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 (%pD) off/no(%Ld/%lu) EINVAL\n",
 128                         iocb->ki_filp, (long long) pos, nr_segs);
 129
 130         return -EINVAL;
 131 #else
 132         VM_BUG_ON(iocb->ki_nbytes != PAGE_SIZE);
 133
 134         if (rw == READ || rw == KERNEL_READ)
 135                 return nfs_file_direct_read(iocb, iov, nr_segs, pos,
 136                                 rw == READ ? true : false);
 137         return nfs_file_direct_write(iocb, iov, nr_segs, pos,
 138                                 rw == WRITE ? true : false);
 139 #endif /* CONFIG_NFS_SWAP */
 140 }
 141
 142 static void nfs_direct_release_pages(struct page **pages, unsigned int npages)
 143 {
 144         unsigned int i;
 145         for (i = 0; i < npages; i++)
 146                 page_cache_release(pages[i]);
 147 }
 148
 149 void nfs_init_cinfo_from_dreq(struct nfs_commit_info *cinfo,
 150                               struct nfs_direct_req *dreq)
 151 {
 152         cinfo->lock = &dreq->lock;
 153         cinfo->mds = &dreq->mds_cinfo;
 154         cinfo->ds = &dreq->ds_cinfo;
 155         cinfo->dreq = dreq;
 156         cinfo->completion_ops = &nfs_direct_commit_completion_ops;
 157 }
 158
 159 static inline struct nfs_direct_req *nfs_direct_req_alloc(void)
 160 {
 161         struct nfs_direct_req *dreq;
 162
 163         dreq = kmem_cache_zalloc(nfs_direct_cachep, GFP_KERNEL);
 164         if (!dreq)
 165                 return NULL;
 166
 167         kref_init(&dreq->kref);
 168         kref_get(&dreq->kref);
 169         init_completion(&dreq->completion);
 170         INIT_LIST_HEAD(&dreq->mds_cinfo.list);
 171         INIT_WORK(&dreq->work, nfs_direct_write_schedule_work);
 172         spin_lock_init(&dreq->lock);
 173
 174         return dreq;
 175 }
 176
 177 static void nfs_direct_req_free(struct kref *kref)
 178 {
 179         struct nfs_direct_req *dreq = container_of(kref, struct nfs_direct_req, kref);
 180
 181         if (dreq->l_ctx != NULL)
 182                 nfs_put_lock_context(dreq->l_ctx);
 183         if (dreq->ctx != NULL)
 184                 put_nfs_open_context(dreq->ctx);
 185         kmem_cache_free(nfs_direct_cachep, dreq);
 186 }
 187
 188 static void nfs_direct_req_release(struct nfs_direct_req *dreq)
 189 {
 190         kref_put(&dreq->kref, nfs_direct_req_free);
 191 }
 192
 193 ssize_t nfs_dreq_bytes_left(struct nfs_direct_req *dreq)
 194 {
 195         return dreq->bytes_left;
 196 }
 197 EXPORT_SYMBOL_GPL(nfs_dreq_bytes_left);
 198
 199 /*
 200  * Collects and returns the final error value/byte-count.
 201  */
 202 static ssize_t nfs_direct_wait(struct nfs_direct_req *dreq)
 203 {
 204         ssize_t result = -EIOCBQUEUED;
 205
 206         /* Async requests don't wait here */
 207         if (dreq->iocb)
 208                 goto out;
 209
 210         result = wait_for_completion_killable(&dreq->completion);
 211
 212         if (!result)
 213                 result = dreq->error;
 214         if (!result)
 215                 result = dreq->count;
 216
 217 out:
 218         return (ssize_t) result;
 219 }
 220
 221 /*
 222  * Synchronous I/O uses a stack-allocated iocb.  Thus we can't trust
 223  * the iocb is still valid here if this is a synchronous request.
