thermal: re-document thermal units
[linux-2.6.git] / fs / direct-io.c
index 538fb04..9e81add 100644 (file)
@@ -27,6 +27,7 @@
 #include <linux/slab.h>
 #include <linux/highmem.h>
 #include <linux/pagemap.h>
+#include <linux/task_io_accounting_ops.h>
 #include <linux/bio.h>
 #include <linux/wait.h>
 #include <linux/err.h>
@@ -121,8 +122,7 @@ struct dio {
 
        /* BIO completion state */
        spinlock_t bio_lock;            /* protects BIO fields below */
-       int bio_count;                  /* nr bios to be completed */
-       int bios_in_flight;             /* nr bios in flight */
+       unsigned long refcount;         /* direct_io_worker() and bios */
        struct bio *bio_list;           /* singly linked via bi_private */
        struct task_struct *waiter;     /* waiting task (NULL if none) */
 
@@ -163,7 +163,7 @@ static int dio_refill_pages(struct dio *dio)
        up_read(&current->mm->mmap_sem);
 
        if (ret < 0 && dio->blocks_available && (dio->rw & WRITE)) {
-               struct page *page = ZERO_PAGE(dio->curr_user_address);
+               struct page *page = ZERO_PAGE(0);
                /*
                 * A memory fault, but the filesystem has some outstanding
                 * mapped blocks.  We need to use those blocks up to avoid
@@ -209,91 +209,81 @@ static struct page *dio_get_page(struct dio *dio)
        return dio->pages[dio->head++];
 }
 
-/*
- * Called when all DIO BIO I/O has been completed - let the filesystem
- * know, if it registered an interest earlier via get_block.  Pass the
- * private field of the map buffer_head so that filesystems can use it
- * to hold additional state between get_block calls and dio_complete.
- */
-static void dio_complete(struct dio *dio, loff_t offset, ssize_t bytes)
-{
-       if (dio->end_io && dio->result)
-               dio->end_io(dio->iocb, offset, bytes, dio->map_bh.b_private);
-       if (dio->lock_type == DIO_LOCKING)
-               up_read(&dio->inode->i_alloc_sem);
-}
-
-/*
- * Called when a BIO has been processed.  If the count goes to zero then IO is
- * complete and we can signal this to the AIO layer.
+/**
+ * dio_complete() - called when all DIO BIO I/O has been completed
+ * @offset: the byte offset in the file of the completed operation
+ *
+ * This releases locks as dictated by the locking type, lets interested parties
+ * know that a DIO operation has completed, and calculates the resulting return
+ * code for the operation.
+ *
+ * It lets the filesystem know if it registered an interest earlier via
+ * get_block.  Pass the private field of the map buffer_head so that
+ * filesystems can use it to hold additional state between get_block calls and
+ * dio_complete.
  */
-static void finished_one_bio(struct dio *dio)
+static int dio_complete(struct dio *dio, loff_t offset, int ret)
 {
-       unsigned long flags;
+       ssize_t transferred = 0;
 
-       spin_lock_irqsave(&dio->bio_lock, flags);
-       if (dio->bio_count == 1) {
-               if (dio->is_async) {
-                       ssize_t transferred;
-                       loff_t offset;
-
-                       /*
-                        * Last reference to the dio is going away.
-                        * Drop spinlock and complete the DIO.
-                        */
-                       spin_unlock_irqrestore(&dio->bio_lock, flags);
+       /*
+        * AIO submission can race with bio completion to get here while
+        * expecting to have the last io completed by bio completion.
+        * In that case -EIOCBQUEUED is in fact not an error we want
+        * to preserve through this call.
+        */
+       if (ret == -EIOCBQUEUED)
+               ret = 0;
 
-                       /* Check for short read case */
-                       transferred = dio->result;
-                       offset = dio->iocb->ki_pos;
+       if (dio->result) {
+               transferred = dio->result;
 
-                       if ((dio->rw == READ) &&
-                           ((offset + transferred) > dio->i_size))
-                               transferred = dio->i_size - offset;
+               /* Check for short read case */
+               if ((dio->rw == READ) && ((offset + transferred) > dio->i_size))
+                       transferred = dio->i_size - offset;
+       }
 
