fs/nfs/direct.c

   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/file.h>
  45 #include <linux/pagemap.h>
  46 #include <linux/kref.h>
  47 #include <linux/slab.h>
  48 #include <linux/task_io_accounting_ops.h>
  49 #include <linux/module.h>
  50
  51 #include <linux/nfs_fs.h>
  52 #include <linux/nfs_page.h>
  53 #include <linux/sunrpc/clnt.h>
  54
  55 #include <asm/uaccess.h>
  56 #include <linux/atomic.h>
  57
  58 #include "internal.h"
  59 #include "iostat.h"
  60 #include "pnfs.h"
  61
  62 #define NFSDBG_FACILITY         NFSDBG_VFS
  63
  64 static struct kmem_cache *nfs_direct_cachep;
  65
  66 /*
  67  * This represents a set of asynchronous requests that we're waiting on
  68  */
  69 struct nfs_direct_req {
  70         struct kref             kref;           /* release manager */
  71
  72         /* I/O parameters */
  73         struct nfs_open_context *ctx;           /* file open context info */
  74         struct nfs_lock_context *l_ctx;         /* Lock context info */
  75         struct kiocb *          iocb;           /* controlling i/o request */
  76         struct inode *          inode;          /* target file of i/o */
  77
  78         /* completion state */
  79         atomic_t                io_count;       /* i/os we're waiting for */
  80         spinlock_t              lock;           /* protect completion state */
  81         ssize_t                 count,          /* bytes actually processed */
  82                                 bytes_left,     /* bytes left to be sent */
  83                                 error;          /* any reported error */
  84         struct completion       completion;     /* wait for i/o completion */
  85
  86         /* commit state */
  87         struct nfs_mds_commit_info mds_cinfo;   /* Storage for cinfo */
  88         struct pnfs_ds_commit_info ds_cinfo;    /* Storage for cinfo */
  89         struct work_struct      work;
  90         int                     flags;
  91 #define NFS_ODIRECT_DO_COMMIT           (1)     /* an unstable reply was received */
  92 #define NFS_ODIRECT_RESCHED_WRITES      (2)     /* write verification failed */
  93         struct nfs_writeverf    verf;           /* unstable write verifier */
  94 };
  95
  96 static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops;
  97 static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops;
  98 static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode);
  99 static void nfs_direct_write_schedule_work(struct work_struct *work);
 100
 101 static inline void get_dreq(struct nfs_direct_req *dreq)
 102 {
 103         atomic_inc(&dreq->io_count);
 104 }
 105
 106 static inline int put_dreq(struct nfs_direct_req *dreq)
 107 {
 108         return atomic_dec_and_test(&dreq->io_count);
 109 }
 110
 111 /**
 112  * nfs_direct_IO - NFS address space operation for direct I/O
 113  * @rw: direction (read or write)
 114  * @iocb: target I/O control block
 115  * @iov: array of vectors that define I/O buffer
 116  * @pos: offset in file to begin the operation
 117  * @nr_segs: size of iovec array
 118  *
 119  * The presence of this routine in the address space ops vector means
 120  * the NFS client supports direct I/O. However, for most direct IO, we
 121  * shunt off direct read and write requests before the VFS gets them,
 122  * so this method is only ever called for swap.
 123  */
 124 ssize_t nfs_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov, loff_t pos, unsigned long nr_segs)
 125 {
 126 #ifndef CONFIG_NFS_SWAP
 127         dprintk("NFS: nfs_direct_IO (%s) off/no(%Ld/%lu) EINVAL\n",
 128                         iocb->ki_filp->f_path.dentry->d_name.name,
 129                         (long long) pos, nr_segs);
 130
 131         return -EINVAL;
 132 #else
 133         VM_BUG_ON(iocb->ki_left != PAGE_SIZE);
 134         VM_BUG_ON(iocb->ki_nbytes != PAGE_SIZE);
 135
 136         if (rw == READ || rw == KERNEL_READ)
 137                 return nfs_file_direct_read(iocb, iov, nr_segs, pos,
 138                                 rw == READ ? true : false);
 139         return nfs_file_direct_write(iocb, iov, nr_segs, pos,
 140                                 rw == WRITE ? true : false);
 141 #endif /* CONFIG_NFS_SWAP */
 142 }
 143
 144 static void nfs_direct_release_pages(struct page **pages, unsigned int npages)
 145 {
 146         unsigned int i;
 147         for (i = 0; i < npages; i++)
 148                 page_cache_release(pages[i]);
 149 }
 150
 151 void nfs_init_cinfo_from_dreq(struct nfs_commit_info *cinfo,
 152                               struct nfs_direct_req *dreq)
 153 {
 154         cinfo->lock = &dreq->lock;
 155         cinfo->mds = &dreq->mds_cinfo;
 156         cinfo->ds = &dreq->ds_cinfo;
 157         cinfo->dreq = dreq;
 158         cinfo->completion_ops = &nfs_direct_commit_completion_ops;
 159 }
 160
 161 static inline struct nfs_direct_req *nfs_direct_req_alloc(void)
 162 {
 163         struct nfs_direct_req *dreq;
 164
 165         dreq = kmem_cache_zalloc(nfs_direct_cachep, GFP_KERNEL);
 166         if (!dreq)
 167                 return NULL;
 168
 169         kref_init(&dreq->kref);
 170         kref_get(&dreq->kref);
 171         init_completion(&dreq->completion);
 172         INIT_LIST_HEAD(&dreq->mds_cinfo.list);
 173         INIT_WORK(&dreq->work, nfs_direct_write_schedule_work);