 224  */
 225 static void nfs_direct_complete(struct nfs_direct_req *dreq, bool write)
 226 {
 227         struct inode *inode = dreq->inode;
 228
 229         if (dreq->iocb) {
 230                 loff_t pos = dreq->iocb->ki_pos + dreq->count;
 231                 long res = (long) dreq->error;
 232                 if (!res)
 233                         res = (long) dreq->count;
 234
 235                 if (write) {
 236                         spin_lock(&inode->i_lock);
 237                         if (i_size_read(inode) < pos)
 238                                 i_size_write(inode, pos);
 239                         spin_unlock(&inode->i_lock);
 240                 }
 241
 242                 aio_complete(dreq->iocb, res, 0);
 243         }
 244         complete_all(&dreq->completion);
 245
 246         nfs_direct_req_release(dreq);
 247 }
 248
 249 static void nfs_direct_readpage_release(struct nfs_page *req)
 250 {
 251         dprintk("NFS: direct read done (%s/%llu %d@%lld)\n",
 252                 req->wb_context->dentry->d_inode->i_sb->s_id,
 253                 (unsigned long long)NFS_FILEID(req->wb_context->dentry->d_inode),
 254                 req->wb_bytes,
 255                 (long long)req_offset(req));
 256         nfs_release_request(req);
 257 }
 258
 259 static void nfs_direct_read_completion(struct nfs_pgio_header *hdr)
 260 {
 261         unsigned long bytes = 0;
 262         struct nfs_direct_req *dreq = hdr->dreq;
 263
 264         if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
 265                 goto out_put;
 266
 267         spin_lock(&dreq->lock);
 268         if (test_bit(NFS_IOHDR_ERROR, &hdr->flags) && (hdr->good_bytes == 0))
 269                 dreq->error = hdr->error;
 270         else
 271                 dreq->count += hdr->good_bytes;
 272         spin_unlock(&dreq->lock);
 273
 274         while (!list_empty(&hdr->pages)) {
 275                 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
 276                 struct page *page = req->wb_page;
 277
 278                 if (!PageCompound(page) && bytes < hdr->good_bytes)
 279                         set_page_dirty(page);
 280                 bytes += req->wb_bytes;
 281                 nfs_list_remove_request(req);
 282                 nfs_direct_readpage_release(req);
 283         }
 284 out_put:
 285         if (put_dreq(dreq))
 286                 nfs_direct_complete(dreq, false);
 287         hdr->release(hdr);
 288 }
 289
 290 static void nfs_read_sync_pgio_error(struct list_head *head)
 291 {
 292         struct nfs_page *req;
 293
 294         while (!list_empty(head)) {
 295                 req = nfs_list_entry(head->next);
 296                 nfs_list_remove_request(req);
 297                 nfs_release_request(req);
 298         }
 299 }
 300
 301 static void nfs_direct_pgio_init(struct nfs_pgio_header *hdr)
 302 {
 303         get_dreq(hdr->dreq);
 304 }
 305
 306 static const struct nfs_pgio_completion_ops nfs_direct_read_completion_ops = {
 307         .error_cleanup = nfs_read_sync_pgio_error,
 308         .init_hdr = nfs_direct_pgio_init,
 309         .completion = nfs_direct_read_completion,
 310 };
 311
 312 /*
 313  * For each rsize'd chunk of the user's buffer, dispatch an NFS READ
 314  * operation.  If nfs_readdata_alloc() or get_user_pages() fails,
 315  * bail and stop sending more reads.  Read length accounting is
 316  * handled automatically by nfs_direct_read_result().  Otherwise, if
 317  * no requests have been sent, just return an error.
 318  */
 319 static ssize_t nfs_direct_read_schedule_segment(struct nfs_pageio_descriptor *desc,
 320                                                 const struct iovec *iov,
 321                                                 loff_t pos, bool uio)
 322 {
 323         struct nfs_direct_req *dreq = desc->pg_dreq;
 324         struct nfs_open_context *ctx = dreq->ctx;
 325         struct inode *inode = ctx->dentry->d_inode;
 326         unsigned long user_addr = (unsigned long)iov->iov_base;
 327         size_t count = iov->iov_len;
 328         size_t rsize = NFS_SERVER(inode)->rsize;
 329         unsigned int pgbase;
 330         int result;
 331         ssize_t started = 0;
 332         struct page **pagevec = NULL;
 333         unsigned int npages;
 334
 335         do {
 336                 size_t bytes;
 337                 int i;
 338
 339                 pgbase = user_addr & ~PAGE_MASK;
 340                 bytes = min(max_t(size_t, rsize, PAGE_SIZE), count);
 341
 342                 result = -ENOMEM;
