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 #define NFS_ODIRECT_MARK_DIRTY          (4)     /* mark read pages dirty */
  94         struct nfs_writeverf    verf;           /* unstable write verifier */
  95 };
  96
  97 static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops;
  98 static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops;
  99 static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode);
 100 static void nfs_direct_write_schedule_work(struct work_struct *work);
 101
 102 static inline void get_dreq(struct nfs_direct_req *dreq)
 103 {
 104         atomic_inc(&dreq->io_count);
 105 }
 106
 107 static inline int put_dreq(struct nfs_direct_req *dreq)
 108 {
 109         return atomic_dec_and_test(&dreq->io_count);
 110 }
 111
 112 /**
 113  * nfs_direct_IO - NFS address space operation for direct I/O
 114  * @rw: direction (read or write)
 115  * @iocb: target I/O control block
 116  * @iter: array of vectors that define I/O buffer
 117  * @pos: offset in file to begin the operation
 118  * @nr_segs: size of iovec array
 119  *
 120  * The presence of this routine in the address space ops vector means
 121  * the NFS client supports direct I/O. However, we shunt off direct
 122  * read and write requests before the VFS gets them, so this method
 123  * should never be called.
 124  */
 125 ssize_t nfs_direct_IO(int rw, struct kiocb *iocb, struct iov_iter *iter,
 126                       loff_t pos)
 127 {
 128         dprintk("NFS: nfs_direct_IO (%s) off/no(%Ld/%lu) EINVAL\n",
 129                         iocb->ki_filp->f_path.dentry->d_name.name,
 130                         (long long) pos, iter->nr_segs);
 131
 132         return -EINVAL;
 133 }
 134
 135 static void nfs_direct_release_pages(struct page **pages, unsigned int npages)
 136 {
 137         unsigned int i;
 138         for (i = 0; i < npages; i++)
 139                 page_cache_release(pages[i]);
 140 }
 141
 142 void nfs_init_cinfo_from_dreq(struct nfs_commit_info *cinfo,
 143                               struct nfs_direct_req *dreq)
 144 {
 145         cinfo->lock = &dreq->lock;
 146         cinfo->mds = &dreq->mds_cinfo;
 147         cinfo->ds = &dreq->ds_cinfo;
 148         cinfo->dreq = dreq;
 149         cinfo->completion_ops = &nfs_direct_commit_completion_ops;
 150 }
 151
 152 static inline struct nfs_direct_req *nfs_direct_req_alloc(void)
 153 {
 154         struct nfs_direct_req *dreq;
 155
 156         dreq = kmem_cache_zalloc(nfs_direct_cachep, GFP_KERNEL);
 157         if (!dreq)
 158                 return NULL;
 159
 160         kref_init(&dreq->kref);
 161         kref_get(&dreq->kref);
 162         init_completion(&dreq->completion);
 163         INIT_LIST_HEAD(&dreq->mds_cinfo.list);
 164         INIT_WORK(&dreq->work, nfs_direct_write_schedule_work);
 165         spin_lock_init(&dreq->lock);
 166
 167         return dreq;
 168 }
 169
 170 static void nfs_direct_req_free(struct kref *kref)
 171 {
 172         struct nfs_direct_req *dreq = container_of(kref, struct nfs_direct_req, kref);
 173
 174         if (dreq->l_ctx != NULL)
 175                 nfs_put_lock_context(dreq->l_ctx);
 176         if (dreq->ctx != NULL)
 177                 put_nfs_open_context(dreq->ctx);
 178         kmem_cache_free(nfs_direct_cachep, dreq);
 179 }
 180
 181 static void nfs_direct_req_release(struct nfs_direct_req *dreq)
 182 {
 183         kref_put(&dreq->kref, nfs_direct_req_free);
 184 }
 185
 186 ssize_t nfs_dreq_bytes_left(struct nfs_direct_req *dreq)
 187 {
 188         return dreq->bytes_left;
 189 }
 190 EXPORT_SYMBOL_GPL(nfs_dreq_bytes_left);
 191
 192 /*
 193  * Collects and returns the final error value/byte-count.
 194  */
 195 static ssize_t nfs_direct_wait(struct nfs_direct_req *dreq)
 196 {
 197         ssize_t result = -EIOCBQUEUED;
 198
 199         /* Async requests don't wait here */
 200         if (dreq->iocb)
 201                 goto out;
 202
 203         result = wait_for_completion_killable(&dreq->completion);
 204
 205         if (!result)
 206                 result = dreq->error;
 207         if (!result)
 208                 result = dreq->count;
 209
 210 out:
 211         return (ssize_t) result;
 212 }
 213
 214 /*
 215  * Synchronous I/O uses a stack-allocated iocb.  Thus we can't trust
 216  * the iocb is still valid here if this is a synchronous request.
