-/* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
- * vim:expandtab:shiftwidth=8:tabstop=8:
- *
+/*
* GPL HEADER START
*
* DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
/*
* Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
* Use is subject to license terms.
+ *
+ * Copyright (c) 2011, 2012, Whamcloud, Inc.
*/
/*
* This file is part of Lustre, http://www.lustre.org/
*/
#define DEBUG_SUBSYSTEM S_CLASS
-#ifndef EXPORT_SYMTAB
-# define EXPORT_SYMTAB
-#endif
#include <obd_class.h>
#include <obd_support.h>
void cl_lock_put(const struct lu_env *env, struct cl_lock *lock)
{
struct cl_object *obj;
- struct cl_object_header *head;
struct cl_site *site;
LINVRNT(cl_lock_invariant(env, lock));
ENTRY;
obj = lock->cll_descr.cld_obj;
LINVRNT(obj != NULL);
- head = cl_object_header(obj);
site = cl_object_site(obj);
CDEBUG(D_TRACE, "releasing reference: %d %p %lu\n",
struct cl_object_header *head;
struct cl_object *obj;
struct cl_lock *lock;
- int ok;
obj = need->cld_obj;
head = cl_object_header(obj);
if (result < 0)
cl_lock_error(env, lock, result);
}
- ok = lock->cll_state == CLS_HELD;
- if (ok) {
+ if (lock->cll_state == CLS_HELD) {
cl_lock_hold_add(env, lock, scope, source);
cl_lock_user_add(env, lock);
+ cl_lock_mutex_put(env, lock);
+ cl_lock_lockdep_acquire(env, lock, 0);
cl_lock_put(env, lock);
- }
- cl_lock_mutex_put(env, lock);
- if (!ok) {
+ } else {
+ cl_lock_mutex_put(env, lock);
cl_lock_put(env, lock);
lock = NULL;
}
EXIT;
}
-
/**
* Waits until lock state is changed.
*
cl_lock_user_del(env, lock);
cl_lock_error(env, lock, result);
}
- LASSERT(ergo(result == 0, lock->cll_state == CLS_ENQUEUED ||
+ LASSERT(ergo(result == 0 && !(enqflags & CEF_AGL),
+ lock->cll_state == CLS_ENQUEUED ||
lock->cll_state == CLS_HELD));
RETURN(result);
}
EXPORT_SYMBOL(cl_lock_at_page);
/**
- * Returns a list of pages protected (only) by a given lock.
- *
- * Scans an extent of page radix tree, corresponding to the \a lock and queues
- * all pages that are not protected by locks other than \a lock into \a queue.
+ * Calculate the page offset at the layer of @lock.
+ * At the time of this writing, @page is top page and @lock is sub lock.
*/
-void cl_lock_page_list_fixup(const struct lu_env *env,
- struct cl_io *io, struct cl_lock *lock,
- struct cl_page_list *queue)
+static pgoff_t pgoff_at_lock(struct cl_page *page, struct cl_lock *lock)
{
- struct cl_page *page;
- struct cl_page *temp;
- struct cl_page_list *plist = &cl_env_info(env)->clt_list;
-
- LINVRNT(cl_lock_invariant(env, lock));
- ENTRY;
+ struct lu_device_type *dtype;
+ const struct cl_page_slice *slice;
- /* No need to fix for WRITE lock because it is exclusive. */
- if (lock->cll_descr.cld_mode >= CLM_WRITE)
- RETURN_EXIT;
+ dtype = lock->cll_descr.cld_obj->co_lu.lo_dev->ld_type;
+ slice = cl_page_at(page, dtype);
+ LASSERT(slice != NULL);
+ return slice->cpl_page->cp_index;
+}
- /* For those pages who are still covered by other PR locks, we should
- * not discard them otherwise a [0, EOF) PR lock will discard all
- * pages.
- */
- cl_page_list_init(plist);
- cl_page_list_for_each_safe(page, temp, queue) {
- pgoff_t idx = page->cp_index;
- struct cl_lock *found;
- struct cl_lock_descr *descr;
-
- /* The algorithm counts on the index-ascending page index. */
- LASSERT(ergo(&temp->cp_batch != &queue->pl_pages,
- page->cp_index < temp->cp_index));
-
- found = cl_lock_at_page(env, lock->cll_descr.cld_obj,
- page, lock, 1, 0);
- if (found == NULL)
- continue;
-
- descr = &found->cll_descr;
- cfs_list_for_each_entry_safe_from(page, temp, &queue->pl_pages,
- cp_batch) {
- idx = page->cp_index;
- if (descr->cld_start > idx || descr->cld_end < idx)
- break;
- cl_page_list_move(plist, queue, page);
+/**
+ * Check if page @page is covered by an extra lock or discard it.