 343                 npages = nfs_page_array_len(pgbase, bytes);
 344                 if (!pagevec)
 345                         pagevec = kmalloc(npages * sizeof(struct page *),
 346                                           GFP_KERNEL);
 347                 if (!pagevec)
 348                         break;
 349                 if (uio) {
 350                         down_read(&current->mm->mmap_sem);
 351                         result = get_user_pages(current, current->mm, user_addr,
 352                                         npages, 1, 0, pagevec, NULL);
 353                         up_read(&current->mm->mmap_sem);
 354                         if (result < 0)
 355                                 break;
 356                 } else {
 357                         WARN_ON(npages != 1);
 358                         result = get_kernel_page(user_addr, 1, pagevec);
 359                         if (WARN_ON(result != 1))
 360                                 break;
 361                 }
 362
 363                 if ((unsigned)result < npages) {
 364                         bytes = result * PAGE_SIZE;
 365                         if (bytes <= pgbase) {
 366                                 nfs_direct_release_pages(pagevec, result);
 367                                 break;
 368                         }
 369                         bytes -= pgbase;
 370                         npages = result;
 371                 }
 372
 373                 for (i = 0; i < npages; i++) {
 374                         struct nfs_page *req;
 375                         unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase);
 376                         /* XXX do we need to do the eof zeroing found in async_filler? */
 377                         req = nfs_create_request(dreq->ctx, dreq->inode,
 378                                                  pagevec[i],
 379                                                  pgbase, req_len);
 380                         if (IS_ERR(req)) {
 381                                 result = PTR_ERR(req);
 382                                 break;
 383                         }
 384                         req->wb_index = pos >> PAGE_SHIFT;
 385                         req->wb_offset = pos & ~PAGE_MASK;
 386                         if (!nfs_pageio_add_request(desc, req)) {
 387                                 result = desc->pg_error;
 388                                 nfs_release_request(req);
 389                                 break;
 390                         }
 391                         pgbase = 0;
 392                         bytes -= req_len;
 393                         started += req_len;
 394                         user_addr += req_len;
 395                         pos += req_len;
 396                         count -= req_len;
 397                         dreq->bytes_left -= req_len;
 398                 }
 399                 /* The nfs_page now hold references to these pages */
 400                 nfs_direct_release_pages(pagevec, npages);
 401         } while (count != 0 && result >= 0);
 402
 403         kfree(pagevec);
 404
 405         if (started)
 406                 return started;
 407         return result < 0 ? (ssize_t) result : -EFAULT;
 408 }
 409
 410 static ssize_t nfs_direct_read_schedule_iovec(struct nfs_direct_req *dreq,
 411                                               const struct iovec *iov,
 412                                               unsigned long nr_segs,
 413                                               loff_t pos, bool uio)
 414 {
 415         struct nfs_pageio_descriptor desc;
 416         ssize_t result = -EINVAL;
 417         size_t requested_bytes = 0;
 418         unsigned long seg;
 419
 420         NFS_PROTO(dreq->inode)->read_pageio_init(&desc, dreq->inode,
 421                              &nfs_direct_read_completion_ops);
 422         get_dreq(dreq);
 423         desc.pg_dreq = dreq;
 424
 425         for (seg = 0; seg < nr_segs; seg++) {
 426                 const struct iovec *vec = &iov[seg];
 427                 result = nfs_direct_read_schedule_segment(&desc, vec, pos, uio);
 428                 if (result < 0)
 429                         break;
 430                 requested_bytes += result;
 431                 if ((size_t)result < vec->iov_len)
 432                         break;
 433                 pos += vec->iov_len;
 434         }
 435
 436         nfs_pageio_complete(&desc);
 437
 438         /*
 439          * If no bytes were started, return the error, and let the
 440          * generic layer handle the completion.