 217  */
 218 static void nfs_direct_complete(struct nfs_direct_req *dreq)
 219 {
 220         if (dreq->iocb) {
 221                 long res = (long) dreq->error;
 222                 if (!res)
 223                         res = (long) dreq->count;
 224                 aio_complete(dreq->iocb, res, 0);
 225         }
 226         complete_all(&dreq->completion);
 227
 228         nfs_direct_req_release(dreq);
 229 }
 230
 231 static void nfs_direct_readpage_release(struct nfs_page *req)
 232 {
 233         dprintk("NFS: direct read done (%s/%lld %d@%lld)\n",
 234                 req->wb_context->dentry->d_inode->i_sb->s_id,
 235                 (long long)NFS_FILEID(req->wb_context->dentry->d_inode),
 236                 req->wb_bytes,
 237                 (long long)req_offset(req));
 238         nfs_release_request(req);
 239 }
 240
 241 static void nfs_direct_read_completion(struct nfs_pgio_header *hdr)
 242 {
 243         unsigned long bytes = 0;
 244         struct nfs_direct_req *dreq = hdr->dreq;
 245
 246         if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
 247                 goto out_put;
 248
 249         spin_lock(&dreq->lock);
 250         if (test_bit(NFS_IOHDR_ERROR, &hdr->flags) && (hdr->good_bytes == 0))
 251                 dreq->error = hdr->error;
 252         else
 253                 dreq->count += hdr->good_bytes;
 254         spin_unlock(&dreq->lock);
 255
 256         while (!list_empty(&hdr->pages)) {
 257                 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
 258                 struct page *page = req->wb_page;
 259
 260                 if ((dreq->flags & NFS_ODIRECT_MARK_DIRTY) &&
 261                     !PageCompound(page) && bytes < hdr->good_bytes)
 262                         set_page_dirty(page);
 263                 bytes += req->wb_bytes;
 264                 nfs_list_remove_request(req);
 265                 nfs_direct_readpage_release(req);
 266         }
 267 out_put:
 268         if (put_dreq(dreq))
 269                 nfs_direct_complete(dreq);
 270         hdr->release(hdr);
 271 }
 272
 273 static void nfs_read_sync_pgio_error(struct list_head *head)
 274 {
 275         struct nfs_page *req;
 276
 277         while (!list_empty(head)) {
 278                 req = nfs_list_entry(head->next);
 279                 nfs_list_remove_request(req);
 280                 nfs_release_request(req);
 281         }
 282 }
 283
 284 static void nfs_direct_pgio_init(struct nfs_pgio_header *hdr)
 285 {
 286         get_dreq(hdr->dreq);
 287 }
 288
 289 static const struct nfs_pgio_completion_ops nfs_direct_read_completion_ops = {
 290         .error_cleanup = nfs_read_sync_pgio_error,
 291         .init_hdr = nfs_direct_pgio_init,
 292         .completion = nfs_direct_read_completion,
 293 };
 294
 295 /*
 296  * For each rsize'd chunk of the user's buffer, dispatch an NFS READ
 297  * operation.  If nfs_readdata_alloc() or get_user_pages() fails,
 298  * bail and stop sending more reads.  Read length accounting is
 299  * handled automatically by nfs_direct_read_result().  Otherwise, if
 300  * no requests have been sent, just return an error.
