Whamcloud - gitweb
LU-8472 scrub: try to avoid recovery during OI scrub
[fs/lustre-release.git] / lustre / target / tgt_main.c
index a920716..e32e558 100644 (file)
@@ -21,7 +21,7 @@
  * GPL HEADER END
  */
 /*
- * Copyright (c) 2012, 2014, Intel Corporation.
+ * Copyright (c) 2012, 2015, Intel Corporation.
  */
 /*
  * lustre/target/tgt_main.c
 #include "tgt_internal.h"
 #include "../ptlrpc/ptlrpc_internal.h"
 
+static spinlock_t uncommitted_slc_locks_guard;
+static struct list_head uncommitted_slc_locks;
+
+/*
+ * Save cross-MDT lock in uncommitted_slc_locks.
+ *
+ * Lock R/W count is not saved, but released in unlock (not canceled remotely),
+ * instead only a refcount is taken, so that the remote MDT where the object
+ * resides can detect conflict with this lock there.
+ *
+ * \param lock cross-MDT lock to save
+ * \param transno when the transaction with this transno is committed, this lock
+ *               can be canceled.
+ */
+void tgt_save_slc_lock(struct ldlm_lock *lock, __u64 transno)
+{
+       spin_lock(&uncommitted_slc_locks_guard);
+       lock_res_and_lock(lock);
+       if (ldlm_is_cbpending(lock)) {
+               /* if it was canceld by server, don't save, because remote MDT
+                * will do Sync-on-Cancel. */
+               LDLM_LOCK_PUT(lock);
+       } else {
+               lock->l_transno = transno;
+               /* if this lock is in the list already, there are two operations
+                * both use this lock, and save it after use, so for the second
+                * one, just put the refcount. */
+               if (list_empty(&lock->l_slc_link))
+                       list_add_tail(&lock->l_slc_link,
+                                     &uncommitted_slc_locks);
+               else
+                       LDLM_LOCK_PUT(lock);
+       }
+       unlock_res_and_lock(lock);
+       spin_unlock(&uncommitted_slc_locks_guard);
+}
+EXPORT_SYMBOL(tgt_save_slc_lock);
+
+/*
+ * Discard cross-MDT lock from uncommitted_slc_locks.
+ *
+ * This is called upon BAST, just remove lock from uncommitted_slc_locks and put
+ * lock refcount. The BAST will cancel this lock.
+ *
+ * \param lock cross-MDT lock to discard
+ */
+void tgt_discard_slc_lock(struct ldlm_lock *lock)
+{
+       spin_lock(&uncommitted_slc_locks_guard);
+       lock_res_and_lock(lock);
+       /* may race with tgt_cancel_slc_locks() */
+       if (lock->l_transno != 0) {
+               LASSERT(!list_empty(&lock->l_slc_link));
+               LASSERT(ldlm_is_cbpending(lock));
+               list_del_init(&lock->l_slc_link);
+               lock->l_transno = 0;
+               LDLM_LOCK_PUT(lock);
+       }
+       unlock_res_and_lock(lock);
+       spin_unlock(&uncommitted_slc_locks_guard);
+}
+EXPORT_SYMBOL(tgt_discard_slc_lock);
+
+/*
+ * Cancel cross-MDT locks upon transaction commit.
+ *
+ * Remove cross-MDT locks from uncommitted_slc_locks, cancel them and put lock
+ * refcount.
+ *
+ * \param transno transaction with this number was committed.
+ */
+void tgt_cancel_slc_locks(__u64 transno)
+{
+       struct ldlm_lock *lock, *next;
+       LIST_HEAD(list);
+       struct lustre_handle lockh;
+
+       spin_lock(&uncommitted_slc_locks_guard);
+       list_for_each_entry_safe(lock, next, &uncommitted_slc_locks,
+                                l_slc_link) {
+               lock_res_and_lock(lock);
+               LASSERT(lock->l_transno != 0);
+               if (lock->l_transno > transno) {
+                       unlock_res_and_lock(lock);
+                       continue;
+               }
+               /* ouch, another operation is using it after it's saved */
+               if (lock->l_readers != 0 || lock->l_writers != 0) {
+                       unlock_res_and_lock(lock);
+                       continue;
+               }
+               /* set CBPENDING so that this lock won't be used again */
+               ldlm_set_cbpending(lock);
+               lock->l_transno = 0;
+               list_move(&lock->l_slc_link, &list);
+               unlock_res_and_lock(lock);
+       }
+       spin_unlock(&uncommitted_slc_locks_guard);
+
+       list_for_each_entry_safe(lock, next, &list, l_slc_link) {
+               list_del_init(&lock->l_slc_link);
+               ldlm_lock2handle(lock, &lockh);
+               ldlm_cli_cancel(&lockh, LCF_ASYNC);
+               LDLM_LOCK_PUT(lock);
+       }
+}
+
 int tgt_init(const struct lu_env *env, struct lu_target *lut,
             struct obd_device *obd, struct dt_device *dt,
             struct tgt_opc_slice *slice, int request_fail_id,
@@ -46,7 +153,7 @@ int tgt_init(const struct lu_env *env, struct lu_target *lut,
        struct lu_attr           attr;
        struct lu_fid            fid;
        struct dt_object        *o;
-       int                      rc = 0;
+       int i, rc = 0;
 
