diff -druN Linux-2.4.16/Documentation/Configure.help linux-2.4.16-odirect/Documentation/Configure.help --- Linux-2.4.16/Documentation/Configure.help Tue Dec 4 13:29:35 2001 +++ linux-2.4.16-odirect/Documentation/Configure.help Fri Dec 14 11:43:25 2001 @@ -14173,6 +14173,30 @@ If unsure, say N. +Allow direct I/O on files in NFS +CONFIG_NFS_DIRECTIO + 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) implement their + own cache coherency protocol that subsumes the NFS 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. + + This option enables applications to perform direct I/O on files in NFS + file systems using the O_DIRECT open() flag. When O_DIRECT is set for + files, their data is not cached in the system's page cache. Direct + read and write operations are aligned to block boundaries. Data is + moved to and from user-level application buffers directly. + + Unless your program is designed to use O_DIRECT properly, you are much + better off allowing the NFS client to manage caching for you. Misusing + O_DIRECT can cause poor server performance or network storms. This + kernel build option defaults OFF to avoid exposing system administrators + unwittingly to a potentially hazardous feature. + + If unsure, say N. + Root file system on NFS CONFIG_ROOT_NFS If you want your Linux box to mount its whole root file system (the diff -druN Linux-2.4.16/fs/Config.in linux-2.4.16-odirect/fs/Config.in --- Linux-2.4.16/fs/Config.in Tue Dec 4 13:29:37 2001 +++ linux-2.4.16-odirect/fs/Config.in Fri Dec 14 11:43:25 2001 @@ -95,6 +95,7 @@ dep_tristate 'InterMezzo file system support (experimental, replicating fs)' CONFIG_INTERMEZZO_FS $CONFIG_INET $CONFIG_EXPERIMENTAL dep_tristate 'NFS file system support' CONFIG_NFS_FS $CONFIG_INET dep_mbool ' Provide NFSv3 client support' CONFIG_NFS_V3 $CONFIG_NFS_FS + dep_bool ' Allow direct I/O on NFS files' CONFIG_NFS_DIRECTIO $CONFIG_NFS_FS dep_bool ' Root file system on NFS' CONFIG_ROOT_NFS $CONFIG_NFS_FS $CONFIG_IP_PNP dep_tristate 'NFS server support' CONFIG_NFSD $CONFIG_INET diff -druN Linux-2.4.16/fs/nfs/Makefile linux-2.4.16-odirect/fs/nfs/Makefile --- Linux-2.4.16/fs/nfs/Makefile Tue Dec 4 13:29:37 2001 +++ linux-2.4.16-odirect/fs/nfs/Makefile Fri Dec 14 11:43:25 2001 @@ -14,6 +14,7 @@ obj-$(CONFIG_ROOT_NFS) += nfsroot.o mount_clnt.o obj-$(CONFIG_NFS_V3) += nfs3proc.o nfs3xdr.o +obj-$(CONFIG_NFS_DIRECTIO) += direct.o obj-m := $(O_TARGET) diff -druN Linux-2.4.16/fs/nfs/direct.c linux-2.4.16-odirect/fs/nfs/direct.c --- Linux-2.4.16/fs/nfs/direct.c Wed Dec 31 19:00:00 1969 +++ linux-2.4.16-odirect/fs/nfs/direct.c Fri Dec 14 11:50:22 2001 @@ -0,0 +1,378 @@ +/* + * linux/fs/nfs/direct.c + * + * High-performance direct I/O for the NFS client + * + * When an application requests uncached I/O, all read and write requests + * are made directly to the server; data stored or fetched via these + * requests is not cached in the Linux page cache. The client does not + * correct unaligned requests from applications. All requested bytes are + * held on permanent storage before a direct write system call returns to + * an application. Applications that manage their own data caching, such + * as databases, make very good use of direct I/O on local file systems. + * + * Solaris implements an uncached I/O facility called directio() that + * is used for backups and sequential I/O to very large files. Solaris + * also supports uncaching whole NFS partitions with "-o forcedirectio," + * an undocumented mount option. + * + * Note that I/O to read in executables (e.g. kernel_read) cannot use + * direct (kiobuf) reads because there is no vma backing the passed-in + * data buffer. + * + * Designed