 301  */
 302 static ssize_t nfs_direct_read_schedule_segment(struct nfs_pageio_descriptor *desc,
 303                                                 const struct iovec *iov,
 304                                                 loff_t pos)
 305 {
 306         struct nfs_direct_req *dreq = desc->pg_dreq;
 307         struct nfs_open_context *ctx = dreq->ctx;
 308         struct inode *inode = ctx->dentry->d_inode;
 309         unsigned long user_addr = (unsigned long)iov->iov_base;
 310         size_t count = iov->iov_len;
 311         size_t rsize = NFS_SERVER(inode)->rsize;
 312         unsigned int pgbase;
 313         int result;
 314         ssize_t started = 0;
 315         struct page **pagevec = NULL;
 316         unsigned int npages;
 317
 318         do {
 319                 size_t bytes;
 320                 int i;
 321
 322                 pgbase = user_addr & ~PAGE_MASK;
 323                 bytes = min(max_t(size_t, rsize, PAGE_SIZE), count);
 324
 325                 result = -ENOMEM;
 326                 npages = nfs_page_array_len(pgbase, bytes);
 327                 if (!pagevec)
 328                         pagevec = kmalloc(npages * sizeof(struct page *),
 329                                           GFP_KERNEL);
 330                 if (!pagevec)
 331                         break;
 332                 down_read(&current->mm->mmap_sem);
 333                 result = get_user_pages(current, current->mm, user_addr,
 334                                         npages, 1, 0, pagevec, NULL);
 335                 up_read(&current->mm->mmap_sem);
 336                 if (result < 0)
 337                         break;
 338                 if ((unsigned)result < npages) {
 339                         bytes = result * PAGE_SIZE;
 340                         if (bytes <= pgbase) {
 341                                 nfs_direct_release_pages(pagevec, result);
 342                                 break;
 343                         }
 344                         bytes -= pgbase;
 345                         npages = result;
 346                 }
 347
 348                 for (i = 0; i < npages; i++) {
 349                         struct nfs_page *req;
 350                         unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase);
 351                         /* XXX do we need to do the eof zeroing found in async_filler? */
 352                         req = nfs_create_request(dreq->ctx, dreq->inode,
 353                                                  pagevec[i],
 354                                                  pgbase, req_len);
 355                         if (IS_ERR(req)) {
 356                                 result = PTR_ERR(req);
 357                                 break;
 358                         }
 359                         req->wb_index = pos >> PAGE_SHIFT;
 360                         req->wb_offset = pos & ~PAGE_MASK;
 361                         if (!nfs_pageio_add_request(desc, req)) {
 362                                 result = desc->pg_error;
 363                                 nfs_release_request(req);
 364                                 break;
 365                         }
 366                         pgbase = 0;
 367                         bytes -= req_len;
 368                         started += req_len;
 369                         user_addr += req_len;
 370                         pos += req_len;
 371                         count -= req_len;
 372                         dreq->bytes_left -= req_len;
 373                 }
 374                 /* The nfs_page now hold references to these pages */
 375                 nfs_direct_release_pages(pagevec, npages);
 376         } while (count != 0 && result >= 0);
 377
 378         kfree(pagevec);
 379
 380         if (started)
 381                 return started;
 382         return result < 0 ? (ssize_t) result : -EFAULT;
 383 }
 384
 385 static ssize_t nfs_direct_do_schedule_read_iovec(
 386                 struct nfs_pageio_descriptor *desc, const struct iovec *iov,
 387                 unsigned long nr_segs, loff_t pos)
 388 {
 389         ssize_t result = -EINVAL;
 390         size_t requested_bytes = 0;
 391         unsigned long seg;
 392
 393         for (seg = 0; seg < nr_segs; seg++) {
 394                 const struct iovec *vec = &iov[seg];
 395                 result = nfs_direct_read_schedule_segment(desc, vec, pos);
 396                 if (result < 0)
 397                         break;
 398                 requested_bytes += result;
 399                 if ((size_t)result < vec->iov_len)
 400                         break;
 401                 pos += vec->iov_len;
 402         }
 403         if (requested_bytes)
 404                 return requested_bytes;
 405
 406         return result < 0 ? result : -EIO;
 407 }
 408
 409 #ifdef CONFIG_BLOCK
 410 static ssize_t nfs_direct_do_schedule_read_bvec(
 411                 struct nfs_pageio_descriptor *desc,
 412                 struct bio_vec *bvec, unsigned long nr_segs, loff_t pos)
 413 {
 414         struct nfs_direct_req *dreq = desc->pg_dreq;
 415         struct nfs_open_context *ctx = dreq->ctx;
 416         struct inode *inode = ctx->dentry->d_inode;
 417         ssize_t result = -EINVAL;
 418         size_t requested_bytes = 0;
 419         unsigned long seg;
 420         struct nfs_page *req;
 421         unsigned int req_len;
 422
 423         for (seg = 0; seg < nr_segs; seg++) {
 424                 result = -EIO;
 425                 req_len = bvec[seg].bv_len;
 426                 req = nfs_create_request(ctx, inode,
 427                                          bvec[seg].bv_page,
 428                                          bvec[seg].bv_offset, req_len);
 429                 if (IS_ERR(req)) {
 430                         result = PTR_ERR(req);
 431                         break;
 432                 }
 433                 req->wb_index = pos >> PAGE_SHIFT;
 434                 req->wb_offset = pos & ~PAGE_MASK;
 435                 if (!nfs_pageio_add_request(desc, req)) {
 436                         result = desc->pg_error;
 437                         nfs_release_request(req);
 438                         break;
 439                 }
 440                 requested_bytes += req_len;
 441                 pos += req_len;
 442         }
 443
 444         if (requested_bytes)
 445                 return requested_bytes;
 446
 447         return result < 0 ? result : -EIO;
 448 }
 449 #endif /* CONFIG_BLOCK */
 450
 451 static ssize_t nfs_direct_read_schedule(struct nfs_direct_req *dreq,
 452                                         struct iov_iter *iter, loff_t pos)
 453 {
 454         struct nfs_pageio_descriptor desc;
 455         ssize_t result;
 456
 457         NFS_PROTO(dreq->inode)->read_pageio_init(&desc, dreq->inode,
 458                              &nfs_direct_read_completion_ops);
 459         get_dreq(dreq);
 460         desc.pg_dreq = dreq;
 461
 462         if (iov_iter_has_iovec(iter)) {
 463                 result = nfs_direct_do_schedule_read_iovec(&desc,
 464                                 iov_iter_iovec(iter), iter->nr_segs, pos);
 465 #ifdef CONFIG_BLOCK
 466         } else if (iov_iter_has_bvec(iter)) {
 467                 result = nfs_direct_do_schedule_read_bvec(&desc,
 468                                 iov_iter_bvec(iter), iter->nr_segs, pos);
 469 #endif
 470         } else
 471                 BUG();
 472
 473         nfs_pageio_complete(&desc);
 474
 475         /*
 476          * If no bytes were started, return the error, and let the
 477          * generic layer handle the completion.