 174         spin_lock_init(&dreq->lock);
 175
 176         return dreq;
 177 }
 178
 179 static void nfs_direct_req_free(struct kref *kref)
 180 {
 181         struct nfs_direct_req *dreq = container_of(kref, struct nfs_direct_req, kref);
 182
 183         if (dreq->l_ctx != NULL)
 184                 nfs_put_lock_context(dreq->l_ctx);
 185         if (dreq->ctx != NULL)
 186                 put_nfs_open_context(dreq->ctx);
 187         kmem_cache_free(nfs_direct_cachep, dreq);
 188 }
 189
 190 static void nfs_direct_req_release(struct nfs_direct_req *dreq)
 191 {
 192         kref_put(&dreq->kref, nfs_direct_req_free);
 193 }
 194
 195 ssize_t nfs_dreq_bytes_left(struct nfs_direct_req *dreq)
 196 {
 197         return dreq->bytes_left;
 198 }
 199 EXPORT_SYMBOL_GPL(nfs_dreq_bytes_left);
 200
 201 /*
 202  * Collects and returns the final error value/byte-count.
 203  */
 204 static ssize_t nfs_direct_wait(struct nfs_direct_req *dreq)
 205 {
 206         ssize_t result = -EIOCBQUEUED;
 207
 208         /* Async requests don't wait here */
 209         if (dreq->iocb)
 210                 goto out;
 211
 212         result = wait_for_completion_killable(&dreq->completion);
 213
 214         if (!result)
 215                 result = dreq->error;
 216         if (!result)
 217                 result = dreq->count;
 218
 219 out:
 220         return (ssize_t) result;
 221 }
 222
 223 /*
 224  * Synchronous I/O uses a stack-allocated iocb.  Thus we can't trust
 225  * the iocb is still valid here if this is a synchronous request.
 226  */
 227 static void nfs_direct_complete(struct nfs_direct_req *dreq)
 228 {
 229         if (dreq->iocb) {
 230                 long res = (long) dreq->error;
 231                 if (!res)
 232                         res = (long) dreq->count;
 233                 aio_complete(dreq->iocb, res, 0);
 234         }
 235         complete_all(&dreq->completion);
 236
 237         nfs_direct_req_release(dreq);
 238 }
 239
 240 static void nfs_direct_readpage_release(struct nfs_page *req)
 241 {
 242         dprintk("NFS: direct read done (%s/%lld %d@%lld)\n",
 243                 req->wb_context->dentry->d_inode->i_sb->s_id,
 244                 (long long)NFS_FILEID(req->wb_context->dentry->d_inode),
 245                 req->wb_bytes,
 246                 (long long)req_offset(req));
 247         nfs_release_request(req);
 248 }
 249
 250 static void nfs_direct_read_completion(struct nfs_pgio_header *hdr)
 251 {
 252         unsigned long bytes = 0;
 253         struct nfs_direct_req *dreq = hdr->dreq;
 254
 255         if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
 256                 goto out_put;
 257
 258         spin_lock(&dreq->lock);
 259         if (test_bit(NFS_IOHDR_ERROR, &hdr->flags) && (hdr->good_bytes == 0))
 260                 dreq->error = hdr->error;
 261         else
 262                 dreq->count += hdr->good_bytes;
 263         spin_unlock(&dreq->lock);
 264
 265         while (!list_empty(&hdr->pages)) {
 266                 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
 267                 struct page *page = req->wb_page;
 268
 269                 if (test_bit(NFS_IOHDR_EOF, &hdr->flags)) {
 270                         if (bytes > hdr->good_bytes)
 271                                 zero_user(page, 0, PAGE_SIZE);
 272                         else if (hdr->good_bytes - bytes < PAGE_SIZE)
 273                                 zero_user_segment(page,
 274                                         hdr->good_bytes & ~PAGE_MASK,
 275                                         PAGE_SIZE);
 276                 }
 277                 if (!PageCompound(page)) {
 278                         if (test_bit(NFS_IOHDR_ERROR, &hdr->flags)) {
 279                                 if (bytes < hdr->good_bytes)
 280                                         set_page_dirty(page);
 281                         } else
 282                                 set_page_dirty(page);
 283                 }
 284                 bytes += req->wb_bytes;
 285                 nfs_list_remove_request(req);
 286                 nfs_direct_readpage_release(req);
 287         }
 288 out_put:
 289         if (put_dreq(dreq))
 290                 nfs_direct_complete(dreq);
 291         hdr->release(hdr);
 292 }
 293
 294 static void nfs_read_sync_pgio_error(struct list_head *head)
 295 {
 296         struct nfs_page *req;
 297
 298         while (!list_empty(head)) {
 299                 req = nfs_list_entry(head->next);
 300                 nfs_list_remove_request(req);
 301                 nfs_release_request(req);
 302         }
 303 }
 304
 305 static void nfs_direct_pgio_init(struct nfs_pgio_header *hdr)
 306 {
 307         get_dreq(hdr->dreq);
 308 }
 309
 310 static const struct nfs_pgio_completion_ops nfs_direct_read_completion_ops = {
 311         .error_cleanup = nfs_read_sync_pgio_error,
 312         .init_hdr = nfs_direct_pgio_init,
 313         .completion = nfs_direct_read_completion,
 314 };
 315
 316 /*
 317  * For each rsize'd chunk of the user's buffer, dispatch an NFS READ
 318  * operation.  If nfs_readdata_alloc() or get_user_pages() fails,
 319  * bail and stop sending more reads.  Read length accounting is
 320  * handled automatically by nfs_direct_read_result().  Otherwise, if
 321  * no requests have been sent, just return an error.
