*
* There are important applications whose performance or correctness
* depends on uncached access to file data. Database clusters
- * (multiple copies of the same instance running on separate hosts)
+ * (multiple copies of the same instance running on separate hosts)
* implement their own cache coherency protocol that subsumes file
- * system cache protocols. Applications that process datasets
- * considerably larger than the client's memory do not always benefit
- * from a local cache. A streaming video server, for instance, has no
+ * system cache protocols. Applications that process datasets
+ * considerably larger than the client's memory do not always benefit
+ * from a local cache. A streaming video server, for instance, has no
* need to cache the contents of a file.
*
* When an application requests uncached I/O, all read and write requests
* 08 Jun 2003 Port to 2.5 APIs --cel
* 31 Mar 2004 Handle direct I/O without VFS support --cel
* 15 Sep 2004 Parallel async reads --cel
+ * 04 May 2005 support O_DIRECT with aio --cel
*
*/
#include <asm/uaccess.h>
#include <asm/atomic.h>
+#include "iostat.h"
+
#define NFSDBG_FACILITY NFSDBG_VFS
-#define MAX_DIRECTIO_SIZE (4096UL << PAGE_SHIFT)
static kmem_cache_t *nfs_direct_cachep;
*/
struct nfs_direct_req {
struct kref kref; /* release manager */
- struct list_head list; /* nfs_read_data structs */
- wait_queue_head_t wait; /* wait for i/o completion */
- struct page ** pages; /* pages in our buffer */
- unsigned int npages; /* count of pages */
- atomic_t complete, /* i/os we're waiting for */
- count, /* bytes actually processed */
+
+ /* I/O parameters */
+ struct list_head list, /* nfs_read/write_data structs */
+ rewrite_list; /* saved nfs_write_data structs */
+ struct nfs_open_context *ctx; /* file open context info */
+ struct kiocb * iocb; /* controlling i/o request */
+ struct inode * inode; /* target file of i/o */
+
+ /* completion state */
+ atomic_t io_count; /* i/os we're waiting for */
+ spinlock_t lock; /* protect completion state */
+ ssize_t count, /* bytes actually processed */
error; /* any reported error */
+ struct completion completion; /* wait for i/o completion */
+
+ /* commit state */
+ struct nfs_write_data * commit_data; /* special write_data for commits */
+ int flags;
+#define NFS_ODIRECT_DO_COMMIT (1) /* an unstable reply was received */
+#define NFS_ODIRECT_RESCHED_WRITES (2) /* write verification failed */
+ struct nfs_writeverf verf; /* unstable write verifier */
};
+static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode);
+static const struct rpc_call_ops nfs_write_direct_ops;
-/**
- * nfs_get_user_pages - find and set up pages underlying user's buffer
- * rw: direction (read or write)
- * user_addr: starting address of this segment of user's buffer
- * count: size of this segment
- * @pages: returned array of page struct pointers underlying user's buffer
+static inline void get_dreq(struct nfs_direct_req *dreq)
+{
+ atomic_inc(&dreq->io_count);
+}
+
+static inline int put_dreq(struct nfs_direct_req *dreq)
+{
+ return atomic_dec_and_test(&dreq->io_count);
+}
+
+/*
+ * "size" is never larger than rsize or wsize.
*/
-static inline int
-nfs_get_user_pages(int rw, unsigned long user_addr, size_t size,
- struct page ***pages)
+static inline int nfs_direct_count_pages(unsigned long user_addr, size_t size)
{
- int result = -ENOMEM;
- unsigned long page_count;
- size_t array_size;
-
- /* set an arbitrary limit to prevent type overflow */
- /* XXX: this can probably be as large as INT_MAX */
- if (size > MAX_DIRECTIO_SIZE) {
- *pages = NULL;
- return -EFBIG;
- }
+ int page_count;
page_count = (user_addr + size + PAGE_SIZE - 1) >> PAGE_SHIFT;
page_count -= user_addr >> PAGE_SHIFT;
+ BUG_ON(page_count < 0);
- array_size = (page_count * sizeof(struct page *));
- *pages = kmalloc(array_size, GFP_KERNEL);
- if (*pages) {
- down_read(¤t->mm->mmap_sem);
- result = get_user_pages(current, current->mm, user_addr,
- page_count, (rw == READ), 0,
- *pages, NULL);
- up_read(¤t->mm->mmap_sem);
- }
- return result;
+ return page_count;
}
/**
- * nfs_free_user_pages - tear down page struct array
- * @pages: array of page struct pointers underlying target buffer
- * @npages: number of pages in the array
- * @do_dirty: dirty the pages as we release them
+ * nfs_direct_IO - NFS address space operation for direct I/O
+ * @rw: direction (read or write)
+ * @iocb: target I/O control block
+ * @iov: array of vectors that define I/O buffer
+ * @pos: offset in file to begin the operation
+ * @nr_segs: size of iovec array
+ *
+ * The presence of this routine in the address space ops vector means
+ * the NFS client supports direct I/O. However, we shunt off direct
+ * read and write requests before the VFS gets them, so this method
+ * should never be called.
