| 1 | /* |
| 2 | * linux/fs/nfs/direct.c |
| 3 | * |
| 4 | * Copyright (C) 2003 by Chuck Lever <cel@netapp.com> |
| 5 | * |
| 6 | * High-performance uncached I/O for the Linux NFS client |
| 7 | * |
| 8 | * There are important applications whose performance or correctness |
| 9 | * depends on uncached access to file data. Database clusters |
| 10 | * (multiple copies of the same instance running on separate hosts) |
| 11 | * implement their own cache coherency protocol that subsumes file |
| 12 | * system cache protocols. Applications that process datasets |
| 13 | * considerably larger than the client's memory do not always benefit |
| 14 | * from a local cache. A streaming video server, for instance, has no |
| 15 | * need to cache the contents of a file. |
| 16 | * |
| 17 | * When an application requests uncached I/O, all read and write requests |
| 18 | * are made directly to the server; data stored or fetched via these |
| 19 | * requests is not cached in the Linux page cache. The client does not |
| 20 | * correct unaligned requests from applications. All requested bytes are |
| 21 | * held on permanent storage before a direct write system call returns to |
| 22 | * an application. |
| 23 | * |
| 24 | * Solaris implements an uncached I/O facility called directio() that |
| 25 | * is used for backups and sequential I/O to very large files. Solaris |
| 26 | * also supports uncaching whole NFS partitions with "-o forcedirectio," |
| 27 | * an undocumented mount option. |
| 28 | * |
| 29 | * Designed by Jeff Kimmel, Chuck Lever, and Trond Myklebust, with |
| 30 | * help from Andrew Morton. |
| 31 | * |
| 32 | * 18 Dec 2001 Initial implementation for 2.4 --cel |
| 33 | * 08 Jul 2002 Version for 2.4.19, with bug fixes --trondmy |
| 34 | * 08 Jun 2003 Port to 2.5 APIs --cel |
| 35 | * 31 Mar 2004 Handle direct I/O without VFS support --cel |
| 36 | * 15 Sep 2004 Parallel async reads --cel |
| 37 | * 04 May 2005 support O_DIRECT with aio --cel |
| 38 | * |
| 39 | */ |
| 40 | |
| 41 | #include <linux/errno.h> |
| 42 | #include <linux/sched.h> |
| 43 | #include <linux/kernel.h> |
| 44 | #include <linux/smp_lock.h> |
| 45 | #include <linux/file.h> |
| 46 | #include <linux/pagemap.h> |
| 47 | #include <linux/kref.h> |
| 48 | |
| 49 | #include <linux/nfs_fs.h> |
| 50 | #include <linux/nfs_page.h> |
| 51 | #include <linux/sunrpc/clnt.h> |
| 52 | |
| 53 | #include <asm/system.h> |
| 54 | #include <asm/uaccess.h> |
| 55 | #include <asm/atomic.h> |
| 56 | |
| 57 | #include "iostat.h" |
| 58 | |
| 59 | #define NFSDBG_FACILITY NFSDBG_VFS |
| 60 | |
| 61 | static struct kmem_cache *nfs_direct_cachep; |
| 62 | |
| 63 | /* |
| 64 | * This represents a set of asynchronous requests that we're waiting on |
| 65 | */ |
| 66 | struct nfs_direct_req { |
| 67 | struct kref kref; /* release manager */ |
| 68 | |
| 69 | /* I/O parameters */ |
| 70 | struct nfs_open_context *ctx; /* file open context info */ |
| 71 | struct kiocb * iocb; /* controlling i/o request */ |
| 72 | struct inode * inode; /* target file of i/o */ |
| 73 | |
| 74 | /* completion state */ |
| 75 | atomic_t io_count; /* i/os we're waiting for */ |
| 76 | spinlock_t lock; /* protect completion state */ |
| 77 | ssize_t count, /* bytes actually processed */ |
| 78 | error; /* any reported error */ |
| 79 | struct completion completion; /* wait for i/o completion */ |
| 80 | |
| 81 | /* commit state */ |
| 82 | struct list_head rewrite_list; /* saved nfs_write_data structs */ |
| 83 | struct nfs_write_data * commit_data; /* special write_data for commits */ |
| 84 | int flags; |
| 85 | #define NFS_ODIRECT_DO_COMMIT (1) /* an unstable reply was received */ |
| 86 | #define NFS_ODIRECT_RESCHED_WRITES (2) /* write verification failed */ |
| 87 | struct nfs_writeverf verf; /* unstable write verifier */ |
| 88 | }; |
| 89 | |
| 90 | static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode); |