 322  */
 323 static ssize_t nfs_direct_read_schedule_segment(struct nfs_pageio_descriptor *desc,
 324                                                 const struct iovec *iov,
 325                                                 loff_t pos, bool uio)
 326 {
 327         struct nfs_direct_req *dreq = desc->pg_dreq;
 328         struct nfs_open_context *ctx = dreq->ctx;
 329         struct inode *inode = ctx->dentry->d_inode;
 330         unsigned long user_addr = (unsigned long)iov->iov_base;
 331         size_t count = iov->iov_len;
 332         size_t rsize = NFS_SERVER(inode)->rsize;
 333         unsigned int pgbase;
 334         int result;
 335         ssize_t started = 0;
 336         struct page **pagevec = NULL;
 337         unsigned int npages;
 338
 339         do {
 340                 size_t bytes;
 341                 int i;
 342
 343                 pgbase = user_addr & ~PAGE_MASK;
 344                 bytes = min(max_t(size_t, rsize, PAGE_SIZE), count);
 345
 346                 result = -ENOMEM;
 347                 npages = nfs_page_array_len(pgbase, bytes);
 348                 if (!pagevec)
 349                         pagevec = kmalloc(npages * sizeof(struct page *),
 350                                           GFP_KERNEL);
 351                 if (!pagevec)
 352                         break;
 353                 if (uio) {
 354                         down_read(&current->mm->mmap_sem);
 355                         result = get_user_pages(current, current->mm, user_addr,
 356                                         npages, 1, 0, pagevec, NULL);
 357                         up_read(&current->mm->mmap_sem);
 358                         if (result < 0)
 359                                 break;
 360                 } else {
 361                         WARN_ON(npages != 1);
 362                         result = get_kernel_page(user_addr, 1, pagevec);
 363                         if (WARN_ON(result != 1))
 364                                 break;
 365                 }
 366
 367                 if ((unsigned)result < npages) {
 368                         bytes = result * PAGE_SIZE;
 369                         if (bytes <= pgbase) {
 370                                 nfs_direct_release_pages(pagevec, result);
 371                                 break;
 372                         }
 373                         bytes -= pgbase;
 374                         npages = result;
 375                 }
 376
 377                 for (i = 0; i < npages; i++) {
 378                         struct nfs_page *req;
 379                         unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase);
 380                         /* XXX do we need to do the eof zeroing found in async_filler? */
 381                         req = nfs_create_request(dreq->ctx, dreq->inode,
 382                                                  pagevec[i],
 383                                                  pgbase, req_len);
 384                         if (IS_ERR(req)) {
 385                                 result = PTR_ERR(req);
 386                                 break;
 387                         }
 388                         req->wb_index = pos >> PAGE_SHIFT;
 389                         req->wb_offset = pos & ~PAGE_MASK;
 390                         if (!nfs_pageio_add_request(desc, req)) {
 391                                 result = desc->pg_error;
 392                                 nfs_release_request(req);
 393                                 break;
 394                         }
 395                         pgbase = 0;
 396                         bytes -= req_len;
 397                         started += req_len;
 398                         user_addr += req_len;
 399                         pos += req_len;
 400                         count -= req_len;
 401                         dreq->bytes_left -= req_len;
 402                 }
 403                 /* The nfs_page now hold references to these pages */
 404                 nfs_direct_release_pages(pagevec, npages);
 405         } while (count != 0 && result >= 0);
 406
 407         kfree(pagevec);
 408
 409         if (started)
 410                 return started;
 411         return result < 0 ? (ssize_t) result : -EFAULT;
 412 }
 413
 414 static ssize_t nfs_direct_read_schedule_iovec(struct nfs_direct_req *dreq,
 415                                               const struct iovec *iov,
 416                                               unsigned long nr_segs,
 417                                               loff_t pos, bool uio)
 418 {
 419         struct nfs_pageio_descriptor desc;
 420         ssize_t result = -EINVAL;
 421         size_t requested_bytes = 0;
 422         unsigned long seg;
 423
 424         NFS_PROTO(dreq->inode)->read_pageio_init(&desc, dreq->inode,
 425                              &nfs_direct_read_completion_ops);
 426         get_dreq(dreq);
 427         desc.pg_dreq = dreq;
 428
 429         for (seg = 0; seg < nr_segs; seg++) {
 430                 const struct iovec *vec = &iov[seg];
 431                 result = nfs_direct_read_schedule_segment(&desc, vec, pos, uio);
 432                 if (result < 0)
 433                         break;
 434                 requested_bytes += result;
 435                 if ((size_t)result < vec->iov_len)
 436                         break;
 437                 pos += vec->iov_len;
 438         }
 439
 440         nfs_pageio_complete(&desc);
 441
 442         /*
 443          * If no bytes were started, return the error, and let the
 444          * generic layer handle the completion.
