fs-writeback: Add helper function to start writeback if idle
[safe/jmp/linux-2.6] / fs / gfs2 / log.c
index 1e1fe8d..4511b08 100644 (file)
@@ -1,6 +1,6 @@
 /*
  * Copyright (C) Sistina Software, Inc.  1997-2003 All rights reserved.
- * Copyright (C) 2004-2006 Red Hat, Inc.  All rights reserved.
+ * Copyright (C) 2004-2007 Red Hat, Inc.  All rights reserved.
  *
  * This copyrighted material is made available to anyone wishing to use,
  * modify, copy, or redistribute it subject to the terms and conditions
 #include <linux/buffer_head.h>
 #include <linux/gfs2_ondisk.h>
 #include <linux/crc32.h>
-#include <linux/lm_interface.h>
 #include <linux/delay.h>
+#include <linux/kthread.h>
+#include <linux/freezer.h>
+#include <linux/bio.h>
 
 #include "gfs2.h"
 #include "incore.h"
@@ -26,6 +28,7 @@
 #include "meta_io.h"
 #include "util.h"
 #include "dir.h"
+#include "trace_gfs2.h"
 
 #define PULL 1
 
@@ -85,13 +88,13 @@ void gfs2_remove_from_ail(struct gfs2_bufdata *bd)
  */
 
 static void gfs2_ail1_start_one(struct gfs2_sbd *sdp, struct gfs2_ail *ai)
+__releases(&sdp->sd_log_lock)
+__acquires(&sdp->sd_log_lock)
 {
        struct gfs2_bufdata *bd, *s;
        struct buffer_head *bh;
        int retry;
 
-       BUG_ON(!spin_is_locked(&sdp->sd_log_lock));
-
        do {
                retry = 0;
 
@@ -118,7 +121,7 @@ static void gfs2_ail1_start_one(struct gfs2_sbd *sdp, struct gfs2_ail *ai)
                        lock_buffer(bh);
                        if (test_clear_buffer_dirty(bh)) {
                                bh->b_end_io = end_buffer_write_sync;
-                               submit_bh(WRITE, bh);
+                               submit_bh(WRITE_SYNC_PLUG, bh);
                        } else {
                                unlock_buffer(bh);
                                brelse(bh);
@@ -208,7 +211,7 @@ static void gfs2_ail1_start(struct gfs2_sbd *sdp, int flags)
        gfs2_log_unlock(sdp);
 }
 
-int gfs2_ail1_empty(struct gfs2_sbd *sdp, int flags)
+static int gfs2_ail1_empty(struct gfs2_sbd *sdp, int flags)
 {
        struct gfs2_ail *ai, *s;
        int ret;
@@ -301,7 +304,7 @@ int gfs2_log_reserve(struct gfs2_sbd *sdp, unsigned int blks)
 
        mutex_lock(&sdp->sd_log_reserve_mutex);
        gfs2_log_lock(sdp);
-       while(sdp->sd_log_blks_free <= (blks + reserved_blks)) {
+       while(atomic_read(&sdp->sd_log_blks_free) <= (blks + reserved_blks)) {
                gfs2_log_unlock(sdp);
                gfs2_ail1_empty(sdp, 0);
                gfs2_log_flush(sdp, NULL);
@@ -310,7 +313,8 @@ int gfs2_log_reserve(struct gfs2_sbd *sdp, unsigned int blks)
                        gfs2_ail1_start(sdp, 0);
                gfs2_log_lock(sdp);
        }
-       sdp->sd_log_blks_free -= blks;
+       atomic_sub(blks, &sdp->sd_log_blks_free);
+       trace_gfs2_log_blocks(sdp, -blks);
        gfs2_log_unlock(sdp);
        mutex_unlock(&sdp->sd_log_reserve_mutex);
 