| 91 | static const struct rpc_call_ops nfs_write_direct_ops; |
| 92 | |
| 93 | static inline void get_dreq(struct nfs_direct_req *dreq) |
| 94 | { |
| 95 | atomic_inc(&dreq->io_count); |
| 96 | } |
| 97 | |
| 98 | static inline int put_dreq(struct nfs_direct_req *dreq) |
| 99 | { |
| 100 | return atomic_dec_and_test(&dreq->io_count); |
| 101 | } |
| 102 | |
| 103 | /** |
| 104 | * nfs_direct_IO - NFS address space operation for direct I/O |
| 105 | * @rw: direction (read or write) |
| 106 | * @iocb: target I/O control block |
| 107 | * @iov: array of vectors that define I/O buffer |
| 108 | * @pos: offset in file to begin the operation |
| 109 | * @nr_segs: size of iovec array |
| 110 | * |
| 111 | * The presence of this routine in the address space ops vector means |
| 112 | * the NFS client supports direct I/O. However, we shunt off direct |
| 113 | * read and write requests before the VFS gets them, so this method |
| 114 | * should never be called. |
| 115 | */ |
| 116 | ssize_t nfs_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov, loff_t pos, unsigned long nr_segs) |
| 117 | { |
| 118 | dprintk("NFS: nfs_direct_IO (%s) off/no(%Ld/%lu) EINVAL\n", |
| 119 | iocb->ki_filp->f_path.dentry->d_name.name, |
| 120 | (long long) pos, nr_segs); |
| 121 | |
| 122 | return -EINVAL; |
| 123 | } |
| 124 | |
| 125 | static void nfs_direct_dirty_pages(struct page **pages, int npages) |
| 126 | { |
| 127 | int i; |
| 128 | for (i = 0; i < npages; i++) { |
| 129 | struct page *page = pages[i]; |
| 130 | if (!PageCompound(page)) |
| 131 | set_page_dirty_lock(page); |
| 132 | } |
| 133 | } |
| 134 | |
| 135 | static void nfs_direct_release_pages(struct page **pages, int npages) |
| 136 | { |
| 137 | int i; |
| 138 | for (i = 0; i < npages; i++) |
| 139 | page_cache_release(pages[i]); |
| 140 | } |
| 141 | |
| 142 | static inline struct nfs_direct_req *nfs_direct_req_alloc(void) |
| 143 | { |
| 144 | struct nfs_direct_req *dreq; |
| 145 | |
| 146 | dreq = kmem_cache_alloc(nfs_direct_cachep, GFP_KERNEL); |
| 147 | if (!dreq) |
| 148 | return NULL; |
| 149 | |
| 150 | kref_init(&dreq->kref); |
| 151 | kref_get(&dreq->kref); |
| 152 | init_completion(&dreq->completion); |
| 153 | INIT_LIST_HEAD(&dreq->rewrite_list); |
| 154 | dreq->iocb = NULL; |
| 155 | dreq->ctx = NULL; |
| 156 | spin_lock_init(&dreq->lock); |
| 157 | atomic_set(&dreq->io_count, 0); |
| 158 | dreq->count = 0; |
| 159 | dreq->error = 0; |
| 160 | dreq->flags = 0; |
| 161 | |
| 162 | return dreq; |
| 163 | } |
| 164 | |
| 165 | static void nfs_direct_req_release(struct kref *kref) |
| 166 | { |
| 167 | struct nfs_direct_req *dreq = container_of(kref, struct nfs_direct_req, kref); |
| 168 | |
| 169 | if (dreq->ctx != NULL) |
| 170 | put_nfs_open_context(dreq->ctx); |
| 171 | kmem_cache_free(nfs_direct_cachep, dreq); |
| 172 | } |
| 173 | |
| 174 | /* |
| 175 | * Collects and returns the final error value/byte-count. |
| 176 | */ |
| 177 | static ssize_t nfs_direct_wait(struct nfs_direct_req *dreq) |
| 178 | { |
| 179 | ssize_t result = -EIOCBQUEUED; |
| 180 | |
| 181 | /* Async requests don't wait here */ |
| 182 | if (dreq->iocb) |
| 183 | goto out; |
| 184 | |
| 185 | result = wait_for_completion_interruptible(&dreq->completion); |
| 186 | |
| 187 | if (!result) |
| 188 | result = dreq->error; |
| 189 | if (!result) |
| 190 | result = dreq->count; |
| 191 | |
| 192 | out: |
| 193 | kref_put(&dreq->kref, nfs_direct_req_release); |
| 194 | return (ssize_t) result; |
| 195 | } |
| 196 | |
| 197 | /* |
| 198 | * Synchronous I/O uses a stack-allocated iocb. Thus we can't trust |
| 199 | * the iocb is still valid here if this is a synchronous request. |
| 200 | */ |
| 201 | static void nfs_direct_complete(struct nfs_direct_req *dreq) |
| 202 | { |
| 203 | if (dreq->iocb) { |
| 204 | long res = (long) dreq->error; |
| 205 | if (!res) |
| 206 | res = (long) dreq->count; |
