#define DEBUG_SUBSYSTEM S_FILTER
-#ifndef AUTOCONF_INCLUDED
-#include <linux/config.h>
-#endif
#include <linux/module.h>
#include <linux/fs.h>
#include <linux/dcache.h>
static int filter_lock_dentry(struct obd_device *obd, struct dentry *dparent)
{
- LOCK_INODE_MUTEX_PARENT(dparent->d_inode);
- return 0;
+ mutex_lock_nested(&dparent->d_inode->i_mutex, I_MUTEX_PARENT);
+ return 0;
}
/* We never dget the object parent, so DON'T dput it either */
/* We never dget the object parent, so DON'T dput it either */
static void filter_parent_unlock(struct dentry *dparent)
{
- UNLOCK_INODE_MUTEX(dparent->d_inode);
+ mutex_unlock(&dparent->d_inode->i_mutex);
}
/* How to get files, dentries, inodes from object id's.
/* don't need dir->i_zombie for 2.4, it is for rename/unlink of dir
* itself we already hold dir->i_mutex for child create/unlink ops */
- LASSERT(dentry->d_inode != NULL);
- LASSERT(TRYLOCK_INODE_MUTEX(dir) == 0);
- LASSERT(TRYLOCK_INODE_MUTEX(dentry->d_inode) == 0);
+ LASSERT(dentry->d_inode != NULL);
+ LASSERT(mutex_trylock(&dir->i_mutex) == 0);
+ LASSERT(mutex_trylock(&dentry->d_inode->i_mutex) == 0);
/* may_delete() */
rc = dir->i_op->unlink(dir, dentry);
out:
- /* need to drop i_mutex before we lose inode reference */
- UNLOCK_INODE_MUTEX(dentry->d_inode);
- if (rc == 0)
- d_delete(dentry);
+ /* need to drop i_mutex before we lose inode reference */
+ mutex_unlock(&dentry->d_inode->i_mutex);
+ if (rc == 0)
+ d_delete(dentry);
- RETURN(rc);
+ RETURN(rc);
}
/* Caller must hold LCK_PW on parent and push us into kernel context.
* at the OST layer there are only (potentially) multiple obd_device of type
* unknown at the time of OST thread creation.
*
- * Instead array of iobuf's is attached to struct filter_obd (->fo_iobuf_pool
- * field). This array has size OST_MAX_THREADS, so that each OST thread uses
- * it's very own iobuf.
+ * We create a cfs_hash for struct filter_obd (->fo_iobuf_hash field) on
+ * initializing, each OST thread will create it's own iobuf on the first
+ * access and insert it into ->fo_iobuf_hash with thread ID as key,
+ * so the iobuf can be found again by thread ID.
*
* Functions below
*
- * filter_kiobuf_pool_init()
+ * filter_iobuf_pool_init()
*
- * filter_kiobuf_pool_done()
+ * filter_iobuf_pool_done()
*
* filter_iobuf_get()
*
*/
static void filter_iobuf_pool_done(struct filter_obd *filter)
{
- struct filter_iobuf **pool;
- int i;
+ ENTRY;
- ENTRY;
-
- pool = filter->fo_iobuf_pool;
- if (pool != NULL) {
- for (i = 0; i < filter->fo_iobuf_count; ++ i) {
- if (pool[i] != NULL)
- filter_free_iobuf(pool[i]);
- }
- OBD_FREE(pool, filter->fo_iobuf_count * sizeof pool[0]);
- filter->fo_iobuf_pool = NULL;
- }
- EXIT;
+ if (filter->fo_iobuf_hash != NULL) {
+ cfs_hash_putref(filter->fo_iobuf_hash);
+ filter->fo_iobuf_hash = NULL;
+ }
+ EXIT;
}
static int filter_adapt_sptlrpc_conf(struct obd_device *obd, int initial)
return 0;
}
-/*
- * pre-allocate pool of iobuf's to be used by filter_{prep,commit}rw_write().