+ */
+static int check_and_discard_cb(const struct lu_env *env, struct cl_io *io,
+ struct cl_page *page, void *cbdata)
+{
+ struct cl_thread_info *info = cl_env_info(env);
+ struct cl_lock *lock = cbdata;
+ pgoff_t index = pgoff_at_lock(page, lock);
+
+ if (index >= info->clt_fn_index) {
+ struct cl_lock *tmp;
+
+ /* refresh non-overlapped index */
+ tmp = cl_lock_at_page(env, lock->cll_descr.cld_obj, page, lock,
+ 1, 0);
+ if (tmp != NULL) {
+ /* Cache the first-non-overlapped index so as to skip
+ * all pages within [index, clt_fn_index). This
+ * is safe because if tmp lock is canceled, it will
+ * discard these pages. */
+ info->clt_fn_index = tmp->cll_descr.cld_end + 1;
+ if (tmp->cll_descr.cld_end == CL_PAGE_EOF)
+ info->clt_fn_index = CL_PAGE_EOF;
+ cl_lock_put(env, tmp);
+ } else if (cl_page_own(env, io, page) == 0) {
+ /* discard the page */
+ cl_page_unmap(env, io, page);
+ cl_page_discard(env, io, page);
+ cl_page_disown(env, io, page);
+ } else {
+ LASSERT(page->cp_state == CPS_FREEING);
}
- cl_lock_put(env, found);
}
- /* The pages in plist are covered by other locks, don't handle them
- * this time.
- */
- if (io != NULL)
- cl_page_list_disown(env, io, plist);
- cl_page_list_fini(env, plist);
- EXIT;
+ info->clt_next_index = index + 1;
+ return CLP_GANG_OKAY;
+}
+
+static int pageout_cb(const struct lu_env *env, struct cl_io *io,
+ struct cl_page *page, void *cbdata)
+{
+ struct cl_thread_info *info = cl_env_info(env);
+ struct cl_page_list *queue = &info->clt_queue.c2_qin;
+ struct cl_lock *lock = cbdata;
+ typeof(cl_page_own) *page_own;
+ int rc = CLP_GANG_OKAY;
+
+ page_own = queue->pl_nr ? cl_page_own_try : cl_page_own;
+ if (page_own(env, io, page) == 0) {
+ cl_page_list_add(queue, page);
+ info->clt_next_index = pgoff_at_lock(page, lock) + 1;
+ } else if (page->cp_state != CPS_FREEING) {
+ /* cl_page_own() won't fail unless
+ * the page is being freed. */
+ LASSERT(queue->pl_nr != 0);
+ rc = CLP_GANG_AGAIN;
+ }
+
+ return rc;
}
-EXPORT_SYMBOL(cl_lock_page_list_fixup);
/**
* Invalidate pages protected by the given lock, sending them out to the
struct cl_io *io = &info->clt_io;
struct cl_2queue *queue = &info->clt_queue;
struct cl_lock_descr *descr = &lock->cll_descr;
- struct lu_device_type *dtype;
+ cl_page_gang_cb_t cb;
long page_count;
- pgoff_t next_index;
int res;
int result;
if (result != 0)
GOTO(out, result);
- dtype = descr->cld_obj->co_lu.lo_dev->ld_type;
- next_index = descr->cld_start;
+ cb = descr->cld_mode == CLM_READ ? check_and_discard_cb : pageout_cb;
+ info->clt_fn_index = info->clt_next_index = descr->cld_start;
do {
- const struct cl_page_slice *slice;
-
cl_2queue_init(queue);
res = cl_page_gang_lookup(env, descr->cld_obj, io,
- next_index, descr->cld_end,
- &queue->c2_qin);
+ info->clt_next_index, descr->cld_end,
+ cb, (void *)lock);
page_count = queue->c2_qin.pl_nr;
- if (page_count == 0)
- break;
-
- /* cl_page_gang_lookup() uses subobj and sublock to look for
- * covered pages, but @queue->c2_qin contains the list of top
- * pages. We have to turn the page back to subpage so as to
- * get `correct' next index. -jay */
- slice = cl_page_at(cl_page_list_last(&queue->c2_qin), dtype);
- next_index = slice->cpl_page->cp_index + 1;
-
- result = cl_page_list_unmap(env, io, &queue->c2_qin);
- if (!discard) {
- long timeout = 600; /* 10 minutes. */
- /* for debug purpose, if this request can't be
- * finished in 10 minutes, we hope it can notify us.