| 207 | aio_complete(dreq->iocb, res, 0); |
| 208 | } |
| 209 | complete_all(&dreq->completion); |
| 210 | |
| 211 | kref_put(&dreq->kref, nfs_direct_req_release); |
| 212 | } |
| 213 | |
| 214 | /* |
| 215 | * We must hold a reference to all the pages in this direct read request |
| 216 | * until the RPCs complete. This could be long *after* we are woken up in |
| 217 | * nfs_direct_wait (for instance, if someone hits ^C on a slow server). |
| 218 | */ |
| 219 | static void nfs_direct_read_result(struct rpc_task *task, void *calldata) |
| 220 | { |
| 221 | struct nfs_read_data *data = calldata; |
| 222 | struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req; |
| 223 | |
| 224 | if (nfs_readpage_result(task, data) != 0) |
| 225 | return; |
| 226 | |
| 227 | nfs_direct_dirty_pages(data->pagevec, data->npages); |
| 228 | nfs_direct_release_pages(data->pagevec, data->npages); |
| 229 | |
| 230 | spin_lock(&dreq->lock); |
| 231 | |
| 232 | if (likely(task->tk_status >= 0)) |
| 233 | dreq->count += data->res.count; |
| 234 | else |
| 235 | dreq->error = task->tk_status; |
| 236 | |
| 237 | spin_unlock(&dreq->lock); |
| 238 | |
| 239 | if (put_dreq(dreq)) |
| 240 | nfs_direct_complete(dreq); |
| 241 | } |
| 242 | |
| 243 | static const struct rpc_call_ops nfs_read_direct_ops = { |
| 244 | .rpc_call_done = nfs_direct_read_result, |
| 245 | .rpc_release = nfs_readdata_release, |
| 246 | }; |
| 247 | |
| 248 | /* |
| 249 | * For each rsize'd chunk of the user's buffer, dispatch an NFS READ |
| 250 | * operation. If nfs_readdata_alloc() or get_user_pages() fails, |
| 251 | * bail and stop sending more reads. Read length accounting is |
| 252 | * handled automatically by nfs_direct_read_result(). Otherwise, if |
| 253 | * no requests have been sent, just return an error. |
| 254 | */ |
| 255 | static ssize_t nfs_direct_read_schedule(struct nfs_direct_req *dreq, unsigned long user_addr, size_t count, loff_t pos) |
| 256 | { |
| 257 | struct nfs_open_context *ctx = dreq->ctx; |
| 258 | struct inode *inode = ctx->dentry->d_inode; |
| 259 | size_t rsize = NFS_SERVER(inode)->rsize; |
| 260 | unsigned int pgbase; |
| 261 | int result; |
| 262 | ssize_t started = 0; |
| 263 | |
| 264 | get_dreq(dreq); |
| 265 | |
| 266 | do { |
| 267 | struct nfs_read_data *data; |
| 268 | size_t bytes; |
| 269 | |
| 270 | pgbase = user_addr & ~PAGE_MASK; |
| 271 | bytes = min(rsize,count); |
| 272 | |
| 273 | result = -ENOMEM; |
| 274 | data = nfs_readdata_alloc(pgbase + bytes); |
| 275 | if (unlikely(!data)) |
| 276 | break; |
| 277 | |
| 278 | down_read(¤t->mm->mmap_sem); |
| 279 | result = get_user_pages(current, current->mm, user_addr, |
| 280 | data->npages, 1, 0, data->pagevec, NULL); |
| 281 | up_read(¤t->mm->mmap_sem); |
| 282 | if (unlikely(result < data->npages)) { |
| 283 | if (result > 0) |
| 284 | nfs_direct_release_pages(data->pagevec, result); |
| 285 | nfs_readdata_release(data); |
| 286 | break; |
| 287 | } |
| 288 | |
| 289 | get_dreq(dreq); |
| 290 | |
| 291 | data->req = (struct nfs_page *) dreq; |
| 292 | data->inode = inode; |
| 293 | data->cred = ctx->cred; |
| 294 | data->args.fh = NFS_FH(inode); |
| 295 | data->args.context = ctx; |
| 296 | data->args.offset = pos; |
| 297 | data->args.pgbase = pgbase; |
| 298 | data->args.pages = data->pagevec; |
| 299 | data->args.count = bytes; |
| 300 | data->res.fattr = &data->fattr; |
| 301 | data->res.eof = 0; |
| 302 | data->res.count = bytes; |
| 303 | |
| 304 | rpc_init_task(&data->task, NFS_CLIENT(inode), RPC_TASK_ASYNC, |
| 305 | &nfs_read_direct_ops, data); |
| 306 | NFS_PROTO(inode)->read_setup(data); |
| 307 | |
| 308 | data->task.tk_cookie = (unsigned long) inode; |
| 309 | |
| 310 | rpc_execute(&data->task); |
| 311 | |
| 312 | dfprintk(VFS, "NFS: %5u initiated direct read call (req %s/%Ld, %zu bytes @ offset %Lu)\n", |
| 313 | data->task.tk_pid, |
| 314 | inode->i_sb->s_id, |
