jbd2: Add buffer triggers
authorJoel Becker <joel.becker@oracle.com>
Thu, 11 Sep 2008 22:35:47 +0000 (15:35 -0700)
committerMark Fasheh <mfasheh@suse.com>
Mon, 5 Jan 2009 16:40:30 +0000 (08:40 -0800)
Filesystems often to do compute intensive operation on some
metadata.  If this operation is repeated many times, it can be very
expensive.  It would be much nicer if the operation could be performed
once before a buffer goes to disk.

This adds triggers to jbd2 buffer heads.  Just before writing a metadata
buffer to the journal, jbd2 will optionally call a commit trigger associated
with the buffer.  If the journal is aborted, an abort trigger will be
called on any dirty buffers as they are dropped from pending
transactions.

ocfs2 will use this feature.

Initially I tried to come up with a more generic trigger that could be
used for non-buffer-related events like transaction completion.  It
doesn't tie nicely, because the information a buffer trigger needs
(specific to a journal_head) isn't the same as what a transaction
trigger needs (specific to a tranaction_t or perhaps journal_t).  So I
implemented a buffer set, with the understanding that
journal/transaction wide triggers should be implemented separately.

There is only one trigger set allowed per buffer.  I can't think of any
reason to attach more than one set.  Contrast this with a journal or
transaction in which multiple places may want to watch the entire
transaction separately.

The trigger sets are considered static allocation from the jbd2
perspective.  ocfs2 will just have one trigger set per block type,
setting the same set on every bh of the same type.

Signed-off-by: Joel Becker <joel.becker@oracle.com>
Cc: "Theodore Ts'o" <tytso@mit.edu>
Cc: <linux-ext4@vger.kernel.org>
Signed-off-by: Mark Fasheh <mfasheh@suse.com>
fs/jbd2/commit.c
fs/jbd2/journal.c
fs/jbd2/transaction.c
include/linux/jbd2.h
include/linux/journal-head.h

index ebc667bc54a8e1ac3c28e8ab6a40ce83f5274ad9..c8a1bace685a83b1b6944de6e7e5ad13faa4f61b 100644 (file)
@@ -509,6 +509,10 @@ void jbd2_journal_commit_transaction(journal_t *journal)
                if (is_journal_aborted(journal)) {
                        clear_buffer_jbddirty(jh2bh(jh));
                        JBUFFER_TRACE(jh, "journal is aborting: refile");
+                       jbd2_buffer_abort_trigger(jh,
+                                                 jh->b_frozen_data ?
+                                                 jh->b_frozen_triggers :
+                                                 jh->b_triggers);
                        jbd2_journal_refile_buffer(journal, jh);
                        /* If that was the last one, we need to clean up
                         * any descriptor buffers which may have been
@@ -844,6 +848,9 @@ restart_loop:
                 * data.
                 *
                 * Otherwise, we can just throw away the frozen data now.
+                *
+                * We also know that the frozen data has already fired
+                * its triggers if they exist, so we can clear that too.
                 */
                if (jh->b_committed_data) {
                        jbd2_free(jh->b_committed_data, bh->b_size);
@@ -851,10 +858,12 @@ restart_loop:
                        if (jh->b_frozen_data) {
                                jh->b_committed_data = jh->b_frozen_data;
                                jh->b_frozen_data = NULL;
+                               jh->b_frozen_triggers = NULL;
                        }
                } else if (jh->b_frozen_data) {
                        jbd2_free(jh->b_frozen_data, bh->b_size);
                        jh->b_frozen_data = NULL;
+                       jh->b_frozen_triggers = NULL;
                }
 
                spin_lock(&journal->j_list_lock);
index e70d657a19f810ad5f3c714e6bfff6e5303d226d..f6bff9d6f8df8193e6fc13f1a2b1763124d29cb6 100644 (file)
@@ -50,6 +50,7 @@ EXPORT_SYMBOL(jbd2_journal_unlock_updates);
 EXPORT_SYMBOL(jbd2_journal_get_write_access);
 EXPORT_SYMBOL(jbd2_journal_get_create_access);
 EXPORT_SYMBOL(jbd2_journal_get_undo_access);
+EXPORT_SYMBOL(jbd2_journal_set_triggers);
 EXPORT_SYMBOL(jbd2_journal_dirty_metadata);
 EXPORT_SYMBOL(jbd2_journal_release_buffer);
 EXPORT_SYMBOL(jbd2_journal_forget);
@@ -290,6 +291,7 @@ int jbd2_journal_write_metadata_buffer(transaction_t *transaction,
        struct page *new_page;
        unsigned int new_offset;
        struct buffer_head *bh_in = jh2bh(jh_in);
+       struct jbd2_buffer_trigger_type *triggers;
 
        /*
         * The buffer really shouldn't be locked: only the current committing
@@ -314,12 +316,22 @@ repeat:
                done_copy_out = 1;
                new_page = virt_to_page(jh_in->b_frozen_data);
                new_offset = offset_in_page(jh_in->b_frozen_data);
+               triggers = jh_in->b_frozen_triggers;
        } else {
                new_page = jh2bh(jh_in)->b_page;
                new_offset = offset_in_page(jh2bh(jh_in)->b_data);
+               triggers = jh_in->b_triggers;
        }
 
        mapped_data = kmap_atomic(new_page, KM_USER0);
+       /*
+        * Fire any commit trigger.  Do this before checking for escaping,
+        * as the trigger may modify the magic offset.  If a copy-out
+        * happens afterwards, it will have the correct data in the buffer.
+        */
+       jbd2_buffer_commit_trigger(jh_in, mapped_data + new_offset,
+                                  triggers);
+
        /*
         * Check for escaping
         */
@@ -352,6 +364,13 @@ repeat:
                new_page = virt_to_page(tmp);
                new_offset = offset_in_page(tmp);
                done_copy_out = 1;
+
+               /*
+                * This isn't strictly necessary, as we're using frozen
+                * data for the escaping, but it keeps consistency with
+                * b_frozen_data usage.
+                */
+               jh_in->b_frozen_triggers = jh_in->b_triggers;
        }
 