-                       /* check for error in completion path */
-                       if (dio->io_error)
-                               transferred = dio->io_error;
+       if (dio->end_io && dio->result)
+               dio->end_io(dio->iocb, offset, transferred,
+                           dio->map_bh.b_private);
+       if (dio->lock_type == DIO_LOCKING)
+               /* lockdep: non-owner release */
+               up_read_non_owner(&dio->inode->i_alloc_sem);
 
-                       dio_complete(dio, offset, transferred);
+       if (ret == 0)
+               ret = dio->page_errors;
+       if (ret == 0)
+               ret = dio->io_error;
+       if (ret == 0)
+               ret = transferred;
 
-                       /* Complete AIO later if falling back to buffered i/o */
-                       if (dio->result == dio->size ||
-                               ((dio->rw == READ) && dio->result)) {
-                               aio_complete(dio->iocb, transferred, 0);
-                               kfree(dio);
-                               return;
-                       } else {
-                               /*
-                                * Falling back to buffered
-                                */
-                               spin_lock_irqsave(&dio->bio_lock, flags);
-                               dio->bio_count--;
-                               if (dio->waiter)
-                                       wake_up_process(dio->waiter);
-                               spin_unlock_irqrestore(&dio->bio_lock, flags);
-                               return;
-                       }
-               }
-       }
-       dio->bio_count--;
-       spin_unlock_irqrestore(&dio->bio_lock, flags);
+       return ret;
 }
 
 static int dio_bio_complete(struct dio *dio, struct bio *bio);
 /*
  * Asynchronous IO callback. 
  */
-static int dio_bio_end_aio(struct bio *bio, unsigned int bytes_done, int error)
+static void dio_bio_end_aio(struct bio *bio, int error)
 {
        struct dio *dio = bio->bi_private;
-
-       if (bio->bi_size)
-               return 1;
+       unsigned long remaining;
+       unsigned long flags;
 
        /* cleanup the bio */
        dio_bio_complete(dio, bio);
-       return 0;
+
+       spin_lock_irqsave(&dio->bio_lock, flags);
+       remaining = --dio->refcount;
+       if (remaining == 1 && dio->waiter)
+               wake_up_process(dio->waiter);
+       spin_unlock_irqrestore(&dio->bio_lock, flags);
+
+       if (remaining == 0) {
+               int ret = dio_complete(dio, dio->iocb->ki_pos, 0);
+               aio_complete(dio->iocb, ret, 0);
+               kfree(dio);
+       }
 }
 
 /*
@@ -303,22 +293,17 @@ static int dio_bio_end_aio(struct bio *bio, unsigned int bytes_done, int error)
  * During I/O bi_private points at the dio.  After I/O, bi_private is used to
  * implement a singly-linked list of completed BIOs, at dio->bio_list.
  */
-static int dio_bio_end_io(struct bio *bio, unsigned int bytes_done, int error)
+static void dio_bio_end_io(struct bio *bio, int error)
 {
        struct dio *dio = bio->bi_private;
        unsigned long flags;
 
-       if (bio->bi_size)
-               return 1;
-
        spin_lock_irqsave(&dio->bio_lock, flags);
        bio->bi_private = dio->bio_list;
        dio->bio_list = bio;
-       dio->bios_in_flight--;
-       if (dio->waiter && dio->bios_in_flight == 0)
+       if (--dio->refcount == 1 && dio->waiter)
                wake_up_process(dio->waiter);
        spin_unlock_irqrestore(&dio->bio_lock, flags);
-       return 0;
 }
 
 static int
@@ -346,6 +331,8 @@ dio_bio_alloc(struct dio *dio, struct block_device *bdev,
  * In the AIO read case we speculatively dirty the pages before starting IO.
  * During IO completion, any of these pages which happen to have been written
  * back will be redirtied by bio_check_pages_dirty().
+ *
+ * bios hold a dio reference between submit_bio and ->end_io.
  */
 static void dio_bio_submit(struct dio *dio)
 {
@@ -353,12 +340,14 @@ static void dio_bio_submit(struct dio *dio)
        unsigned long flags;
 
        bio->bi_private = dio;
+
        spin_lock_irqsave(&dio->bio_lock, flags);
-       dio->bio_count++;
-       dio->bios_in_flight++;
+       dio->refcount++;
        spin_unlock_irqrestore(&dio->bio_lock, flags);
+
        if (dio->is_async && dio->rw == READ)
                bio_set_pages_dirty(bio);
+
        submit_bio(dio->rw, bio);
 
        dio->bio = NULL;
@@ -375,28 +364,37 @@ static void dio_cleanup(struct dio *dio)
 }
 