 445          */
 446         if (requested_bytes == 0) {
 447                 nfs_direct_req_release(dreq);
 448                 return result < 0 ? result : -EIO;
 449         }
 450
 451         if (put_dreq(dreq))
 452                 nfs_direct_complete(dreq);
 453         return 0;
 454 }
 455
 456 static ssize_t nfs_direct_read(struct kiocb *iocb, const struct iovec *iov,
 457                                unsigned long nr_segs, loff_t pos, bool uio)
 458 {
 459         ssize_t result = -ENOMEM;
 460         struct inode *inode = iocb->ki_filp->f_mapping->host;
 461         struct nfs_direct_req *dreq;
 462         struct nfs_lock_context *l_ctx;
 463
 464         dreq = nfs_direct_req_alloc();
 465         if (dreq == NULL)
 466                 goto out;
 467
 468         dreq->inode = inode;
 469         dreq->bytes_left = iov_length(iov, nr_segs);
 470         dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
 471         l_ctx = nfs_get_lock_context(dreq->ctx);
 472         if (IS_ERR(l_ctx)) {
 473                 result = PTR_ERR(l_ctx);
 474                 goto out_release;
 475         }
 476         dreq->l_ctx = l_ctx;
 477         if (!is_sync_kiocb(iocb))
 478                 dreq->iocb = iocb;
 479
 480         NFS_I(inode)->read_io += iov_length(iov, nr_segs);
 481         result = nfs_direct_read_schedule_iovec(dreq, iov, nr_segs, pos, uio);
 482         if (!result)
 483                 result = nfs_direct_wait(dreq);
 484 out_release:
 485         nfs_direct_req_release(dreq);
 486 out:
 487         return result;
 488 }
 489
 490 static void nfs_inode_dio_write_done(struct inode *inode)
 491 {
 492         nfs_zap_mapping(inode, inode->i_mapping);
 493         inode_dio_done(inode);
 494 }
 495
 496 #if IS_ENABLED(CONFIG_NFS_V3) || IS_ENABLED(CONFIG_NFS_V4)
 497 static void nfs_direct_write_reschedule(struct nfs_direct_req *dreq)
 498 {
 499         struct nfs_pageio_descriptor desc;
 500         struct nfs_page *req, *tmp;
 501         LIST_HEAD(reqs);
 502         struct nfs_commit_info cinfo;
 503         LIST_HEAD(failed);
 504
 505         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 506         pnfs_recover_commit_reqs(dreq->inode, &reqs, &cinfo);
 507         spin_lock(cinfo.lock);
 508         nfs_scan_commit_list(&cinfo.mds->list, &reqs, &cinfo, 0);
 509         spin_unlock(cinfo.lock);
 510
 511         dreq->count = 0;
 512         get_dreq(dreq);
 513
 514         NFS_PROTO(dreq->inode)->write_pageio_init(&desc, dreq->inode, FLUSH_STABLE,
 515                               &nfs_direct_write_completion_ops);
 516         desc.pg_dreq = dreq;
 517
 518         list_for_each_entry_safe(req, tmp, &reqs, wb_list) {
 519                 if (!nfs_pageio_add_request(&desc, req)) {
 520                         nfs_list_remove_request(req);
 521                         nfs_list_add_request(req, &failed);
 522                         spin_lock(cinfo.lock);
 523                         dreq->flags = 0;
 524                         dreq->error = -EIO;
 525                         spin_unlock(cinfo.lock);
 526                 }
 527                 nfs_release_request(req);
 528         }
 529         nfs_pageio_complete(&desc);
 530
 531         while (!list_empty(&failed)) {
 532                 req = nfs_list_entry(failed.next);
 533                 nfs_list_remove_request(req);
 534                 nfs_unlock_and_release_request(req);
 535         }
 536
 537         if (put_dreq(dreq))
 538                 nfs_direct_write_complete(dreq, dreq->inode);
 539 }
 540
 541 static void nfs_direct_commit_complete(struct nfs_commit_data *data)
 542 {
 543         struct nfs_direct_req *dreq = data->dreq;
 544         struct nfs_commit_info cinfo;
 545         struct nfs_page *req;
 546         int status = data->task.tk_status;
 547