 441          */
 442         if (requested_bytes == 0) {
 443                 nfs_direct_req_release(dreq);
 444                 return result < 0 ? result : -EIO;
 445         }
 446
 447         if (put_dreq(dreq))
 448                 nfs_direct_complete(dreq, false);
 449         return 0;
 450 }
 451
 452 static ssize_t nfs_direct_read(struct kiocb *iocb, const struct iovec *iov,
 453                                unsigned long nr_segs, loff_t pos, bool uio)
 454 {
 455         ssize_t result = -ENOMEM;
 456         struct inode *inode = iocb->ki_filp->f_mapping->host;
 457         struct nfs_direct_req *dreq;
 458         struct nfs_lock_context *l_ctx;
 459
 460         dreq = nfs_direct_req_alloc();
 461         if (dreq == NULL)
 462                 goto out;
 463
 464         dreq->inode = inode;
 465         dreq->bytes_left = iov_length(iov, nr_segs);
 466         dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
 467         l_ctx = nfs_get_lock_context(dreq->ctx);
 468         if (IS_ERR(l_ctx)) {
 469                 result = PTR_ERR(l_ctx);
 470                 goto out_release;
 471         }
 472         dreq->l_ctx = l_ctx;
 473         if (!is_sync_kiocb(iocb))
 474                 dreq->iocb = iocb;
 475
 476         NFS_I(inode)->read_io += iov_length(iov, nr_segs);
 477         result = nfs_direct_read_schedule_iovec(dreq, iov, nr_segs, pos, uio);
 478         if (!result)
 479                 result = nfs_direct_wait(dreq);
 480 out_release:
 481         nfs_direct_req_release(dreq);
 482 out:
 483         return result;
 484 }
 485
 486 static void nfs_inode_dio_write_done(struct inode *inode)
 487 {
 488         nfs_zap_mapping(inode, inode->i_mapping);
 489         inode_dio_done(inode);
 490 }
 491
 492 #if IS_ENABLED(CONFIG_NFS_V3) || IS_ENABLED(CONFIG_NFS_V4)
 493 static void nfs_direct_write_reschedule(struct nfs_direct_req *dreq)
 494 {
 495         struct nfs_pageio_descriptor desc;
 496         struct nfs_page *req, *tmp;
 497         LIST_HEAD(reqs);
 498         struct nfs_commit_info cinfo;
 499         LIST_HEAD(failed);
 500
 501         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 502         pnfs_recover_commit_reqs(dreq->inode, &reqs, &cinfo);
 503         spin_lock(cinfo.lock);
 504         nfs_scan_commit_list(&cinfo.mds->list, &reqs, &cinfo, 0);
 505         spin_unlock(cinfo.lock);
 506
 507         dreq->count = 0;
 508         get_dreq(dreq);
 509
 510         NFS_PROTO(dreq->inode)->write_pageio_init(&desc, dreq->inode, FLUSH_STABLE,
 511                               &nfs_direct_write_completion_ops);
 512         desc.pg_dreq = dreq;
 513
 514         list_for_each_entry_safe(req, tmp, &reqs, wb_list) {
 515                 if (!nfs_pageio_add_request(&desc, req)) {
 516                         nfs_list_remove_request(req);
 517                         nfs_list_add_request(req, &failed);
 518                         spin_lock(cinfo.lock);
 519                         dreq->flags = 0;
 520                         dreq->error = -EIO;
 521                         spin_unlock(cinfo.lock);
 522                 }
 523                 nfs_release_request(req);
 524         }
 525         nfs_pageio_complete(&desc);
 526
 527         while (!list_empty(&failed)) {
 528                 req = nfs_list_entry(failed.next);
 529                 nfs_list_remove_request(req);
 530                 nfs_unlock_and_release_request(req);
 531         }
 532
 533         if (put_dreq(dreq))
 534                 nfs_direct_write_complete(dreq, dreq->inode);
 535 }
 536
 537 static void nfs_direct_commit_complete(struct nfs_commit_data *data)
 538 {
 539         struct nfs_direct_req *dreq = data->dreq;
 540         struct nfs_commit_info cinfo;
 541         struct nfs_page *req;
 542         int status = data->task.tk_status;
 543
 544         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 545         if (status < 0) {
 546                 dprintk("NFS: %5u commit failed with error %d.\n",
 547                         data->task.tk_pid, status);
 548                 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 549         } else if (memcmp(&dreq->verf, &data->verf, sizeof(data->verf))) {
 550                 dprintk("NFS: %5u commit verify failed\n", data->task.tk_pid);
 551                 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 552         }
 553
 554         dprintk("NFS: %5u commit returned %d\n", data->task.tk_pid, status);
 555         while (!list_empty(&data->pages)) {
 556                 req = nfs_list_entry(data->pages.next);
 557                 nfs_list_remove_request(req);
 558                 if (dreq->flags == NFS_ODIRECT_RESCHED_WRITES) {
 559                         /* Note the rewrite will go through mds */
 560                         nfs_mark_request_commit(req, NULL, &cinfo);
 561                 } else