 /*
- * Wait for the next BIO to complete.  Remove it and return it.
+ * Wait for the next BIO to complete.  Remove it and return it.  NULL is
+ * returned once all BIOs have been completed.  This must only be called once
+ * all bios have been issued so that dio->refcount can only decrease.  This
+ * requires that that the caller hold a reference on the dio.
  */
 static struct bio *dio_await_one(struct dio *dio)
 {
        unsigned long flags;
-       struct bio *bio;
+       struct bio *bio = NULL;
 
        spin_lock_irqsave(&dio->bio_lock, flags);
-       while (dio->bio_list == NULL) {
-               set_current_state(TASK_UNINTERRUPTIBLE);
-               if (dio->bio_list == NULL) {
-                       dio->waiter = current;
-                       spin_unlock_irqrestore(&dio->bio_lock, flags);
-                       blk_run_address_space(dio->inode->i_mapping);
-                       io_schedule();
-                       spin_lock_irqsave(&dio->bio_lock, flags);
-                       dio->waiter = NULL;
-               }
-               set_current_state(TASK_RUNNING);
+
+       /*
+        * Wait as long as the list is empty and there are bios in flight.  bio
+        * completion drops the count, maybe adds to the list, and wakes while
+        * holding the bio_lock so we don't need set_current_state()'s barrier
+        * and can call it after testing our condition.
+        */
+       while (dio->refcount > 1 && dio->bio_list == NULL) {
+               __set_current_state(TASK_UNINTERRUPTIBLE);
+               dio->waiter = current;
+               spin_unlock_irqrestore(&dio->bio_lock, flags);
+               io_schedule();
+               /* wake up sets us TASK_RUNNING */
+               spin_lock_irqsave(&dio->bio_lock, flags);
+               dio->waiter = NULL;
+       }
+       if (dio->bio_list) {
+               bio = dio->bio_list;
+               dio->bio_list = bio->bi_private;
        }
-       bio = dio->bio_list;
-       dio->bio_list = bio->bi_private;
        spin_unlock_irqrestore(&dio->bio_lock, flags);
        return bio;
 }
@@ -425,34 +423,24 @@ static int dio_bio_complete(struct dio *dio, struct bio *bio)
                }
                bio_put(bio);
        }
-       finished_one_bio(dio);
        return uptodate ? 0 : -EIO;
 }
 
 /*
- * Wait on and process all in-flight BIOs.
+ * Wait on and process all in-flight BIOs.  This must only be called once
+ * all bios have been issued so that the refcount can only decrease.
+ * This just waits for all bios to make it through dio_bio_complete.  IO
+ * errors are propagated through dio->io_error and should be propagated via
+ * dio_complete().
  */
-static int dio_await_completion(struct dio *dio)
+static void dio_await_completion(struct dio *dio)
 {
-       int ret = 0;
-
-       if (dio->bio)
-               dio_bio_submit(dio);
-
-       /*
-        * The bio_lock is not held for the read of bio_count.
-        * This is ok since it is the dio_bio_complete() that changes
-        * bio_count.
-        */
-       while (dio->bio_count) {
-               struct bio *bio = dio_await_one(dio);
-               int ret2;
-
-               ret2 = dio_bio_complete(dio, bio);
-               if (ret == 0)
-                       ret = ret2;
-       }
-       return ret;
+       struct bio *bio;
+       do {
+               bio = dio_await_one(dio);
+               if (bio)
+                       dio_bio_complete(dio, bio);
+       } while (bio);
 }
 
 /*
@@ -674,6 +662,13 @@ submit_page_section(struct dio *dio, struct page *page,
 {
        int ret = 0;
 
+       if (dio->rw & WRITE) {
+               /*
+                * Read accounting is performed in submit_bio()
+                */
+               task_io_account_write(len);
+       }
+
        /*
         * Can we just grow the current page's presence in the dio?
         */
@@ -768,7 +763,7 @@ static void dio_zero_block(struct dio *dio, int end)
 
        this_chunk_bytes = this_chunk_blocks << dio->blkbits;
 