*/
-static void
-nfs_free_user_pages(struct page **pages, int npages, int do_dirty)
+ssize_t nfs_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov, loff_t pos, unsigned long nr_segs)
+{
+ dprintk("NFS: nfs_direct_IO (%s) off/no(%Ld/%lu) EINVAL\n",
+ iocb->ki_filp->f_dentry->d_name.name,
+ (long long) pos, nr_segs);
+
+ return -EINVAL;
+}
+
+static void nfs_direct_dirty_pages(struct page **pages, int npages)
{
int i;
for (i = 0; i < npages; i++) {
- if (do_dirty)
- set_page_dirty_lock(pages[i]);
- page_cache_release(pages[i]);
+ struct page *page = pages[i];
+ if (!PageCompound(page))
+ set_page_dirty_lock(page);
}
- kfree(pages);
}
-/**
- * nfs_direct_req_release - release nfs_direct_req structure for direct read
- * @kref: kref object embedded in an nfs_direct_req structure
- *
- */
+static void nfs_direct_release_pages(struct page **pages, int npages)
+{
+ int i;
+ for (i = 0; i < npages; i++)
+ page_cache_release(pages[i]);
+}
+
+static inline struct nfs_direct_req *nfs_direct_req_alloc(void)
+{
+ struct nfs_direct_req *dreq;
+
+ dreq = kmem_cache_alloc(nfs_direct_cachep, SLAB_KERNEL);
+ if (!dreq)
+ return NULL;
+
+ kref_init(&dreq->kref);
+ init_completion(&dreq->completion);
+ INIT_LIST_HEAD(&dreq->list);
+ INIT_LIST_HEAD(&dreq->rewrite_list);
+ dreq->iocb = NULL;
+ dreq->ctx = NULL;
+ spin_lock_init(&dreq->lock);
+ atomic_set(&dreq->io_count, 0);
+ dreq->count = 0;
+ dreq->error = 0;
+ dreq->flags = 0;
+
+ return dreq;
+}
+
static void nfs_direct_req_release(struct kref *kref)
{
struct nfs_direct_req *dreq = container_of(kref, struct nfs_direct_req, kref);
+
+ if (dreq->ctx != NULL)
+ put_nfs_open_context(dreq->ctx);
kmem_cache_free(nfs_direct_cachep, dreq);
}
-/**
- * nfs_direct_read_alloc - allocate nfs_read_data structures for direct read
- * @count: count of bytes for the read request
- * @rsize: local rsize setting
- *
+/*
+ * Collects and returns the final error value/byte-count.
+ */
+static ssize_t nfs_direct_wait(struct nfs_direct_req *dreq)
+{
+ ssize_t result = -EIOCBQUEUED;
+
+ /* Async requests don't wait here */
+ if (dreq->iocb)
+ goto out;
+
+ result = wait_for_completion_interruptible(&dreq->completion);
+
+ if (!result)
+ result = dreq->error;
+ if (!result)
+ result = dreq->count;
+
+out:
+ kref_put(&dreq->kref, nfs_direct_req_release);
+ return (ssize_t) result;
+}
+
+/*
+ * Synchronous I/O uses a stack-allocated iocb. Thus we can't trust
+ * the iocb is still valid here if this is a synchronous request.
+ */
+static void nfs_direct_complete(struct nfs_direct_req *dreq)
+{
+ if (dreq->iocb) {
+ long res = (long) dreq->error;
+ if (!res)
+ res = (long) dreq->count;
+ aio_complete(dreq->iocb, res, 0);
+ }
+ complete_all(&dreq->completion);
+
+ kref_put(&dreq->kref, nfs_direct_req_release);
+}
+
+/*
* Note we also set the number of requests we have in the dreq when we are
* done. This prevents races with I/O completion so we will always wait
* until all requests have been dispatched and completed.
*/
-static struct nfs_direct_req *nfs_direct_read_alloc(size_t nbytes, unsigned int rsize)
+static struct nfs_direct_req *nfs_direct_read_alloc(size_t nbytes, size_t rsize)
{
struct list_head *list;
struct nfs_direct_req *dreq;
- unsigned int reads = 0;
unsigned int rpages = (rsize + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
- dreq = kmem_cache_alloc(nfs_direct_cachep, SLAB_KERNEL);
+ dreq = nfs_direct_req_alloc();
if (!dreq)
return NULL;
- kref_init(&dreq->kref);
- init_waitqueue_head(&dreq->wait);
- INIT_LIST_HEAD(&dreq->list);
- atomic_set(&dreq->count, 0);
- atomic_set(&dreq->error, 0);
-
list = &dreq->list;
for(;;) {
struct nfs_read_data *data = nfs_readdata_alloc(rpages);
list_add(&data->pages, list);
data->req = (struct nfs_page *) dreq;
- reads++;
+ get_dreq(dreq);
if (nbytes <= rsize)
break;
nbytes -= rsize;
}
kref_get(&dreq->kref);
- atomic_set(&dreq->complete, reads);
return dreq;
}
-/**
- * nfs_direct_read_result - handle a read reply for a direct read request
- * @data: address of NFS READ operation control block
- * @status: status of this NFS READ operation
- *
+/*
* We must hold a reference to all the pages in this direct read request
* until the RPCs complete. This could be long *after* we are woken up in
- * nfs_direct_read_wait (for instance, if someone hits ^C on a slow server).
+ * nfs_direct_wait (for instance, if someone hits ^C on a slow server).
*/
-static void nfs_direct_read_result(struct nfs_read_data *data, int status)
+static void nfs_direct_read_result(struct rpc_task *task, void *calldata)
{
+ struct nfs_read_data *data = calldata;
struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
- if (likely(status >= 0))
- atomic_add(data->res.count, &dreq->count);
+ if (nfs_readpage_result(task, data) != 0)
+ return;
+
+ nfs_direct_dirty_pages(data->pagevec, data->npages);
+ nfs_direct_release_pages(data->pagevec, data->npages);
+
+ spin_lock(&dreq->lock);
+
+ if (likely(task->tk_status >= 0))
+ dreq->count += data->res.count;
else
- atomic_set(&dreq->error, status);
+ dreq->error = task->tk_status;
- if (unlikely(atomic_dec_and_test(&dreq->complete))) {
- nfs_free_user_pages(dreq->pages, dreq->npages, 1);
- wake_up(&dreq->wait);
- kref_put(&dreq->kref, nfs_direct_req_release);
- }
+ spin_unlock(&dreq->lock);
+
+ if (put_dreq(dreq))
+ nfs_direct_complete(dreq);
}
-/**
- * nfs_direct_read_schedule - dispatch NFS READ operations for a direct read
- * @dreq: address of nfs_direct_req struct for this request
- * @inode: target inode
- * @ctx: target file open context
- * @user_addr: starting address of this segment of user's buffer
- * @count: size of this segment
- * @file_offset: offset in file to begin the operation
- *
+static const struct rpc_call_ops nfs_read_direct_ops = {
+ .rpc_call_done = nfs_direct_read_result,
+ .rpc_release = nfs_readdata_release,
+};
+
+/*
* For each nfs_read_data struct that was allocated on the list, dispatch
- * an NFS READ operation
+ * an NFS READ operation. If get_user_pages() fails, we stop sending reads.