- */
-static int filter_iobuf_pool_init(struct filter_obd *filter)
+static unsigned
+filter_iobuf_hop_hash(cfs_hash_t *hs, const void *key, unsigned mask)
{
- void **pool;
+ __u64 val = *((__u64 *)key);
- ENTRY;
+ return cfs_hash_long(val, hs->hs_cur_bits);
+}
+static void *
+filter_iobuf_hop_key(cfs_hlist_node_t *hnode)
+{
+ struct filter_iobuf *pool;
- OBD_ALLOC_GFP(filter->fo_iobuf_pool, OSS_THREADS_MAX * sizeof(*pool),
- GFP_KERNEL);
- if (filter->fo_iobuf_pool == NULL)
- RETURN(-ENOMEM);
+ pool = cfs_hlist_entry(hnode, struct filter_iobuf, dr_hlist);
+ return &pool->dr_hkey;
+}
- filter->fo_iobuf_count = OSS_THREADS_MAX;
+static int
+filter_iobuf_hop_keycmp(const void *key, cfs_hlist_node_t *hnode)
+{
+ struct filter_iobuf *pool;
- RETURN(0);
+ pool = cfs_hlist_entry(hnode, struct filter_iobuf, dr_hlist);
+ return pool->dr_hkey == *((__u64 *)key);
}
-/* Return iobuf allocated for @thread_id. We don't know in advance how
- * many threads there will be so we allocate a large empty array and only
- * fill in those slots that are actually in use.
- * If we haven't allocated a pool entry for this thread before, do so now. */
-void *filter_iobuf_get(struct filter_obd *filter, struct obd_trans_info *oti)
+static void *
+filter_iobuf_hop_object(cfs_hlist_node_t *hnode)
{
- int thread_id = (oti && oti->oti_thread) ?
- oti->oti_thread->t_id : -1;
- struct filter_iobuf *pool = NULL;
- struct filter_iobuf **pool_place = NULL;
+ return cfs_hlist_entry(hnode, struct filter_iobuf, dr_hlist);
+}
- if (thread_id >= 0) {
- LASSERT(thread_id < filter->fo_iobuf_count);
- pool = *(pool_place = &filter->fo_iobuf_pool[thread_id]);
- }
+static void
+filter_iobuf_hop_get(cfs_hash_t *hs, cfs_hlist_node_t *hnode)
+{
+ /* dummy, required by cfs_hash */
+}
- if (unlikely(pool == NULL)) {
- pool = filter_alloc_iobuf(filter, OBD_BRW_WRITE,
- PTLRPC_MAX_BRW_PAGES);
- if (pool_place != NULL)
- *pool_place = pool;
- }
+static void
+filter_iobuf_hop_put_locked(cfs_hash_t *hs, cfs_hlist_node_t *hnode)
+{
+ /* dummy, required by cfs_hash */
+}
+
+static void
+filter_iobuf_hop_exit(cfs_hash_t *hs, cfs_hlist_node_t *hnode)
+{
+ struct filter_iobuf *pool;
+
+ pool = cfs_hlist_entry(hnode, struct filter_iobuf, dr_hlist);
+ filter_free_iobuf(pool);
+}
+
+static struct cfs_hash_ops filter_iobuf_hops = {
+ .hs_hash = filter_iobuf_hop_hash,
+ .hs_key = filter_iobuf_hop_key,
+ .hs_keycmp = filter_iobuf_hop_keycmp,
+ .hs_object = filter_iobuf_hop_object,
+ .hs_get = filter_iobuf_hop_get,
+ .hs_put_locked = filter_iobuf_hop_put_locked,
+ .hs_exit = filter_iobuf_hop_exit
+};
- return pool;
+#define FILTER_IOBUF_HASH_BITS 9
+#define FILTER_IOBUF_HBKT_BITS 4
+
+/*
+ * pre-allocate pool of iobuf's to be used by filter_{prep,commit}rw_write().
+ */
+static int filter_iobuf_pool_init(struct filter_obd *filter)
+{
+ filter->fo_iobuf_hash = cfs_hash_create("filter_iobuf",
+ FILTER_IOBUF_HASH_BITS,
+ FILTER_IOBUF_HASH_BITS,
+ FILTER_IOBUF_HBKT_BITS, 0,
+ CFS_HASH_MIN_THETA,
+ CFS_HASH_MAX_THETA,
+ &filter_iobuf_hops,
+ CFS_HASH_RW_BKTLOCK |
+ CFS_HASH_NO_ITEMREF);
+
+ return filter->fo_iobuf_hash != NULL ? 0 : -ENOMEM;
+}
+
+/* Return iobuf allocated for @thread_id.
