]> www.pilppa.org Git - linux-2.6-omap-h63xx.git/commitdiff
Btrfs: Add ordered async work queues
authorChris Mason <chris.mason@oracle.com>
Fri, 7 Nov 2008 03:03:00 +0000 (22:03 -0500)
committerChris Mason <chris.mason@oracle.com>
Fri, 7 Nov 2008 03:03:00 +0000 (22:03 -0500)
Btrfs uses kernel threads to create async work queues for cpu intensive
operations such as checksumming and decompression.  These work well,
but they make it difficult to keep IO order intact.

A single writepages call from pdflush or fsync will turn into a number
of bios, and each bio is checksummed in parallel.  Once the checksum is
computed, the bio is sent down to the disk, and since we don't control
the order in which the parallel operations happen, they might go down to
the disk in almost any order.

The code deals with this somewhat by having deep work queues for a single
kernel thread, making it very likely that a single thread will process all
the bios for a single inode.

This patch introduces an explicitly ordered work queue.  As work structs
are placed into the queue they are put onto the tail of a list.  They have
three callbacks:

->func (cpu intensive processing here)
->ordered_func (order sensitive processing here)
->ordered_free (free the work struct, all processing is done)

The work struct has three callbacks.  The func callback does the cpu intensive
work, and when it completes the work struct is marked as done.

Every time a work struct completes, the list is checked to see if the head
is marked as done.  If so the ordered_func callback is used to do the
order sensitive processing and the ordered_free callback is used to do
any cleanup.  Then we loop back and check the head of the list again.

This patch also changes the checksumming code to use the ordered workqueues.
One a 4 drive array, it increases streaming writes from 280MB/s to 350MB/s.

Signed-off-by: Chris Mason <chris.mason@oracle.com>
fs/btrfs/async-thread.c
fs/btrfs/async-thread.h
fs/btrfs/disk-io.c
fs/btrfs/disk-io.h
fs/btrfs/inode.c

index d82efd722a481d174c9787834ab1427098cb3165..e1e49715459ee9d45617bf67879afe5317257050 100644 (file)
 # include <linux/freezer.h>
 #include "async-thread.h"
 
+#define WORK_QUEUED_BIT 0
+#define WORK_DONE_BIT 1
+#define WORK_ORDER_DONE_BIT 2
+
 /*
  * container for the kthread task pointer and the list of pending work
  * One of these is allocated per thread.
@@ -88,6 +92,47 @@ static void check_busy_worker(struct btrfs_worker_thread *worker)
        }
 }
 
+static noinline int run_ordered_completions(struct btrfs_workers *workers,
+                                           struct btrfs_work *work)
+{
+       unsigned long flags;
+
+       if (!workers->ordered)
+               return 0;
+
+       set_bit(WORK_DONE_BIT, &work->flags);
+
+       spin_lock_irqsave(&workers->lock, flags);
+
+       while(!list_empty(&workers->order_list)) {
+               work = list_entry(workers->order_list.next,
+                                 struct btrfs_work, order_list);
+
+               if (!test_bit(WORK_DONE_BIT, &work->flags))
+                       break;
+
+               /* we are going to call the ordered done function, but
+                * we leave the work item on the list as a barrier so
+                * that later work items that are done don't have their
+                * functions called before this one returns
+                */
+               if (test_and_set_bit(WORK_ORDER_DONE_BIT, &work->flags))
+                       break;
+
+               spin_unlock_irqrestore(&workers->lock, flags);
+
+               work->ordered_func(work);
+
+               /* now take the lock again and call the freeing code */
+               spin_lock_irqsave(&workers->lock, flags);
+               list_del(&work->order_list);
+               work->ordered_free(work);
+       }
+
+       spin_unlock_irqrestore(&workers->lock, flags);
+       return 0;
+}
+
 /*
  * main loop for servicing work items
  */
@@ -102,7 +147,7 @@ static int worker_loop(void *arg)
                        cur = worker->pending.next;
                        work = list_entry(cur, struct btrfs_work, list);
                        list_del(&work->list);
-                       clear_bit(0, &work->flags);
+                       clear_bit(WORK_QUEUED_BIT, &work->flags);
 
                        work->worker = worker;
                        spin_unlock_irq(&worker->lock);
@@ -110,8 +155,15 @@ static int worker_loop(void *arg)
                        work->func(work);
 