 478          */
 479         if (result < 0) {
 480                 nfs_direct_req_release(dreq);
 481                 return result;
 482         }
 483
 484         if (put_dreq(dreq))
 485                 nfs_direct_complete(dreq);
 486         return 0;
 487 }
 488
 489 static ssize_t nfs_direct_read(struct kiocb *iocb, struct iov_iter *iter,
 490                                loff_t pos)
 491 {
 492         ssize_t result = -ENOMEM;
 493         struct inode *inode = iocb->ki_filp->f_mapping->host;
 494         struct nfs_direct_req *dreq;
 495         struct nfs_lock_context *l_ctx;
 496
 497         dreq = nfs_direct_req_alloc();
 498         if (dreq == NULL)
 499                 goto out;
 500
 501         dreq->inode = inode;
 502         dreq->bytes_left = iov_iter_count(iter);
 503         dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
 504         l_ctx = nfs_get_lock_context(dreq->ctx);
 505         if (IS_ERR(l_ctx)) {
 506                 result = PTR_ERR(l_ctx);
 507                 goto out_release;
 508         }
 509         dreq->l_ctx = l_ctx;
 510         if (!is_sync_kiocb(iocb))
 511                 dreq->iocb = iocb;
 512
 513         NFS_I(inode)->read_io += iov_iter_count(iter);
 514         result = nfs_direct_read_schedule(dreq, iter, pos);
 515         if (!result)
 516                 result = nfs_direct_wait(dreq);
 517 out_release:
 518         nfs_direct_req_release(dreq);
 519 out:
 520         return result;
 521 }
 522
 523 static void nfs_inode_dio_write_done(struct inode *inode)
 524 {
 525         nfs_zap_mapping(inode, inode->i_mapping);
 526         inode_dio_done(inode);
 527 }
 528
 529 #if IS_ENABLED(CONFIG_NFS_V3) || IS_ENABLED(CONFIG_NFS_V4)
 530 static void nfs_direct_write_reschedule(struct nfs_direct_req *dreq)
 531 {
 532         struct nfs_pageio_descriptor desc;
 533         struct nfs_page *req, *tmp;
 534         LIST_HEAD(reqs);
 535         struct nfs_commit_info cinfo;
 536         LIST_HEAD(failed);
 537
 538         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 539         pnfs_recover_commit_reqs(dreq->inode, &reqs, &cinfo);
 540         spin_lock(cinfo.lock);
 541         nfs_scan_commit_list(&cinfo.mds->list, &reqs, &cinfo, 0);
 542         spin_unlock(cinfo.lock);
 543
 544         dreq->count = 0;
 545         get_dreq(dreq);
 546
 547         NFS_PROTO(dreq->inode)->write_pageio_init(&desc, dreq->inode, FLUSH_STABLE,
 548                               &nfs_direct_write_completion_ops);
 549         desc.pg_dreq = dreq;
 550
 551         list_for_each_entry_safe(req, tmp, &reqs, wb_list) {
 552                 if (!nfs_pageio_add_request(&desc, req)) {
 553                         nfs_list_remove_request(req);
 554                         nfs_list_add_request(req, &failed);
 555                         spin_lock(cinfo.lock);
 556                         dreq->flags = 0;
 557                         dreq->error = -EIO;
 558                         spin_unlock(cinfo.lock);
 559                 }
 560                 nfs_release_request(req);
 561         }
 562         nfs_pageio_complete(&desc);
 563
 564         while (!list_empty(&failed)) {
 565                 req = nfs_list_entry(failed.next);
 566                 nfs_list_remove_request(req);
 567                 nfs_unlock_and_release_request(req);
 568         }
 569
 570         if (put_dreq(dreq))
 571                 nfs_direct_write_complete(dreq, dreq->inode);
 572 }
 573
 574 static void nfs_direct_commit_complete(struct nfs_commit_data *data)
 575 {
 576         struct nfs_direct_req *dreq = data->dreq;
 577         struct nfs_commit_info cinfo;
 578         struct nfs_page *req;
 579         int status = data->task.tk_status;
 580
 581         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 582         if (status < 0) {