+ * Read length accounting is handled by nfs_direct_read_result().
+ * Otherwise, if no requests have been sent, just return an error.
*/
-static void nfs_direct_read_schedule(struct nfs_direct_req *dreq,
- struct inode *inode, struct nfs_open_context *ctx,
- unsigned long user_addr, size_t count, loff_t file_offset)
+static ssize_t nfs_direct_read_schedule(struct nfs_direct_req *dreq, unsigned long user_addr, size_t count, loff_t pos)
{
+ struct nfs_open_context *ctx = dreq->ctx;
+ struct inode *inode = ctx->dentry->d_inode;
struct list_head *list = &dreq->list;
- struct page **pages = dreq->pages;
- unsigned int curpage, pgbase;
- unsigned int rsize = NFS_SERVER(inode)->rsize;
+ size_t rsize = NFS_SERVER(inode)->rsize;
+ unsigned int pgbase;
+ int result;
+ ssize_t started = 0;
+ struct nfs_read_data *data;
- curpage = 0;
pgbase = user_addr & ~PAGE_MASK;
do {
- struct nfs_read_data *data;
- unsigned int bytes;
+ size_t bytes;
bytes = rsize;
if (count < rsize)
bytes = count;
+ BUG_ON(list_empty(list));
data = list_entry(list->next, struct nfs_read_data, pages);
list_del_init(&data->pages);
+ data->npages = nfs_direct_count_pages(user_addr, bytes);
+ down_read(¤t->mm->mmap_sem);
+ result = get_user_pages(current, current->mm, user_addr,
+ data->npages, 1, 0, data->pagevec, NULL);
+ up_read(¤t->mm->mmap_sem);
+ if (unlikely(result < data->npages))
+ goto out_err;
+
data->inode = inode;
data->cred = ctx->cred;
data->args.fh = NFS_FH(inode);
data->args.context = ctx;
- data->args.offset = file_offset;
+ data->args.offset = pos;
data->args.pgbase = pgbase;
- data->args.pages = &pages[curpage];
+ data->args.pages = data->pagevec;
data->args.count = bytes;
data->res.fattr = &data->fattr;
data->res.eof = 0;
data->res.count = bytes;
+ rpc_init_task(&data->task, NFS_CLIENT(inode), RPC_TASK_ASYNC,
+ &nfs_read_direct_ops, data);
NFS_PROTO(inode)->read_setup(data);
data->task.tk_cookie = (unsigned long) inode;
- data->complete = nfs_direct_read_result;
lock_kernel();
rpc_execute(&data->task);
unlock_kernel();
- dfprintk(VFS, "NFS: %4d initiated direct read call (req %s/%Ld, %u bytes @ offset %Lu)\n",
+ dfprintk(VFS, "NFS: %5u initiated direct read call (req %s/%Ld, %zu bytes @ offset %Lu)\n",
data->task.tk_pid,
inode->i_sb->s_id,
(long long)NFS_FILEID(inode),
bytes,
(unsigned long long)data->args.offset);
- file_offset += bytes;
+ started += bytes;
+ user_addr += bytes;
+ pos += bytes;
pgbase += bytes;
- curpage += pgbase >> PAGE_SHIFT;
pgbase &= ~PAGE_MASK;
count -= bytes;
} while (count != 0);
-}
+ BUG_ON(!list_empty(list));
+ return 0;
-/**
- * nfs_direct_read_wait - wait for I/O completion for direct reads
- * @dreq: request on which we are to wait
- * @intr: whether or not this wait can be interrupted
- *
- * Collects and returns the final error value/byte-count.
- */
-static ssize_t nfs_direct_read_wait(struct nfs_direct_req *dreq, int intr)
-{
- int result = 0;
+out_err:
+ if (result > 0)
+ nfs_direct_release_pages(data->pagevec, result);
- if (intr) {
- result = wait_event_interruptible(dreq->wait,
- (atomic_read(&dreq->complete) == 0));
- } else {
- wait_event(dreq->wait, (atomic_read(&dreq->complete) == 0));
+ list_add(&data->pages, list);
+ while (!list_empty(list)) {
+ data = list_entry(list->next, struct nfs_read_data, pages);
+ list_del(&data->pages);
+ nfs_readdata_free(data);
+ if (put_dreq(dreq))
+ nfs_direct_complete(dreq);
}
- if (!result)
- result = atomic_read(&dreq->error);
- if (!result)
- result = atomic_read(&dreq->count);
-
- kref_put(&dreq->kref, nfs_direct_req_release);
- return (ssize_t) result;
+ if (started)
+ return 0;
+ return result < 0 ? (ssize_t) result : -EFAULT;
}
-/**
- * nfs_direct_read_seg - Read in one iov segment. Generate separate
- * read RPCs for each "rsize" bytes.
