/** super class */
struct cl_io_slice oi_cl;
/** true if this io is lockless. */
- unsigned int oi_lockless;
+ unsigned int oi_lockless:1,
+ /** true if this io is counted as active IO */
+ oi_is_active:1;
/** how many LRU pages are reserved for this IO */
unsigned long oi_lru_reserved;
struct list_head *ext_list, int cmd);
long osc_lru_shrink(const struct lu_env *env, struct client_obd *cli,
long target, bool force);
-long osc_lru_reclaim(struct client_obd *cli, unsigned long npages);
+unsigned long osc_lru_reserve(struct client_obd *cli, unsigned long npages);
+void osc_lru_unreserve(struct client_obd *cli, unsigned long npages);
unsigned long osc_ldlm_weigh_ast(struct ldlm_lock *dlmlock);
spin_lock(&imp->imp_lock);
if (likely(!imp->imp_invalid)) {
+ struct osc_io *oio = osc_env_io(env);
+
atomic_inc(&osc->oo_nr_ios);
+ oio->oi_is_active = 1;
rc = 0;
}
spin_unlock(&imp->imp_lock);
struct cl_io *io = ios->cis_io;
struct osc_io *oio = osc_env_io(env);
struct osc_object *osc = cl2osc(ios->cis_obj);
- struct client_obd *cli = osc_cli(osc);
- unsigned long c;
unsigned long npages;
- unsigned long max_pages;
if (cl_io_is_append(io))
return osc_io_iter_init(env, ios);
if (io->u.ci_rw.crw_pos & ~PAGE_MASK)
++npages;
- max_pages = cli->cl_max_pages_per_rpc * cli->cl_max_rpcs_in_flight;
- if (npages > max_pages)
- npages = max_pages;
-
- c = atomic_long_read(cli->cl_lru_left);
- if (c < npages && osc_lru_reclaim(cli, npages) > 0)
- c = atomic_long_read(cli->cl_lru_left);
- while (c >= npages) {
- if (c == atomic_long_cmpxchg(cli->cl_lru_left, c, c - npages)) {
- oio->oi_lru_reserved = npages;
- break;
- }
- c = atomic_long_read(cli->cl_lru_left);
- }
- if (atomic_long_read(cli->cl_lru_left) < max_pages) {
- /*
- * If there aren't enough pages in the per-OSC LRU then
- * wake up the LRU thread to try and clear out space, so
- * we don't block if pages are being dirtied quickly.
- */
- CDEBUG(D_CACHE, "%s: queue LRU, left: %lu/%ld.\n",
- cli_name(cli), atomic_long_read(cli->cl_lru_left),
- max_pages);
- (void)ptlrpcd_queue_work(cli->cl_lru_work);
- }
+ oio->oi_lru_reserved = osc_lru_reserve(osc_cli(osc), npages);
return osc_io_iter_init(env, ios);
}
static void osc_io_iter_fini(const struct lu_env *env,
const struct cl_io_slice *ios)
{
- struct osc_object *osc = cl2osc(ios->cis_obj);
+ struct osc_io *oio = osc_env_io(env);
- LASSERT(atomic_read(&osc->oo_nr_ios) > 0);
- if (atomic_dec_and_test(&osc->oo_nr_ios))
- wake_up_all(&osc->oo_io_waitq);
+ if (oio->oi_is_active) {
+ struct osc_object *osc = cl2osc(ios->cis_obj);
+
+ oio->oi_is_active = 0;
+ LASSERT(atomic_read(&osc->oo_nr_ios) > 0);
+ if (atomic_dec_and_test(&osc->oo_nr_ios))
+ wake_up_all(&osc->oo_io_waitq);
+ }
}
static void osc_io_write_iter_fini(const struct lu_env *env,
{
struct osc_io *oio = osc_env_io(env);
struct osc_object *osc = cl2osc(ios->cis_obj);
- struct client_obd *cli = osc_cli(osc);
if (oio->oi_lru_reserved > 0) {
- atomic_long_add(oio->oi_lru_reserved, cli->cl_lru_left);
+ osc_lru_unreserve(osc_cli(osc), oio->oi_lru_reserved);
oio->oi_lru_reserved = 0;
}
oio->oi_write_osclock = NULL;
spin_unlock(&oscl->ols_lock);
}
-static void osc_lock_enqueue_wait(const struct lu_env *env,
- struct osc_object *obj,
- struct osc_lock *oscl)