 583                 dprintk("NFS: %5u commit failed with error %d.\n",
 584                         data->task.tk_pid, status);
 585                 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 586         } else if (memcmp(&dreq->verf, &data->verf, sizeof(data->verf))) {
 587                 dprintk("NFS: %5u commit verify failed\n", data->task.tk_pid);
 588                 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 589         }
 590
 591         dprintk("NFS: %5u commit returned %d\n", data->task.tk_pid, status);
 592         while (!list_empty(&data->pages)) {
 593                 req = nfs_list_entry(data->pages.next);
 594                 nfs_list_remove_request(req);
 595                 if (dreq->flags == NFS_ODIRECT_RESCHED_WRITES) {
 596                         /* Note the rewrite will go through mds */
 597                         nfs_mark_request_commit(req, NULL, &cinfo);
 598                 } else
 599                         nfs_release_request(req);
 600                 nfs_unlock_and_release_request(req);
 601         }
 602
 603         if (atomic_dec_and_test(&cinfo.mds->rpcs_out))
 604                 nfs_direct_write_complete(dreq, data->inode);
 605 }
 606
 607 static void nfs_direct_error_cleanup(struct nfs_inode *nfsi)
 608 {
 609         /* There is no lock to clear */
 610 }
 611
 612 static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops = {
 613         .completion = nfs_direct_commit_complete,
 614         .error_cleanup = nfs_direct_error_cleanup,
 615 };
 616
 617 static void nfs_direct_commit_schedule(struct nfs_direct_req *dreq)
 618 {
 619         int res;
 620         struct nfs_commit_info cinfo;
 621         LIST_HEAD(mds_list);
 622
 623         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 624         nfs_scan_commit(dreq->inode, &mds_list, &cinfo);
 625         res = nfs_generic_commit_list(dreq->inode, &mds_list, 0, &cinfo);
 626         if (res < 0) /* res == -ENOMEM */
 627                 nfs_direct_write_reschedule(dreq);
 628 }
 629
 630 static void nfs_direct_write_schedule_work(struct work_struct *work)
 631 {
 632         struct nfs_direct_req *dreq = container_of(work, struct nfs_direct_req, work);
 633         int flags = dreq->flags;
 634
 635         dreq->flags = 0;
 636         switch (flags) {
 637                 case NFS_ODIRECT_DO_COMMIT:
 638                         nfs_direct_commit_schedule(dreq);
 639                         break;
 640                 case NFS_ODIRECT_RESCHED_WRITES:
 641                         nfs_direct_write_reschedule(dreq);
 642                         break;
 643                 default:
 644                         nfs_inode_dio_write_done(dreq->inode);
 645                         nfs_direct_complete(dreq);
 646         }
 647 }
 648
 649 static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode)
 650 {
 651         schedule_work(&dreq->work); /* Calls nfs_direct_write_schedule_work */
 652 }
 653
 654 #else
 655 static void nfs_direct_write_schedule_work(struct work_struct *work)
 656 {
 657 }
 658
 659 static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode)
 660 {
 661         nfs_inode_dio_write_done(inode);
 662         nfs_direct_complete(dreq);
 663 }
 664 #endif
 665
 666 /*
 667  * NB: Return the value of the first error return code.  Subsequent
 668  *     errors after the first one are ignored.
 669  */
 670 /*
 671  * For each wsize'd chunk of the user's buffer, dispatch an NFS WRITE
 672  * operation.  If nfs_writedata_alloc() or get_user_pages() fails,
 673  * bail and stop sending more writes.  Write length accounting is
 674  * handled automatically by nfs_direct_write_result().  Otherwise, if
 675  * no requests have been sent, just return an error.