| 315 | (long long)NFS_FILEID(inode), |
| 316 | bytes, |
| 317 | (unsigned long long)data->args.offset); |
| 318 | |
| 319 | started += bytes; |
| 320 | user_addr += bytes; |
| 321 | pos += bytes; |
| 322 | /* FIXME: Remove this unnecessary math from final patch */ |
| 323 | pgbase += bytes; |
| 324 | pgbase &= ~PAGE_MASK; |
| 325 | BUG_ON(pgbase != (user_addr & ~PAGE_MASK)); |
| 326 | |
| 327 | count -= bytes; |
| 328 | } while (count != 0); |
| 329 | |
| 330 | if (put_dreq(dreq)) |
| 331 | nfs_direct_complete(dreq); |
| 332 | |
| 333 | if (started) |
| 334 | return 0; |
| 335 | return result < 0 ? (ssize_t) result : -EFAULT; |
| 336 | } |
| 337 | |
| 338 | static ssize_t nfs_direct_read(struct kiocb *iocb, unsigned long user_addr, size_t count, loff_t pos) |
| 339 | { |
| 340 | ssize_t result = 0; |
| 341 | sigset_t oldset; |
| 342 | struct inode *inode = iocb->ki_filp->f_mapping->host; |
| 343 | struct rpc_clnt *clnt = NFS_CLIENT(inode); |
| 344 | struct nfs_direct_req *dreq; |
| 345 | |
| 346 | dreq = nfs_direct_req_alloc(); |
| 347 | if (!dreq) |
| 348 | return -ENOMEM; |
| 349 | |
| 350 | dreq->inode = inode; |
| 351 | dreq->ctx = get_nfs_open_context((struct nfs_open_context *)iocb->ki_filp->private_data); |
| 352 | if (!is_sync_kiocb(iocb)) |
| 353 | dreq->iocb = iocb; |
| 354 | |
| 355 | nfs_add_stats(inode, NFSIOS_DIRECTREADBYTES, count); |
| 356 | rpc_clnt_sigmask(clnt, &oldset); |
| 357 | result = nfs_direct_read_schedule(dreq, user_addr, count, pos); |
| 358 | if (!result) |
| 359 | result = nfs_direct_wait(dreq); |
| 360 | rpc_clnt_sigunmask(clnt, &oldset); |
| 361 | |
| 362 | return result; |
| 363 | } |
| 364 | |
| 365 | static void nfs_direct_free_writedata(struct nfs_direct_req *dreq) |
| 366 | { |
| 367 | while (!list_empty(&dreq->rewrite_list)) { |
| 368 | struct nfs_write_data *data = list_entry(dreq->rewrite_list.next, struct nfs_write_data, pages); |
| 369 | list_del(&data->pages); |
| 370 | nfs_direct_release_pages(data->pagevec, data->npages); |
| 371 | nfs_writedata_release(data); |
| 372 | } |
| 373 | } |
| 374 | |
| 375 | #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4) |
| 376 | static void nfs_direct_write_reschedule(struct nfs_direct_req *dreq) |
| 377 | { |
| 378 | struct inode *inode = dreq->inode; |
| 379 | struct list_head *p; |
| 380 | struct nfs_write_data *data; |
| 381 | |
| 382 | dreq->count = 0; |
| 383 | get_dreq(dreq); |
| 384 | |
| 385 | list_for_each(p, &dreq->rewrite_list) { |
| 386 | data = list_entry(p, struct nfs_write_data, pages); |
| 387 | |
| 388 | get_dreq(dreq); |
| 389 | |
| 390 | /* |
| 391 | * Reset data->res. |
| 392 | */ |
| 393 | nfs_fattr_init(&data->fattr); |
| 394 | data->res.count = data->args.count; |
| 395 | memset(&data->verf, 0, sizeof(data->verf)); |
| 396 | |
| 397 | /* |
| 398 | * Reuse data->task; data->args should not have changed |
| 399 | * since the original request was sent. |
| 400 | */ |
| 401 | rpc_init_task(&data->task, NFS_CLIENT(inode), RPC_TASK_ASYNC, |
| 402 | &nfs_write_direct_ops, data); |
| 403 | NFS_PROTO(inode)->write_setup(data, FLUSH_STABLE); |
| 404 | |
| 405 | data->task.tk_priority = RPC_PRIORITY_NORMAL; |
| 406 | data->task.tk_cookie = (unsigned long) inode; |
| 407 | |
| 408 | /* |
| 409 | * We're called via an RPC callback, so BKL is already held. |
| 410 | */ |
| 411 | rpc_execute(&data->task); |
| 412 | |
| 413 | dprintk("NFS: %5u rescheduled direct write call (req %s/%Ld, %u bytes @ offset %Lu)\n", |
| 414 | data->task.tk_pid, |
| 415 | inode->i_sb->s_id, |
| 416 | (long long)NFS_FILEID(inode), |
| 417 | data->args.count, |
| 418 | (unsigned long long)data->args.offset); |
| 419 | } |
| 420 | |
| 421 | if (put_dreq(dreq)) |
| 422 | nfs_direct_write_complete(dreq, inode); |
| 423 | } |
| 424 | |
| 425 | static void nfs_direct_commit_result(struct rpc_task *task, void *calldata) |
| 426 | { |
| 427 | struct nfs_write_data *data = calldata; |