by Jeff Kimmel, Chuck Lever, and Trond Myklebust. + * + * Initial implementation: 12/2001 by Chuck Lever + * + * TODO: + * + * 1. Use concurrent asynchronous network requests rather than + * serialized synchronous network requests for normal (non-sync) + * direct I/O. + */ + +#include +#include +#include +#include +#include +#include +#include +#include + +#include +#include + +#define NFSDBG_FACILITY (NFSDBG_PAGECACHE | NFSDBG_VFS) +#define VERF_SIZE (2 * sizeof(__u32)) + +static /* inline */ int +nfs_direct_read_rpc(struct file *file, struct nfs_readargs *arg) +{ + int result; + struct inode * inode = file->f_dentry->d_inode; + struct nfs_fattr fattr; + struct rpc_message msg; + struct nfs_readres res = { &fattr, arg->count, 0 }; + +#ifdef CONFIG_NFS_V3 + msg.rpc_proc = (NFS_PROTO(inode)->version == 3) ? + NFS3PROC_READ : NFSPROC_READ; +#else + msg.rpc_proc = NFSPROC_READ; +#endif + msg.rpc_argp = arg; + msg.rpc_resp = &res; + + lock_kernel(); + msg.rpc_cred = nfs_file_cred(file); + fattr.valid = 0; + result = rpc_call_sync(NFS_CLIENT(inode), &msg, 0); + nfs_refresh_inode(inode, &fattr); + unlock_kernel(); + + return result; +} + +static /* inline */ int +nfs_direct_write_rpc(struct file *file, struct nfs_writeargs *arg, + struct nfs_writeverf *verf) +{ + int result; + struct inode *inode = file->f_dentry->d_inode; + struct nfs_fattr fattr; + struct rpc_message msg; + struct nfs_writeres res = { &fattr, verf, 0 }; + +#ifdef CONFIG_NFS_V3 + msg.rpc_proc = (NFS_PROTO(inode)->version == 3) ? + NFS3PROC_WRITE : NFSPROC_WRITE; +#else + msg.rpc_proc = NFSPROC_WRITE; +#endif + msg.rpc_argp = arg; + msg.rpc_resp = &res; + + lock_kernel(); + msg.rpc_cred = get_rpccred(nfs_file_cred(file)); + fattr.valid = 0; + result = rpc_call_sync(NFS_CLIENT(inode), &msg, 0); + nfs_write_attributes(inode, &fattr); + put_rpccred(msg.rpc_cred); + unlock_kernel(); + + if (result > 0) { + if ((arg->stable == NFS_FILE_SYNC) && + (verf->committed != NFS_FILE_SYNC)) { + printk(KERN_ERR __FUNCTION__ + ": server didn't sync stable write request\n"); + return -EIO; + } + + if (result != arg->count) + printk(KERN_INFO __FUNCTION__ + ": short write, count=%u, result=%d\n", + arg->count, result); + } + + return result; +} + +#ifdef CONFIG_NFS_V3 +static /* inline */ int +nfs_direct_commit_rpc(struct inode *inode, loff_t offset, size_t count, + struct nfs_writeverf *verf) +{ + int result; + struct nfs_fattr fattr; + struct nfs_writeargs arg = { NFS_FH(inode), offset, count, 0, 0, + {{0, 0}, {0,0}, {0,0}, {0,0}, + {0,0}, {0,0}, {0,0}, {0,0}} }; + struct nfs_writeres res = { &fattr, verf, 0 }; + struct rpc_message msg = { NFS3PROC_COMMIT, &arg, &res, NULL }; + + fattr.valid = 0; + + lock_kernel(); + result = rpc_call_sync(NFS_CLIENT(inode), &msg, 0); + nfs_write_attributes(inode, &fattr); + unlock_kernel(); + + return result; +} +#else +static inline int +nfs_direct_commit_rpc(struct inode *inode, loff_t offset, size_t count, + struct nfs_writeverf *verf) +{ + return 0; +} +#endif + +/* + * Walk through the iobuf and create an iovec for each "rsize" bytes. + */ +static int +nfs_direct_read(struct file *file, struct kiobuf *iobuf, loff_t offset, + size_t count) +{ + int curpage, total; + struct inode *inode = file->f_dentry->d_inode; + int rsize = NFS_SERVER(inode)->rsize; + struct nfs_readargs args = { NFS_FH(inode), 0, 0, 0 }; + + total = 0; + curpage = 0; + while (count) { + int starting_offset, request, result, first, last, i; + struct iovec *iovec = args.iov; + + request = count; + if (count > rsize) + request = rsize; + args.count = request; + args.offset = offset; + + starting_offset = iobuf->offset; + first = last = curpage; + while (curpage < iobuf->nr_pages) { + struct page *page = iobuf->maplist[curpage]; + + if (!page) + return -EFAULT; + + iovec->iov_base = kmap(page) + starting_offset; + iovec->iov_len = (PAGE_SIZE - starting_offset); + if ((starting_offset + request) < PAGE_SIZE) + iovec->iov_len = request; + + request -= iovec->iov_len; + starting_offset = 0; /* zero after the first page */ + last = curpage; + curpage++; + iovec++; + args.nriov++; + } + + result = nfs_direct_read_rpc(file, &args); + + for (i = first; i < last; i++) { + flush_dcache_page(iobuf->maplist[i]); + kunmap(iobuf->maplist[i]); + } + + if (result < 0) { + if (result == -EISDIR) + total = -EINVAL; + else + total = result; + break; + } + + total += result; + count -= result; + offset += result; + + if (result < args.count) /* NFSv2ism */ + break; + }; + return total; +} + +/* + * Walk through the iobuf and create an iovec for each "wsize" bytes. + * If only one network write is necessary, or if the O_SYNC flag or + * 'sync' mount option are present, or if this is a V2 inode, use + * FILE_SYNC. Otherwise, use UNSTABLE and finish with a COMMIT. + * + * The mechanics of this function are much the same as nfs_direct_read, + * with the added complexity of committing unstable writes. + */ +static int +nfs_direct_write(struct file *file, struct kiobuf *iobuf, + loff_t offset, size_t count) +{ + int curpage, total; + int need_commit = 0; + loff_t save_offset = offset; + struct inode *inode = file->f_dentry->d_inode; + int wsize = NFS_SERVER(inode)->wsize; + struct nfs_writeverf first_verf, ret_verf; + struct nfs_writeargs args = { NFS_FH(inode), 0, 0, NFS_FILE_SYNC, 0 }; + +#ifdef CONFIG_NFS_V3 + if ((NFS_PROTO(inode)->version == 3) && (count > wsize) && + (!IS_SYNC(inode))) + args.stable = NFS_UNSTABLE; +#endif + +retry: + total = 0; + curpage = 0; + while (count) { + int starting_offset, request, result, first, last, i; + struct iovec *iovec = args.iov; + + request = count; + if (count > wsize) + request = wsize; + args.count = request; + args.offset = offset; + + starting_offset = iobuf->offset; + first = last = curpage; + while (curpage < iobuf->nr_pages) { + struct page *page = iobuf->maplist[curpage]; + + if (!page) + return -EFAULT; + + iovec->iov_base = kmap(page) + starting_offset; + iovec->iov_len = (PAGE_SIZE - starting_offset); + if ((starting_offset + request) < PAGE_SIZE) + iovec->iov_len = request; + + request -= iovec->iov_len; + starting_offset = 0; /* zero after the first page */ + last = curpage; + curpage++; + iovec++; + args.nriov++; + } + + result = nfs_direct_write_rpc(file, &args, &ret_verf); + + for (i = first; i < last; i++) + kunmap(iobuf->maplist[i]); + + if (result < 0) { + total = result; + break; + } + + if (!total) + memcpy(&first_verf.verifier, &ret_verf.verifier, + VERF_SIZE); + if (ret_verf.committed != NFS_FILE_SYNC) { + need_commit = 1; + if (memcmp(&first_verf.verifier, &ret_verf.verifier, + VERF_SIZE)) + goto print_retry; + } + + total += result; + count -= result; + offset += result; + }; + + /* + * Commit data written so far, even in the event of an error + */ + if (need_commit) { + if (nfs_direct_commit_rpc(inode, save_offset, + iobuf->length - count, &ret_verf)) + goto print_retry; + if (memcmp(&first_verf.verifier, &ret_verf.verifier, + VERF_SIZE)) + goto print_retry; + } + + return total; + +print_retry: + printk(KERN_INFO __FUNCTION__ + ": detected server restart; retrying with FILE_SYNC\n"); + args.stable = NFS_FILE_SYNC; + offset = save_offset; + count = iobuf->length; + goto retry; +} + +/* + * Read or write data, moving the data directly to/from the + * application's buffer without caching in the page cache. + * + * Rules