 548         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 549         if (status < 0) {
 550                 dprintk("NFS: %5u commit failed with error %d.\n",
 551                         data->task.tk_pid, status);
 552                 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 553         } else if (memcmp(&dreq->verf, &data->verf, sizeof(data->verf))) {
 554                 dprintk("NFS: %5u commit verify failed\n", data->task.tk_pid);
 555                 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 556         }
 557
 558         dprintk("NFS: %5u commit returned %d\n", data->task.tk_pid, status);
 559         while (!list_empty(&data->pages)) {
 560                 req = nfs_list_entry(data->pages.next);
 561                 nfs_list_remove_request(req);
 562                 if (dreq->flags == NFS_ODIRECT_RESCHED_WRITES) {
 563                         /* Note the rewrite will go through mds */
 564                         nfs_mark_request_commit(req, NULL, &cinfo);
 565                 } else
 566                         nfs_release_request(req);
 567                 nfs_unlock_and_release_request(req);
 568         }
 569
 570         if (atomic_dec_and_test(&cinfo.mds->rpcs_out))
 571                 nfs_direct_write_complete(dreq, data->inode);
 572 }
 573
 574 static void nfs_direct_error_cleanup(struct nfs_inode *nfsi)
 575 {
 576         /* There is no lock to clear */
 577 }
 578
 579 static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops = {
 580         .completion = nfs_direct_commit_complete,
 581         .error_cleanup = nfs_direct_error_cleanup,
 582 };
 583
 584 static void nfs_direct_commit_schedule(struct nfs_direct_req *dreq)
 585 {
 586         int res;
 587         struct nfs_commit_info cinfo;
 588         LIST_HEAD(mds_list);
 589
 590         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 591         nfs_scan_commit(dreq->inode, &mds_list, &cinfo);
 592         res = nfs_generic_commit_list(dreq->inode, &mds_list, 0, &cinfo);
 593         if (res < 0) /* res == -ENOMEM */
 594                 nfs_direct_write_reschedule(dreq);
 595 }
 596
 597 static void nfs_direct_write_schedule_work(struct work_struct *work)
 598 {
 599         struct nfs_direct_req *dreq = container_of(work, struct nfs_direct_req, work);
 600         int flags = dreq->flags;
 601
 602         dreq->flags = 0;
 603         switch (flags) {
 604                 case NFS_ODIRECT_DO_COMMIT:
 605                         nfs_direct_commit_schedule(dreq);
 606                         break;
 607                 case NFS_ODIRECT_RESCHED_WRITES:
 608                         nfs_direct_write_reschedule(dreq);
 609                         break;
 610                 default:
 611                         nfs_inode_dio_write_done(dreq->inode);
 612                         nfs_direct_complete(dreq);
 613         }
 614 }
 615
 616 static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode)
 617 {
 618         schedule_work(&dreq->work); /* Calls nfs_direct_write_schedule_work */
 619 }
 620
 621 #else
 622 static void nfs_direct_write_schedule_work(struct work_struct *work)
 623 {
 624 }
 625
 626 static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode)
 627 {
 628         nfs_inode_dio_write_done(inode);
 629         nfs_direct_complete(dreq);
 630 }
 631 #endif
 632
 633 /*
 634  * NB: Return the value of the first error return code.  Subsequent
 635  *     errors after the first one are ignored.
 636  */
 637 /*
 638  * For each wsize'd chunk of the user's buffer, dispatch an NFS WRITE
 639  * operation.  If nfs_writedata_alloc() or get_user_pages() fails,
 640  * bail and stop sending more writes.  Write length accounting is
 641  * handled automatically by nfs_direct_write_result().  Otherwise, if
 642  * no requests have been sent, just return an error.