-       page = ZERO_PAGE(dio->curr_user_address);
+       page = ZERO_PAGE(0);
        if (submit_page_section(dio, page, 0, this_chunk_bytes, 
                                dio->next_block_for_io))
                return;
@@ -863,7 +858,6 @@ static int do_direct_IO(struct dio *dio)
 do_holes:
                        /* Handle holes */
                        if (!buffer_mapped(map_bh)) {
-                               char *kaddr;
                                loff_t i_size_aligned;
 
                                /* AKPM: eargh, -ENOTBLK is a hack */
@@ -884,11 +878,8 @@ do_holes:
                                        page_cache_release(page);
                                        goto out;
                                }
-                               kaddr = kmap_atomic(page, KM_USER0);
-                               memset(kaddr + (block_in_page << blkbits),
-                                               0, 1 << blkbits);
-                               flush_dcache_page(page);
-                               kunmap_atomic(kaddr, KM_USER0);
+                               zero_user(page, block_in_page << blkbits,
+                                               1 << blkbits);
                                dio->block_in_file++;
                                block_in_page++;
                                goto next_block;
@@ -952,49 +943,28 @@ direct_io_worker(int rw, struct kiocb *iocb, struct inode *inode,
        struct dio *dio)
 {
        unsigned long user_addr; 
+       unsigned long flags;
        int seg;
        ssize_t ret = 0;
        ssize_t ret2;
        size_t bytes;
 
-       dio->bio = NULL;
        dio->inode = inode;
        dio->rw = rw;
        dio->blkbits = blkbits;
        dio->blkfactor = inode->i_blkbits - blkbits;
-       dio->start_zero_done = 0;
-       dio->size = 0;
        dio->block_in_file = offset >> blkbits;
-       dio->blocks_available = 0;
-       dio->cur_page = NULL;
 
-       dio->boundary = 0;
-       dio->reap_counter = 0;
        dio->get_block = get_block;
        dio->end_io = end_io;
-       dio->map_bh.b_private = NULL;
        dio->final_block_in_bio = -1;
        dio->next_block_for_io = -1;
 
-       dio->page_errors = 0;
-       dio->io_error = 0;
-       dio->result = 0;
        dio->iocb = iocb;
        dio->i_size = i_size_read(inode);
 
-       /*
-        * BIO completion state.
-        *
-        * ->bio_count starts out at one, and we decrement it to zero after all
-        * BIOs are submitted.  This to avoid the situation where a really fast
-        * (or synchronous) device could take the count to zero while we're
-        * still submitting BIOs.
-        */
-       dio->bio_count = 1;
-       dio->bios_in_flight = 0;
        spin_lock_init(&dio->bio_lock);
-       dio->bio_list = NULL;
-       dio->waiter = NULL;
+       dio->refcount = 1;
 
        /*
         * In case of non-aligned buffers, we may need 2 more
@@ -1002,8 +972,6 @@ direct_io_worker(int rw, struct kiocb *iocb, struct inode *inode,
         */
        if (unlikely(dio->blkfactor))
                dio->pages_in_io = 2;
-       else
-               dio->pages_in_io = 0;
 
        for (seg = 0; seg < nr_segs; seg++) {
                user_addr = (unsigned long)iov[seg].iov_base;
@@ -1068,6 +1036,9 @@ direct_io_worker(int rw, struct kiocb *iocb, struct inode *inode,
        if (dio->bio)
                dio_bio_submit(dio);
 
+       /* All IO is now issued, send it on its way */
+       blk_run_address_space(inode->i_mapping);
+
        /*
         * It is possible that, we return short IO due to end of file.
         * In that case, we need to release all the pages we got hold on.
@@ -1083,74 +1054,41 @@ direct_io_worker(int rw, struct kiocb *iocb, struct inode *inode,
                mutex_unlock(&dio->inode->i_mutex);
 
        /*
-        * OK, all BIOs are submitted, so we can decrement bio_count to truly
-        * reflect the number of to-be-processed BIOs.
+        * The only time we want to leave bios in flight is when a successful
+        * partial aio read or full aio write have been setup.  In that case
+        * bio completion will call aio_complete.  The only time it's safe to
+        * call aio_complete is when we return -EIOCBQUEUED, so we key on that.
+        * This had *better* be the only place that raises -EIOCBQUEUED.
         */
-       if (dio->is_async) {
-               int should_wait = 0;
+       BUG_ON(ret == -EIOCBQUEUED);
+       if (dio->is_async && ret == 0 && dio->result &&
+           ((rw & READ) || (dio->result == dio->size)))
+               ret = -EIOCBQUEUED;
 