| 428 | struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req; |
| 429 | |
| 430 | /* Call the NFS version-specific code */ |
| 431 | if (NFS_PROTO(data->inode)->commit_done(task, data) != 0) |
| 432 | return; |
| 433 | if (unlikely(task->tk_status < 0)) { |
| 434 | dreq->error = task->tk_status; |
| 435 | dreq->flags = NFS_ODIRECT_RESCHED_WRITES; |
| 436 | } |
| 437 | if (memcmp(&dreq->verf, &data->verf, sizeof(data->verf))) { |
| 438 | dprintk("NFS: %5u commit verify failed\n", task->tk_pid); |
| 439 | dreq->flags = NFS_ODIRECT_RESCHED_WRITES; |
| 440 | } |
| 441 | |
| 442 | dprintk("NFS: %5u commit returned %d\n", task->tk_pid, task->tk_status); |
| 443 | nfs_direct_write_complete(dreq, data->inode); |
| 444 | } |
| 445 | |
| 446 | static const struct rpc_call_ops nfs_commit_direct_ops = { |
| 447 | .rpc_call_done = nfs_direct_commit_result, |
| 448 | .rpc_release = nfs_commit_release, |
| 449 | }; |
| 450 | |
| 451 | static void nfs_direct_commit_schedule(struct nfs_direct_req *dreq) |
| 452 | { |
| 453 | struct nfs_write_data *data = dreq->commit_data; |
| 454 | |
| 455 | data->inode = dreq->inode; |
| 456 | data->cred = dreq->ctx->cred; |
| 457 | |
| 458 | data->args.fh = NFS_FH(data->inode); |
| 459 | data->args.offset = 0; |
| 460 | data->args.count = 0; |
| 461 | data->res.count = 0; |
| 462 | data->res.fattr = &data->fattr; |
| 463 | data->res.verf = &data->verf; |
| 464 | |
| 465 | rpc_init_task(&data->task, NFS_CLIENT(dreq->inode), RPC_TASK_ASYNC, |
| 466 | &nfs_commit_direct_ops, data); |
| 467 | NFS_PROTO(data->inode)->commit_setup(data, 0); |
| 468 | |
| 469 | data->task.tk_priority = RPC_PRIORITY_NORMAL; |
| 470 | data->task.tk_cookie = (unsigned long)data->inode; |
| 471 | /* Note: task.tk_ops->rpc_release will free dreq->commit_data */ |
| 472 | dreq->commit_data = NULL; |
| 473 | |
| 474 | dprintk("NFS: %5u initiated commit call\n", data->task.tk_pid); |
| 475 | |
| 476 | rpc_execute(&data->task); |
| 477 | } |
| 478 | |
| 479 | static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode) |
| 480 | { |
| 481 | int flags = dreq->flags; |
| 482 | |
| 483 | dreq->flags = 0; |
| 484 | switch (flags) { |
| 485 | case NFS_ODIRECT_DO_COMMIT: |
| 486 | nfs_direct_commit_schedule(dreq); |
| 487 | break; |
| 488 | case NFS_ODIRECT_RESCHED_WRITES: |
| 489 | nfs_direct_write_reschedule(dreq); |
| 490 | break; |
| 491 | default: |
| 492 | nfs_end_data_update(inode); |
| 493 | if (dreq->commit_data != NULL) |
| 494 | nfs_commit_free(dreq->commit_data); |
| 495 | nfs_direct_free_writedata(dreq); |
| 496 | nfs_zap_mapping(inode, inode->i_mapping); |
| 497 | nfs_direct_complete(dreq); |
| 498 | } |
| 499 | } |
| 500 | |
| 501 | static void nfs_alloc_commit_data(struct nfs_direct_req *dreq) |
| 502 | { |
| 503 | dreq->commit_data = nfs_commit_alloc(); |
| 504 | if (dreq->commit_data != NULL) |
| 505 | dreq->commit_data->req = (struct nfs_page *) dreq; |
| 506 | } |
| 507 | #else |
| 508 | static inline void nfs_alloc_commit_data(struct nfs_direct_req *dreq) |
| 509 | { |
| 510 | dreq->commit_data = NULL; |
| 511 | } |
| 512 | |
| 513 | static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode) |
| 514 | { |
| 515 | nfs_end_data_update(inode); |
| 516 | nfs_direct_free_writedata(dreq); |
| 517 | nfs_zap_mapping(inode, inode->i_mapping); |
| 518 | nfs_direct_complete(dreq); |
| 519 | } |
| 520 | #endif |
| 521 | |
| 522 | static void nfs_direct_write_result(struct rpc_task *task, void *calldata) |
| 523 | { |
| 524 | struct nfs_write_data *data = calldata; |
| 525 | struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req; |
| 526 | int status = task->tk_status; |
| 527 | |
| 528 | if (nfs_writeback_done(task, data) != 0) |
| 529 | return; |
| 530 | |
| 531 | spin_lock(&dreq->lock); |
| 532 | |
| 533 | if (unlikely(status < 0)) { |
| 534 | dreq->error = status; |
| 535 | goto out_unlock; |
| 536 | } |