 643  */
 644 static ssize_t nfs_direct_write_schedule_segment(struct nfs_pageio_descriptor *desc,
 645                                                  const struct iovec *iov,
 646                                                  loff_t pos, bool uio)
 647 {
 648         struct nfs_direct_req *dreq = desc->pg_dreq;
 649         struct nfs_open_context *ctx = dreq->ctx;
 650         struct inode *inode = ctx->dentry->d_inode;
 651         unsigned long user_addr = (unsigned long)iov->iov_base;
 652         size_t count = iov->iov_len;
 653         size_t wsize = NFS_SERVER(inode)->wsize;
 654         unsigned int pgbase;
 655         int result;
 656         ssize_t started = 0;
 657         struct page **pagevec = NULL;
 658         unsigned int npages;
 659
 660         do {
 661                 size_t bytes;
 662                 int i;
 663
 664                 pgbase = user_addr & ~PAGE_MASK;
 665                 bytes = min(max_t(size_t, wsize, PAGE_SIZE), count);
 666
 667                 result = -ENOMEM;
 668                 npages = nfs_page_array_len(pgbase, bytes);
 669                 if (!pagevec)
 670                         pagevec = kmalloc(npages * sizeof(struct page *), GFP_KERNEL);
 671                 if (!pagevec)
 672                         break;
 673
 674                 if (uio) {
 675                         down_read(&current->mm->mmap_sem);
 676                         result = get_user_pages(current, current->mm, user_addr,
 677                                                 npages, 0, 0, pagevec, NULL);
 678                         up_read(&current->mm->mmap_sem);
 679                         if (result < 0)
 680                                 break;
 681                 } else {
 682                         WARN_ON(npages != 1);
 683                         result = get_kernel_page(user_addr, 0, pagevec);
 684                         if (WARN_ON(result != 1))
 685                                 break;
 686                 }
 687
 688                 if ((unsigned)result < npages) {
 689                         bytes = result * PAGE_SIZE;
 690                         if (bytes <= pgbase) {
 691                                 nfs_direct_release_pages(pagevec, result);
 692                                 break;
 693                         }
 694                         bytes -= pgbase;
 695                         npages = result;
 696                 }
 697
 698                 for (i = 0; i < npages; i++) {
 699                         struct nfs_page *req;
 700                         unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase);
 701
 702                         req = nfs_create_request(dreq->ctx, dreq->inode,
 703                                                  pagevec[i],
 704                                                  pgbase, req_len);
 705                         if (IS_ERR(req)) {
 706                                 result = PTR_ERR(req);
 707                                 break;
 708                         }
 709                         nfs_lock_request(req);
 710                         req->wb_index = pos >> PAGE_SHIFT;
 711                         req->wb_offset = pos & ~PAGE_MASK;
 712                         if (!nfs_pageio_add_request(desc, req)) {
 713                                 result = desc->pg_error;
 714                                 nfs_unlock_and_release_request(req);
 715                                 break;
 716                         }
 717                         pgbase = 0;
 718                         bytes -= req_len;
 719                         started += req_len;
 720                         user_addr += req_len;
 721                         pos += req_len;
 722                         count -= req_len;
 723                         dreq->bytes_left -= req_len;
 724                 }
 725                 /* The nfs_page now hold references to these pages */
 726                 nfs_direct_release_pages(pagevec, npages);
 727         } while (count != 0 && result >= 0);
 728
 729         kfree(pagevec);
 730
 731         if (started)
 732                 return started;
 733         return result < 0 ? (ssize_t) result : -EFAULT;
 734 }
 735
 736 static void nfs_direct_write_completion(struct nfs_pgio_header *hdr)
 737 {
 738         struct nfs_direct_req *dreq = hdr->dreq;
 739         struct nfs_commit_info cinfo;
 740         int bit = -1;
 741         struct nfs_page *req = nfs_list_entry(hdr->pages.next);
 742
 743         if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
 744                 goto out_put;
 745
 746         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 747
 748         spin_lock(&dreq->lock);
 749
 750         if (test_bit(NFS_IOHDR_ERROR, &hdr->flags)) {
 751                 dreq->flags = 0;
 752                 dreq->error = hdr->error;
 753         }
 754         if (dreq->error != 0)
 755                 bit = NFS_IOHDR_ERROR;
 756         else {
 757                 dreq->count += hdr->good_bytes;
 758                 if (test_bit(NFS_IOHDR_NEED_RESCHED, &hdr->flags)) {
 759                         dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 760                         bit = NFS_IOHDR_NEED_RESCHED;
 761                 } else if (test_bit(NFS_IOHDR_NEED_COMMIT, &hdr->flags)) {
 762                         if (dreq->flags == NFS_ODIRECT_RESCHED_WRITES)
 763                                 bit = NFS_IOHDR_NEED_RESCHED;
 764                         else if (dreq->flags == 0) {
 765                                 memcpy(&dreq->verf, hdr->verf,
 766                                        sizeof(dreq->verf));
 767                                 bit = NFS_IOHDR_NEED_COMMIT;
 768                                 dreq->flags = NFS_ODIRECT_DO_COMMIT;
 769                         } else if (dreq->flags == NFS_ODIRECT_DO_COMMIT) {