+static int osc_lock_enqueue_wait(const struct lu_env *env,
+ struct osc_object *obj,
+ struct osc_lock *oscl)
{
struct osc_lock *tmp_oscl;
struct cl_lock_descr *need = &oscl->ols_cl.cls_lock->cll_descr;
struct cl_sync_io *waiter = &osc_env_info(env)->oti_anchor;
+ int rc = 0;
spin_lock(&obj->oo_ol_spin);
list_add_tail(&oscl->ols_nextlock_oscobj, &obj->oo_ol_list);
spin_unlock(&tmp_oscl->ols_lock);
spin_unlock(&obj->oo_ol_spin);
- (void)cl_sync_io_wait(env, waiter, 0);
-
+ rc = cl_sync_io_wait(env, waiter, 0);
spin_lock(&obj->oo_ol_spin);
+ if (rc < 0)
+ break;
+
oscl->ols_owner = NULL;
goto restart;
}
spin_unlock(&obj->oo_ol_spin);
+
+ return rc;
}
/**
goto enqueue_base;
}
- osc_lock_enqueue_wait(env, osc, oscl);
+ result = osc_lock_enqueue_wait(env, osc, oscl);
+ if (result < 0)
+ goto out;
/* we can grant lockless lock right after all conflicting locks
* are canceled.
* osc_lock.
*/
ostid_build_res_name(&osc->oo_oinfo->loi_oi, resname);
- osc_lock_build_einfo(env, lock, osc, &oscl->ols_einfo);
osc_lock_build_policy(env, lock, policy);
if (oscl->ols_agl) {
oscl->ols_einfo.ei_cbdata = NULL;
upcall, cookie,
&oscl->ols_einfo, PTLRPCD_SET, async,
oscl->ols_agl);
- if (result != 0) {
- oscl->ols_state = OLS_CANCELLED;
- osc_lock_wake_waiters(env, osc, oscl);
-
- /* hide error for AGL lock. */
- if (oscl->ols_agl) {
- cl_object_put(env, osc2cl(osc));
- result = 0;
- }
- if (anchor)
- cl_sync_io_note(env, anchor, result);
- } else {
+ if (!result) {
if (osc_lock_is_lockless(oscl)) {
oio->oi_lockless = 1;
} else if (!async) {
LASSERT(oscl->ols_hold);
LASSERT(oscl->ols_dlmlock);
}
+ } else if (oscl->ols_agl) {
+ cl_object_put(env, osc2cl(osc));
+ result = 0;
+ }
+
+out:
+ if (result < 0) {
+ oscl->ols_state = OLS_CANCELLED;
+ osc_lock_wake_waiters(env, osc, oscl);
+
+ if (anchor)
+ cl_sync_io_note(env, anchor, result);
}
return result;
}
oscl->ols_flags |= LDLM_FL_BLOCK_GRANTED;
oscl->ols_glimpse = 1;
}
+ osc_lock_build_einfo(env, lock, cl2osc(obj), &oscl->ols_einfo);
cl_lock_slice_add(lock, &oscl->ols_cl, obj, &osc_lock_ops);
l_wait_event(osc->oo_io_waitq, !atomic_read(&osc->oo_nr_ios), &lwi);
- /* Discard all pages of this object. */
+ /* Discard all dirty pages of this object. */
osc_cache_truncate_start(env, osc, 0, NULL);
+ /* Discard all caching pages */
+ osc_lock_discard_pages(env, osc, 0, CL_PAGE_EOF, CLM_WRITE);
+
+ /* Clear ast data of dlm lock. Do this after discarding all pages */
+ osc_object_prune(env, osc2cl(osc));
+
return 0;
}
static void osc_lru_del(struct client_obd *cli, struct osc_page *opg);
static void osc_lru_use(struct client_obd *cli, struct osc_page *opg);
-static int osc_lru_reserve(const struct lu_env *env, struct osc_object *obj,
- struct osc_page *opg);
+static int osc_lru_alloc(const struct lu_env *env, struct client_obd *cli,
+ struct osc_page *opg);
/** \addtogroup osc
* @{
/* reserve an LRU space for this page */
if (page->cp_type == CPT_CACHEABLE && result == 0) {
- result = osc_lru_reserve(env, osc, opg);
+ result = osc_lru_alloc(env, osc_cli(osc), opg);
if (result == 0) {
spin_lock(&osc->oo_tree_lock);
result = radix_tree_insert(&osc->oo_tree, index, opg);
* LRU pages in batch. Therefore, the actual number is adjusted at least
* max_pages_per_rpc.