-               if (dio->result < dio->size && (rw & WRITE)) {
-                       dio->waiter = current;
-                       should_wait = 1;
-               }
-               if (ret == 0)
-                       ret = dio->result;
-               finished_one_bio(dio);          /* This can free the dio */
-               blk_run_address_space(inode->i_mapping);
-               if (should_wait) {
-                       unsigned long flags;
-                       /*
-                        * Wait for already issued I/O to drain out and
-                        * release its references to user-space pages
-                        * before returning to fallback on buffered I/O
-                        */
-
-                       spin_lock_irqsave(&dio->bio_lock, flags);
-                       set_current_state(TASK_UNINTERRUPTIBLE);
-                       while (dio->bio_count) {
-                               spin_unlock_irqrestore(&dio->bio_lock, flags);
-                               io_schedule();
-                               spin_lock_irqsave(&dio->bio_lock, flags);
-                               set_current_state(TASK_UNINTERRUPTIBLE);
-                       }
-                       spin_unlock_irqrestore(&dio->bio_lock, flags);
-                       set_current_state(TASK_RUNNING);
-                       kfree(dio);
-               }
-       } else {
-               ssize_t transferred = 0;
-
-               finished_one_bio(dio);
-               ret2 = dio_await_completion(dio);
-               if (ret == 0)
-                       ret = ret2;
-               if (ret == 0)
-                       ret = dio->page_errors;
-               if (dio->result) {
-                       loff_t i_size = i_size_read(inode);
+       if (ret != -EIOCBQUEUED)
+               dio_await_completion(dio);
 
-                       transferred = dio->result;
-                       /*
-                        * Adjust the return value if the read crossed a
-                        * non-block-aligned EOF.
-                        */
-                       if (rw == READ && (offset + transferred > i_size))
-                               transferred = i_size - offset;
-               }
-               dio_complete(dio, offset, transferred);
-               if (ret == 0)
-                       ret = transferred;
+       /*
+        * Sync will always be dropping the final ref and completing the
+        * operation.  AIO can if it was a broken operation described above or
+        * in fact if all the bios race to complete before we get here.  In
+        * that case dio_complete() translates the EIOCBQUEUED into the proper
+        * return code that the caller will hand to aio_complete().
+        *
+        * This is managed by the bio_lock instead of being an atomic_t so that
+        * completion paths can drop their ref and use the remaining count to
+        * decide to wake the submission path atomically.
+        */
+       spin_lock_irqsave(&dio->bio_lock, flags);
+       ret2 = --dio->refcount;
+       spin_unlock_irqrestore(&dio->bio_lock, flags);
 
-               /* We could have also come here on an AIO file extend */
-               if (!is_sync_kiocb(iocb) && (rw & WRITE) &&
-                   ret >= 0 && dio->result == dio->size)
-                       /*
-                        * For AIO writes where we have completed the
-                        * i/o, we have to mark the the aio complete.
-                        */
-                       aio_complete(iocb, ret, 0);
+       if (ret2 == 0) {
+               ret = dio_complete(dio, offset, ret);
                kfree(dio);
-       }
+       } else
+               BUG_ON(ret != -EIOCBQUEUED);
+
        return ret;
 }
 
@@ -1221,7 +1159,7 @@ __blockdev_direct_IO(int rw, struct kiocb *iocb, struct inode *inode,
                }
        }
 
-       dio = kmalloc(sizeof(*dio), GFP_KERNEL);
+       dio = kzalloc(sizeof(*dio), GFP_KERNEL);
        retval = -ENOMEM;
        if (!dio)
                goto out;
@@ -1261,7 +1199,8 @@ __blockdev_direct_IO(int rw, struct kiocb *iocb, struct inode *inode,
                }
 
                if (dio_lock_type == DIO_LOCKING)
-                       down_read(&inode->i_alloc_sem);
+                       /* lockdep: not the owner will release it */
+                       down_read_non_owner(&inode->i_alloc_sem);
        }
 
        /*