 676  */
 677 static ssize_t nfs_direct_write_schedule_segment(struct nfs_pageio_descriptor *desc,
 678                                                  const struct iovec *iov,
 679                                                  loff_t pos)
 680 {
 681         struct nfs_direct_req *dreq = desc->pg_dreq;
 682         struct nfs_open_context *ctx = dreq->ctx;
 683         struct inode *inode = ctx->dentry->d_inode;
 684         unsigned long user_addr = (unsigned long)iov->iov_base;
 685         size_t count = iov->iov_len;
 686         size_t wsize = NFS_SERVER(inode)->wsize;
 687         unsigned int pgbase;
 688         int result;
 689         ssize_t started = 0;
 690         struct page **pagevec = NULL;
 691         unsigned int npages;
 692
 693         do {
 694                 size_t bytes;
 695                 int i;
 696
 697                 pgbase = user_addr & ~PAGE_MASK;
 698                 bytes = min(max_t(size_t, wsize, PAGE_SIZE), count);
 699
 700                 result = -ENOMEM;
 701                 npages = nfs_page_array_len(pgbase, bytes);
 702                 if (!pagevec)
 703                         pagevec = kmalloc(npages * sizeof(struct page *), GFP_KERNEL);
 704                 if (!pagevec)
 705                         break;
 706
 707                 down_read(&current->mm->mmap_sem);
 708                 result = get_user_pages(current, current->mm, user_addr,
 709                                         npages, 0, 0, pagevec, NULL);
 710                 up_read(&current->mm->mmap_sem);
 711                 if (result < 0)
 712                         break;
 713
 714                 if ((unsigned)result < npages) {
 715                         bytes = result * PAGE_SIZE;
 716                         if (bytes <= pgbase) {
 717                                 nfs_direct_release_pages(pagevec, result);
 718                                 break;
 719                         }
 720                         bytes -= pgbase;
 721                         npages = result;
 722                 }
 723
 724                 for (i = 0; i < npages; i++) {
 725                         struct nfs_page *req;
 726                         unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase);
 727
 728                         req = nfs_create_request(dreq->ctx, dreq->inode,
 729                                                  pagevec[i],
 730                                                  pgbase, req_len);
 731                         if (IS_ERR(req)) {
 732                                 result = PTR_ERR(req);
 733                                 break;
 734                         }
 735                         nfs_lock_request(req);
 736                         req->wb_index = pos >> PAGE_SHIFT;
 737                         req->wb_offset = pos & ~PAGE_MASK;
 738                         if (!nfs_pageio_add_request(desc, req)) {
 739                                 result = desc->pg_error;
 740                                 nfs_unlock_and_release_request(req);
 741                                 break;
 742                         }
 743                         pgbase = 0;
 744                         bytes -= req_len;
 745                         started += req_len;
 746                         user_addr += req_len;
 747                         pos += req_len;
 748                         count -= req_len;
 749                         dreq->bytes_left -= req_len;
 750                 }
 751                 /* The nfs_page now hold references to these pages */
 752                 nfs_direct_release_pages(pagevec, npages);
 753         } while (count != 0 && result >= 0);
 754
 755         kfree(pagevec);
 756
 757         if (started)
 758                 return started;
 759         return result < 0 ? (ssize_t) result : -EFAULT;
 760 }
 761
 762 static void nfs_direct_write_completion(struct nfs_pgio_header *hdr)
 763 {
 764         struct nfs_direct_req *dreq = hdr->dreq;
 765         struct nfs_commit_info cinfo;
 766         int bit = -1;
 767         struct nfs_page *req = nfs_list_entry(hdr->pages.next);
 768
 769         if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
 770                 goto out_put;
 771
 772         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 773
 774         spin_lock(&dreq->lock);
 775
 776         if (test_bit(NFS_IOHDR_ERROR, &hdr->flags)) {
 777                 dreq->flags = 0;
 778                 dreq->error = hdr->error;
 779         }
 780         if (dreq->error != 0)
 781                 bit = NFS_IOHDR_ERROR;
 782         else {
 783                 dreq->count += hdr->good_bytes;
 784                 if (test_bit(NFS_IOHDR_NEED_RESCHED, &hdr->flags)) {
 785                         dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 786                         bit = NFS_IOHDR_NEED_RESCHED;
 787                 } else if (test_bit(NFS_IOHDR_NEED_COMMIT, &hdr->flags)) {
 788                         if (dreq->flags == NFS_ODIRECT_RESCHED_WRITES)
 789                                 bit = NFS_IOHDR_NEED_RESCHED;
 790                         else if (dreq->flags == 0) {
 791                                 memcpy(&dreq->verf, hdr->verf,
 792                                        sizeof(dreq->verf));
 793                                 bit = NFS_IOHDR_NEED_COMMIT;
 794                                 dreq->flags = NFS_ODIRECT_DO_COMMIT;
 795                         } else if (dreq->flags == NFS_ODIRECT_DO_COMMIT) {
 796                                 if (memcmp(&dreq->verf, hdr->verf, sizeof(dreq->verf))) {