 562                         nfs_release_request(req);
 563                 nfs_unlock_and_release_request(req);
 564         }
 565
 566         if (atomic_dec_and_test(&cinfo.mds->rpcs_out))
 567                 nfs_direct_write_complete(dreq, data->inode);
 568 }
 569
 570 static void nfs_direct_error_cleanup(struct nfs_inode *nfsi)
 571 {
 572         /* There is no lock to clear */
 573 }
 574
 575 static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops = {
 576         .completion = nfs_direct_commit_complete,
 577         .error_cleanup = nfs_direct_error_cleanup,
 578 };
 579
 580 static void nfs_direct_commit_schedule(struct nfs_direct_req *dreq)
 581 {
 582         int res;
 583         struct nfs_commit_info cinfo;
 584         LIST_HEAD(mds_list);
 585
 586         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 587         nfs_scan_commit(dreq->inode, &mds_list, &cinfo);
 588         res = nfs_generic_commit_list(dreq->inode, &mds_list, 0, &cinfo);
 589         if (res < 0) /* res == -ENOMEM */
 590                 nfs_direct_write_reschedule(dreq);
 591 }
 592
 593 static void nfs_direct_write_schedule_work(struct work_struct *work)
 594 {
 595         struct nfs_direct_req *dreq = container_of(work, struct nfs_direct_req, work);
 596         int flags = dreq->flags;
 597
 598         dreq->flags = 0;
 599         switch (flags) {
 600                 case NFS_ODIRECT_DO_COMMIT:
 601                         nfs_direct_commit_schedule(dreq);
 602                         break;
 603                 case NFS_ODIRECT_RESCHED_WRITES:
 604                         nfs_direct_write_reschedule(dreq);
 605                         break;
 606                 default:
 607                         nfs_inode_dio_write_done(dreq->inode);
 608                         nfs_direct_complete(dreq, true);
 609         }
 610 }
 611
 612 static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode)
 613 {
 614         schedule_work(&dreq->work); /* Calls nfs_direct_write_schedule_work */
 615 }
 616
 617 #else
 618 static void nfs_direct_write_schedule_work(struct work_struct *work)
 619 {
 620 }
 621
 622 static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode)
 623 {
 624         nfs_inode_dio_write_done(inode);
 625         nfs_direct_complete(dreq, true);
 626 }
 627 #endif
 628
 629 /*
 630  * NB: Return the value of the first error return code.  Subsequent
 631  *     errors after the first one are ignored.
 632  */
 633 /*
 634  * For each wsize'd chunk of the user's buffer, dispatch an NFS WRITE
 635  * operation.  If nfs_writedata_alloc() or get_user_pages() fails,
 636  * bail and stop sending more writes.  Write length accounting is
 637  * handled automatically by nfs_direct_write_result().  Otherwise, if
 638  * no requests have been sent, just return an error.
 639  */
 640 static ssize_t nfs_direct_write_schedule_segment(struct nfs_pageio_descriptor *desc,
 641                                                  const struct iovec *iov,
 642                                                  loff_t pos, bool uio)
 643 {
 644         struct nfs_direct_req *dreq = desc->pg_dreq;
 645         struct nfs_open_context *ctx = dreq->ctx;
 646         struct inode *inode = ctx->dentry->d_inode;
 647         unsigned long user_addr = (unsigned long)iov->iov_base;
 648         size_t count = iov->iov_len;
 649         size_t wsize = NFS_SERVER(inode)->wsize;
 650         unsigned int pgbase;
 651         int result;
 652         ssize_t started = 0;
 653         struct page **pagevec = NULL;
 654         unsigned int npages;
 655
 656         do {
 657                 size_t bytes;
 658                 int i;
 659
 660                 pgbase = user_addr & ~PAGE_MASK;
 661                 bytes = min(max_t(size_t, wsize, PAGE_SIZE), count);
 662
 663                 result = -ENOMEM;
 664                 npages = nfs_page_array_len(pgbase, bytes);
 665                 if (!pagevec)
 666                         pagevec = kmalloc(npages * sizeof(struct page *), GFP_KERNEL);
 667                 if (!pagevec)
 668                         break;
 669
 670                 if (uio) {
 671                         down_read(&current->mm->mmap_sem);
 672                         result = get_user_pages(current, current->mm, user_addr,
 673                                                 npages, 0, 0, pagevec, NULL);
 674                         up_read(&current->mm->mmap_sem);
 675                         if (result < 0)
 676                                 break;
 677                 } else {
 678                         WARN_ON(npages != 1);
 679                         result = get_kernel_page(user_addr, 0, pagevec);
 680                         if (WARN_ON(result != 1))
 681                                 break;
 682                 }
 683
 684                 if ((unsigned)result < npages) {