                        atomic_dec(&worker->num_pending);
+                       /*
+                        * unless this is an ordered work queue,
+                        * 'work' was probably freed by func above.
+                        */
+                       run_ordered_completions(worker->workers, work);
+
                        spin_lock_irq(&worker->lock);
                        check_idle_worker(worker);
+
                }
                worker->working = 0;
                if (freezing(current)) {
@@ -154,10 +206,12 @@ void btrfs_init_workers(struct btrfs_workers *workers, char *name, int max)
        workers->num_workers = 0;
        INIT_LIST_HEAD(&workers->worker_list);
        INIT_LIST_HEAD(&workers->idle_list);
+       INIT_LIST_HEAD(&workers->order_list);
        spin_lock_init(&workers->lock);
        workers->max_workers = max;
        workers->idle_thresh = 32;
        workers->name = name;
+       workers->ordered = 0;
 }
 
 /*
@@ -296,7 +350,7 @@ int btrfs_requeue_work(struct btrfs_work *work)
        struct btrfs_worker_thread *worker = work->worker;
        unsigned long flags;
 
-       if (test_and_set_bit(0, &work->flags))
+       if (test_and_set_bit(WORK_QUEUED_BIT, &work->flags))
                goto out;
 
        spin_lock_irqsave(&worker->lock, flags);
@@ -330,10 +384,17 @@ int btrfs_queue_worker(struct btrfs_workers *workers, struct btrfs_work *work)
        int wake = 0;
 
        /* don't requeue something already on a list */
-       if (test_and_set_bit(0, &work->flags))
+       if (test_and_set_bit(WORK_QUEUED_BIT, &work->flags))
                goto out;
 
        worker = find_worker(workers);
+       if (workers->ordered) {
+               spin_lock_irqsave(&workers->lock, flags);
+               list_add_tail(&work->order_list, &workers->order_list);
+               spin_unlock_irqrestore(&workers->lock, flags);
+       } else {
+               INIT_LIST_HEAD(&work->order_list);
+       }
 
        spin_lock_irqsave(&worker->lock, flags);
        atomic_inc(&worker->num_pending);
index 4ec9a2ee0f9d8c17a2b72da514b20151154aa019..31be4ed8b63ea0aec80b2c31781fad2ef57786f6 100644 (file)
@@ -37,10 +37,16 @@ struct btrfs_worker_thread;
  */
 struct btrfs_work {
        /*
-        * only func should be set to the function you want called
+        * func should be set to the function you want called
         * your work struct is passed as the only arg
+        *
+        * ordered_func must be set for work sent to an ordered work queue,
+        * and it is called to complete a given work item in the same
+        * order they were sent to the queue.
         */
        void (*func)(struct btrfs_work *work);
+       void (*ordered_func)(struct btrfs_work *work);
+       void (*ordered_free)(struct btrfs_work *work);
 
        /*
         * flags should be set to zero.  It is used to make sure the
@@ -51,6 +57,7 @@ struct btrfs_work {
        /* don't touch these */
        struct btrfs_worker_thread *worker;
        struct list_head list;
+       struct list_head order_list;
 };
 
 struct btrfs_workers {
@@ -63,6 +70,9 @@ struct btrfs_workers {
        /* once a worker has this many requests or fewer, it is idle */
        int idle_thresh;
 
+       /* force completions in the order they were queued */
+       int ordered;
+
        /* list with all the work threads.  The workers on the idle thread
         * may be actively servicing jobs, but they haven't yet hit the
         * idle thresh limit above.
@@ -70,6 +80,12 @@ struct btrfs_workers {
        struct list_head worker_list;
        struct list_head idle_list;
 
+       /*
+        * when operating in ordered mode, this maintains the list
+        * of work items waiting for completion
+        */
+       struct list_head order_list;
+
        /* lock for finding the next worker thread to queue on */
        spinlock_t lock;
 
index 94b4e50f6b2cbadf1726f2856dab15678c6f0a5e..e0a28f705a6464052a2d87998ee16f79a33131b3 100644 (file)
@@ -80,7 +80,8 @@ struct async_submit_bio {
        struct inode *inode;
        struct bio *bio;
        struct list_head list;
-       extent_submit_bio_hook_t *submit_bio_hook;
+       extent_submit_bio_hook_t *submit_bio_start;
+       extent_submit_bio_hook_t *submit_bio_done;
        int rw;
        int mirror_num;
        unsigned long bio_flags;
@@ -452,7 +453,18 @@ int btrfs_congested_async(struct btrfs_fs_info *info, int iodone)
                btrfs_async_submit_limit(info);
 }
 