- * @inode: target inode
- * @ctx: target file open context
- * @user_addr: starting address of this segment of user's buffer
- * @count: size of this segment
- * @file_offset: offset in file to begin the operation
- * @pages: array of addresses of page structs defining user's buffer
- * @nr_pages: number of pages in the array
- *
- */
-static ssize_t nfs_direct_read_seg(struct inode *inode,
- struct nfs_open_context *ctx, unsigned long user_addr,
- size_t count, loff_t file_offset, struct page **pages,
- unsigned int nr_pages)
+static ssize_t nfs_direct_read(struct kiocb *iocb, unsigned long user_addr, size_t count, loff_t pos)
{
ssize_t result;
sigset_t oldset;
+ struct inode *inode = iocb->ki_filp->f_mapping->host;
struct rpc_clnt *clnt = NFS_CLIENT(inode);
struct nfs_direct_req *dreq;
if (!dreq)
return -ENOMEM;
- dreq->pages = pages;
- dreq->npages = nr_pages;
+ dreq->inode = inode;
+ dreq->ctx = get_nfs_open_context((struct nfs_open_context *)iocb->ki_filp->private_data);
+ if (!is_sync_kiocb(iocb))
+ dreq->iocb = iocb;
+ nfs_add_stats(inode, NFSIOS_DIRECTREADBYTES, count);
rpc_clnt_sigmask(clnt, &oldset);
- nfs_direct_read_schedule(dreq, inode, ctx, user_addr, count,
- file_offset);
- result = nfs_direct_read_wait(dreq, clnt->cl_intr);
+ result = nfs_direct_read_schedule(dreq, user_addr, count, pos);
+ if (!result)
+ result = nfs_direct_wait(dreq);
rpc_clnt_sigunmask(clnt, &oldset);
return result;
}
-/**
- * nfs_direct_read - For each iov segment, map the user's buffer
- * then generate read RPCs.
- * @inode: target inode
- * @ctx: target file open context
- * @iov: array of vectors that define I/O buffer
- * file_offset: offset in file to begin the operation
- * nr_segs: size of iovec array
- *
- * We've already pushed out any non-direct writes so that this read
- * will see them when we read from the server.
- */
-static ssize_t
-nfs_direct_read(struct inode *inode, struct nfs_open_context *ctx,
- const struct iovec *iov, loff_t file_offset,
- unsigned long nr_segs)
+static void nfs_direct_free_writedata(struct nfs_direct_req *dreq)
{
- ssize_t tot_bytes = 0;
- unsigned long seg = 0;
-
- while ((seg < nr_segs) && (tot_bytes >= 0)) {
- ssize_t result;
- int page_count;
- struct page **pages;
- const struct iovec *vec = &iov[seg++];
- unsigned long user_addr = (unsigned long) vec->iov_base;
- size_t size = vec->iov_len;
-
- page_count = nfs_get_user_pages(READ, user_addr, size, &pages);
- if (page_count < 0) {
- nfs_free_user_pages(pages, 0, 0);
- if (tot_bytes > 0)
- break;
- return page_count;
- }
+ list_splice_init(&dreq->rewrite_list, &dreq->list);
+ while (!list_empty(&dreq->list)) {
+ struct nfs_write_data *data = list_entry(dreq->list.next, struct nfs_write_data, pages);
+ list_del(&data->pages);
+ nfs_direct_release_pages(data->pagevec, data->npages);
+ nfs_writedata_release(data);
+ }
+}
- result = nfs_direct_read_seg(inode, ctx, user_addr, size,
- file_offset, pages, page_count);
+#if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
+static void nfs_direct_write_reschedule(struct nfs_direct_req *dreq)
+{
+ struct inode *inode = dreq->inode;
+ struct list_head *p;
+ struct nfs_write_data *data;
+
+ dreq->count = 0;
+ get_dreq(dreq);
+
+ list_for_each(p, &dreq->rewrite_list) {
+ data = list_entry(p, struct nfs_write_data, pages);
+
+ get_dreq(dreq);
+
+ /*
+ * Reset data->res.
+ */
+ nfs_fattr_init(&data->fattr);
+ data->res.count = data->args.count;
+ memset(&data->verf, 0, sizeof(data->verf));
+
+ /*
+ * Reuse data->task; data->args should not have changed
+ * since the original request was sent.
+ */
+ rpc_init_task(&data->task, NFS_CLIENT(inode), RPC_TASK_ASYNC,
+ &nfs_write_direct_ops, data);
+ NFS_PROTO(inode)->write_setup(data, FLUSH_STABLE);
+
+ data->task.tk_priority = RPC_PRIORITY_NORMAL;
+ data->task.tk_cookie = (unsigned long) inode;
- if (result <= 0) {
- if (tot_bytes > 0)
- break;
- return result;
- }
- tot_bytes += result;
- file_offset += result;
- if (result < size)
- break;
+ /*
+ * We're called via an RPC callback, so BKL is already held.
+ */
+ rpc_execute(&data->task);
+
+ dprintk("NFS: %5u rescheduled direct write call (req %s/%Ld, %u bytes @ offset %Lu)\n",
+ data->task.tk_pid,
+ inode->i_sb->s_id,
+ (long long)NFS_FILEID(inode),
+ data->args.count,
+ (unsigned long long)data->args.offset);
}
- return tot_bytes;
+ if (put_dreq(dreq))
+ nfs_direct_write_complete(dreq, inode);
}
-/**
- * nfs_direct_write_seg - Write out one iov segment. Generate separate
- * write RPCs for each "wsize" bytes, then commit.