@@ -330,27 +334,24 @@ void gfs2_log_release(struct gfs2_sbd *sdp, unsigned int blks)
 {
 
        gfs2_log_lock(sdp);
-       sdp->sd_log_blks_free += blks;
+       atomic_add(blks, &sdp->sd_log_blks_free);
+       trace_gfs2_log_blocks(sdp, blks);
        gfs2_assert_withdraw(sdp,
-                            sdp->sd_log_blks_free <= sdp->sd_jdesc->jd_blocks);
+                            atomic_read(&sdp->sd_log_blks_free) <= sdp->sd_jdesc->jd_blocks);
        gfs2_log_unlock(sdp);
        up_read(&sdp->sd_log_flush_lock);
 }
 
 static u64 log_bmap(struct gfs2_sbd *sdp, unsigned int lbn)
 {
-       struct inode *inode = sdp->sd_jdesc->jd_inode;
-       int error;
-       struct buffer_head bh_map = { .b_state = 0, .b_blocknr = 0 };
-
-       bh_map.b_size = 1 << inode->i_blkbits;
-       error = gfs2_block_map(inode, lbn, 0, &bh_map);
-       if (error || !bh_map.b_blocknr)
-               printk(KERN_INFO "error=%d, dbn=%llu lbn=%u", error,
-                      (unsigned long long)bh_map.b_blocknr, lbn);
-       gfs2_assert_withdraw(sdp, !error && bh_map.b_blocknr);
-
-       return bh_map.b_blocknr;
+       struct gfs2_journal_extent *je;
+
+       list_for_each_entry(je, &sdp->sd_jdesc->extent_list, extent_list) {
+               if (lbn >= je->lblock && lbn < je->lblock + je->blocks)
+                       return je->dblock + lbn - je->lblock;
+       }
+
+       return -1;
 }
 
 /**
@@ -559,8 +560,9 @@ static void log_pull_tail(struct gfs2_sbd *sdp, unsigned int new_tail)
        ail2_empty(sdp, new_tail);
 
        gfs2_log_lock(sdp);
-       sdp->sd_log_blks_free += dist;
-       gfs2_assert_withdraw(sdp, sdp->sd_log_blks_free <= sdp->sd_jdesc->jd_blocks);
+       atomic_add(dist, &sdp->sd_log_blks_free);
+       trace_gfs2_log_blocks(sdp, dist);
+       gfs2_assert_withdraw(sdp, atomic_read(&sdp->sd_log_blks_free) <= sdp->sd_jdesc->jd_blocks);
        gfs2_log_unlock(sdp);
 
        sdp->sd_log_tail = new_tail;
@@ -586,7 +588,6 @@ static void log_write_header(struct gfs2_sbd *sdp, u32 flags, int pull)
        memset(bh->b_data, 0, bh->b_size);
        set_buffer_uptodate(bh);
        clear_buffer_dirty(bh);
-       unlock_buffer(bh);
 
        gfs2_ail1_empty(sdp, 0);
        tail = current_tail(sdp);
@@ -595,7 +596,9 @@ static void log_write_header(struct gfs2_sbd *sdp, u32 flags, int pull)
        memset(lh, 0, sizeof(struct gfs2_log_header));
        lh->lh_header.mh_magic = cpu_to_be32(GFS2_MAGIC);
        lh->lh_header.mh_type = cpu_to_be32(GFS2_METATYPE_LH);
+       lh->lh_header.__pad0 = cpu_to_be64(0);
        lh->lh_header.mh_format = cpu_to_be32(GFS2_FORMAT_LH);
+       lh->lh_header.mh_jid = cpu_to_be32(sdp->sd_jdesc->jd_jid);
        lh->lh_sequence = cpu_to_be64(sdp->sd_log_sequence++);
        lh->lh_flags = cpu_to_be32(flags);
        lh->lh_tail = cpu_to_be32(tail);
@@ -603,8 +606,23 @@ static void log_write_header(struct gfs2_sbd *sdp, u32 flags, int pull)
        hash = gfs2_disk_hash(bh->b_data, sizeof(struct gfs2_log_header));
        lh->lh_hash = cpu_to_be32(hash);
 
-       set_buffer_dirty(bh);
-       if (sync_dirty_buffer(bh))
+       bh->b_end_io = end_buffer_write_sync;
+       if (test_bit(SDF_NOBARRIERS, &sdp->sd_flags))
+               goto skip_barrier;
+       get_bh(bh);
+       submit_bh(WRITE_SYNC | (1 << BIO_RW_BARRIER) | (1 << BIO_RW_META), bh);
+       wait_on_buffer(bh);
+       if (buffer_eopnotsupp(bh)) {
+               clear_buffer_eopnotsupp(bh);
+               set_buffer_uptodate(bh);
+               set_bit(SDF_NOBARRIERS, &sdp->sd_flags);
+               lock_buffer(bh);
+skip_barrier:
+               get_bh(bh);
+               submit_bh(WRITE_SYNC | (1 << BIO_RW_META), bh);
+               wait_on_buffer(bh);
+       }
+       if (!buffer_uptodate(bh))
                gfs2_io_error_bh(sdp, bh);
        brelse(bh);
 