 685                         bytes = result * PAGE_SIZE;
 686                         if (bytes <= pgbase) {
 687                                 nfs_direct_release_pages(pagevec, result);
 688                                 break;
 689                         }
 690                         bytes -= pgbase;
 691                         npages = result;
 692                 }
 693
 694                 for (i = 0; i < npages; i++) {
 695                         struct nfs_page *req;
 696                         unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase);
 697
 698                         req = nfs_create_request(dreq->ctx, dreq->inode,
 699                                                  pagevec[i],
 700                                                  pgbase, req_len);
 701                         if (IS_ERR(req)) {
 702                                 result = PTR_ERR(req);
 703                                 break;
 704                         }
 705                         nfs_lock_request(req);
 706                         req->wb_index = pos >> PAGE_SHIFT;
 707                         req->wb_offset = pos & ~PAGE_MASK;
 708                         if (!nfs_pageio_add_request(desc, req)) {
 709                                 result = desc->pg_error;
 710                                 nfs_unlock_and_release_request(req);
 711                                 break;
 712                         }
 713                         pgbase = 0;
 714                         bytes -= req_len;
 715                         started += req_len;
 716                         user_addr += req_len;
 717                         pos += req_len;
 718                         count -= req_len;
 719                         dreq->bytes_left -= req_len;
 720                 }
 721                 /* The nfs_page now hold references to these pages */
 722                 nfs_direct_release_pages(pagevec, npages);
 723         } while (count != 0 && result >= 0);
 724
 725         kfree(pagevec);
 726
 727         if (started)
 728                 return started;
 729         return result < 0 ? (ssize_t) result : -EFAULT;
 730 }
 731
 732 static void nfs_direct_write_completion(struct nfs_pgio_header *hdr)
 733 {
 734         struct nfs_direct_req *dreq = hdr->dreq;
 735         struct nfs_commit_info cinfo;
 736         int bit = -1;
 737         struct nfs_page *req = nfs_list_entry(hdr->pages.next);
 738
 739         if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
 740                 goto out_put;
 741
 742         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 743
 744         spin_lock(&dreq->lock);
 745
 746         if (test_bit(NFS_IOHDR_ERROR, &hdr->flags)) {
 747                 dreq->flags = 0;
 748                 dreq->error = hdr->error;
 749         }
 750         if (dreq->error != 0)
 751                 bit = NFS_IOHDR_ERROR;
 752         else {
 753                 dreq->count += hdr->good_bytes;
 754                 if (test_bit(NFS_IOHDR_NEED_RESCHED, &hdr->flags)) {
 755                         dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 756                         bit = NFS_IOHDR_NEED_RESCHED;
 757                 } else if (test_bit(NFS_IOHDR_NEED_COMMIT, &hdr->flags)) {
 758                         if (dreq->flags == NFS_ODIRECT_RESCHED_WRITES)
 759                                 bit = NFS_IOHDR_NEED_RESCHED;
 760                         else if (dreq->flags == 0) {
 761                                 memcpy(&dreq->verf, hdr->verf,
 762                                        sizeof(dreq->verf));
 763                                 bit = NFS_IOHDR_NEED_COMMIT;
 764                                 dreq->flags = NFS_ODIRECT_DO_COMMIT;
 765                         } else if (dreq->flags == NFS_ODIRECT_DO_COMMIT) {
 766                                 if (memcmp(&dreq->verf, hdr->verf, sizeof(dreq->verf))) {
 767                                         dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 768                                         bit = NFS_IOHDR_NEED_RESCHED;
 769                                 } else
 770                                         bit = NFS_IOHDR_NEED_COMMIT;
 771                         }
 772                 }
 773         }
 774         spin_unlock(&dreq->lock);
 775
 776         while (!list_empty(&hdr->pages)) {
 777                 req = nfs_list_entry(hdr->pages.next);
 778                 nfs_list_remove_request(req);
 779                 switch (bit) {
 780                 case NFS_IOHDR_NEED_RESCHED:
 781                 case NFS_IOHDR_NEED_COMMIT:
 782                         kref_get(&req->wb_kref);
 783                         nfs_mark_request_commit(req, hdr->lseg, &cinfo);
 784                 }
 785                 nfs_unlock_and_release_request(req);
 786         }
 787
 788 out_put:
 789         if (put_dreq(dreq))