 770                                 if (memcmp(&dreq->verf, hdr->verf, sizeof(dreq->verf))) {
 771                                         dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 772                                         bit = NFS_IOHDR_NEED_RESCHED;
 773                                 } else
 774                                         bit = NFS_IOHDR_NEED_COMMIT;
 775                         }
 776                 }
 777         }
 778         spin_unlock(&dreq->lock);
 779
 780         while (!list_empty(&hdr->pages)) {
 781                 req = nfs_list_entry(hdr->pages.next);
 782                 nfs_list_remove_request(req);
 783                 switch (bit) {
 784                 case NFS_IOHDR_NEED_RESCHED:
 785                 case NFS_IOHDR_NEED_COMMIT:
 786                         kref_get(&req->wb_kref);
 787                         nfs_mark_request_commit(req, hdr->lseg, &cinfo);
 788                 }
 789                 nfs_unlock_and_release_request(req);
 790         }
 791
 792 out_put:
 793         if (put_dreq(dreq))
 794                 nfs_direct_write_complete(dreq, hdr->inode);
 795         hdr->release(hdr);
 796 }
 797
 798 static void nfs_write_sync_pgio_error(struct list_head *head)
 799 {
 800         struct nfs_page *req;
 801
 802         while (!list_empty(head)) {
 803                 req = nfs_list_entry(head->next);
 804                 nfs_list_remove_request(req);
 805                 nfs_unlock_and_release_request(req);
 806         }
 807 }
 808
 809 static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops = {
 810         .error_cleanup = nfs_write_sync_pgio_error,
 811         .init_hdr = nfs_direct_pgio_init,
 812         .completion = nfs_direct_write_completion,
 813 };
 814
 815 static ssize_t nfs_direct_write_schedule_iovec(struct nfs_direct_req *dreq,
 816                                                const struct iovec *iov,
 817                                                unsigned long nr_segs,
 818                                                loff_t pos, bool uio)
 819 {
 820         struct nfs_pageio_descriptor desc;
 821         struct inode *inode = dreq->inode;
 822         ssize_t result = 0;
 823         size_t requested_bytes = 0;
 824         unsigned long seg;
 825
 826         NFS_PROTO(inode)->write_pageio_init(&desc, inode, FLUSH_COND_STABLE,
 827                               &nfs_direct_write_completion_ops);
 828         desc.pg_dreq = dreq;
 829         get_dreq(dreq);
 830         atomic_inc(&inode->i_dio_count);
 831
 832         NFS_I(dreq->inode)->write_io += iov_length(iov, nr_segs);
 833         for (seg = 0; seg < nr_segs; seg++) {
 834                 const struct iovec *vec = &iov[seg];
 835                 result = nfs_direct_write_schedule_segment(&desc, vec, pos, uio);
 836                 if (result < 0)
 837                         break;
 838                 requested_bytes += result;
 839                 if ((size_t)result < vec->iov_len)
 840                         break;
 841                 pos += vec->iov_len;
 842         }
 843         nfs_pageio_complete(&desc);
 844
 845         /*
 846          * If no bytes were started, return the error, and let the
 847          * generic layer handle the completion.
 848          */
 849         if (requested_bytes == 0) {
 850                 inode_dio_done(inode);
 851                 nfs_direct_req_release(dreq);
 852                 return result < 0 ? result : -EIO;
 853         }
 854
 855         if (put_dreq(dreq))
 856                 nfs_direct_write_complete(dreq, dreq->inode);
 857         return 0;
 858 }
 859
 860 static ssize_t nfs_direct_write(struct kiocb *iocb, const struct iovec *iov,
 861                                 unsigned long nr_segs, loff_t pos,
 862                                 size_t count, bool uio)
 863 {
 864         ssize_t result = -ENOMEM;
 865         struct inode *inode = iocb->ki_filp->f_mapping->host;
 866         struct nfs_direct_req *dreq;
 867         struct nfs_lock_context *l_ctx;
 868
 869         dreq = nfs_direct_req_alloc();
 870         if (!dreq)
 871                 goto out;
 872
 873         dreq->inode = inode;
 874         dreq->bytes_left = count;
 875         dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
 876         l_ctx = nfs_get_lock_context(dreq->ctx);
 877         if (IS_ERR(l_ctx)) {
 878                 result = PTR_ERR(l_ctx);
 879                 goto out_release;
 880         }
 881         dreq->l_ctx = l_ctx;
 882         if (!is_sync_kiocb(iocb))
 883                 dreq->iocb = iocb;
 884
 885         result = nfs_direct_write_schedule_iovec(dreq, iov, nr_segs, pos, uio);
 886         if (!result)
 887                 result = nfs_direct_wait(dreq);
 888 out_release:
 889         nfs_direct_req_release(dreq);
 890 out:
 891         return result;
 892 }
 893
 894 /**
 895  * nfs_file_direct_read - file direct read operation for NFS files
 896  * @iocb: target I/O control block
 897  * @iov: vector of user buffers into which to read data
 898  * @nr_segs: size of iov vector
 899  * @pos: byte offset in file where reading starts
 900  *
 901  * We use this function for direct reads instead of calling
 902  * generic_file_aio_read() in order to avoid gfar's check to see if
 903  * the request starts before the end of the file.  For that check
 904  * to work, we must generate a GETATTR before each direct read, and
 905  * even then there is a window between the GETATTR and the subsequent
 906  * READ where the file size could change.  Our preference is simply
 907  * to do all reads the application wants, and the server will take
 908  * care of managing the end of file boundary.
 909  *
 910  * This function also eliminates unnecessarily updating the file's
 911  * atime locally, as the NFS server sets the file's atime, and this
 912  * client must read the updated atime from the server back into its
 913  * cache.