- */
- result = cl_io_submit_sync(env, io, CRT_WRITE, queue,
- CRP_CANCEL, timeout);
- if (result)
- CWARN("Writing %lu pages error: %d\n",
- page_count, result);
+ if (page_count > 0) {
+ /* must be writeback case */
+ LASSERTF(descr->cld_mode >= CLM_WRITE, "lock mode %s\n",
+ cl_lock_mode_name(descr->cld_mode));
+
+ result = cl_page_list_unmap(env, io, &queue->c2_qin);
+ if (!discard) {
+ long timeout = 600; /* 10 minutes. */
+ /* for debug purpose, if this request can't be
+ * finished in 10 minutes, we hope it can
+ * notify us.
+ */
+ result = cl_io_submit_sync(env, io, CRT_WRITE,
+ queue, CRP_CANCEL,
+ timeout);
+ if (result)
+ CWARN("Writing %lu pages error: %d\n",
+ page_count, result);
+ }
+ cl_2queue_discard(env, io, queue);
+ cl_2queue_disown(env, io, queue);
+ cl_2queue_fini(env, queue);
}
- cl_lock_page_list_fixup(env, io, lock, &queue->c2_qout);
- cl_2queue_discard(env, io, queue);
- cl_2queue_disown(env, io, queue);
- cl_2queue_fini(env, queue);
- if (next_index > descr->cld_end)
+ if (info->clt_next_index > descr->cld_end)
break;
if (res == CLP_GANG_RESCHED)
cl_lock_get_trust(lock);
cfs_spin_unlock(&head->coh_lock_guard);
lu_ref_add(&lock->cll_reference, "prune", cfs_current());
+
+again:
cl_lock_mutex_get(env, lock);
if (lock->cll_state < CLS_FREEING) {
LASSERT(lock->cll_holds == 0);
- LASSERT(lock->cll_users == 0);
+ LASSERT(lock->cll_users <= 1);
+ if (unlikely(lock->cll_users == 1)) {
+ struct l_wait_info lwi = { 0 };
+
+ cl_lock_mutex_put(env, lock);
+ l_wait_event(lock->cll_wq,
+ lock->cll_users == 0,
+ &lwi);
+ goto again;
+ }
+
if (cancel)
cl_lock_cancel(env, lock);
cl_lock_delete(env, lock);
ENTRY;
do {
lock = cl_lock_hold_mutex(env, io, need, scope, source);
- if (!IS_ERR(lock)) {
- rc = cl_enqueue_locked(env, lock, io, enqflags);
- if (rc == 0) {
- if (cl_lock_fits_into(env, lock, need, io)) {
+ if (IS_ERR(lock))
+ break;
+
+ rc = cl_enqueue_locked(env, lock, io, enqflags);
+ if (rc == 0) {
+ if (cl_lock_fits_into(env, lock, need, io)) {
+ if (!(enqflags & CEF_AGL)) {
cl_lock_mutex_put(env, lock);
- cl_lock_lockdep_acquire(env,
- lock, enqflags);
+ cl_lock_lockdep_acquire(env, lock,
+ enqflags);
break;
}
- cl_unuse_locked(env, lock);
+ rc = 1;
}
- cl_lock_trace(D_DLMTRACE, env, "enqueue failed", lock);
- cl_lock_hold_release(env, lock, scope, source);
- cl_lock_mutex_put(env, lock);
- lu_ref_del(&lock->cll_reference, scope, source);
- cl_lock_put(env, lock);
+ cl_unuse_locked(env, lock);
+ }
+ cl_lock_trace(D_DLMTRACE, env,
+ rc <= 0 ? "enqueue failed" : "agl succeed", lock);
+ cl_lock_hold_release(env, lock, scope, source);
+ cl_lock_mutex_put(env, lock);
+ lu_ref_del(&lock->cll_reference, scope, source);
+ cl_lock_put(env, lock);
+ if (rc > 0) {
+ LASSERT(enqflags & CEF_AGL);
+ lock = NULL;
+ } else if (rc != 0) {
lock = ERR_PTR(rc);
- } else
- rc = PTR_ERR(lock);
+ }
} while (rc == 0);
RETURN(lock);
}
}
EXPORT_SYMBOL(cl_lock_user_add);
-int cl_lock_user_del(const struct lu_env *env, struct cl_lock *lock)
+void cl_lock_user_del(const struct lu_env *env, struct cl_lock *lock)
{
LINVRNT(cl_lock_is_mutexed(lock));
LINVRNT(cl_lock_invariant(env, lock));
ENTRY;
cl_lock_used_mod(env, lock, -1);
- RETURN(lock->cll_users == 0);
+ if (lock->cll_users == 0)
+ cfs_waitq_broadcast(&lock->cll_wq);
+ EXIT;
}
EXPORT_SYMBOL(cl_lock_user_del);