 790                 nfs_direct_write_complete(dreq, hdr->inode);
 791         hdr->release(hdr);
 792 }
 793
 794 static void nfs_write_sync_pgio_error(struct list_head *head)
 795 {
 796         struct nfs_page *req;
 797
 798         while (!list_empty(head)) {
 799                 req = nfs_list_entry(head->next);
 800                 nfs_list_remove_request(req);
 801                 nfs_unlock_and_release_request(req);
 802         }
 803 }
 804
 805 static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops = {
 806         .error_cleanup = nfs_write_sync_pgio_error,
 807         .init_hdr = nfs_direct_pgio_init,
 808         .completion = nfs_direct_write_completion,
 809 };
 810
 811 static ssize_t nfs_direct_write_schedule_iovec(struct nfs_direct_req *dreq,
 812                                                const struct iovec *iov,
 813                                                unsigned long nr_segs,
 814                                                loff_t pos, bool uio)
 815 {
 816         struct nfs_pageio_descriptor desc;
 817         struct inode *inode = dreq->inode;
 818         ssize_t result = 0;
 819         size_t requested_bytes = 0;
 820         unsigned long seg;
 821
 822         NFS_PROTO(inode)->write_pageio_init(&desc, inode, FLUSH_COND_STABLE,
 823                               &nfs_direct_write_completion_ops);
 824         desc.pg_dreq = dreq;
 825         get_dreq(dreq);
 826         atomic_inc(&inode->i_dio_count);
 827
 828         NFS_I(dreq->inode)->write_io += iov_length(iov, nr_segs);
 829         for (seg = 0; seg < nr_segs; seg++) {
 830                 const struct iovec *vec = &iov[seg];
 831                 result = nfs_direct_write_schedule_segment(&desc, vec, pos, uio);
 832                 if (result < 0)
 833                         break;
 834                 requested_bytes += result;
 835                 if ((size_t)result < vec->iov_len)
 836                         break;
 837                 pos += vec->iov_len;
 838         }
 839         nfs_pageio_complete(&desc);
 840
 841         /*
 842          * If no bytes were started, return the error, and let the
 843          * generic layer handle the completion.
 844          */
 845         if (requested_bytes == 0) {
 846                 inode_dio_done(inode);
 847                 nfs_direct_req_release(dreq);
 848                 return result < 0 ? result : -EIO;
 849         }
 850
 851         if (put_dreq(dreq))
 852                 nfs_direct_write_complete(dreq, dreq->inode);
 853         return 0;
 854 }
 855
 856 static ssize_t nfs_direct_write(struct kiocb *iocb, const struct iovec *iov,
 857                                 unsigned long nr_segs, loff_t pos,
 858                                 size_t count, bool uio)
 859 {
 860         ssize_t result = -ENOMEM;
 861         struct inode *inode = iocb->ki_filp->f_mapping->host;
 862         struct nfs_direct_req *dreq;
 863         struct nfs_lock_context *l_ctx;
 864
 865         dreq = nfs_direct_req_alloc();
 866         if (!dreq)
 867                 goto out;
 868
 869         dreq->inode = inode;
 870         dreq->bytes_left = count;
 871         dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
 872         l_ctx = nfs_get_lock_context(dreq->ctx);
 873         if (IS_ERR(l_ctx)) {
 874                 result = PTR_ERR(l_ctx);
 875                 goto out_release;
 876         }
 877         dreq->l_ctx = l_ctx;
 878         if (!is_sync_kiocb(iocb))
 879                 dreq->iocb = iocb;
 880
 881         result = nfs_direct_write_schedule_iovec(dreq, iov, nr_segs, pos, uio);
 882         if (!result)
 883                 result = nfs_direct_wait(dreq);
 884 out_release:
 885         nfs_direct_req_release(dreq);
 886 out:
 887         return result;
 888 }
 889
 890 /**
 891  * nfs_file_direct_read - file direct read operation for NFS files
 892  * @iocb: target I/O control block
 893  * @iov: vector of user buffers into which to read data
 894  * @nr_segs: size of iov vector
 895  * @pos: byte offset in file where reading starts
 896  *
 897  * We use this function for direct reads instead of calling
 898  * generic_file_aio_read() in order to avoid gfar's check to see if
 899  * the request starts before the end of the file.  For that check
 900  * to work, we must generate a GETATTR before each direct read, and
 901  * even then there is a window between the GETATTR and the subsequent
 902  * READ where the file size could change.  Our preference is simply
 903  * to do all reads the application wants, and the server will take
 904  * care of managing the end of file boundary.
 905  *
 906  * This function also eliminates unnecessarily updating the file's
 907  * atime locally, as the NFS server sets the file's atime, and this
 908  * client must read the updated atime from the server back into its
 909  * cache.