-static void run_one_async_submit(struct btrfs_work *work)
+static void run_one_async_start(struct btrfs_work *work)
+{
+       struct btrfs_fs_info *fs_info;
+       struct async_submit_bio *async;
+
+       async = container_of(work, struct  async_submit_bio, work);
+       fs_info = BTRFS_I(async->inode)->root->fs_info;
+       async->submit_bio_start(async->inode, async->rw, async->bio,
+                              async->mirror_num, async->bio_flags);
+}
+
+static void run_one_async_done(struct btrfs_work *work)
 {
        struct btrfs_fs_info *fs_info;
        struct async_submit_bio *async;
@@ -470,15 +482,23 @@ static void run_one_async_submit(struct btrfs_work *work)
            waitqueue_active(&fs_info->async_submit_wait))
                wake_up(&fs_info->async_submit_wait);
 
-       async->submit_bio_hook(async->inode, async->rw, async->bio,
+       async->submit_bio_done(async->inode, async->rw, async->bio,
                               async->mirror_num, async->bio_flags);
+}
+
+static void run_one_async_free(struct btrfs_work *work)
+{
+       struct async_submit_bio *async;
+
+       async = container_of(work, struct  async_submit_bio, work);
        kfree(async);
 }
 
 int btrfs_wq_submit_bio(struct btrfs_fs_info *fs_info, struct inode *inode,
                        int rw, struct bio *bio, int mirror_num,
                        unsigned long bio_flags,
-                       extent_submit_bio_hook_t *submit_bio_hook)
+                       extent_submit_bio_hook_t *submit_bio_start,
+                       extent_submit_bio_hook_t *submit_bio_done)
 {
        struct async_submit_bio *async;
        int limit = btrfs_async_submit_limit(fs_info);
@@ -491,8 +511,13 @@ int btrfs_wq_submit_bio(struct btrfs_fs_info *fs_info, struct inode *inode,
        async->rw = rw;
        async->bio = bio;
        async->mirror_num = mirror_num;
-       async->submit_bio_hook = submit_bio_hook;
-       async->work.func = run_one_async_submit;
+       async->submit_bio_start = submit_bio_start;
+       async->submit_bio_done = submit_bio_done;
+
+       async->work.func = run_one_async_start;
+       async->work.ordered_func = run_one_async_done;
+       async->work.ordered_free = run_one_async_free;
+
        async->work.flags = 0;
        async->bio_flags = bio_flags;
 
@@ -533,29 +558,25 @@ static int btree_csum_one_bio(struct bio *bio)
        return 0;
 }
 
-static int __btree_submit_bio_hook(struct inode *inode, int rw, struct bio *bio,
-                                int mirror_num, unsigned long bio_flags)
+static int __btree_submit_bio_start(struct inode *inode, int rw,
+                                   struct bio *bio, int mirror_num,
+                                   unsigned long bio_flags)
 {
-       struct btrfs_root *root = BTRFS_I(inode)->root;
-       int ret;
-
        /*
         * when we're called for a write, we're already in the async
         * submission context.  Just jump into btrfs_map_bio
         */
-       if (rw & (1 << BIO_RW)) {
-               btree_csum_one_bio(bio);
-               return btrfs_map_bio(BTRFS_I(inode)->root, rw, bio,
-                                    mirror_num, 1);
-       }
+       btree_csum_one_bio(bio);
+       return 0;
+}
 
+static int __btree_submit_bio_done(struct inode *inode, int rw, struct bio *bio,
+                                int mirror_num, unsigned long bio_flags)
+{
        /*
-        * called for a read, do the setup so that checksum validation
-        * can happen in the async kernel threads
+        * when we're called for a write, we're already in the async
+        * submission context.  Just jump into btrfs_map_bio
         */
-       ret = btrfs_bio_wq_end_io(root->fs_info, bio, 1);
-       BUG_ON(ret);
-
        return btrfs_map_bio(BTRFS_I(inode)->root, rw, bio, mirror_num, 1);
 }
 