- * @inode: target inode
- * @ctx: target file open context
- * user_addr: starting address of this segment of user's buffer
- * count: size of this segment
- * file_offset: offset in file to begin the operation
- * @pages: array of addresses of page structs defining user's buffer
- * nr_pages: size of pages array
- */
-static ssize_t nfs_direct_write_seg(struct inode *inode,
- struct nfs_open_context *ctx, unsigned long user_addr,
- size_t count, loff_t file_offset, struct page **pages,
- int nr_pages)
+static void nfs_direct_commit_result(struct rpc_task *task, void *calldata)
{
- const unsigned int wsize = NFS_SERVER(inode)->wsize;
- size_t request;
- int curpage, need_commit;
- ssize_t result, tot_bytes;
- struct nfs_writeverf first_verf;
- struct nfs_write_data *wdata;
-
- wdata = nfs_writedata_alloc(NFS_SERVER(inode)->wpages);
- if (!wdata)
- return -ENOMEM;
+ struct nfs_write_data *data = calldata;
+ struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
- wdata->inode = inode;
- wdata->cred = ctx->cred;
- wdata->args.fh = NFS_FH(inode);
- wdata->args.context = ctx;
- wdata->args.stable = NFS_UNSTABLE;
- if (IS_SYNC(inode) || NFS_PROTO(inode)->version == 2 || count <= wsize)
- wdata->args.stable = NFS_FILE_SYNC;
- wdata->res.fattr = &wdata->fattr;
- wdata->res.verf = &wdata->verf;
+ /* Call the NFS version-specific code */
+ if (NFS_PROTO(data->inode)->commit_done(task, data) != 0)
+ return;
+ if (unlikely(task->tk_status < 0)) {
+ dreq->error = task->tk_status;
+ dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
+ }
+ if (memcmp(&dreq->verf, &data->verf, sizeof(data->verf))) {
+ dprintk("NFS: %5u commit verify failed\n", task->tk_pid);
+ dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
+ }
- nfs_begin_data_update(inode);
-retry:
- need_commit = 0;
- tot_bytes = 0;
- curpage = 0;
- request = count;
- wdata->args.pgbase = user_addr & ~PAGE_MASK;
- wdata->args.offset = file_offset;
- do {
- wdata->args.count = request;
- if (wdata->args.count > wsize)
- wdata->args.count = wsize;
- wdata->args.pages = &pages[curpage];
+ dprintk("NFS: %5u commit returned %d\n", task->tk_pid, task->tk_status);
+ nfs_direct_write_complete(dreq, data->inode);
+}
- dprintk("NFS: direct write: c=%u o=%Ld ua=%lu, pb=%u, cp=%u\n",
- wdata->args.count, (long long) wdata->args.offset,
- user_addr + tot_bytes, wdata->args.pgbase, curpage);
+static const struct rpc_call_ops nfs_commit_direct_ops = {
+ .rpc_call_done = nfs_direct_commit_result,
+ .rpc_release = nfs_commit_release,
+};
- lock_kernel();
- result = NFS_PROTO(inode)->write(wdata);
- unlock_kernel();
+static void nfs_direct_commit_schedule(struct nfs_direct_req *dreq)
+{
+ struct nfs_write_data *data = dreq->commit_data;
- if (result <= 0) {
- if (tot_bytes > 0)
- break;
- goto out;
- }
+ data->inode = dreq->inode;
+ data->cred = dreq->ctx->cred;
- if (tot_bytes == 0)
- memcpy(&first_verf.verifier, &wdata->verf.verifier,
- sizeof(first_verf.verifier));
- if (wdata->verf.committed != NFS_FILE_SYNC) {
- need_commit = 1;
- if (memcmp(&first_verf.verifier, &wdata->verf.verifier,
- sizeof(first_verf.verifier)));
- goto sync_retry;
- }
+ data->args.fh = NFS_FH(data->inode);
+ data->args.offset = 0;
+ data->args.count = 0;
+ data->res.count = 0;
+ data->res.fattr = &data->fattr;
+ data->res.verf = &data->verf;
- tot_bytes += result;
+ rpc_init_task(&data->task, NFS_CLIENT(dreq->inode), RPC_TASK_ASYNC,
+ &nfs_commit_direct_ops, data);
+ NFS_PROTO(data->inode)->commit_setup(data, 0);
- /* in case of a short write: stop now, let the app recover */
- if (result < wdata->args.count)
- break;
+ data->task.tk_priority = RPC_PRIORITY_NORMAL;
+ data->task.tk_cookie = (unsigned long)data->inode;
+ /* Note: task.tk_ops->rpc_release will free dreq->commit_data */
+ dreq->commit_data = NULL;
- wdata->args.offset += result;
- wdata->args.pgbase += result;
- curpage += wdata->args.pgbase >> PAGE_SHIFT;
- wdata->args.pgbase &= ~PAGE_MASK;
- request -= result;
- } while (request != 0);
+ dprintk("NFS: %5u initiated commit call\n", data->task.tk_pid);
- /*
- * Commit data written so far, even in the event of an error
- */
- if (need_commit) {
- wdata->args.count = tot_bytes;
- wdata->args.offset = file_offset;
+ lock_kernel();
+ rpc_execute(&data->task);
+ unlock_kernel();
+}
- lock_kernel();
- result = NFS_PROTO(inode)->commit(wdata);
- unlock_kernel();
+static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode)
+{
+ int flags = dreq->flags;
- if (result < 0 || memcmp(&first_verf.verifier,
- &wdata->verf.verifier,
- sizeof(first_verf.verifier)) != 0)