 910  */
 911 ssize_t nfs_file_direct_read(struct kiocb *iocb, const struct iovec *iov,
 912                                 unsigned long nr_segs, loff_t pos, bool uio)
 913 {
 914         ssize_t retval = -EINVAL;
 915         struct file *file = iocb->ki_filp;
 916         struct address_space *mapping = file->f_mapping;
 917         size_t count;
 918
 919         count = iov_length(iov, nr_segs);
 920         nfs_add_stats(mapping->host, NFSIOS_DIRECTREADBYTES, count);
 921
 922         dfprintk(FILE, "NFS: direct read(%pD2, %zd@%Ld)\n",
 923                 file, count, (long long) pos);
 924
 925         retval = 0;
 926         if (!count)
 927                 goto out;
 928
 929         retval = nfs_sync_mapping(mapping);
 930         if (retval)
 931                 goto out;
 932
 933         task_io_account_read(count);
 934
 935         retval = nfs_direct_read(iocb, iov, nr_segs, pos, uio);
 936         if (retval > 0)
 937                 iocb->ki_pos = pos + retval;
 938
 939 out:
 940         return retval;
 941 }
 942
 943 /**
 944  * nfs_file_direct_write - file direct write operation for NFS files
 945  * @iocb: target I/O control block
 946  * @iov: vector of user buffers from which to write data
 947  * @nr_segs: size of iov vector
 948  * @pos: byte offset in file where writing starts
 949  *
 950  * We use this function for direct writes instead of calling
 951  * generic_file_aio_write() in order to avoid taking the inode
 952  * semaphore and updating the i_size.  The NFS server will set
 953  * the new i_size and this client must read the updated size
 954  * back into its cache.  We let the server do generic write
 955  * parameter checking and report problems.
 956  *
 957  * We eliminate local atime updates, see direct read above.
 958  *
 959  * We avoid unnecessary page cache invalidations for normal cached
 960  * readers of this file.
 961  *
 962  * Note that O_APPEND is not supported for NFS direct writes, as there
 963  * is no atomic O_APPEND write facility in the NFS protocol.
 964  */
 965 ssize_t nfs_file_direct_write(struct kiocb *iocb, const struct iovec *iov,
 966                                 unsigned long nr_segs, loff_t pos, bool uio)
 967 {
 968         ssize_t retval = -EINVAL;
 969         struct file *file = iocb->ki_filp;
 970         struct address_space *mapping = file->f_mapping;
 971         size_t count;
 972
 973         count = iov_length(iov, nr_segs);
 974         nfs_add_stats(mapping->host, NFSIOS_DIRECTWRITTENBYTES, count);
 975
 976         dfprintk(FILE, "NFS: direct write(%pD2, %zd@%Ld)\n",
 977                 file, count, (long long) pos);
 978
 979         retval = generic_write_checks(file, &pos, &count, 0);
 980         if (retval)
 981                 goto out;
 982
 983         retval = -EINVAL;
 984         if ((ssize_t) count < 0)
 985                 goto out;
 986         retval = 0;
 987         if (!count)
 988                 goto out;
 989
 990         retval = nfs_sync_mapping(mapping);
 991         if (retval)
 992                 goto out;
 993
 994         task_io_account_write(count);
 995
 996         retval = nfs_direct_write(iocb, iov, nr_segs, pos, count, uio);
 997         if (retval > 0) {
 998                 struct inode *inode = mapping->host;
 999
1000                 iocb->ki_pos = pos + retval;
1001                 spin_lock(&inode->i_lock);
1002                 if (i_size_read(inode) < iocb->ki_pos)
1003                         i_size_write(inode, iocb->ki_pos);
1004                 spin_unlock(&inode->i_lock);
1005         }
1006 out:
1007         return retval;
1008 }
1009
1010 /**
1011  * nfs_init_directcache - create a slab cache for nfs_direct_req structures
1012  *
1013  */
1014 int __init nfs_init_directcache(void)
1015 {
1016         nfs_direct_cachep = kmem_cache_create("nfs_direct_cache",
1017                                                 sizeof(struct nfs_direct_req),
1018                                                 0, (SLAB_RECLAIM_ACCOUNT|
1019                                                         SLAB_MEM_SPREAD),
1020                                                 NULL);
1021         if (nfs_direct_cachep == NULL)
1022                 return -ENOMEM;
1023
1024         return 0;
1025 }
1026
1027 /**
1028  * nfs_destroy_directcache - destroy the slab cache for nfs_direct_req structures
1029  *
1030  */
1031 void nfs_destroy_directcache(void)
1032 {
1033         kmem_cache_destroy(nfs_direct_cachep);
1034 }