@@ -652,7 +670,7 @@ static void gfs2_ordered_write(struct gfs2_sbd *sdp)
                lock_buffer(bh);
                if (buffer_mapped(bh) && test_clear_buffer_dirty(bh)) {
                        bh->b_end_io = end_buffer_write_sync;
-                       submit_bh(WRITE, bh);
+                       submit_bh(WRITE_SYNC_PLUG, bh);
                } else {
                        unlock_buffer(bh);
                        brelse(bh);
@@ -692,21 +710,18 @@ static void gfs2_ordered_wait(struct gfs2_sbd *sdp)
  *
  */
 
-void gfs2_log_flush(struct gfs2_sbd *sdp, struct gfs2_glock *gl)
+void __gfs2_log_flush(struct gfs2_sbd *sdp, struct gfs2_glock *gl)
 {
        struct gfs2_ail *ai;
 
        down_write(&sdp->sd_log_flush_lock);
 
-       if (gl) {
-               gfs2_log_lock(sdp);
-               if (list_empty(&gl->gl_le.le_list)) {
-                       gfs2_log_unlock(sdp);
-                       up_write(&sdp->sd_log_flush_lock);
-                       return;
-               }
-               gfs2_log_unlock(sdp);
+       /* Log might have been flushed while we waited for the flush lock */
+       if (gl && !test_bit(GLF_LFLUSH, &gl->gl_flags)) {
+               up_write(&sdp->sd_log_flush_lock);
+               return;
        }
+       trace_gfs2_log_flush(sdp, 1);
 
        ai = kzalloc(sizeof(struct gfs2_ail), GFP_NOFS | __GFP_NOFAIL);
        INIT_LIST_HEAD(&ai->ai_ail1_list);
@@ -737,7 +752,8 @@ void gfs2_log_flush(struct gfs2_sbd *sdp, struct gfs2_glock *gl)
                log_flush_commit(sdp);
        else if (sdp->sd_log_tail != current_tail(sdp) && !sdp->sd_log_idle){
                gfs2_log_lock(sdp);
-               sdp->sd_log_blks_free--; /* Adjust for unreserved buffer */
+               atomic_dec(&sdp->sd_log_blks_free); /* Adjust for unreserved buffer */
+               trace_gfs2_log_blocks(sdp, -1);
                gfs2_log_unlock(sdp);
                log_write_header(sdp, 0, PULL);
        }
@@ -755,8 +771,7 @@ void gfs2_log_flush(struct gfs2_sbd *sdp, struct gfs2_glock *gl)
                ai = NULL;
        }
        gfs2_log_unlock(sdp);
-
-       sdp->sd_vfs->s_dirt = 0;
+       trace_gfs2_log_flush(sdp, 0);
        up_write(&sdp->sd_log_flush_lock);
 
        kfree(ai);
@@ -765,7 +780,7 @@ void gfs2_log_flush(struct gfs2_sbd *sdp, struct gfs2_glock *gl)
 static void log_refund(struct gfs2_sbd *sdp, struct gfs2_trans *tr)
 {
        unsigned int reserved;
-       unsigned int old;
+       unsigned int unused;
 
        gfs2_log_lock(sdp);
 
@@ -777,19 +792,32 @@ static void log_refund(struct gfs2_sbd *sdp, struct gfs2_trans *tr)
        sdp->sd_log_commited_revoke += tr->tr_num_revoke - tr->tr_num_revoke_rm;
        gfs2_assert_withdraw(sdp, ((int)sdp->sd_log_commited_revoke) >= 0);
        reserved = calc_reserved(sdp);
-       old = sdp->sd_log_blks_free;
-       sdp->sd_log_blks_free += tr->tr_reserved -
-                                (reserved - sdp->sd_log_blks_reserved);
-
-       gfs2_assert_withdraw(sdp, sdp->sd_log_blks_free >= old);
-       gfs2_assert_withdraw(sdp, sdp->sd_log_blks_free <=
+       gfs2_assert_withdraw(sdp, sdp->sd_log_blks_reserved + tr->tr_reserved >= reserved);
+       unused = sdp->sd_log_blks_reserved - reserved + tr->tr_reserved;
+       atomic_add(unused, &sdp->sd_log_blks_free);
+       trace_gfs2_log_blocks(sdp, unused);
+       gfs2_assert_withdraw(sdp, atomic_read(&sdp->sd_log_blks_free) <=
                             sdp->sd_jdesc->jd_blocks);
-
        sdp->sd_log_blks_reserved = reserved;
 