| 537 | |
| 538 | dreq->count += data->res.count; |
| 539 | |
| 540 | if (data->res.verf->committed != NFS_FILE_SYNC) { |
| 541 | switch (dreq->flags) { |
| 542 | case 0: |
| 543 | memcpy(&dreq->verf, &data->verf, sizeof(dreq->verf)); |
| 544 | dreq->flags = NFS_ODIRECT_DO_COMMIT; |
| 545 | break; |
| 546 | case NFS_ODIRECT_DO_COMMIT: |
| 547 | if (memcmp(&dreq->verf, &data->verf, sizeof(dreq->verf))) { |
| 548 | dprintk("NFS: %5u write verify failed\n", task->tk_pid); |
| 549 | dreq->flags = NFS_ODIRECT_RESCHED_WRITES; |
| 550 | } |
| 551 | } |
| 552 | } |
| 553 | out_unlock: |
| 554 | spin_unlock(&dreq->lock); |
| 555 | } |
| 556 | |
| 557 | /* |
| 558 | * NB: Return the value of the first error return code. Subsequent |
| 559 | * errors after the first one are ignored. |
| 560 | */ |
| 561 | static void nfs_direct_write_release(void *calldata) |
| 562 | { |
| 563 | struct nfs_write_data *data = calldata; |
| 564 | struct nfs_direct_req *dreq = (struct nfs_direct_req *) data->req; |
| 565 | |
| 566 | if (put_dreq(dreq)) |
| 567 | nfs_direct_write_complete(dreq, data->inode); |
| 568 | } |
| 569 | |
| 570 | static const struct rpc_call_ops nfs_write_direct_ops = { |
| 571 | .rpc_call_done = nfs_direct_write_result, |
| 572 | .rpc_release = nfs_direct_write_release, |
| 573 | }; |
| 574 | |
| 575 | /* |
| 576 | * For each wsize'd chunk of the user's buffer, dispatch an NFS WRITE |
| 577 | * operation. If nfs_writedata_alloc() or get_user_pages() fails, |
| 578 | * bail and stop sending more writes. Write length accounting is |
| 579 | * handled automatically by nfs_direct_write_result(). Otherwise, if |
| 580 | * no requests have been sent, just return an error. |
| 581 | */ |
| 582 | static ssize_t nfs_direct_write_schedule(struct nfs_direct_req *dreq, unsigned long user_addr, size_t count, loff_t pos, int sync) |
| 583 | { |
| 584 | struct nfs_open_context *ctx = dreq->ctx; |
| 585 | struct inode *inode = ctx->dentry->d_inode; |
| 586 | size_t wsize = NFS_SERVER(inode)->wsize; |
| 587 | unsigned int pgbase; |
| 588 | int result; |
| 589 | ssize_t started = 0; |
| 590 | |
| 591 | get_dreq(dreq); |
| 592 | |
| 593 | do { |
| 594 | struct nfs_write_data *data; |
| 595 | size_t bytes; |
| 596 | |
| 597 | pgbase = user_addr & ~PAGE_MASK; |
| 598 | bytes = min(wsize,count); |
| 599 | |
| 600 | result = -ENOMEM; |
| 601 | data = nfs_writedata_alloc(pgbase + bytes); |
| 602 | if (unlikely(!data)) |
| 603 | break; |
| 604 | |
| 605 | down_read(¤t->mm->mmap_sem); |
| 606 | result = get_user_pages(current, current->mm, user_addr, |
| 607 | data->npages, 0, 0, data->pagevec, NULL); |
| 608 | up_read(¤t->mm->mmap_sem); |
| 609 | if (unlikely(result < data->npages)) { |
| 610 | if (result > 0) |
| 611 | nfs_direct_release_pages(data->pagevec, result); |
| 612 | nfs_writedata_release(data); |
| 613 | break; |
| 614 | } |
| 615 | |
| 616 | get_dreq(dreq); |
| 617 | |
| 618 | list_move_tail(&data->pages, &dreq->rewrite_list); |
| 619 | |
| 620 | data->req = (struct nfs_page *) dreq; |
| 621 | data->inode = inode; |
| 622 | data->cred = ctx->cred; |
| 623 | data->args.fh = NFS_FH(inode); |
| 624 | data->args.context = ctx; |
| 625 | data->args.offset = pos; |
| 626 | data->args.pgbase = pgbase; |
| 627 | data->args.pages = data->pagevec; |
| 628 | data->args.count = bytes; |
| 629 | data->res.fattr = &data->fattr; |
| 630 | data->res.count = bytes; |
| 631 | data->res.verf = &data->verf; |
| 632 | |
| 633 | rpc_init_task(&data->task, NFS_CLIENT(inode), RPC_TASK_ASYNC, |
| 634 | &nfs_write_direct_ops, data); |
| 635 | NFS_PROTO(inode)->write_setup(data, sync); |
| 636 | |
| 637 | data->task.tk_priority = RPC_PRIORITY_NORMAL; |
| 638 | data->task.tk_cookie = (unsigned long) inode; |
| 639 | |
| 640 | rpc_execute(&data->task); |
| 641 | |
| 642 | dfprintk(VFS, "NFS: %5u initiated direct write call (req %s/%Ld, %zu bytes @ offset %Lu)\n", |