+ * If we haven't allocated a pool entry for this thread before, do so now and
+ * insert it into fo_iobuf_hash, otherwise we can find it from fo_iobuf_hash */
+void *filter_iobuf_get(struct filter_obd *filter, struct obd_trans_info *oti)
+{
+ struct filter_iobuf *pool = NULL;
+ __u64 key = 0;
+ int thread_id;
+ int rc;
+
+ thread_id = (oti && oti->oti_thread) ? oti->oti_thread->t_id : -1;
+ if (thread_id >= 0) {
+ struct ptlrpc_service_part *svcpt;
+
+ svcpt = oti->oti_thread->t_svcpt;
+ LASSERT(svcpt != NULL);
+
+ key = (__u64)(svcpt->scp_cpt) << 32 | thread_id;
+ pool = cfs_hash_lookup(filter->fo_iobuf_hash, &key);
+ if (pool != NULL)
+ return pool;
+ }
+
+ pool = filter_alloc_iobuf(filter, OBD_BRW_WRITE, PTLRPC_MAX_BRW_PAGES);
+ if (pool == NULL)
+ return NULL;
+
+ if (thread_id >= 0) {
+ pool->dr_hkey = key;
+ rc = cfs_hash_add_unique(filter->fo_iobuf_hash,
+ &key, &pool->dr_hlist);
+ /* ptlrpc service thould guarantee thread ID is unique */
+ LASSERT(rc != -EALREADY);
+ }
+
+ return pool;
}
/* mount the file system (secretly). lustre_cfg parameters are:
* We actually do sync in disconnect time, but disconnect
* may not come being marked rq_no_resend = 1.
*/
- llog_sync(ctxt, NULL);
+ llog_sync(ctxt, NULL, OBD_LLOG_FL_EXIT);
/*
* Balance class_import_get() in llog_receptor_accept().
class_import_put(ctxt->loc_imp);
ctxt->loc_imp = NULL;
}
+
+ if (filter->fo_lcm) {
+ llog_recov_thread_fini(filter->fo_lcm, obd->obd_force);
+ filter->fo_lcm = NULL;
+ }
+
cfs_mutex_unlock(&ctxt->loc_mutex);
llog_ctxt_put(ctxt);
}
- if (filter->fo_lcm) {
- cfs_mutex_lock(&ctxt->loc_mutex);
- llog_recov_thread_fini(filter->fo_lcm, obd->obd_force);
- filter->fo_lcm = NULL;
- cfs_mutex_unlock(&ctxt->loc_mutex);
- }
RETURN(filter_olg_fini(&obd->obd_olg));
}
/* The client set in ocd_cksum_types the checksum types it
* supports. We have to mask off the algorithms that we don't
* support */
- data->ocd_cksum_types &= cksum_types_supported();
+ data->ocd_cksum_types &= cksum_types_supported_server();
- /* 1.6.4- only support CRC32 and didn't set ocd_cksum_types */
- if (unlikely(data->ocd_cksum_types == 0))
- data->ocd_cksum_types = OBD_CKSUM_CRC32;
+ /* 1.6.4 clients are not supported any more */
CDEBUG(D_RPCTRACE, "%s: cli %s supports cksum type %x, return "
"%x\n", exp->exp_obd->obd_name,
ctxt = llog_group_get_ctxt(olg_min,
LLOG_MDS_OST_REPL_CTXT);
if (ctxt) {
- err = llog_sync(ctxt, olg_min->olg_exp);
+ err = llog_sync(ctxt, olg_min->olg_exp, 0);
llog_ctxt_put(ctxt);
if (err) {
CERROR("error flushing logs to MDS: "
RETURN(ERR_PTR(-ENOENT));
}
-#if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2,7,50,0)
+#if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2, 7, 50, 0)
/* Try to correct for a bug in 2.1.0 (LU-221) that caused negative
* timestamps to appear to be in the far future, due old timestamp
* being stored on disk as an unsigned value. This fixes up any
if (fcc != NULL)
*fcc = oa->o_lcookie;
}
- if (ia_valid & (ATTR_SIZE | ATTR_UID | ATTR_GID)) {
- unsigned long now = jiffies;
- /* Filter truncates and writes are serialized by