        gfs2_log_unlock(sdp);
 }
 
+static void buf_lo_incore_commit(struct gfs2_sbd *sdp, struct gfs2_trans *tr)
+{
+       struct list_head *head = &tr->tr_list_buf;
+       struct gfs2_bufdata *bd;
+
+       gfs2_log_lock(sdp);
+       while (!list_empty(head)) {
+               bd = list_entry(head->next, struct gfs2_bufdata, bd_list_tr);
+               list_del_init(&bd->bd_list_tr);
+               tr->tr_num_buf--;
+       }
+       gfs2_log_unlock(sdp);
+       gfs2_assert_warn(sdp, !tr->tr_num_buf);
+}
+
 /**
  * gfs2_log_commit - Commit a transaction to the log
  * @sdp: the filesystem
@@ -801,9 +829,8 @@ static void log_refund(struct gfs2_sbd *sdp, struct gfs2_trans *tr)
 void gfs2_log_commit(struct gfs2_sbd *sdp, struct gfs2_trans *tr)
 {
        log_refund(sdp, tr);
-       lops_incore_commit(sdp, tr);
+       buf_lo_incore_commit(sdp, tr);
 
-       sdp->sd_vfs->s_dirt = 1;
        up_read(&sdp->sd_log_flush_lock);
 
        gfs2_log_lock(sdp);
@@ -823,7 +850,6 @@ void gfs2_log_shutdown(struct gfs2_sbd *sdp)
        down_write(&sdp->sd_log_flush_lock);
 
        gfs2_assert_withdraw(sdp, !sdp->sd_log_blks_reserved);
-       gfs2_assert_withdraw(sdp, !sdp->sd_log_num_gl);
        gfs2_assert_withdraw(sdp, !sdp->sd_log_num_buf);
        gfs2_assert_withdraw(sdp, !sdp->sd_log_num_revoke);
        gfs2_assert_withdraw(sdp, !sdp->sd_log_num_rg);
@@ -836,7 +862,7 @@ void gfs2_log_shutdown(struct gfs2_sbd *sdp)
        log_write_header(sdp, GFS2_LOG_HEAD_UNMOUNT,
                         (sdp->sd_log_tail == current_tail(sdp)) ? 0 : PULL);
 
-       gfs2_assert_warn(sdp, sdp->sd_log_blks_free == sdp->sd_jdesc->jd_blocks);
+       gfs2_assert_warn(sdp, atomic_read(&sdp->sd_log_blks_free) == sdp->sd_jdesc->jd_blocks);
        gfs2_assert_warn(sdp, sdp->sd_log_head == sdp->sd_log_tail);
        gfs2_assert_warn(sdp, list_empty(&sdp->sd_ail2_list));
 
@@ -864,3 +890,42 @@ void gfs2_meta_syncfs(struct gfs2_sbd *sdp)
        }
 }
 
+
+/**
+ * gfs2_logd - Update log tail as Active Items get flushed to in-place blocks
+ * @sdp: Pointer to GFS2 superblock
+ *
+ * Also, periodically check to make sure that we're using the most recent
+ * journal index.
+ */
+
+int gfs2_logd(void *data)
+{
+       struct gfs2_sbd *sdp = data;
+       unsigned long t;
+       int need_flush;
+
+       while (!kthread_should_stop()) {
+               /* Advance the log tail */
+
+               t = sdp->sd_log_flush_time +
+                   gfs2_tune_get(sdp, gt_log_flush_secs) * HZ;
+
+               gfs2_ail1_empty(sdp, DIO_ALL);
+               gfs2_log_lock(sdp);
+               need_flush = sdp->sd_log_num_buf > gfs2_tune_get(sdp, gt_incore_log_blocks);
+               gfs2_log_unlock(sdp);
+               if (need_flush || time_after_eq(jiffies, t)) {
+                       gfs2_log_flush(sdp, NULL);
+                       sdp->sd_log_flush_time = jiffies;
+               }
+
+               t = gfs2_tune_get(sdp, gt_logd_secs) * HZ;
+               if (freezing(current))
+                       refrigerator();
+               schedule_timeout_interruptible(t);
+       }
+
+       return 0;
+}
+