for direct I/O + * + * 1. block size = 512 bytes or more + * 2. file byte offset is block aligned + * 3. byte count is a multiple of block size + * 4. user buffer is not aligned + * 5. user buffer is faulted in and pinned + * + * These are verified before we get here. + */ +int +nfs_direct_IO(int rw, struct file *file, struct kiobuf *iobuf, + unsigned long blocknr, int blocksize) +{ + int result = -EINVAL; + size_t count = iobuf->length; + struct dentry *dentry = file->f_dentry; + struct inode *inode = dentry->d_inode; + loff_t offset = blocknr << inode->i_blkbits; + + switch (rw) { + case READ: + dfprintk(VFS, + "NFS: direct_IO(READ) (%s/%s) off/cnt(%Lu/%d)\n", + dentry->d_parent->d_name.name, + dentry->d_name.name, offset, count); + + result = nfs_direct_read(file, iobuf, offset, count); + break; + case WRITE: + dfprintk(VFS, + "NFS: direct_IO(WRITE) (%s/%s) off/cnt(%Lu/%d)\n", + dentry->d_parent->d_name.name, + dentry->d_name.name, offset, count); + + result = nfs_direct_write(file, iobuf, offset, count); + break; + default: + break; + } + + dfprintk(VFS, "NFS: direct_IO result = %d\n", result); + return result; +} diff -druN Linux-2.4.16/fs/nfs/file.c linux-2.4.16-odirect/fs/nfs/file.c --- Linux-2.4.16/fs/nfs/file.c Tue Dec 4 13:29:24 2001 +++ linux-2.4.16-odirect/fs/nfs/file.c Fri Dec 14 11:43:25 2001 @@ -204,7 +204,10 @@ sync_page: nfs_sync_page, writepage: nfs_writepage, prepare_write: nfs_prepare_write, - commit_write: nfs_commit_write + commit_write: nfs_commit_write, +#ifdef CONFIG_NFS_DIRECTIO + direct_IO: nfs_direct_IO, +#endif }; /* diff -druN Linux-2.4.16/fs/nfs/write.c linux-2.4.16-odirect/fs/nfs/write.c --- Linux-2.4.16/fs/nfs/write.c Tue Dec 4 13:29:37 2001 +++ linux-2.4.16-odirect/fs/nfs/write.c Fri Dec 14 11:43:25 2001 @@ -121,23 +121,6 @@ } /* - * This function will be used to simulate weak cache consistency - * under NFSv2 when the NFSv3 attribute patch is included. - * For the moment, we just call nfs_refresh_inode(). - */ -static __inline__ int -nfs_write_attributes(struct inode *inode, struct nfs_fattr *fattr) -{ - if ((fattr->valid & NFS_ATTR_FATTR) && !(fattr->valid & NFS_ATTR_WCC)) { - fattr->pre_size = NFS_CACHE_ISIZE(inode); - fattr->pre_mtime = NFS_CACHE_MTIME(inode); - fattr->pre_ctime = NFS_CACHE_CTIME(inode); - fattr->valid |= NFS_ATTR_WCC; - } - return nfs_refresh_inode(inode, fattr); -} - -/* * Write a page synchronously. * Offset is the data offset within the page. */ diff -druN Linux-2.4.16/include/linux/nfs_fs.h linux-2.4.16-odirect/include/linux/nfs_fs.h --- Linux-2.4.16/include/linux/nfs_fs.h Tue Dec 4 13:29:37 2001 +++ linux-2.4.16-odirect/include/linux/nfs_fs.h Fri Dec 14 11:47:15 2001 @@ -266,6 +266,11 @@ extern int nfs_scan_lru_read_timeout(struct nfs_server *, struct list_head *); /* + * linux/fs/nfs/direct.c + */ +extern int nfs_direct_IO(int, struct file *, struct kiobuf *, unsigned long, int); + +/* * linux/fs/mount_clnt.c * (Used only by nfsroot module) */ @@ -289,6 +294,23 @@ if ((fattr->valid & NFS_ATTR_FATTR) == 0) return 0; return __nfs_refresh_inode(inode,fattr); +} + +/* + * This function will be used to simulate weak cache consistency + * under NFSv2 when the NFSv3 attribute patch is included. + * For the moment, we just call nfs_refresh_inode(). + */ +static __inline__ int +nfs_write_attributes(struct inode *inode, struct nfs_fattr *fattr) +{ + if ((fattr->valid & NFS_ATTR_FATTR) && !(fattr->valid & NFS_ATTR_WCC)) { + fattr->pre_size = NFS_CACHE_ISIZE(inode); + fattr->pre_mtime = NFS_CACHE_MTIME(inode); + fattr->pre_ctime = NFS_CACHE_CTIME(inode); + fattr->valid |= NFS_ATTR_WCC; + } + return nfs_refresh_inode(inode, fattr); } static inline loff_t