- * i_alloc_sem, see the comment in
- * filter_preprw_write.*/
- if (ia_valid & ATTR_SIZE)
- down_write(&inode->i_alloc_sem);
- LOCK_INODE_MUTEX(inode);
- fsfilt_check_slow(exp->exp_obd, now, "i_alloc_sem and i_mutex");
- old_size = i_size_read(inode);
- }
+ if (ia_valid & (ATTR_SIZE | ATTR_UID | ATTR_GID)) {
+ unsigned long now = jiffies;
+ /* Filter truncates and writes are serialized by
+ * i_alloc_sem, see the comment in
+ * filter_preprw_write.*/
+ if (ia_valid & ATTR_SIZE)
+ down_write(&inode->i_alloc_sem);
+ mutex_lock(&inode->i_mutex);
+ fsfilt_check_slow(exp->exp_obd, now, "i_alloc_sem and i_mutex");
+ old_size = i_size_read(inode);
+ }
/* VBR: version recovery check */
rc = filter_version_get_check(exp, oti, inode);
if (page)
page_cache_release(page);
- if (ia_valid & (ATTR_SIZE | ATTR_UID | ATTR_GID))
- UNLOCK_INODE_MUTEX(inode);
- if (ia_valid & ATTR_SIZE)
- up_write(&inode->i_alloc_sem);
- if (fcc)
- OBD_FREE(fcc, sizeof(*fcc));
+ if (ia_valid & (ATTR_SIZE | ATTR_UID | ATTR_GID))
+ mutex_unlock(&inode->i_mutex);
+ if (ia_valid & ATTR_SIZE)
+ up_write(&inode->i_alloc_sem);
+ if (fcc)
+ OBD_FREE(fcc, sizeof(*fcc));
/* trigger quota release */
if (ia_valid & (ATTR_SIZE | ATTR_UID | ATTR_GID)) {
CDEBUG(D_RPCTRACE, "filter_last_id() = "LPU64" -> diff = %d\n",
filter_last_id(filter, group), diff);
- LASSERTF(diff >= 0,"%s: "LPU64" - "LPU64" = %d\n",obd->obd_name,
- oa->o_id, filter_last_id(filter, group), diff);
+ /*
+ * Check obd->obd_recovering to handle the race condition
+ * while recreating missing precreated objects through
+ * filter_preprw_write() and mds_lov_clear_orphans()
+ * at the same time.
+ */
+ LASSERTF(ergo(!obd->obd_recovering, diff >= 0),
+ "%s: "LPU64" - "LPU64" = %d\n", obd->obd_name,
+ oa->o_id, filter_last_id(filter, group), diff);
}
if (diff > 0) {
}
fcc = &oa->o_lcookie;
ctxt = llog_group_get_ctxt(olg, fcc->lgc_subsys + 1);
- llog_cancel(ctxt, NULL, 1, fcc, 0);
+ llog_cancel(NULL, ctxt, NULL, 1, fcc, 0);
llog_ctxt_put(ctxt);
fcc = NULL; /* we didn't allocate fcc, don't free it */
}
* between page lock, i_mutex & starting new journal handle.
* (see bug 20321) -johann
*/
- now = jiffies;
- down_write(&dchild->d_inode->i_alloc_sem);
- LOCK_INODE_MUTEX(dchild->d_inode);
- fsfilt_check_slow(exp->exp_obd, now, "i_alloc_sem and i_mutex");
-
- /* VBR: version recovery check */
- rc = filter_version_get_check(exp, oti, dchild->d_inode);
- if (rc) {
- UNLOCK_INODE_MUTEX(dchild->d_inode);
- up_write(&dchild->d_inode->i_alloc_sem);
- GOTO(cleanup, rc);
- }
-
- handle = fsfilt_start_log(obd, dchild->d_inode, FSFILT_OP_SETATTR,
- NULL, 1);
- if (IS_ERR(handle)) {
- UNLOCK_INODE_MUTEX(dchild->d_inode);
- up_write(&dchild->d_inode->i_alloc_sem);
- GOTO(cleanup, rc = PTR_ERR(handle));
- }
-
- /* Locking order: i_mutex -> journal_lock -> dqptr_sem. LU-952 */
- ll_vfs_dq_init(dchild->d_inode);
-
- iattr.ia_valid = ATTR_SIZE;
- iattr.ia_size = 0;
- rc = fsfilt_setattr(obd, dchild, handle, &iattr, 1);
- rc2 = fsfilt_commit(obd, dchild->d_inode, handle, 0);
- UNLOCK_INODE_MUTEX(dchild->d_inode);