 914  */
 915 ssize_t nfs_file_direct_read(struct kiocb *iocb, const struct iovec *iov,
 916                                 unsigned long nr_segs, loff_t pos, bool uio)
 917 {
 918         ssize_t retval = -EINVAL;
 919         struct file *file = iocb->ki_filp;
 920         struct address_space *mapping = file->f_mapping;
 921         size_t count;
 922
 923         count = iov_length(iov, nr_segs);
 924         nfs_add_stats(mapping->host, NFSIOS_DIRECTREADBYTES, count);
 925
 926         dfprintk(FILE, "NFS: direct read(%s/%s, %zd@%Ld)\n",
 927                 file->f_path.dentry->d_parent->d_name.name,
 928                 file->f_path.dentry->d_name.name,
 929                 count, (long long) pos);
 930
 931         retval = 0;
 932         if (!count)
 933                 goto out;
 934
 935         retval = nfs_sync_mapping(mapping);
 936         if (retval)
 937                 goto out;
 938
 939         task_io_account_read(count);
 940
 941         retval = nfs_direct_read(iocb, iov, nr_segs, pos, uio);
 942         if (retval > 0)
 943                 iocb->ki_pos = pos + retval;
 944
 945 out:
 946         return retval;
 947 }
 948
 949 /**
 950  * nfs_file_direct_write - file direct write operation for NFS files
 951  * @iocb: target I/O control block
 952  * @iov: vector of user buffers from which to write data
 953  * @nr_segs: size of iov vector
 954  * @pos: byte offset in file where writing starts
 955  *
 956  * We use this function for direct writes instead of calling
 957  * generic_file_aio_write() in order to avoid taking the inode
 958  * semaphore and updating the i_size.  The NFS server will set
 959  * the new i_size and this client must read the updated size
 960  * back into its cache.  We let the server do generic write
 961  * parameter checking and report problems.
 962  *
 963  * We eliminate local atime updates, see direct read above.
 964  *
 965  * We avoid unnecessary page cache invalidations for normal cached
 966  * readers of this file.
 967  *
 968  * Note that O_APPEND is not supported for NFS direct writes, as there
 969  * is no atomic O_APPEND write facility in the NFS protocol.
 970  */
 971 ssize_t nfs_file_direct_write(struct kiocb *iocb, const struct iovec *iov,
 972                                 unsigned long nr_segs, loff_t pos, bool uio)
 973 {
 974         ssize_t retval = -EINVAL;
 975         struct file *file = iocb->ki_filp;
 976         struct address_space *mapping = file->f_mapping;
 977         size_t count;
 978
 979         count = iov_length(iov, nr_segs);
 980         nfs_add_stats(mapping->host, NFSIOS_DIRECTWRITTENBYTES, count);
 981
 982         dfprintk(FILE, "NFS: direct write(%s/%s, %zd@%Ld)\n",
 983                 file->f_path.dentry->d_parent->d_name.name,
 984                 file->f_path.dentry->d_name.name,
 985                 count, (long long) pos);
 986
 987         retval = generic_write_checks(file, &pos, &count, 0);
 988         if (retval)
 989                 goto out;
 990
 991         retval = -EINVAL;
 992         if ((ssize_t) count < 0)
 993                 goto out;
 994         retval = 0;
 995         if (!count)
 996                 goto out;
 997
 998         retval = nfs_sync_mapping(mapping);
 999         if (retval)
1000                 goto out;
1001
1002         task_io_account_write(count);
1003
1004         retval = nfs_direct_write(iocb, iov, nr_segs, pos, count, uio);
1005         if (retval > 0) {
1006                 struct inode *inode = mapping->host;
1007
1008                 iocb->ki_pos = pos + retval;
1009                 spin_lock(&inode->i_lock);
1010                 if (i_size_read(inode) < iocb->ki_pos)
1011                         i_size_write(inode, iocb->ki_pos);
1012                 spin_unlock(&inode->i_lock);
1013         }
1014 out:
1015         return retval;
1016 }
1017
1018 /**
1019  * nfs_init_directcache - create a slab cache for nfs_direct_req structures
1020  *
1021  */
1022 int __init nfs_init_directcache(void)
1023 {
1024         nfs_direct_cachep = kmem_cache_create("nfs_direct_cache",
1025                                                 sizeof(struct nfs_direct_req),
1026                                                 0, (SLAB_RECLAIM_ACCOUNT|
1027                                                         SLAB_MEM_SPREAD),
1028                                                 NULL);
1029         if (nfs_direct_cachep == NULL)
1030                 return -ENOMEM;
1031
1032         return 0;
1033 }
1034
1035 /**
1036  * nfs_destroy_directcache - destroy the slab cache for nfs_direct_req structures
1037  *
1038  */
1039 void nfs_destroy_directcache(void)
1040 {
1041         kmem_cache_destroy(nfs_direct_cachep);
1042 }