- goto sync_retry;
+ dreq->flags = 0;
+ switch (flags) {
+ case NFS_ODIRECT_DO_COMMIT:
+ nfs_direct_commit_schedule(dreq);
+ break;
+ case NFS_ODIRECT_RESCHED_WRITES:
+ nfs_direct_write_reschedule(dreq);
+ break;
+ default:
+ nfs_end_data_update(inode);
+ if (dreq->commit_data != NULL)
+ nfs_commit_free(dreq->commit_data);
+ nfs_direct_free_writedata(dreq);
+ nfs_direct_complete(dreq);
}
- result = tot_bytes;
+}
-out:
+static void nfs_alloc_commit_data(struct nfs_direct_req *dreq)
+{
+ dreq->commit_data = nfs_commit_alloc(0);
+ if (dreq->commit_data != NULL)
+ dreq->commit_data->req = (struct nfs_page *) dreq;
+}
+#else
+static inline void nfs_alloc_commit_data(struct nfs_direct_req *dreq)
+{
+ dreq->commit_data = NULL;
+}
+
+static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode)
+{
nfs_end_data_update(inode);
- nfs_writedata_free(wdata);
- return result;
+ nfs_direct_free_writedata(dreq);
+ nfs_direct_complete(dreq);
+}
+#endif
+
+static struct nfs_direct_req *nfs_direct_write_alloc(size_t nbytes, size_t wsize)
+{
+ struct list_head *list;
+ struct nfs_direct_req *dreq;
+ unsigned int wpages = (wsize + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
+
+ dreq = nfs_direct_req_alloc();
+ if (!dreq)
+ return NULL;
+
+ list = &dreq->list;
+ for(;;) {
+ struct nfs_write_data *data = nfs_writedata_alloc(wpages);
+
+ if (unlikely(!data)) {
+ while (!list_empty(list)) {
+ data = list_entry(list->next,
+ struct nfs_write_data, pages);
+ list_del(&data->pages);
+ nfs_writedata_free(data);
+ }
+ kref_put(&dreq->kref, nfs_direct_req_release);
+ return NULL;
+ }
+
+ INIT_LIST_HEAD(&data->pages);
+ list_add(&data->pages, list);
+
+ data->req = (struct nfs_page *) dreq;
+ get_dreq(dreq);
+ if (nbytes <= wsize)
+ break;
+ nbytes -= wsize;
+ }
-sync_retry:
- wdata->args.stable = NFS_FILE_SYNC;
- goto retry;
+ nfs_alloc_commit_data(dreq);
+
+ kref_get(&dreq->kref);
+ return dreq;
}
-/**
- * nfs_direct_write - For each iov segment, map the user's buffer
- * then generate write and commit RPCs.
- * @inode: target inode
- * @ctx: target file open context
- * @iov: array of vectors that define I/O buffer
- * file_offset: offset in file to begin the operation
- * nr_segs: size of iovec array
- *
- * Upon return, generic_file_direct_IO invalidates any cached pages
- * that non-direct readers might access, so they will pick up these
- * writes immediately.
- */
-static ssize_t nfs_direct_write(struct inode *inode,
- struct nfs_open_context *ctx, const struct iovec *iov,
- loff_t file_offset, unsigned long nr_segs)
+static void nfs_direct_write_result(struct rpc_task *task, void *calldata)
{
- ssize_t tot_bytes = 0;
- unsigned long seg = 0;
-
- while ((seg < nr_segs) && (tot_bytes >= 0)) {
- ssize_t result;
- int page_count;
- struct page **pages;
- const struct iovec *vec = &iov[seg++];
- unsigned long user_addr = (unsigned long) vec->iov_base;
- size_t size = vec->iov_len;
-
- page_count = nfs_get_user_pages(WRITE, user_addr, size, &pages);
- if (page_count < 0) {
- nfs_free_user_pages(pages, 0, 0);
- if (tot_bytes > 0)
- break;
- return page_count;
- }
+ struct nfs_write_data *data = calldata;
+ struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
+ int status = task->tk_status;
+
+ if (nfs_writeback_done(task, data) != 0)
+ return;
- result = nfs_direct_write_seg(inode, ctx, user_addr, size,
- file_offset, pages, page_count);
- nfs_free_user_pages(pages, page_count, 0);
+ spin_lock(&dreq->lock);
- if (result <= 0) {
- if (tot_bytes > 0)
+ if (likely(status >= 0))
+ dreq->count += data->res.count;
+ else
+ dreq->error = task->tk_status;
+
+ if (data->res.verf->committed != NFS_FILE_SYNC) {
+ switch (dreq->flags) {
+ case 0:
+ memcpy(&dreq->verf, &data->verf, sizeof(dreq->verf));
+ dreq->flags = NFS_ODIRECT_DO_COMMIT;
break;
- return result;
+ case NFS_ODIRECT_DO_COMMIT:
+ if (memcmp(&dreq->verf, &data->verf, sizeof(dreq->verf))) {
+ dprintk("NFS: %5u write verify failed\n", task->tk_pid);
+ dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
+ }
}
- tot_bytes += result;
- file_offset += result;
- if (result < size)
- break;
}
- return tot_bytes;
+
+ spin_unlock(&dreq->lock);
}
-/**
- * nfs_direct_IO - NFS address space operation for direct I/O
- * rw: direction (read or write)
- * @iocb: target I/O control block
- * @iov: array of vectors that define I/O buffer
- * file_offset: offset in file to begin the operation
- * nr_segs: size of iovec array
- *
+/*
+ * NB: Return the value of the first error return code. Subsequent
+ * errors after the first one are ignored.
*/
-ssize_t
-nfs_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov,
- loff_t file_offset, unsigned long nr_segs)
+static void nfs_direct_write_release(void *calldata)
{
- ssize_t result = -EINVAL;
- struct file *file = iocb->ki_filp;
- struct nfs_open_context *ctx;
- struct dentry *dentry = file->f_dentry;
- struct inode *inode = dentry->d_inode;
+ struct nfs_write_data *data = calldata;
+ struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req;
- /*
- * No support for async yet
- */
- if (!is_sync_kiocb(iocb))
- return result;
-
- ctx = (struct nfs_open_context *)file->private_data;
- switch (rw) {
- case READ:
- dprintk("NFS: direct_IO(read) (%s) off/no(%Lu/%lu)\n",
- dentry->d_name.name, file_offset, nr_segs);
-
- result = nfs_direct_read(inode, ctx, iov,
- file_offset, nr_segs);
- break;
- case WRITE:
- dprintk("NFS: direct_IO(write) (%s) off/no(%Lu/%lu)\n",
- dentry->d_name.name, file_offset, nr_segs);
-
- result = nfs_direct_write(inode, ctx, iov,
- file_offset, nr_segs);
- break;
- default:
- break;
+ if (put_dreq(dreq))
+ nfs_direct_write_complete(dreq, data->inode);
+}
+
+static const struct rpc_call_ops nfs_write_direct_ops = {
+ .rpc_call_done = nfs_direct_write_result,
+ .rpc_release = nfs_direct_write_release,
+};
+
+/*
+ * For each nfs_write_data struct that was allocated on the list, dispatch
+ * an NFS WRITE operation. If get_user_pages() fails, we stop sending writes.
+ * Write length accounting is handled by nfs_direct_write_result().
+ * Otherwise, if no requests have been sent, just return an error.
+ */
+static ssize_t nfs_direct_write_schedule(struct nfs_direct_req *dreq, unsigned long user_addr, size_t count, loff_t pos, int sync)
+{
+ struct nfs_open_context *ctx = dreq->ctx;
+ struct inode *inode = ctx->dentry->d_inode;
+ struct list_head *list = &dreq->list;
+ size_t wsize = NFS_SERVER(inode)->wsize;
+ unsigned int pgbase;
+ int result;
+ ssize_t started = 0;
+ struct nfs_write_data *data;
+
+ pgbase = user_addr & ~PAGE_MASK;
+ do {
+ size_t bytes;
+
+ bytes = wsize;
+ if (count < wsize)
+ bytes = count;
+
+ BUG_ON(list_empty(list));
+ data = list_entry(list->next, struct nfs_write_data, pages);
+
+ data->npages = nfs_direct_count_pages(user_addr, bytes);
+ down_read(¤t->mm->mmap_sem);
+ result = get_user_pages(current, current->mm, user_addr,
+ data->npages, 0, 0, data->pagevec, NULL);
+ up_read(¤t->mm->mmap_sem);
+ if (unlikely(result < data->npages))
+ goto out_err;
+
+ list_move_tail(&data->pages, &dreq->rewrite_list);
+
+ data->inode = inode;
+ data->cred = ctx->cred;
+ data->args.fh = NFS_FH(inode);
+ data->args.context = ctx;
+ data->args.offset = pos;
+ data->args.pgbase = pgbase;
+ data->args.pages = data->pagevec;
+ data->args.count = bytes;
+ data->res.fattr = &data->fattr;
+ data->res.count = bytes;
+ data->res.verf = &data->verf;
+
+ rpc_init_task(&data->task, NFS_CLIENT(inode), RPC_TASK_ASYNC,
+ &nfs_write_direct_ops, data);
+ NFS_PROTO(inode)->write_setup(data, sync);
+
+ data->task.tk_priority = RPC_PRIORITY_NORMAL;
+ data->task.tk_cookie = (unsigned long) inode;
+
+ lock_kernel();
+ rpc_execute(&data->task);
+ unlock_kernel();
+
+ dfprintk(VFS, "NFS: %5u initiated direct write call (req %s/%Ld, %zu bytes @ offset %Lu)\n",
+ data->task.tk_pid,
+ inode->i_sb->s_id,
+ (long long)NFS_FILEID(inode),
+ bytes,
+ (unsigned long long)data->args.offset);
+
+ started += bytes;
+ user_addr += bytes;
+ pos += bytes;
+ pgbase += bytes;
+ pgbase &= ~PAGE_MASK;
+
+ count -= bytes;
+ } while (count != 0);
+ BUG_ON(!list_empty(list));
+ return 0;
+
+out_err:
+ if (result > 0)
+ nfs_direct_release_pages(data->pagevec, result);
+
+ list_add(&data->pages, list);
+ while (!list_empty(list)) {
+ data = list_entry(list->next, struct nfs_write_data, pages);
+ list_del(&data->pages);
+ nfs_writedata_free(data);
+ if (put_dreq(dreq))
+ nfs_direct_write_complete(dreq, inode);
}
+
+ if (started)
+ return 0;
+ return result < 0 ? (ssize_t) result : -EFAULT;
+}
+
+static ssize_t nfs_direct_write(struct kiocb *iocb, unsigned long user_addr, size_t count, loff_t pos)
+{
+ ssize_t result;
+ sigset_t oldset;
+ struct inode *inode = iocb->ki_filp->f_mapping->host;
+ struct rpc_clnt *clnt = NFS_CLIENT(inode);
+ struct nfs_direct_req *dreq;
+ size_t wsize = NFS_SERVER(inode)->wsize;
+ int sync = 0;
+
+ dreq = nfs_direct_write_alloc(count, wsize);
+ if (!dreq)
+ return -ENOMEM;
+ if (dreq->commit_data == NULL || count < wsize)
+ sync = FLUSH_STABLE;
+
+ dreq->inode = inode;
+ dreq->ctx = get_nfs_open_context((struct nfs_open_context *)iocb->ki_filp->private_data);
+ if (!is_sync_kiocb(iocb))
+ dreq->iocb = iocb;
+
+ nfs_add_stats(inode, NFSIOS_DIRECTWRITTENBYTES, count);
+
+ nfs_begin_data_update(inode);
+
+ rpc_clnt_sigmask(clnt, &oldset);
+ result = nfs_direct_write_schedule(dreq, user_addr, count, pos, sync);
+ if (!result)
+ result = nfs_direct_wait(dreq);
+ rpc_clnt_sigunmask(clnt, &oldset);
+
return result;
}
* nfs_file_direct_read - file direct read operation for NFS files
* @iocb: target I/O control block
* @buf: user's buffer into which to read data
- * count: number of bytes to read
- * pos: byte offset in file where reading starts
+ * @count: number of bytes to read
+ * @pos: byte offset in file where reading starts
*
* We use this function for direct reads instead of calling
* generic_file_aio_read() in order to avoid gfar's check to see if
* the request starts before the end of the file. For that check
* to work, we must generate a GETATTR before each direct read, and
* even then there is a window between the GETATTR and the subsequent
- * READ where the file size could change. So our preference is simply
+ * READ where the file size could change. Our preference is simply
* to do all reads the application wants, and the server will take
* care of managing the end of file boundary.
- *
+ *
* This function also eliminates unnecessarily updating the file's
* atime locally, as the NFS server sets the file's atime, and this
* client must read the updated atime from the server back into its
* cache.
*/
-ssize_t
-nfs_file_direct_read(struct kiocb *iocb, char __user *buf, size_t count, loff_t pos)
+ssize_t nfs_file_direct_read(struct kiocb *iocb, char __user *buf, size_t count, loff_t pos)
{
ssize_t retval = -EINVAL;
- loff_t *ppos = &iocb->ki_pos;
struct file *file = iocb->ki_filp;
- struct nfs_open_context *ctx =
- (struct nfs_open_context *) file->private_data;
struct address_space *mapping = file->f_mapping;
- struct inode *inode = mapping->host;
- struct iovec iov = {
- .iov_base = buf,
- .iov_len = count,
- };
dprintk("nfs: direct read(%s/%s, %lu@%Ld)\n",
file->f_dentry->d_parent->d_name.name,
file->f_dentry->d_name.name,
(unsigned long) count, (long long) pos);
- if (!is_sync_kiocb(iocb))
- goto out;
if (count < 0)
goto out;
retval = -EFAULT;
- if (!access_ok(VERIFY_WRITE, iov.iov_base, iov.iov_len))
+ if (!access_ok(VERIFY_WRITE, buf, count))
goto out;
retval = 0;
if (!count)
if (retval)
goto out;
- retval = nfs_direct_read(inode, ctx, &iov, pos, 1);
+ retval = nfs_direct_read(iocb, (unsigned long) buf, count, pos);
if (retval > 0)
- *ppos = pos + retval;
+ iocb->ki_pos = pos + retval;
out:
return retval;
* nfs_file_direct_write - file direct write operation for NFS files
* @iocb: target I/O control block
* @buf: user's buffer from which to write data
- * count: number of bytes to write
- * pos: byte offset in file where writing starts
+ * @count: number of bytes to write
+ * @pos: byte offset in file where writing starts
*
* We use this function for direct writes instead of calling
* generic_file_aio_write() in order to avoid taking the inode
* Note that O_APPEND is not supported for NFS direct writes, as there
* is no atomic O_APPEND write facility in the NFS protocol.
*/
-ssize_t
-nfs_file_direct_write(struct kiocb *iocb, const char __user *buf, size_t count, loff_t pos)
+ssize_t nfs_file_direct_write(struct kiocb *iocb, const char __user *buf, size_t count, loff_t pos)
{
ssize_t retval;
struct file *file = iocb->ki_filp;
- struct nfs_open_context *ctx =
- (struct nfs_open_context *) file->private_data;
struct address_space *mapping = file->f_mapping;
- struct inode *inode = mapping->host;
- struct iovec iov = {
- .iov_base = (char __user *)buf,
- };
dfprintk(VFS, "nfs: direct write(%s/%s, %lu@%Ld)\n",
file->f_dentry->d_parent->d_name.name,
file->f_dentry->d_name.name,
(unsigned long) count, (long long) pos);
- retval = -EINVAL;
- if (!is_sync_kiocb(iocb))
- goto out;
-
retval = generic_write_checks(file, &pos, &count, 0);
if (retval)
goto out;
retval = 0;
if (!count)
goto out;
- iov.iov_len = count,
retval = -EFAULT;
- if (!access_ok(VERIFY_READ, iov.iov_base, iov.iov_len))
+ if (!access_ok(VERIFY_READ, buf, count))
goto out;
retval = nfs_sync_mapping(mapping);
if (retval)
goto out;
- retval = nfs_direct_write(inode, ctx, &iov, pos, 1);
+ retval = nfs_direct_write(iocb, (unsigned long) buf, count, pos);
+
+ /*
+ * XXX: nfs_end_data_update() already ensures this file's
+ * cached data is subsequently invalidated. Do we really
+ * need to call invalidate_inode_pages2() again here?
+ *
+ * For aio writes, this invalidation will almost certainly
+ * occur before the writes complete. Kind of racey.
+ */
if (mapping->nrpages)
invalidate_inode_pages2(mapping);
+
if (retval > 0)
iocb->ki_pos = pos + retval;
return retval;
}
-int nfs_init_directcache(void)
+/**
+ * nfs_init_directcache - create a slab cache for nfs_direct_req structures
+ *
+ */
+int __init nfs_init_directcache(void)
{
nfs_direct_cachep = kmem_cache_create("nfs_direct_cache",
sizeof(struct nfs_direct_req),
- 0, SLAB_RECLAIM_ACCOUNT,
+ 0, (SLAB_RECLAIM_ACCOUNT|
+ SLAB_MEM_SPREAD),
NULL, NULL);
if (nfs_direct_cachep == NULL)
return -ENOMEM;
return 0;
}
-void nfs_destroy_directcache(void)
+/**
+ * nfs_destroy_directcache - destroy the slab cache for nfs_direct_req structures
+ *
+ */
+void __exit nfs_destroy_directcache(void)
{
if (kmem_cache_destroy(nfs_direct_cachep))
printk(KERN_INFO "nfs_direct_cache: not all structures were freed\n");