- up_write(&dchild->d_inode->i_alloc_sem);
- if (rc)
- GOTO(cleanup, rc);
- if (rc2)
- GOTO(cleanup, rc = rc2);
+ now = jiffies;
+ down_write(&dchild->d_inode->i_alloc_sem);
+ mutex_lock(&dchild->d_inode->i_mutex);
+ fsfilt_check_slow(exp->exp_obd, now, "i_alloc_sem and i_mutex");
+
+ /* VBR: version recovery check */
+ rc = filter_version_get_check(exp, oti, dchild->d_inode);
+ if (rc) {
+ mutex_unlock(&dchild->d_inode->i_mutex);
+ up_write(&dchild->d_inode->i_alloc_sem);
+ GOTO(cleanup, rc);
+ }
+
+ handle = fsfilt_start_log(obd, dchild->d_inode, FSFILT_OP_SETATTR,
+ NULL, 1);
+ if (IS_ERR(handle)) {
+ mutex_unlock(&dchild->d_inode->i_mutex);
+ up_write(&dchild->d_inode->i_alloc_sem);
+ GOTO(cleanup, rc = PTR_ERR(handle));
+ }
+
+ /* Locking order: i_mutex -> journal_lock -> dqptr_sem. LU-952 */
+ ll_vfs_dq_init(dchild->d_inode);
+
+ iattr.ia_valid = ATTR_SIZE;
+ iattr.ia_size = 0;
+ rc = fsfilt_setattr(obd, dchild, handle, &iattr, 1);
+ rc2 = fsfilt_commit(obd, dchild->d_inode, handle, 0);
+ mutex_unlock(&dchild->d_inode->i_mutex);
+ up_write(&dchild->d_inode->i_alloc_sem);
+ if (rc)
+ GOTO(cleanup, rc);
+ if (rc2)
+ GOTO(cleanup, rc = rc2);
/* We don't actually need to lock the parent until we are unlinking
* here, and not while truncating above. That avoids holding the
GOTO(cleanup, rc = PTR_ERR(dparent));
cleanup_phase = 3; /* filter_parent_unlock */
- LOCK_INODE_MUTEX(dchild->d_inode);
- handle = fsfilt_start_log(obd, dparent->d_inode,FSFILT_OP_UNLINK,oti,1);
- if (IS_ERR(handle)) {
- UNLOCK_INODE_MUTEX(dchild->d_inode);
- GOTO(cleanup, rc = PTR_ERR(handle));
- }
- cleanup_phase = 4; /* fsfilt_commit */
+ mutex_lock(&dchild->d_inode->i_mutex);
+ handle = fsfilt_start_log(obd, dparent->d_inode,
+ FSFILT_OP_UNLINK, oti, 1);
+ if (IS_ERR(handle)) {
+ mutex_unlock(&dchild->d_inode->i_mutex);
+ GOTO(cleanup, rc = PTR_ERR(handle));
+ }
+ cleanup_phase = 4; /* fsfilt_commit */
/* Quota release need uid/gid of inode */
obdo_from_inode(oa, dchild->d_inode, OBD_MD_FLUID | OBD_MD_FLGID);
push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
- LOCK_INODE_MUTEX(dentry->d_inode);
+ mutex_lock(&dentry->d_inode->i_mutex);
- rc = filemap_fdatawrite(dentry->d_inode->i_mapping);
- if (rc == 0) {
- /* just any file to grab fsync method - "file" arg unused */
- struct file *file = obt->obt_rcvd_filp;
+ rc = filemap_fdatawrite(dentry->d_inode->i_mapping);
+ if (rc == 0) {
+ /* just any file to grab fsync method - "file" arg unused */
+ struct file *file = obt->obt_rcvd_filp;
- if (file->f_op && file->f_op->fsync)
- rc = file->f_op->fsync(NULL, dentry, 1);
+ if (file->f_op && file->f_op->fsync)
+ rc = file->f_op->fsync(NULL, dentry, 1);
- rc2 = filemap_fdatawait(dentry->d_inode->i_mapping);
- if (!rc)
- rc = rc2;
- }
- UNLOCK_INODE_MUTEX(dentry->d_inode);
+ rc2 = filemap_fdatawait(dentry->d_inode->i_mapping);
+ if (!rc)
+ rc = rc2;
+ }
+ mutex_unlock(&dentry->d_inode->i_mutex);
oinfo->oi_oa->o_valid = OBD_MD_FLID;
obdo_from_inode(oinfo->oi_oa, dentry->d_inode, FILTER_VALID_FLAGS);
git://git.whamcloud.com / fs / lustre-release.git / blobdiff