| 643 | data->task.tk_pid, |
| 644 | inode->i_sb->s_id, |
| 645 | (long long)NFS_FILEID(inode), |
| 646 | bytes, |
| 647 | (unsigned long long)data->args.offset); |
| 648 | |
| 649 | started += bytes; |
| 650 | user_addr += bytes; |
| 651 | pos += bytes; |
| 652 | |
| 653 | /* FIXME: Remove this useless math from the final patch */ |
| 654 | pgbase += bytes; |
| 655 | pgbase &= ~PAGE_MASK; |
| 656 | BUG_ON(pgbase != (user_addr & ~PAGE_MASK)); |
| 657 | |
| 658 | count -= bytes; |
| 659 | } while (count != 0); |
| 660 | |
| 661 | if (put_dreq(dreq)) |
| 662 | nfs_direct_write_complete(dreq, inode); |
| 663 | |
| 664 | if (started) |
| 665 | return 0; |
| 666 | return result < 0 ? (ssize_t) result : -EFAULT; |
| 667 | } |
| 668 | |
| 669 | static ssize_t nfs_direct_write(struct kiocb *iocb, unsigned long user_addr, size_t count, loff_t pos) |
| 670 | { |
| 671 | ssize_t result = 0; |
| 672 | sigset_t oldset; |
| 673 | struct inode *inode = iocb->ki_filp->f_mapping->host; |
| 674 | struct rpc_clnt *clnt = NFS_CLIENT(inode); |
| 675 | struct nfs_direct_req *dreq; |
| 676 | size_t wsize = NFS_SERVER(inode)->wsize; |
| 677 | int sync = 0; |
| 678 | |
| 679 | dreq = nfs_direct_req_alloc(); |
| 680 | if (!dreq) |
| 681 | return -ENOMEM; |
| 682 | nfs_alloc_commit_data(dreq); |
| 683 | |
| 684 | if (dreq->commit_data == NULL || count < wsize) |
| 685 | sync = FLUSH_STABLE; |
| 686 | |
| 687 | dreq->inode = inode; |
| 688 | dreq->ctx = get_nfs_open_context((struct nfs_open_context *)iocb->ki_filp->private_data); |
| 689 | if (!is_sync_kiocb(iocb)) |
| 690 | dreq->iocb = iocb; |
| 691 | |
| 692 | nfs_add_stats(inode, NFSIOS_DIRECTWRITTENBYTES, count); |
| 693 | |
| 694 | nfs_begin_data_update(inode); |
| 695 | |
| 696 | rpc_clnt_sigmask(clnt, &oldset); |
| 697 | result = nfs_direct_write_schedule(dreq, user_addr, count, pos, sync); |
| 698 | if (!result) |
| 699 | result = nfs_direct_wait(dreq); |
| 700 | rpc_clnt_sigunmask(clnt, &oldset); |
| 701 | |
| 702 | return result; |
| 703 | } |
| 704 | |
| 705 | /** |
| 706 | * nfs_file_direct_read - file direct read operation for NFS files |
| 707 | * @iocb: target I/O control block |
| 708 | * @iov: vector of user buffers into which to read data |
| 709 | * @nr_segs: size of iov vector |
| 710 | * @pos: byte offset in file where reading starts |
| 711 | * |
| 712 | * We use this function for direct reads instead of calling |
| 713 | * generic_file_aio_read() in order to avoid gfar's check to see if |
| 714 | * the request starts before the end of the file. For that check |
| 715 | * to work, we must generate a GETATTR before each direct read, and |
| 716 | * even then there is a window between the GETATTR and the subsequent |
| 717 | * READ where the file size could change. Our preference is simply |
| 718 | * to do all reads the application wants, and the server will take |
| 719 | * care of managing the end of file boundary. |
| 720 | * |
| 721 | * This function also eliminates unnecessarily updating the file's |
| 722 | * atime locally, as the NFS server sets the file's atime, and this |
| 723 | * client must read the updated atime from the server back into its |
| 724 | * cache. |
| 725 | */ |
| 726 | ssize_t nfs_file_direct_read(struct kiocb *iocb, const struct iovec *iov, |
| 727 | unsigned long nr_segs, loff_t pos) |
| 728 | { |
| 729 | ssize_t retval = -EINVAL; |
| 730 | struct file *file = iocb->ki_filp; |
| 731 | struct address_space *mapping = file->f_mapping; |
| 732 | /* XXX: temporary */ |
| 733 | const char __user *buf = iov[0].iov_base; |
| 734 | size_t count = iov[0].iov_len; |
| 735 | |
| 736 | dprintk("nfs: direct read(%s/%s, %lu@%Ld)\n", |
| 737 | file->f_path.dentry->d_parent->d_name.name, |
| 738 | file->f_path.dentry->d_name.name, |
| 739 | (unsigned long) count, (long long) pos); |
| 740 | |
| 741 | if (nr_segs != 1) |
| 742 | return -EINVAL; |
| 743 | |
| 744 | if (count < 0) |