 797                                         dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 798                                         bit = NFS_IOHDR_NEED_RESCHED;
 799                                 } else
 800                                         bit = NFS_IOHDR_NEED_COMMIT;
 801                         }
 802                 }
 803         }
 804         spin_unlock(&dreq->lock);
 805
 806         while (!list_empty(&hdr->pages)) {
 807                 req = nfs_list_entry(hdr->pages.next);
 808                 nfs_list_remove_request(req);
 809                 switch (bit) {
 810                 case NFS_IOHDR_NEED_RESCHED:
 811                 case NFS_IOHDR_NEED_COMMIT:
 812                         kref_get(&req->wb_kref);
 813                         nfs_mark_request_commit(req, hdr->lseg, &cinfo);
 814                 }
 815                 nfs_unlock_and_release_request(req);
 816         }
 817
 818 out_put:
 819         if (put_dreq(dreq))
 820                 nfs_direct_write_complete(dreq, hdr->inode);
 821         hdr->release(hdr);
 822 }
 823
 824 static void nfs_write_sync_pgio_error(struct list_head *head)
 825 {
 826         struct nfs_page *req;
 827
 828         while (!list_empty(head)) {
 829                 req = nfs_list_entry(head->next);
 830                 nfs_list_remove_request(req);
 831                 nfs_unlock_and_release_request(req);
 832         }
 833 }
 834
 835 static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops = {
 836         .error_cleanup = nfs_write_sync_pgio_error,
 837         .init_hdr = nfs_direct_pgio_init,
 838         .completion = nfs_direct_write_completion,
 839 };
 840
 841 static ssize_t nfs_direct_do_schedule_write_iovec(
 842                 struct nfs_pageio_descriptor *desc, const struct iovec *iov,
 843                 unsigned long nr_segs, loff_t pos)
 844 {
 845         ssize_t result = -EINVAL;
 846         size_t requested_bytes = 0;
 847         unsigned long seg;
 848
 849         for (seg = 0; seg < nr_segs; seg++) {
 850                 const struct iovec *vec = &iov[seg];
 851                 result = nfs_direct_write_schedule_segment(desc, vec,
 852                                                            pos);
 853                 if (result < 0)
 854                         break;
 855                 requested_bytes += result;
 856                 if ((size_t)result < vec->iov_len)
 857                         break;
 858                 pos += vec->iov_len;
 859         }
 860
 861         if (requested_bytes)
 862                 return requested_bytes;
 863
 864         return result < 0 ? result : -EIO;
 865 }
 866
 867 #ifdef CONFIG_BLOCK
 868 static ssize_t nfs_direct_do_schedule_write_bvec(
 869                 struct nfs_pageio_descriptor *desc,
 870                 struct bio_vec *bvec, unsigned long nr_segs, loff_t pos)
 871 {
 872         struct nfs_direct_req *dreq = desc->pg_dreq;
 873         struct nfs_open_context *ctx = dreq->ctx;
 874         struct inode *inode = dreq->inode;
 875         ssize_t result = 0;
 876         size_t requested_bytes = 0;
 877         unsigned long seg;
 878         struct nfs_page *req;
 879         unsigned int req_len;
 880
 881         for (seg = 0; seg < nr_segs; seg++) {
 882                 req_len = bvec[seg].bv_len;
 883
 884                 req = nfs_create_request(ctx, inode, bvec[seg].bv_page,
 885                                          bvec[seg].bv_offset, req_len);
 886                 if (IS_ERR(req)) {
 887                         result = PTR_ERR(req);
 888                         break;
 889                 }
 890                 nfs_lock_request(req);
 891                 req->wb_index = pos >> PAGE_SHIFT;
 892                 req->wb_offset = pos & ~PAGE_MASK;
 893                 if (!nfs_pageio_add_request(desc, req)) {
 894                         result = desc->pg_error;
 895                         nfs_unlock_and_release_request(req);
 896                         break;
 897                 }
 898                 requested_bytes += req_len;
 899                 pos += req_len;
 900         }
 901
 902         if (requested_bytes)
 903                 return requested_bytes;
 904
 905         return result < 0 ? result : -EIO;
 906 }
 907 #endif /* CONFIG_BLOCK */
 908
 909 static ssize_t nfs_direct_write_schedule(struct nfs_direct_req *dreq,
 910                                          struct iov_iter *iter, loff_t pos)
 911 {
 912         struct nfs_pageio_descriptor desc;
 913         struct inode *inode = dreq->inode;
 914         ssize_t result = 0;
 915
 916         NFS_PROTO(inode)->write_pageio_init(&desc, inode, FLUSH_COND_STABLE,
 917                               &nfs_direct_write_completion_ops);
 918         desc.pg_dreq = dreq;
 919         get_dreq(dreq);
 920         atomic_inc(&inode->i_dio_count);
 921
 922         NFS_I(dreq->inode)->write_io += iov_iter_count(iter);
 923
 924         if (iov_iter_has_iovec(iter)) {
 925                 result = nfs_direct_do_schedule_write_iovec(&desc,
 926                                 iov_iter_iovec(iter), iter->nr_segs, pos);
 927 #ifdef CONFIG_BLOCK
 928         } else if (iov_iter_has_bvec(iter)) {
 929                 result = nfs_direct_do_schedule_write_bvec(&desc,
 930                                 iov_iter_bvec(iter), iter->nr_segs, pos);
 931 #endif
 932         } else
 933                 BUG();
 934
 935         nfs_pageio_complete(&desc);
 936
 937         /*
 938          * If no bytes were started, return the error, and let the
 939          * generic layer handle the completion.