        /*
index 39b7805a599a79fd80c0150073cbf0f473fe8bcf..4f925a4f3d05051ec7b1edd65c50ecdabe27000a 100644 (file)
@@ -741,6 +741,12 @@ done:
                source = kmap_atomic(page, KM_USER0);
                memcpy(jh->b_frozen_data, source+offset, jh2bh(jh)->b_size);
                kunmap_atomic(source, KM_USER0);
+
+               /*
+                * Now that the frozen data is saved off, we need to store
+                * any matching triggers.
+                */
+               jh->b_frozen_triggers = jh->b_triggers;
        }
        jbd_unlock_bh_state(bh);
 
@@ -943,6 +949,47 @@ out:
        return err;
 }
 
+/**
+ * void jbd2_journal_set_triggers() - Add triggers for commit writeout
+ * @bh: buffer to trigger on
+ * @type: struct jbd2_buffer_trigger_type containing the trigger(s).
+ *
+ * Set any triggers on this journal_head.  This is always safe, because
+ * triggers for a committing buffer will be saved off, and triggers for
+ * a running transaction will match the buffer in that transaction.
+ *
+ * Call with NULL to clear the triggers.
+ */
+void jbd2_journal_set_triggers(struct buffer_head *bh,
+                              struct jbd2_buffer_trigger_type *type)
+{
+       struct journal_head *jh = bh2jh(bh);
+
+       jh->b_triggers = type;
+}
+
+void jbd2_buffer_commit_trigger(struct journal_head *jh, void *mapped_data,
+                               struct jbd2_buffer_trigger_type *triggers)
+{
+       struct buffer_head *bh = jh2bh(jh);
+
+       if (!triggers || !triggers->t_commit)
+               return;
+
+       triggers->t_commit(triggers, bh, mapped_data, bh->b_size);
+}
+
+void jbd2_buffer_abort_trigger(struct journal_head *jh,
+                              struct jbd2_buffer_trigger_type *triggers)
+{
+       if (!triggers || !triggers->t_abort)
+               return;
+
+       triggers->t_abort(triggers, jh2bh(jh));
+}
+
+
+
 /**
  * int jbd2_journal_dirty_metadata() -  mark a buffer as containing dirty metadata
  * @handle: transaction to add buffer to.
index f36645745489313cc09e21edfdc7f6d11bfde9a8..34456476e761330686a4e6d6afa98cf570fc94de 100644 (file)
@@ -1008,6 +1008,35 @@ int __jbd2_journal_clean_checkpoint_list(journal_t *journal);
 int __jbd2_journal_remove_checkpoint(struct journal_head *);
 void __jbd2_journal_insert_checkpoint(struct journal_head *, transaction_t *);
 
+
+/*
+ * Triggers
+ */
+
+struct jbd2_buffer_trigger_type {
+       /*
+        * Fired just before a buffer is written to the journal.
+        * mapped_data is a mapped buffer that is the frozen data for
+        * commit.
+        */
+       void (*t_commit)(struct jbd2_buffer_trigger_type *type,
+                        struct buffer_head *bh, void *mapped_data,
+                        size_t size);
+
+       /*
+        * Fired during journal abort for dirty buffers that will not be
+        * committed.
+        */
+       void (*t_abort)(struct jbd2_buffer_trigger_type *type,
+                       struct buffer_head *bh);
+};
+
+extern void jbd2_buffer_commit_trigger(struct journal_head *jh,
+                                      void *mapped_data,
+                                      struct jbd2_buffer_trigger_type *triggers);
+extern void jbd2_buffer_abort_trigger(struct journal_head *jh,
+                                     struct jbd2_buffer_trigger_type *triggers);
+
 /* Buffer IO */
 extern int
 jbd2_journal_write_metadata_buffer(transaction_t         *transaction,
@@ -1046,6 +1075,8 @@ extern int         jbd2_journal_extend (handle_t *, int nblocks);
 extern int      jbd2_journal_get_write_access(handle_t *, struct buffer_head *);
 extern int      jbd2_journal_get_create_access (handle_t *, struct buffer_head *);
 extern int      jbd2_journal_get_undo_access(handle_t *, struct buffer_head *);
+void            jbd2_journal_set_triggers(struct buffer_head *,
+                                          struct jbd2_buffer_trigger_type *type);
 extern int      jbd2_journal_dirty_metadata (handle_t *, struct buffer_head *);
 extern void     jbd2_journal_release_buffer (handle_t *, struct buffer_head *);
 extern int      jbd2_journal_forget (handle_t *, struct buffer_head *);
index bb70ebb6a2d5bd9f7435387bb2eb4fa91cbce166..525aac3c97dfeca86b1a6801d07aec59c9e8599d 100644 (file)
@@ -12,6 +12,8 @@
 
 typedef unsigned int           tid_t;          /* Unique transaction ID */
 typedef struct transaction_s   transaction_t;  /* Compound transaction type */
+
+
 struct buffer_head;
 
 struct journal_head {
@@ -87,6 +89,12 @@ struct journal_head {
         * [j_list_lock]
         */
        struct journal_head *b_cpnext, *b_cpprev;
+
+       /* Trigger type */
+       struct jbd2_buffer_trigger_type *b_triggers;
+
+       /* Trigger type for the committing transaction's frozen data */
+       struct jbd2_buffer_trigger_type *b_frozen_triggers;
 };
 
 #endif         /* JOURNAL_HEAD_H_INCLUDED */