*/
-long osc_lru_reclaim(struct client_obd *cli, unsigned long npages)
+static long osc_lru_reclaim(struct client_obd *cli, unsigned long npages)
{
struct lu_env *env;
struct cl_client_cache *cache = cli->cl_cache;
}
/**
- * osc_lru_reserve() is called to reserve an LRU slot for a cl_page.
+ * osc_lru_alloc() is called to reserve an LRU slot for a cl_page.
*
* Usually the LRU slots are reserved in osc_io_iter_rw_init().
* Only in the case that the LRU slots are in extreme shortage, it should
* have reserved enough slots for an IO.
*/
-static int osc_lru_reserve(const struct lu_env *env, struct osc_object *obj,
- struct osc_page *opg)
+static int osc_lru_alloc(const struct lu_env *env, struct client_obd *cli,
+ struct osc_page *opg)
{
struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
struct osc_io *oio = osc_env_io(env);
- struct client_obd *cli = osc_cli(obj);
int rc = 0;
if (!cli->cl_cache) /* shall not be in LRU */
return rc;
}
+/**
+ * osc_lru_reserve() is called to reserve enough LRU slots for I/O.
+ *
+ * The benefit of doing this is to reduce contention against atomic counter
+ * cl_lru_left by changing it from per-page access to per-IO access.
+ */
+unsigned long osc_lru_reserve(struct client_obd *cli, unsigned long npages)
+{
+ unsigned long reserved = 0;
+ unsigned long max_pages;
+ unsigned long c;
+
+ /*
+ * reserve a full RPC window at most to avoid that a thread accidentally
+ * consumes too many LRU slots
+ */
+ max_pages = cli->cl_max_pages_per_rpc * cli->cl_max_rpcs_in_flight;
+ if (npages > max_pages)
+ npages = max_pages;
+
+ c = atomic_long_read(cli->cl_lru_left);
+ if (c < npages && osc_lru_reclaim(cli, npages) > 0)
+ c = atomic_long_read(cli->cl_lru_left);
+ while (c >= npages) {
+ if (c == atomic_long_cmpxchg(cli->cl_lru_left, c, c - npages)) {
+ reserved = npages;
+ break;
+ }
+ c = atomic_long_read(cli->cl_lru_left);
+ }
+ if (atomic_long_read(cli->cl_lru_left) < max_pages) {
+ /*
+ * If there aren't enough pages in the per-OSC LRU then
+ * wake up the LRU thread to try and clear out space, so
+ * we don't block if pages are being dirtied quickly.
+ */
+ CDEBUG(D_CACHE, "%s: queue LRU, left: %lu/%ld.\n",
+ cli_name(cli), atomic_long_read(cli->cl_lru_left),
+ max_pages);
+ (void)ptlrpcd_queue_work(cli->cl_lru_work);
+ }
+
+ return reserved;
+}
+
+/**
+ * osc_lru_unreserve() is called to unreserve LRU slots.
+ *
+ * LRU slots reserved by osc_lru_reserve() may have entries left due to several
+ * reasons such as page already existing or I/O error. Those reserved slots
+ * should be freed by calling this function.
+ */
+void osc_lru_unreserve(struct client_obd *cli, unsigned long npages)
+{
+ atomic_long_add(npages, cli->cl_lru_left);
+ wake_up_all(&osc_lru_waitq);
+}
+
/**
* Atomic operations are expensive. We accumulate the accounting for the
* same page pgdat to get better performance.
projects / karo-tx-linux.git / commitdiff