| 745 | goto out; |
| 746 | retval = -EFAULT; |
| 747 | if (!access_ok(VERIFY_WRITE, buf, count)) |
| 748 | goto out; |
| 749 | retval = 0; |
| 750 | if (!count) |
| 751 | goto out; |
| 752 | |
| 753 | retval = nfs_sync_mapping(mapping); |
| 754 | if (retval) |
| 755 | goto out; |
| 756 | |
| 757 | retval = nfs_direct_read(iocb, (unsigned long) buf, count, pos); |
| 758 | if (retval > 0) |
| 759 | iocb->ki_pos = pos + retval; |
| 760 | |
| 761 | out: |
| 762 | return retval; |
| 763 | } |
| 764 | |
| 765 | /** |
| 766 | * nfs_file_direct_write - file direct write operation for NFS files |
| 767 | * @iocb: target I/O control block |
| 768 | * @iov: vector of user buffers from which to write data |
| 769 | * @nr_segs: size of iov vector |
| 770 | * @pos: byte offset in file where writing starts |
| 771 | * |
| 772 | * We use this function for direct writes instead of calling |
| 773 | * generic_file_aio_write() in order to avoid taking the inode |
| 774 | * semaphore and updating the i_size. The NFS server will set |
| 775 | * the new i_size and this client must read the updated size |
| 776 | * back into its cache. We let the server do generic write |
| 777 | * parameter checking and report problems. |
| 778 | * |
| 779 | * We also avoid an unnecessary invocation of generic_osync_inode(), |
| 780 | * as it is fairly meaningless to sync the metadata of an NFS file. |
| 781 | * |
| 782 | * We eliminate local atime updates, see direct read above. |
| 783 | * |
| 784 | * We avoid unnecessary page cache invalidations for normal cached |
| 785 | * readers of this file. |
| 786 | * |
| 787 | * Note that O_APPEND is not supported for NFS direct writes, as there |
| 788 | * is no atomic O_APPEND write facility in the NFS protocol. |
| 789 | */ |
| 790 | ssize_t nfs_file_direct_write(struct kiocb *iocb, const struct iovec *iov, |
| 791 | unsigned long nr_segs, loff_t pos) |
| 792 | { |
| 793 | ssize_t retval; |
| 794 | struct file *file = iocb->ki_filp; |
| 795 | struct address_space *mapping = file->f_mapping; |
| 796 | /* XXX: temporary */ |
| 797 | const char __user *buf = iov[0].iov_base; |
| 798 | size_t count = iov[0].iov_len; |
| 799 | |
| 800 | dfprintk(VFS, "nfs: direct write(%s/%s, %lu@%Ld)\n", |
| 801 | file->f_path.dentry->d_parent->d_name.name, |
| 802 | file->f_path.dentry->d_name.name, |
| 803 | (unsigned long) count, (long long) pos); |
| 804 | |
| 805 | if (nr_segs != 1) |
| 806 | return -EINVAL; |
| 807 | |
| 808 | retval = generic_write_checks(file, &pos, &count, 0); |
| 809 | if (retval) |
| 810 | goto out; |
| 811 | |
| 812 | retval = -EINVAL; |
| 813 | if ((ssize_t) count < 0) |
| 814 | goto out; |
| 815 | retval = 0; |
| 816 | if (!count) |
| 817 | goto out; |
| 818 | |
| 819 | retval = -EFAULT; |
| 820 | if (!access_ok(VERIFY_READ, buf, count)) |
| 821 | goto out; |
| 822 | |
| 823 | retval = nfs_sync_mapping(mapping); |
| 824 | if (retval) |
| 825 | goto out; |
| 826 | |
| 827 | retval = nfs_direct_write(iocb, (unsigned long) buf, count, pos); |
| 828 | |
| 829 | if (retval > 0) |
| 830 | iocb->ki_pos = pos + retval; |
| 831 | |
| 832 | out: |
| 833 | return retval; |
| 834 | } |
| 835 | |
| 836 | /** |
| 837 | * nfs_init_directcache - create a slab cache for nfs_direct_req structures |
| 838 | * |
| 839 | */ |
| 840 | int __init nfs_init_directcache(void) |
| 841 | { |
| 842 | nfs_direct_cachep = kmem_cache_create("nfs_direct_cache", |
| 843 | sizeof(struct nfs_direct_req), |
| 844 | 0, (SLAB_RECLAIM_ACCOUNT| |
| 845 | SLAB_MEM_SPREAD), |
| 846 | NULL, NULL); |
| 847 | if (nfs_direct_cachep == NULL) |
| 848 | return -ENOMEM; |
| 849 | |
| 850 | return 0; |
| 851 | } |
| 852 | |
| 853 | /** |
| 854 | * nfs_destroy_directcache - destroy the slab cache for nfs_direct_req structures |
| 855 | * |
| 856 | */ |
| 857 | void nfs_destroy_directcache(void) |
| 858 | { |
| 859 | kmem_cache_destroy(nfs_direct_cachep); |
| 860 | } |