
int result;
int i;
- LASSERT(slice->cls_lock->cll_state == CLS_CACHED);
+ LASSERT(slice->cls_lock->cll_state == CLS_INTRANSIT);
ENTRY;
for (result = 0, i = 0; i < lck->lls_nr; ++i) {
struct lov_lock_sub *lls;
struct lov_sublock_env *subenv;
- if (slice->cls_lock->cll_state != CLS_CACHED) {
- /* see comment in lov_lock_enqueue(). */
- LASSERT(i > 0 && result != 0);
- break;
- }
- /*
- * if a sub-lock was destroyed while top-lock was in
- * CLS_CACHED state, top-lock would have been moved into
- * CLS_NEW state, so all sub-locks have to be in place.
- */
+ LASSERT(slice->cls_lock->cll_state == CLS_INTRANSIT);
+
lls = &lck->lls_sub[i];
sub = lls->sub_lock;
- LASSERT(sub != NULL);
+ if (sub == NULL) {
+ /*
+ * Sub-lock might have been canceled, while top-lock was
+ * cached.
+ */
+ result = -ESTALE;
+ break;
+ }
+
sublock = sub->lss_cl.cls_lock;
- rc = lov_sublock_lock(env, lls, closure, &subenv);
+ rc = lov_sublock_lock(env, lck, lls, closure, &subenv);
if (rc == 0) {
LASSERT(sublock->cll_state != CLS_FREEING);
lov_sublock_hold(env, lck, i);
if (sublock->cll_state == CLS_CACHED) {
- rc = cl_use_try(subenv->lse_env, sublock);
+ rc = cl_use_try(subenv->lse_env, sublock, 0);
if (rc != 0)
rc = lov_sublock_release(env, lck,
i, 1, rc);
- } else
- rc = 0;
+ }
lov_sublock_unlock(env, sub, closure, subenv);
}
result = lov_subresult(result, rc);
- if (result < 0)
+ if (result != 0)
break;
}
+
+ if (lck->lls_cancel_race) {
+ /*
+ * If there is unlocking happened at the same time, then
+ * sublock_lock state should be FREEING, and lov_sublock_lock
+ * should return CLO_REPEAT. In this case, it should return
+ * ESTALE, and up layer should reset the lock state to be NEW.
+ */
+ lck->lls_cancel_race = 0;
+ LASSERT(result != 0);
+ result = -ESTALE;
+ }
cl_lock_closure_fini(closure);
RETURN(result);
}
LASSERT(cl_lock_is_mutexed(sub->lss_cl.cls_lock));
ENTRY;
- list_del_init(&link->lll_list);
+ cfs_list_del_init(&link->lll_list);
LASSERT(lck->lls_sub[link->lll_idx].sub_lock == sub);
/* yank this sub-lock from parent's array */
lck->lls_sub[link->lll_idx].sub_lock = NULL;
LASSERT(cl_lock_is_mutexed(sub->lss_cl.cls_lock));
ENTRY;
- list_for_each_entry(scan, &sub->lss_parents, lll_list) {
+ cfs_list_for_each_entry(scan, &sub->lss_parents, lll_list) {
if (scan->lll_super == lck)
RETURN(scan);
}
continue;
sublock = lsl->lss_cl.cls_lock;
- rc = lov_sublock_lock(env, lls, closure, NULL);
+ rc = lov_sublock_lock(env, lck, lls, closure, NULL);
if (rc == 0) {
- if (lck->lls_sub[i].sub_flags & LSF_HELD)
+ if (lls->sub_flags & LSF_HELD)
lov_sublock_release(env, lck, i, 1, 0);
if (sublock->cll_state < CLS_FREEING) {
struct lov_lock_link *link;
.clo_wait = lov_lock_wait,
.clo_use = lov_lock_use,
.clo_unuse = lov_lock_unuse,
+ .clo_cancel = lov_lock_cancel,
.clo_fits_into = lov_lock_fits_into,
.clo_delete = lov_lock_delete,
.clo_print = lov_lock_print
struct cl_lock_closure *closure;
closure = &lov_env_info(env)->lti_closure;
- LASSERT(list_empty(&closure->clc_list));
+ LASSERT(cfs_list_empty(&closure->clc_list));
cl_lock_closure_init(env, closure, parent, 1);
return closure;
}