        ENTRY;
 
@@ -80,6 +187,7 @@ int tgt_init(const struct lu_env *env, struct lu_target *lut,
                RETURN(0);
 
        spin_lock_init(&lut->lut_translock);
+       spin_lock_init(&lut->lut_client_bitmap_lock);
 
        OBD_ALLOC(lut->lut_client_bitmap, LR_MAX_CLIENTS >> 3);
        if (lut->lut_client_bitmap == NULL)
@@ -97,13 +205,13 @@ int tgt_init(const struct lu_env *env, struct lu_target *lut,
                rc = PTR_ERR(o);
                CERROR("%s: cannot open LAST_RCVD: rc = %d\n", tgt_name(lut),
                       rc);
-               GOTO(out, rc);
+               GOTO(out_put, rc);
        }
 
        lut->lut_last_rcvd = o;
        rc = tgt_server_data_init(env, lut);
        if (rc < 0)
-               GOTO(out, rc);
+               GOTO(out_put, rc);
 
        /* prepare transactions callbacks */
        lut->lut_txn_cb.dtc_txn_start = tgt_txn_start_cb;
@@ -123,7 +231,7 @@ int tgt_init(const struct lu_env *env, struct lu_target *lut,
        OBD_ALLOC(lut->lut_reply_bitmap,
                  LUT_REPLY_SLOTS_MAX_CHUNKS * sizeof(unsigned long *));
        if (lut->lut_reply_bitmap == NULL)
-               GOTO(out, rc);
+               GOTO(out, rc = -ENOMEM);
 
        memset(&attr, 0, sizeof(attr));
        attr.la_valid = LA_MODE;
@@ -145,22 +253,36 @@ int tgt_init(const struct lu_env *env, struct lu_target *lut,
        if (rc < 0)
                GOTO(out, rc);
 
+       atomic_set(&lut->lut_sync_count, 0);
+
        RETURN(0);
+
 out:
+       dt_txn_callback_del(lut->lut_bottom, &lut->lut_txn_cb);
+out_put:
+       obd->u.obt.obt_magic = 0;
+       obd->u.obt.obt_lut = NULL;
        if (lut->lut_last_rcvd != NULL) {
                lu_object_put(env, &lut->lut_last_rcvd->do_lu);
-               dt_txn_callback_del(lut->lut_bottom, &lut->lut_txn_cb);
+               lut->lut_last_rcvd = NULL;
        }
-       lut->lut_last_rcvd = NULL;
        if (lut->lut_client_bitmap != NULL)
                OBD_FREE(lut->lut_client_bitmap, LR_MAX_CLIENTS >> 3);
        lut->lut_client_bitmap = NULL;
        if (lut->lut_reply_data != NULL)
                lu_object_put(env, &lut->lut_reply_data->do_lu);
        lut->lut_reply_data = NULL;
-       if (lut->lut_reply_bitmap != NULL)
+       if (lut->lut_reply_bitmap != NULL) {
+               for (i = 0; i < LUT_REPLY_SLOTS_MAX_CHUNKS; i++) {
+                       if (lut->lut_reply_bitmap[i] != NULL)
+                               OBD_FREE_LARGE(lut->lut_reply_bitmap[i],
+                                   BITS_TO_LONGS(LUT_REPLY_SLOTS_PER_CHUNK) *
+                                   sizeof(long));
+                       lut->lut_reply_bitmap[i] = NULL;
+               }
                OBD_FREE(lut->lut_reply_bitmap,
                         LUT_REPLY_SLOTS_MAX_CHUNKS * sizeof(unsigned long *));
+       }
        lut->lut_reply_bitmap = NULL;
        return rc;
 }
@@ -192,7 +314,7 @@ void tgt_fini(const struct lu_env *env, struct lu_target *lut)
        if (lut->lut_reply_bitmap != NULL) {
                for (i = 0; i < LUT_REPLY_SLOTS_MAX_CHUNKS; i++) {
                        if (lut->lut_reply_bitmap[i] != NULL)
-                               OBD_FREE(lut->lut_reply_bitmap[i],
+                               OBD_FREE_LARGE(lut->lut_reply_bitmap[i],
                                    BITS_TO_LONGS(LUT_REPLY_SLOTS_PER_CHUNK) *
                                    sizeof(long));
                        lut->lut_reply_bitmap[i] = NULL;
@@ -280,7 +402,7 @@ int tgt_mod_init(void)
 {
        ENTRY;
 
-       tgt_page_to_corrupt = alloc_page(GFP_IOFS);
+       tgt_page_to_corrupt = alloc_page(GFP_KERNEL);
 
        tgt_key_init_generic(&tgt_thread_key, NULL);
        lu_context_key_register_many(&tgt_thread_key, NULL);
@@ -290,13 +412,16 @@ int tgt_mod_init(void)
 
        update_info_init();
 
+       spin_lock_init(&uncommitted_slc_locks_guard);
+       INIT_LIST_HEAD(&uncommitted_slc_locks);
+
        RETURN(0);
 }
 
 void tgt_mod_exit(void)
 {
        if (tgt_page_to_corrupt != NULL)
-               page_cache_release(tgt_page_to_corrupt);
+               put_page(tgt_page_to_corrupt);
 
        lu_context_key_degister(&tgt_thread_key);
        lu_context_key_degister(&tgt_session_key);