@@ -567,11 +588,22 @@ static int btree_submit_bio_hook(struct inode *inode, int rw, struct bio *bio,
         * can happen in parallel across all CPUs
         */
        if (!(rw & (1 << BIO_RW))) {
-               return __btree_submit_bio_hook(inode, rw, bio, mirror_num, 0);
+               int ret;
+               /*
+                * called for a read, do the setup so that checksum validation
+                * can happen in the async kernel threads
+                */
+               ret = btrfs_bio_wq_end_io(BTRFS_I(inode)->root->fs_info,
+                                         bio, 1);
+               BUG_ON(ret);
+
+               return btrfs_map_bio(BTRFS_I(inode)->root, rw, bio,
+                                    mirror_num, 1);
        }
        return btrfs_wq_submit_bio(BTRFS_I(inode)->root->fs_info,
                                   inode, rw, bio, mirror_num, 0,
-                                  __btree_submit_bio_hook);
+                                  __btree_submit_bio_start,
+                                  __btree_submit_bio_done);
 }
 
 static int btree_writepage(struct page *page, struct writeback_control *wbc)
@@ -1534,7 +1566,8 @@ struct btrfs_root *open_ctree(struct super_block *sb,
         * were sent by the writeback daemons, improving overall locality
         * of the IO going down the pipe.
         */
-       fs_info->workers.idle_thresh = 128;
+       fs_info->workers.idle_thresh = 8;
+       fs_info->workers.ordered = 1;
 
        btrfs_init_workers(&fs_info->fixup_workers, "fixup", 1);
        btrfs_init_workers(&fs_info->endio_workers, "endio",
index 4eb1f1408d21285bd75a7abe58dfbe571aeb94e6..b8d5948fa279135b1d1b26fc2e0d2f0a0ea4bf06 100644 (file)
@@ -72,7 +72,9 @@ int btrfs_bio_wq_end_io(struct btrfs_fs_info *info, struct bio *bio,
 int btrfs_wq_submit_bio(struct btrfs_fs_info *fs_info, struct inode *inode,
                        int rw, struct bio *bio, int mirror_num,
                        unsigned long bio_flags,
-                       extent_submit_bio_hook_t *submit_bio_hook);
+                       extent_submit_bio_hook_t *submit_bio_start,
+                       extent_submit_bio_hook_t *submit_bio_done);
+
 int btrfs_congested_async(struct btrfs_fs_info *info, int iodone);
 unsigned long btrfs_async_submit_limit(struct btrfs_fs_info *info);
 int btrfs_write_tree_block(struct extent_buffer *buf);
index 806708dd7e3847c10207403505180367c8feb472..3df0ffad976eb338a8a8293f256ffc0cde3f2efd 100644 (file)
@@ -881,7 +881,7 @@ int btrfs_merge_bio_hook(struct page *page, unsigned long offset,
  * At IO completion time the cums attached on the ordered extent record
  * are inserted into the btree
  */
-int __btrfs_submit_bio_hook(struct inode *inode, int rw, struct bio *bio,
+int __btrfs_submit_bio_start(struct inode *inode, int rw, struct bio *bio,
                          int mirror_num, unsigned long bio_flags)
 {
        struct btrfs_root *root = BTRFS_I(inode)->root;
@@ -889,7 +889,21 @@ int __btrfs_submit_bio_hook(struct inode *inode, int rw, struct bio *bio,
 
        ret = btrfs_csum_one_bio(root, inode, bio);
        BUG_ON(ret);
+       return 0;
+}
 
+/*
+ * in order to insert checksums into the metadata in large chunks,
+ * we wait until bio submission time.   All the pages in the bio are
+ * checksummed and sums are attached onto the ordered extent record.
+ *
+ * At IO completion time the cums attached on the ordered extent record
+ * are inserted into the btree
+ */
+int __btrfs_submit_bio_done(struct inode *inode, int rw, struct bio *bio,
+                         int mirror_num, unsigned long bio_flags)
+{
+       struct btrfs_root *root = BTRFS_I(inode)->root;
        return btrfs_map_bio(root, rw, bio, mirror_num, 1);
 }
 
@@ -922,7 +936,8 @@ int btrfs_submit_bio_hook(struct inode *inode, int rw, struct bio *bio,
                /* we're doing a write, do the async checksumming */
                return btrfs_wq_submit_bio(BTRFS_I(inode)->root->fs_info,
                                   inode, rw, bio, mirror_num,
-                                  bio_flags, __btrfs_submit_bio_hook);
+                                  bio_flags, __btrfs_submit_bio_start,
+                                  __btrfs_submit_bio_done);
        }
 
 mapit: