From 91d5b47023b608227d605d1e916b29dd0215bff7 Mon Sep 17 00:00:00 2001 From: Chuck Lever Date: Mon, 20 Mar 2006 13:44:14 -0500 Subject: NFS: add I/O performance counters Invoke the byte and event counter macros where we want to count bytes and events. Clean-up: fix a possible NULL dereference in nfs_lock, and simplify nfs_file_open. Test-plan: fsx and iozone on UP and SMP systems, with and without pre-emption. Watch for memory overwrite bugs, and performance loss (significantly more CPU required per op). Signed-off-by: Chuck Lever Signed-off-by: Trond Myklebust --- fs/nfs/direct.c | 7 +++++++ 1 file changed, 7 insertions(+) (limited to 'fs/nfs/direct.c') diff --git a/fs/nfs/direct.c b/fs/nfs/direct.c index 4e9b3a1b36c..fc07ce4885d 100644 --- a/fs/nfs/direct.c +++ b/fs/nfs/direct.c @@ -54,6 +54,8 @@ #include #include +#include "iostat.h" + #define NFSDBG_FACILITY NFSDBG_VFS #define MAX_DIRECTIO_SIZE (4096UL << PAGE_SHIFT) @@ -67,6 +69,7 @@ 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 inode * inode; /* target file of I/O */ struct page ** pages; /* pages in our buffer */ unsigned int npages; /* count of pages */ atomic_t complete, /* i/os we're waiting for */ @@ -357,7 +360,9 @@ static ssize_t nfs_direct_read_seg(struct inode *inode, dreq->pages = pages; dreq->npages = nr_pages; + dreq->inode = inode; + nfs_add_stats(inode, NFSIOS_DIRECTREADBYTES, count); rpc_clnt_sigmask(clnt, &oldset); nfs_direct_read_schedule(dreq, inode, ctx, user_addr, count, file_offset); @@ -572,6 +577,7 @@ static ssize_t nfs_direct_write(struct inode *inode, return page_count; } + nfs_add_stats(inode, NFSIOS_DIRECTWRITTENBYTES, size); result = nfs_direct_write_seg(inode, ctx, user_addr, size, file_offset, pages, page_count); nfs_free_user_pages(pages, page_count, 0); @@ -581,6 +587,7 @@ static ssize_t nfs_direct_write(struct inode *inode, break; return result; } + nfs_add_stats(inode, NFSIOS_SERVERWRITTENBYTES, result); tot_bytes += result; file_offset += result; if (result < size) -- cgit v1.2.3-18-g5258 From ec06c096edec0755534c7126f4caded69de131c2 Mon Sep 17 00:00:00 2001 From: Trond Myklebust Date: Mon, 20 Mar 2006 13:44:27 -0500 Subject: NFS: Cleanup of NFS read code Same callback hierarchy inversion as for the NFS write calls. This patch is not strictly speaking needed by the O_DIRECT code, but avoids confusing differences between the asynchronous read and write code. Signed-off-by: Trond Myklebust --- fs/nfs/direct.c | 17 +++++++++++++---- 1 file changed, 13 insertions(+), 4 deletions(-) (limited to 'fs/nfs/direct.c') diff --git a/fs/nfs/direct.c b/fs/nfs/direct.c index fc07ce4885d..3f87a72bd13 100644 --- a/fs/nfs/direct.c +++ b/fs/nfs/direct.c @@ -218,14 +218,17 @@ static struct nfs_direct_req *nfs_direct_read_alloc(size_t nbytes, unsigned int * 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). */ -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)) + if (nfs_readpage_result(task, data) != 0) + return; + if (likely(task->tk_status >= 0)) atomic_add(data->res.count, &dreq->count); else - atomic_set(&dreq->error, status); + atomic_set(&dreq->error, task->tk_status); if (unlikely(atomic_dec_and_test(&dreq->complete))) { nfs_free_user_pages(dreq->pages, dreq->npages, 1); @@ -234,6 +237,11 @@ static void nfs_direct_read_result(struct nfs_read_data *data, int status) } } +static const struct rpc_call_ops nfs_read_direct_ops = { + .rpc_call_done = nfs_direct_read_result, + .rpc_release = nfs_readdata_release, +}; + /** * nfs_direct_read_schedule - dispatch NFS READ operations for a direct read * @dreq: address of nfs_direct_req struct for this request @@ -280,10 +288,11 @@ static void nfs_direct_read_schedule(struct nfs_direct_req *dreq, 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); -- cgit v1.2.3-18-g5258 From b8a32e2b8b7fefff994c89d398b6ac920a195b43 Mon Sep 17 00:00:00 2001 From: Chuck Lever Date: Mon, 20 Mar 2006 13:44:28 -0500 Subject: NFS: clean up NFS client's a_ops->direct_IO method The NFS client's a_ops->direct_IO method, nfs_direct_IO, is required to be present to allow NFS files to be opened with O_DIRECT, but is never called because the NFS client shunts reads and writes to files opened with O_DIRECT directly to its own routines. Gut the nfs_direct_IO function. This eliminates the only part of the NFS client's direct I/O path that requires support for multi-segment iovs, allowing further simplification in subsequent patches. Test plan: Compile the kernel with CONFIG_NFS and CONFIG_NFS_DIRECTIO enabled. Millions of fsx-odirect ops. OraSim. Signed-off-by: Chuck Lever Signed-off-by: Trond Myklebust --- fs/nfs/direct.c | 70 +++++++++++++++++++-------------------------------------- 1 file changed, 23 insertions(+), 47 deletions(-) (limited to 'fs/nfs/direct.c') diff --git a/fs/nfs/direct.c b/fs/nfs/direct.c index 3f87a72bd13..8096d326bd7 100644 --- a/fs/nfs/direct.c +++ b/fs/nfs/direct.c @@ -78,6 +78,29 @@ struct nfs_direct_req { }; +/** + * 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. + */ +ssize_t nfs_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov, loff_t pos, unsigned long nr_segs) +{ + struct dentry *dentry = iocb->ki_filp->f_dentry; + + dprintk("NFS: nfs_direct_IO (%s) off/no(%Ld/%lu) EINVAL\n", + dentry->d_name.name, (long long) pos, nr_segs); + + return -EINVAL; +} + /** * nfs_get_user_pages - find and set up pages underlying user's buffer * rw: direction (read or write) @@ -605,53 +628,6 @@ static ssize_t nfs_direct_write(struct inode *inode, return tot_bytes; } -/** - * 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 - * - */ -ssize_t -nfs_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov, - loff_t file_offset, unsigned long nr_segs) -{ - 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; - - /* - * 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; - } - return result; -} - /** * nfs_file_direct_read - file direct read operation for NFS files * @iocb: target I/O control block -- cgit v1.2.3-18-g5258 From d4cc948ba97980c55a308eab167a695109796456 Mon Sep 17 00:00:00 2001 From: Chuck Lever Date: Mon, 20 Mar 2006 13:44:28 -0500 Subject: NFS: update comments and function definitions in fs/nfs/direct.c Update to latest coding style standards. Remove block comments on statically defined functions, and place function definitions all on one line. Test plan: Compile kernel with CONFIG_NFS and CONFIG_NFS_DIRECTIO. Signed-off-by: Chuck Lever Signed-off-by: Trond Myklebust --- fs/nfs/direct.c | 125 +++++++------------------------------------------------- 1 file changed, 15 insertions(+), 110 deletions(-) (limited to 'fs/nfs/direct.c') diff --git a/fs/nfs/direct.c b/fs/nfs/direct.c index 8096d326bd7..d38c3dc052a 100644 --- a/fs/nfs/direct.c +++ b/fs/nfs/direct.c @@ -101,16 +101,7 @@ ssize_t nfs_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov, loff_ return -EINVAL; } -/** - * 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 int -nfs_get_user_pages(int rw, unsigned long user_addr, size_t size, - struct page ***pages) +static inline int nfs_get_user_pages(int rw, unsigned long user_addr, size_t size, struct page ***pages) { int result = -ENOMEM; unsigned long page_count; @@ -147,14 +138,7 @@ nfs_get_user_pages(int rw, unsigned long user_addr, size_t size, return result; } -/** - * 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 - */ -static void -nfs_free_user_pages(struct page **pages, int npages, int do_dirty) +static void nfs_free_user_pages(struct page **pages, int npages, int do_dirty) { int i; for (i = 0; i < npages; i++) { @@ -166,22 +150,13 @@ nfs_free_user_pages(struct page **pages, int npages, int do_dirty) 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_req_release(struct kref *kref) { struct nfs_direct_req *dreq = container_of(kref, struct nfs_direct_req, kref); 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 - * +/* * 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. @@ -232,11 +207,7 @@ static struct nfs_direct_req *nfs_direct_read_alloc(size_t nbytes, unsigned int 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). @@ -265,21 +236,11 @@ static const struct rpc_call_ops nfs_read_direct_ops = { .rpc_release = nfs_readdata_release, }; -/** - * 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 - * +/* * For each nfs_read_data struct that was allocated on the list, dispatch * an NFS READ operation */ -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 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) { struct list_head *list = &dreq->list; struct page **pages = dreq->pages; @@ -337,11 +298,7 @@ static void nfs_direct_read_schedule(struct nfs_direct_req *dreq, } while (count != 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) @@ -364,22 +321,7 @@ static ssize_t nfs_direct_read_wait(struct nfs_direct_req *dreq, int intr) return (ssize_t) result; } -/** - * 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_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) { ssize_t result; sigset_t oldset; @@ -404,22 +346,11 @@ static ssize_t nfs_direct_read_seg(struct inode *inode, 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 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) { ssize_t tot_bytes = 0; unsigned long seg = 0; @@ -457,21 +388,7 @@ nfs_direct_read(struct inode *inode, struct nfs_open_context *ctx, return tot_bytes; } -/** - * 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 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) { const unsigned int wsize = NFS_SERVER(inode)->wsize; size_t request; @@ -573,22 +490,12 @@ sync_retry: goto retry; } -/** - * 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 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) { ssize_t tot_bytes = 0; unsigned long seg = 0; @@ -649,8 +556,7 @@ static ssize_t nfs_direct_write(struct inode *inode, * 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; @@ -717,8 +623,7 @@ out: * 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; -- cgit v1.2.3-18-g5258 From 5dd602f20688e08c85ac91e0451c4e6321ed25d7 Mon Sep 17 00:00:00 2001 From: Chuck Lever Date: Mon, 20 Mar 2006 13:44:29 -0500 Subject: NFS: use size_t type for holding rsize bytes in NFS O_DIRECT read path size_t is used for holding byte counts, so use it for variables storing rsize. Note that the write path will be updated as we add support for async O_DIRECT writes. Test plan: Need to verify that existing comparisons against new size_t variables behave correctly. Signed-off-by: Chuck Lever Signed-off-by: Trond Myklebust --- fs/nfs/direct.c | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-) (limited to 'fs/nfs/direct.c') diff --git a/fs/nfs/direct.c b/fs/nfs/direct.c index d38c3dc052a..8f5d2dfd5a8 100644 --- a/fs/nfs/direct.c +++ b/fs/nfs/direct.c @@ -161,7 +161,7 @@ static void nfs_direct_req_release(struct kref *kref) * 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; @@ -244,14 +244,14 @@ static void nfs_direct_read_schedule(struct nfs_direct_req *dreq, struct inode * { struct list_head *list = &dreq->list; struct page **pages = dreq->pages; + size_t rsize = NFS_SERVER(inode)->rsize; unsigned int curpage, pgbase; - unsigned int rsize = NFS_SERVER(inode)->rsize; curpage = 0; pgbase = user_addr & ~PAGE_MASK; do { struct nfs_read_data *data; - unsigned int bytes; + size_t bytes; bytes = rsize; if (count < rsize) -- cgit v1.2.3-18-g5258 From 0cdd80d07fb0f558dfdb30f6e0b9905f5e5475f1 Mon Sep 17 00:00:00 2001 From: Chuck Lever Date: Mon, 20 Mar 2006 13:44:29 -0500 Subject: NFS: remove support for multi-segment iovs in the direct read path Eliminate the persistent use of automatic storage in all parts of the NFS client's direct read path to pave the way for introducing support for aio against files opened with the O_DIRECT flag. Test plan: Compile the kernel with CONFIG_NFS and CONFIG_NFS_DIRECTIO enabled. Millions of fsx-odirect ops. OraSim. Signed-off-by: Chuck Lever Signed-off-by: Trond Myklebust --- fs/nfs/direct.c | 66 +++++++++++++-------------------------------------------- 1 file changed, 15 insertions(+), 51 deletions(-) (limited to 'fs/nfs/direct.c') diff --git a/fs/nfs/direct.c b/fs/nfs/direct.c index 8f5d2dfd5a8..6ecde9602f9 100644 --- a/fs/nfs/direct.c +++ b/fs/nfs/direct.c @@ -321,7 +321,7 @@ static ssize_t nfs_direct_read_wait(struct nfs_direct_req *dreq, int intr) return (ssize_t) result; } -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 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) { ssize_t result; sigset_t oldset; @@ -346,48 +346,6 @@ static ssize_t nfs_direct_read_seg(struct inode *inode, struct nfs_open_context return result; } -/* - * 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) -{ - 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; - } - - result = nfs_direct_read_seg(inode, ctx, user_addr, size, - file_offset, pages, page_count); - - if (result <= 0) { - if (tot_bytes > 0) - break; - return result; - } - tot_bytes += result; - file_offset += result; - if (result < size) - break; - } - - return tot_bytes; -} - 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) { const unsigned int wsize = NFS_SERVER(inode)->wsize; @@ -559,16 +517,13 @@ static ssize_t nfs_direct_write(struct inode *inode, struct nfs_open_context *ct 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; + int page_count; + struct page **pages; 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, @@ -580,7 +535,7 @@ ssize_t nfs_file_direct_read(struct kiocb *iocb, char __user *buf, size_t count, 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) @@ -590,9 +545,18 @@ ssize_t nfs_file_direct_read(struct kiocb *iocb, char __user *buf, size_t count, if (retval) goto out; - retval = nfs_direct_read(inode, ctx, &iov, pos, 1); + page_count = nfs_get_user_pages(READ, (unsigned long) buf, + count, &pages); + if (page_count < 0) { + nfs_free_user_pages(pages, 0, 0); + retval = page_count; + goto out; + } + + retval = nfs_direct_read(inode, ctx, (unsigned long) buf, count, pos, + pages, page_count); if (retval > 0) - *ppos = pos + retval; + iocb->ki_pos = pos + retval; out: return retval; -- cgit v1.2.3-18-g5258 From 99514f8fdda2beef1ca922b7f9d89c1a2c57fec0 Mon Sep 17 00:00:00 2001 From: Chuck Lever Date: Mon, 20 Mar 2006 13:44:30 -0500 Subject: NFS: make iocb available everywhere in direct read path Pass the iocb argument all the way down to the direct read request scheduler, and make it available in nfs_direct_read_result. Test plan: Compile the kernel with CONFIG_NFS and CONFIG_NFS_DIRECTIO enabled. Millions of fsx-odirect ops. OraSim. Signed-off-by: Chuck Lever Signed-off-by: Trond Myklebust --- fs/nfs/direct.c | 20 ++++++++++++-------- 1 file changed, 12 insertions(+), 8 deletions(-) (limited to 'fs/nfs/direct.c') diff --git a/fs/nfs/direct.c b/fs/nfs/direct.c index 6ecde9602f9..6cbddc51acb 100644 --- a/fs/nfs/direct.c +++ b/fs/nfs/direct.c @@ -68,6 +68,8 @@ static kmem_cache_t *nfs_direct_cachep; struct nfs_direct_req { struct kref kref; /* release manager */ struct list_head list; /* nfs_read_data structs */ + struct file * filp; /* file descriptor */ + struct kiocb * iocb; /* controlling i/o request */ wait_queue_head_t wait; /* wait for i/o completion */ struct inode * inode; /* target file of I/O */ struct page ** pages; /* pages in our buffer */ @@ -240,8 +242,12 @@ static const struct rpc_call_ops nfs_read_direct_ops = { * For each nfs_read_data struct that was allocated on the list, dispatch * an NFS READ operation */ -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 void nfs_direct_read_schedule(struct nfs_direct_req *dreq, unsigned long user_addr, size_t count, loff_t file_offset) { + struct file *file = dreq->filp; + struct inode *inode = file->f_mapping->host; + struct nfs_open_context *ctx = (struct nfs_open_context *) + file->private_data; struct list_head *list = &dreq->list; struct page **pages = dreq->pages; size_t rsize = NFS_SERVER(inode)->rsize; @@ -321,10 +327,11 @@ static ssize_t nfs_direct_read_wait(struct nfs_direct_req *dreq, int intr) return (ssize_t) result; } -static ssize_t nfs_direct_read(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 file_offset, struct page **pages, unsigned int nr_pages) { 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; @@ -335,11 +342,11 @@ static ssize_t nfs_direct_read(struct inode *inode, struct nfs_open_context *ctx dreq->pages = pages; dreq->npages = nr_pages; dreq->inode = inode; + dreq->filp = iocb->ki_filp; nfs_add_stats(inode, NFSIOS_DIRECTREADBYTES, count); rpc_clnt_sigmask(clnt, &oldset); - nfs_direct_read_schedule(dreq, inode, ctx, user_addr, count, - file_offset); + nfs_direct_read_schedule(dreq, user_addr, count, file_offset); result = nfs_direct_read_wait(dreq, clnt->cl_intr); rpc_clnt_sigunmask(clnt, &oldset); @@ -520,10 +527,7 @@ ssize_t nfs_file_direct_read(struct kiocb *iocb, char __user *buf, size_t count, int page_count; struct page **pages; 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; dprintk("nfs: direct read(%s/%s, %lu@%Ld)\n", file->f_dentry->d_parent->d_name.name, @@ -553,7 +557,7 @@ ssize_t nfs_file_direct_read(struct kiocb *iocb, char __user *buf, size_t count, goto out; } - retval = nfs_direct_read(inode, ctx, (unsigned long) buf, count, pos, + retval = nfs_direct_read(iocb, (unsigned long) buf, count, pos, pages, page_count); if (retval > 0) iocb->ki_pos = pos + retval; -- cgit v1.2.3-18-g5258 From 487b83723ed4d4eaafd5109f36560da4f15c6578 Mon Sep 17 00:00:00 2001 From: Chuck Lever Date: Mon, 20 Mar 2006 13:44:30 -0500 Subject: NFS: support EIOCBQUEUED return in direct read path For async iocb's, the NFS direct read path should return EIOCBQUEUED and call aio_complete when all the requested reads are finished. The synchronous part of the NFS direct read path behaves exactly as it was before. Test plan: aio-stress with "-O". OraSim. Signed-off-by: Chuck Lever Signed-off-by: Trond Myklebust --- fs/nfs/direct.c | 25 +++++++++++++++++++++---- 1 file changed, 21 insertions(+), 4 deletions(-) (limited to 'fs/nfs/direct.c') diff --git a/fs/nfs/direct.c b/fs/nfs/direct.c index 6cbddc51acb..094456c3df9 100644 --- a/fs/nfs/direct.c +++ b/fs/nfs/direct.c @@ -177,6 +177,7 @@ static struct nfs_direct_req *nfs_direct_read_alloc(size_t nbytes, size_t rsize) kref_init(&dreq->kref); init_waitqueue_head(&dreq->wait); INIT_LIST_HEAD(&dreq->list); + dreq->iocb = NULL; atomic_set(&dreq->count, 0); atomic_set(&dreq->error, 0); @@ -213,6 +214,10 @@ static struct nfs_direct_req *nfs_direct_read_alloc(size_t nbytes, size_t rsize) * 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). + * + * In addition, 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. If the waiter is woken prematurely, the iocb is long gone. */ static void nfs_direct_read_result(struct rpc_task *task, void *calldata) { @@ -228,7 +233,13 @@ static void nfs_direct_read_result(struct rpc_task *task, void *calldata) if (unlikely(atomic_dec_and_test(&dreq->complete))) { nfs_free_user_pages(dreq->pages, dreq->npages, 1); - wake_up(&dreq->wait); + if (dreq->iocb) { + long res = atomic_read(&dreq->error); + if (!res) + res = atomic_read(&dreq->count); + aio_complete(dreq->iocb, res, 0); + } else + wake_up(&dreq->wait); kref_put(&dreq->kref, nfs_direct_req_release); } } @@ -309,8 +320,13 @@ static void nfs_direct_read_schedule(struct nfs_direct_req *dreq, unsigned long */ static ssize_t nfs_direct_read_wait(struct nfs_direct_req *dreq, int intr) { - int result = 0; + int result = -EIOCBQUEUED; + + /* Async requests don't wait here */ + if (dreq->iocb) + goto out; + result = 0; if (intr) { result = wait_event_interruptible(dreq->wait, (atomic_read(&dreq->complete) == 0)); @@ -323,6 +339,7 @@ static ssize_t nfs_direct_read_wait(struct nfs_direct_req *dreq, int intr) if (!result) result = atomic_read(&dreq->count); +out: kref_put(&dreq->kref, nfs_direct_req_release); return (ssize_t) result; } @@ -343,6 +360,8 @@ static ssize_t nfs_direct_read(struct kiocb *iocb, unsigned long user_addr, size dreq->npages = nr_pages; dreq->inode = inode; dreq->filp = iocb->ki_filp; + if (!is_sync_kiocb(iocb)) + dreq->iocb = iocb; nfs_add_stats(inode, NFSIOS_DIRECTREADBYTES, count); rpc_clnt_sigmask(clnt, &oldset); @@ -534,8 +553,6 @@ ssize_t nfs_file_direct_read(struct kiocb *iocb, char __user *buf, size_t count, 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; -- cgit v1.2.3-18-g5258 From bc0fb201b34b12e2d16e8cbd5bb078c1db936304 Mon Sep 17 00:00:00 2001 From: Chuck Lever Date: Mon, 20 Mar 2006 13:44:31 -0500 Subject: NFS: create common routine for waiting for direct I/O to complete We're about to add asynchrony to the NFS direct write path. Begin by abstracting out the common pieces in the read path. The first piece is nfs_direct_read_wait, which works the same whether the process is waiting for a read or a write. Test plan: Compile kernel with CONFIG_NFS and CONFIG_NFS_DIRECTIO enabled. Signed-off-by: Chuck Lever Signed-off-by: Trond Myklebust --- fs/nfs/direct.c | 57 ++++++++++++++++++++++++++------------------------------- 1 file changed, 26 insertions(+), 31 deletions(-) (limited to 'fs/nfs/direct.c') diff --git a/fs/nfs/direct.c b/fs/nfs/direct.c index 094456c3df9..2593f47eaff 100644 --- a/fs/nfs/direct.c +++ b/fs/nfs/direct.c @@ -158,6 +158,30 @@ static void nfs_direct_req_release(struct kref *kref) kmem_cache_free(nfs_direct_cachep, dreq); } +/* + * Collects and returns the final error value/byte-count. + */ +static ssize_t nfs_direct_wait(struct nfs_direct_req *dreq) +{ + int result = -EIOCBQUEUED; + + /* Async requests don't wait here */ + if (dreq->iocb) + goto out; + + result = wait_event_interruptible(dreq->wait, + (atomic_read(&dreq->complete) == 0)); + + if (!result) + result = atomic_read(&dreq->error); + if (!result) + result = atomic_read(&dreq->count); + +out: + kref_put(&dreq->kref, nfs_direct_req_release); + return (ssize_t) result; +} + /* * 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 @@ -213,7 +237,7 @@ static struct nfs_direct_req *nfs_direct_read_alloc(size_t nbytes, size_t rsize) /* * 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). * * In addition, synchronous I/O uses a stack-allocated iocb. Thus we * can't trust the iocb is still valid here if this is a synchronous @@ -315,35 +339,6 @@ static void nfs_direct_read_schedule(struct nfs_direct_req *dreq, unsigned long } while (count != 0); } -/* - * 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 = -EIOCBQUEUED; - - /* Async requests don't wait here */ - if (dreq->iocb) - goto out; - - result = 0; - if (intr) { - result = wait_event_interruptible(dreq->wait, - (atomic_read(&dreq->complete) == 0)); - } else { - wait_event(dreq->wait, (atomic_read(&dreq->complete) == 0)); - } - - if (!result) - result = atomic_read(&dreq->error); - if (!result) - result = atomic_read(&dreq->count); - -out: - kref_put(&dreq->kref, nfs_direct_req_release); - return (ssize_t) result; -} - static ssize_t nfs_direct_read(struct kiocb *iocb, unsigned long user_addr, size_t count, loff_t file_offset, struct page **pages, unsigned int nr_pages) { ssize_t result; @@ -366,7 +361,7 @@ static ssize_t nfs_direct_read(struct kiocb *iocb, unsigned long user_addr, size nfs_add_stats(inode, NFSIOS_DIRECTREADBYTES, count); rpc_clnt_sigmask(clnt, &oldset); nfs_direct_read_schedule(dreq, user_addr, count, file_offset); - result = nfs_direct_read_wait(dreq, clnt->cl_intr); + result = nfs_direct_wait(dreq); rpc_clnt_sigunmask(clnt, &oldset); return result; -- cgit v1.2.3-18-g5258 From 93619e5989173614bef0013b0bb8a3fe3dbd5a95 Mon Sep 17 00:00:00 2001 From: Chuck Lever Date: Mon, 20 Mar 2006 13:44:31 -0500 Subject: NFS: create common routine for allocating nfs_direct_req Factor out a small common piece of the path that allocate nfs_direct_req structures. Test plan: Compile kernel with CONFIG_NFS and CONFIG_NFS_DIRECTIO enabled. Signed-off-by: Chuck Lever Signed-off-by: Trond Myklebust --- fs/nfs/direct.c | 27 +++++++++++++++++++-------- 1 file changed, 19 insertions(+), 8 deletions(-) (limited to 'fs/nfs/direct.c') diff --git a/fs/nfs/direct.c b/fs/nfs/direct.c index 2593f47eaff..489f736d0f5 100644 --- a/fs/nfs/direct.c +++ b/fs/nfs/direct.c @@ -152,6 +152,24 @@ static void nfs_free_user_pages(struct page **pages, int npages, int do_dirty) kfree(pages); } +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_waitqueue_head(&dreq->wait); + INIT_LIST_HEAD(&dreq->list); + dreq->iocb = NULL; + atomic_set(&dreq->count, 0); + atomic_set(&dreq->error, 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); @@ -194,17 +212,10 @@ static struct nfs_direct_req *nfs_direct_read_alloc(size_t nbytes, size_t rsize) 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); - dreq->iocb = NULL; - atomic_set(&dreq->count, 0); - atomic_set(&dreq->error, 0); - list = &dreq->list; for(;;) { struct nfs_read_data *data = nfs_readdata_alloc(rpages); -- cgit v1.2.3-18-g5258 From 63ab46abc70b01cb0711301f5ddb08c1c0bb9b1c Mon Sep 17 00:00:00 2001 From: Chuck Lever Date: Mon, 20 Mar 2006 13:44:31 -0500 Subject: NFS: create common routine for handling direct I/O completion Factor out the common piece of completing an NFS direct I/O request. Test plan: Compile kernel with CONFIG_NFS and CONFIG_NFS_DIRECTIO enabled. Signed-off-by: Chuck Lever Signed-off-by: Trond Myklebust --- fs/nfs/direct.c | 46 ++++++++++++++++++++++++++-------------------- 1 file changed, 26 insertions(+), 20 deletions(-) (limited to 'fs/nfs/direct.c') diff --git a/fs/nfs/direct.c b/fs/nfs/direct.c index 489f736d0f5..4df21ce28e1 100644 --- a/fs/nfs/direct.c +++ b/fs/nfs/direct.c @@ -200,6 +200,30 @@ out: return (ssize_t) result; } +/* + * 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_wait (for instance, if someone hits ^C on a slow server). + * + * In addition, 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. If the waiter is woken prematurely, the iocb is long gone. + */ +static void nfs_direct_complete(struct nfs_direct_req *dreq) +{ + nfs_free_user_pages(dreq->pages, dreq->npages, 1); + + if (dreq->iocb) { + long res = atomic_read(&dreq->error); + if (!res) + res = atomic_read(&dreq->count); + aio_complete(dreq->iocb, res, 0); + } else + wake_up(&dreq->wait); + + 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 @@ -245,15 +269,6 @@ static struct nfs_direct_req *nfs_direct_read_alloc(size_t nbytes, size_t rsize) return dreq; } -/* - * 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_wait (for instance, if someone hits ^C on a slow server). - * - * In addition, 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. If the waiter is woken prematurely, the iocb is long gone. - */ static void nfs_direct_read_result(struct rpc_task *task, void *calldata) { struct nfs_read_data *data = calldata; @@ -266,17 +281,8 @@ static void nfs_direct_read_result(struct rpc_task *task, void *calldata) else atomic_set(&dreq->error, task->tk_status); - if (unlikely(atomic_dec_and_test(&dreq->complete))) { - nfs_free_user_pages(dreq->pages, dreq->npages, 1); - if (dreq->iocb) { - long res = atomic_read(&dreq->error); - if (!res) - res = atomic_read(&dreq->count); - aio_complete(dreq->iocb, res, 0); - } else - wake_up(&dreq->wait); - kref_put(&dreq->kref, nfs_direct_req_release); - } + if (unlikely(atomic_dec_and_test(&dreq->complete))) + nfs_direct_complete(dreq); } static const struct rpc_call_ops nfs_read_direct_ops = { -- cgit v1.2.3-18-g5258 From 462d5b3296b56289efec426499a83faad4c08d9e Mon Sep 17 00:00:00 2001 From: Chuck Lever Date: Mon, 20 Mar 2006 13:44:32 -0500 Subject: NFS: make direct write path generate write requests concurrently Duplicate infrastructure from direct read path that will allow write path to generate multiple write requests concurrently. This will enable us to add support for aio in this path. Temporarily we will lose the ability to do UNSTABLE writes followed by a COMMIT in the direct write path. However, all applications I am aware of that use NFS O_DIRECT currently write in relatively small chunks, so this should not be inconvenient in any way. Test plan: Millions of fsx-odirect ops. OraSim. Signed-off-by: Chuck Lever Signed-off-by: Trond Myklebust --- fs/nfs/direct.c | 240 +++++++++++++++++++++++++++++++++++++------------------- 1 file changed, 159 insertions(+), 81 deletions(-) (limited to 'fs/nfs/direct.c') diff --git a/fs/nfs/direct.c b/fs/nfs/direct.c index 4df21ce28e1..dea3239cdde 100644 --- a/fs/nfs/direct.c +++ b/fs/nfs/direct.c @@ -384,106 +384,185 @@ static ssize_t nfs_direct_read(struct kiocb *iocb, unsigned long user_addr, size return result; } -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 struct nfs_direct_req *nfs_direct_write_alloc(size_t nbytes, size_t wsize) { - 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 list_head *list; + struct nfs_direct_req *dreq; + unsigned int writes = 0; + unsigned int wpages = (wsize + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT; - 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; + 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; + writes++; + if (nbytes <= wsize) + break; + nbytes -= wsize; + } + kref_get(&dreq->kref); + atomic_set(&dreq->complete, writes); + return dreq; +} + +/* + * Collects and returns the final error value/byte-count. + */ +static ssize_t nfs_direct_write_wait(struct nfs_direct_req *dreq, int intr) +{ + int result = 0; + + if (intr) { + result = wait_event_interruptible(dreq->wait, + (atomic_read(&dreq->complete) == 0)); + } else { + wait_event(dreq->wait, (atomic_read(&dreq->complete) == 0)); + } + + 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; +} + +static void nfs_direct_write_result(struct rpc_task *task, void *calldata) +{ + 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; + /* If the server fell back to an UNSTABLE write, it's an error. */ + if (unlikely(data->res.verf->committed != NFS_FILE_SYNC)) + status = -EIO; + + if (likely(status >= 0)) + atomic_add(data->res.count, &dreq->count); + else + atomic_set(&dreq->error, status); + + if (unlikely(atomic_dec_and_test(&dreq->complete))) + nfs_direct_complete(dreq); +} + +static const struct rpc_call_ops nfs_write_direct_ops = { + .rpc_call_done = nfs_direct_write_result, + .rpc_release = nfs_writedata_release, +}; + +/* + * For each nfs_write_data struct that was allocated on the list, dispatch + * an NFS WRITE operation + * + * XXX: For now, support only FILE_SYNC writes. Later we may add + * support for UNSTABLE + COMMIT. + */ +static void nfs_direct_write_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) +{ + struct list_head *list = &dreq->list; + struct page **pages = dreq->pages; + size_t wsize = NFS_SERVER(inode)->wsize; + unsigned int curpage, pgbase; - 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; + pgbase = user_addr & ~PAGE_MASK; do { - wdata->args.count = request; - if (wdata->args.count > wsize) - wdata->args.count = wsize; - wdata->args.pages = &pages[curpage]; + struct nfs_write_data *data; + size_t bytes; + + bytes = wsize; + if (count < wsize) + bytes = count; + + data = list_entry(list->next, struct nfs_write_data, pages); + list_del_init(&data->pages); + + data->inode = inode; + data->cred = ctx->cred; + data->args.fh = NFS_FH(inode); + data->args.context = ctx; + data->args.offset = file_offset; + data->args.pgbase = pgbase; + data->args.pages = &pages[curpage]; + data->args.count = bytes; + data->res.fattr = &data->fattr; + data->res.count = bytes; + + rpc_init_task(&data->task, NFS_CLIENT(inode), RPC_TASK_ASYNC, + &nfs_write_direct_ops, data); + NFS_PROTO(inode)->write_setup(data, FLUSH_STABLE); - 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); + data->task.tk_priority = RPC_PRIORITY_NORMAL; + data->task.tk_cookie = (unsigned long) inode; lock_kernel(); - result = NFS_PROTO(inode)->write(wdata); + rpc_execute(&data->task); unlock_kernel(); - if (result <= 0) { - if (tot_bytes > 0) - break; - goto out; - } + dfprintk(VFS, "NFS: %4d initiated 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), + bytes, + (unsigned long long)data->args.offset); - 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; - } + file_offset += bytes; + pgbase += bytes; + curpage += pgbase >> PAGE_SHIFT; + pgbase &= ~PAGE_MASK; - tot_bytes += result; + count -= bytes; + } while (count != 0); +} - /* in case of a short write: stop now, let the app recover */ - if (result < wdata->args.count) - break; +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) +{ + ssize_t result; + sigset_t oldset; + struct rpc_clnt *clnt = NFS_CLIENT(inode); + struct nfs_direct_req *dreq; - wdata->args.offset += result; - wdata->args.pgbase += result; - curpage += wdata->args.pgbase >> PAGE_SHIFT; - wdata->args.pgbase &= ~PAGE_MASK; - request -= result; - } while (request != 0); + dreq = nfs_direct_write_alloc(count, NFS_SERVER(inode)->wsize); + if (!dreq) + return -ENOMEM; - /* - * 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; + dreq->pages = pages; + dreq->npages = nr_pages; - lock_kernel(); - result = NFS_PROTO(inode)->commit(wdata); - unlock_kernel(); + nfs_begin_data_update(inode); - if (result < 0 || memcmp(&first_verf.verifier, - &wdata->verf.verifier, - sizeof(first_verf.verifier)) != 0) - goto sync_retry; - } - result = tot_bytes; + rpc_clnt_sigmask(clnt, &oldset); + nfs_direct_write_schedule(dreq, inode, ctx, user_addr, count, + file_offset); + result = nfs_direct_write_wait(dreq, clnt->cl_intr); + rpc_clnt_sigunmask(clnt, &oldset); -out: nfs_end_data_update(inode); - nfs_writedata_free(wdata); - return result; -sync_retry: - wdata->args.stable = NFS_FILE_SYNC; - goto retry; + return result; } /* @@ -515,7 +594,6 @@ static ssize_t nfs_direct_write(struct inode *inode, struct nfs_open_context *ct nfs_add_stats(inode, NFSIOS_DIRECTWRITTENBYTES, size); result = nfs_direct_write_seg(inode, ctx, user_addr, size, file_offset, pages, page_count); - nfs_free_user_pages(pages, page_count, 0); if (result <= 0) { if (tot_bytes > 0) -- cgit v1.2.3-18-g5258 From 47989d7454398827500d0e73766270986a3b488f Mon Sep 17 00:00:00 2001 From: Chuck Lever Date: Mon, 20 Mar 2006 13:44:32 -0500 Subject: NFS: remove support for multi-segment iovs in the direct write path Eliminate the persistent use of automatic storage in all parts of the NFS client's direct write path to pave the way for introducing support for aio against files opened with the O_DIRECT flag. Test plan: Compile the kernel with CONFIG_NFS and CONFIG_NFS_DIRECTIO enabled. Millions of fsx-odirect ops. OraSim. Signed-off-by: Chuck Lever Signed-off-by: Trond Myklebust --- fs/nfs/direct.c | 68 +++++++++++++++------------------------------------------ 1 file changed, 17 insertions(+), 51 deletions(-) (limited to 'fs/nfs/direct.c') diff --git a/fs/nfs/direct.c b/fs/nfs/direct.c index dea3239cdde..9a7d4590705 100644 --- a/fs/nfs/direct.c +++ b/fs/nfs/direct.c @@ -510,6 +510,7 @@ static void nfs_direct_write_schedule(struct nfs_direct_req *dreq, struct inode 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); @@ -538,7 +539,7 @@ static void nfs_direct_write_schedule(struct nfs_direct_req *dreq, struct inode } while (count != 0); } -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 ssize_t nfs_direct_write(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) { ssize_t result; sigset_t oldset; @@ -552,6 +553,8 @@ static ssize_t nfs_direct_write_seg(struct inode *inode, struct nfs_open_context dreq->pages = pages; dreq->npages = nr_pages; + nfs_add_stats(inode, NFSIOS_DIRECTWRITTENBYTES, count); + nfs_begin_data_update(inode); rpc_clnt_sigmask(clnt, &oldset); @@ -565,50 +568,6 @@ static ssize_t nfs_direct_write_seg(struct inode *inode, struct nfs_open_context return result; } -/* - * 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) -{ - 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; - } - - nfs_add_stats(inode, NFSIOS_DIRECTWRITTENBYTES, size); - result = nfs_direct_write_seg(inode, ctx, user_addr, size, - file_offset, pages, page_count); - - if (result <= 0) { - if (tot_bytes > 0) - break; - return result; - } - nfs_add_stats(inode, NFSIOS_SERVERWRITTENBYTES, result); - tot_bytes += result; - file_offset += result; - if (result < size) - break; - } - return tot_bytes; -} - /** * nfs_file_direct_read - file direct read operation for NFS files * @iocb: target I/O control block @@ -701,14 +660,13 @@ out: ssize_t nfs_file_direct_write(struct kiocb *iocb, const char __user *buf, size_t count, loff_t pos) { ssize_t retval; + int page_count; + struct page **pages; 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, @@ -729,17 +687,25 @@ ssize_t nfs_file_direct_write(struct kiocb *iocb, const char __user *buf, size_t 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); + page_count = nfs_get_user_pages(WRITE, (unsigned long) buf, + count, &pages); + if (page_count < 0) { + nfs_free_user_pages(pages, 0, 0); + retval = page_count; + goto out; + } + + retval = nfs_direct_write(inode, ctx, (unsigned long) buf, count, + pos, pages, page_count); if (mapping->nrpages) invalidate_inode_pages2(mapping); if (retval > 0) -- cgit v1.2.3-18-g5258 From c89f2ee5f9223b864725f7344f24a037dfa76568 Mon Sep 17 00:00:00 2001 From: Chuck Lever Date: Mon, 20 Mar 2006 13:44:33 -0500 Subject: NFS: make iocb available everywhere in direct write path Pass the iocb argument all the way down to the direct write request scheduler, and make it available in nfs_direct_write_result. Test plan: Compile the kernel with CONFIG_NFS and CONFIG_NFS_DIRECTIO enabled. Millions of fsx-odirect ops. OraSim. Signed-off-by: Chuck Lever Signed-off-by: Trond Myklebust --- fs/nfs/direct.c | 46 ++++++++++++++-------------------------------- 1 file changed, 14 insertions(+), 32 deletions(-) (limited to 'fs/nfs/direct.c') diff --git a/fs/nfs/direct.c b/fs/nfs/direct.c index 9a7d4590705..9d57a299824 100644 --- a/fs/nfs/direct.c +++ b/fs/nfs/direct.c @@ -424,29 +424,6 @@ static struct nfs_direct_req *nfs_direct_write_alloc(size_t nbytes, size_t wsize return dreq; } -/* - * Collects and returns the final error value/byte-count. - */ -static ssize_t nfs_direct_write_wait(struct nfs_direct_req *dreq, int intr) -{ - int result = 0; - - if (intr) { - result = wait_event_interruptible(dreq->wait, - (atomic_read(&dreq->complete) == 0)); - } else { - wait_event(dreq->wait, (atomic_read(&dreq->complete) == 0)); - } - - 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; -} - static void nfs_direct_write_result(struct rpc_task *task, void *calldata) { struct nfs_write_data *data = calldata; @@ -480,8 +457,12 @@ static const struct rpc_call_ops nfs_write_direct_ops = { * XXX: For now, support only FILE_SYNC writes. Later we may add * support for UNSTABLE + COMMIT. */ -static void nfs_direct_write_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 void nfs_direct_write_schedule(struct nfs_direct_req *dreq, unsigned long user_addr, size_t count, loff_t file_offset) { + struct file *file = dreq->filp; + struct inode *inode = file->f_mapping->host; + struct nfs_open_context *ctx = (struct nfs_open_context *) + file->private_data; struct list_head *list = &dreq->list; struct page **pages = dreq->pages; size_t wsize = NFS_SERVER(inode)->wsize; @@ -539,10 +520,11 @@ static void nfs_direct_write_schedule(struct nfs_direct_req *dreq, struct inode } while (count != 0); } -static ssize_t nfs_direct_write(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 ssize_t nfs_direct_write(struct kiocb *iocb, unsigned long user_addr, size_t count, loff_t file_offset, struct page **pages, int nr_pages) { 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; @@ -552,15 +534,18 @@ static ssize_t nfs_direct_write(struct inode *inode, struct nfs_open_context *ct dreq->pages = pages; dreq->npages = nr_pages; + dreq->inode = inode; + dreq->filp = iocb->ki_filp; + 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); - nfs_direct_write_schedule(dreq, inode, ctx, user_addr, count, - file_offset); - result = nfs_direct_write_wait(dreq, clnt->cl_intr); + nfs_direct_write_schedule(dreq, user_addr, count, file_offset); + result = nfs_direct_wait(dreq); rpc_clnt_sigunmask(clnt, &oldset); nfs_end_data_update(inode); @@ -663,10 +648,7 @@ ssize_t nfs_file_direct_write(struct kiocb *iocb, const char __user *buf, size_t int page_count; struct page **pages; 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; dfprintk(VFS, "nfs: direct write(%s/%s, %lu@%Ld)\n", file->f_dentry->d_parent->d_name.name, @@ -704,7 +686,7 @@ ssize_t nfs_file_direct_write(struct kiocb *iocb, const char __user *buf, size_t goto out; } - retval = nfs_direct_write(inode, ctx, (unsigned long) buf, count, + retval = nfs_direct_write(iocb, (unsigned long) buf, count, pos, pages, page_count); if (mapping->nrpages) invalidate_inode_pages2(mapping); -- cgit v1.2.3-18-g5258 From 9eafa8cc521b489f205bf7b0634c99e34e046606 Mon Sep 17 00:00:00 2001 From: Chuck Lever Date: Mon, 20 Mar 2006 13:44:33 -0500 Subject: NFS: support EIOCBQUEUED return in direct write path For async iocb's, the NFS direct write path now returns EIOCBQUEUED, and calls aio_complete when all the requested writes are finished. The synchronous part of the NFS direct write path behaves exactly as it was before. Shared mapped NFS files will have some coherency difficulties when accessed concurrently with aio+dio. Will need to explore how this is handled in the local file system case. Test plan: aio-stress with "-O". OraSim. Signed-off-by: Chuck Lever Signed-off-by: Trond Myklebust --- fs/nfs/direct.c | 20 +++++++++++++------- 1 file changed, 13 insertions(+), 7 deletions(-) (limited to 'fs/nfs/direct.c') diff --git a/fs/nfs/direct.c b/fs/nfs/direct.c index 9d57a299824..df86e526702 100644 --- a/fs/nfs/direct.c +++ b/fs/nfs/direct.c @@ -441,8 +441,10 @@ static void nfs_direct_write_result(struct rpc_task *task, void *calldata) else atomic_set(&dreq->error, status); - if (unlikely(atomic_dec_and_test(&dreq->complete))) + if (unlikely(atomic_dec_and_test(&dreq->complete))) { + nfs_end_data_update(data->inode); nfs_direct_complete(dreq); + } } static const struct rpc_call_ops nfs_write_direct_ops = { @@ -548,8 +550,6 @@ static ssize_t nfs_direct_write(struct kiocb *iocb, unsigned long user_addr, siz result = nfs_direct_wait(dreq); rpc_clnt_sigunmask(clnt, &oldset); - nfs_end_data_update(inode); - return result; } @@ -655,10 +655,6 @@ ssize_t nfs_file_direct_write(struct kiocb *iocb, const char __user *buf, size_t 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; @@ -688,8 +684,18 @@ ssize_t nfs_file_direct_write(struct kiocb *iocb, const char __user *buf, size_t retval = nfs_direct_write(iocb, (unsigned long) buf, count, pos, pages, page_count); + + /* + * 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; -- cgit v1.2.3-18-g5258 From 88467055f7654302c12df74e5fe4d12516656a39 Mon Sep 17 00:00:00 2001 From: Chuck Lever Date: Mon, 20 Mar 2006 13:44:34 -0500 Subject: NFS: clean up comments and tab damage in direct.c Clean up tab damage and comments. Replace "file_offset" with more commonly used "pos". Test plan: Compile with CONFIG_NFS and CONFIG_NFS_DIRECTIO enabled. Signed-off-by: Chuck Lever Signed-off-by: Trond Myklebust --- fs/nfs/direct.c | 54 +++++++++++++++++++++++++++++++----------------------- 1 file changed, 31 insertions(+), 23 deletions(-) (limited to 'fs/nfs/direct.c') diff --git a/fs/nfs/direct.c b/fs/nfs/direct.c index df86e526702..bcbc213b403 100644 --- a/fs/nfs/direct.c +++ b/fs/nfs/direct.c @@ -7,11 +7,11 @@ * * 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 @@ -34,6 +34,7 @@ * 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 * */ @@ -67,11 +68,11 @@ static kmem_cache_t *nfs_direct_cachep; */ struct nfs_direct_req { struct kref kref; /* release manager */ - struct list_head list; /* nfs_read_data structs */ + struct list_head list; /* nfs_read/write_data structs */ struct file * filp; /* file descriptor */ struct kiocb * iocb; /* controlling i/o request */ wait_queue_head_t wait; /* wait for i/o completion */ - struct inode * inode; /* target file of I/O */ + struct inode * inode; /* target file of i/o */ struct page ** pages; /* pages in our buffer */ unsigned int npages; /* count of pages */ atomic_t complete, /* i/os we're waiting for */ @@ -110,7 +111,6 @@ static inline int nfs_get_user_pages(int rw, unsigned long user_addr, size_t siz 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; @@ -294,7 +294,7 @@ static const struct rpc_call_ops nfs_read_direct_ops = { * For each nfs_read_data struct that was allocated on the list, dispatch * an NFS READ operation */ -static void