 940          */
 941         if (result < 0) {
 942                 inode_dio_done(inode);
 943                 nfs_direct_req_release(dreq);
 944                 return result;
 945         }
 946
 947         if (put_dreq(dreq))
 948                 nfs_direct_write_complete(dreq, dreq->inode);
 949         return 0;
 950 }
 951
 952 static ssize_t nfs_direct_write(struct kiocb *iocb, struct iov_iter *iter,
 953                                 loff_t pos)
 954 {
 955         ssize_t result = -ENOMEM;
 956         struct inode *inode = iocb->ki_filp->f_mapping->host;
 957         struct nfs_direct_req *dreq;
 958         struct nfs_lock_context *l_ctx;
 959
 960         dreq = nfs_direct_req_alloc();
 961         if (!dreq)
 962                 goto out;
 963
 964         dreq->inode = inode;
 965         dreq->bytes_left = iov_iter_count(iter);
 966         dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
 967         l_ctx = nfs_get_lock_context(dreq->ctx);
 968         if (IS_ERR(l_ctx)) {
 969                 result = PTR_ERR(l_ctx);
 970                 goto out_release;
 971         }
 972         dreq->l_ctx = l_ctx;
 973         if (!is_sync_kiocb(iocb))
 974                 dreq->iocb = iocb;
 975
 976         result = nfs_direct_write_schedule(dreq, iter, pos);
 977         if (!result)
 978                 result = nfs_direct_wait(dreq);
 979 out_release:
 980         nfs_direct_req_release(dreq);
 981 out:
 982         return result;
 983 }
 984
 985 /**
 986  * nfs_file_direct_read - file direct read operation for NFS files
 987  * @iocb: target I/O control block
 988  * @iter: vector of buffers into which to read data
 989  * @pos: byte offset in file where reading starts
 990  *
 991  * We use this function for direct reads instead of calling
 992  * generic_file_read_iter() in order to avoid gfar's check to see if
 993  * the request starts before the end of the file.  For that check
 994  * to work, we must generate a GETATTR before each direct read, and
 995  * even then there is a window between the GETATTR and the subsequent
 996  * READ where the file size could change.  Our preference is simply
 997  * to do all reads the application wants, and the server will take
 998  * care of managing the end of file boundary.
 999  *
1000  * This function also eliminates unnecessarily updating the file's
1001  * atime locally, as the NFS server sets the file's atime, and this
1002  * client must read the updated atime from the server back into its
1003  * cache.
1004  */
1005 ssize_t nfs_file_direct_read(struct kiocb *iocb, struct iov_iter *iter,
1006                              loff_t pos)
1007 {
1008         ssize_t retval = -EINVAL;
1009         struct file *file = iocb->ki_filp;
1010         struct address_space *mapping = file->f_mapping;
1011         size_t count;
1012
1013         count = iov_iter_count(iter);
1014         nfs_add_stats(mapping->host, NFSIOS_DIRECTREADBYTES, count);
1015
1016         dfprintk(FILE, "NFS: direct read(%s/%s, %zd@%Ld)\n",
1017                 file->f_path.dentry->d_parent->d_name.name,
1018                 file->f_path.dentry->d_name.name,
1019                 count, (long long) pos);
1020
1021         retval = 0;
1022         if (!count)
1023                 goto out;
1024
1025         retval = nfs_sync_mapping(mapping);
1026         if (retval)
1027                 goto out;
1028
1029         task_io_account_read(count);
1030
1031         retval = nfs_direct_read(iocb, iter, pos);
1032         if (retval > 0)
1033                 iocb->ki_pos = pos + retval;
1034
1035 out:
1036         return retval;
1037 }
1038
1039 /**
1040  * nfs_file_direct_write - file direct write operation for NFS files
1041  * @iocb: target I/O control block
1042  * @iter: vector of buffers from which to write data
1043  * @pos: byte offset in file where writing starts
1044  *
1045  * We use this function for direct writes instead of calling
1046  * generic_file_write_iter() in order to avoid taking the inode
1047  * semaphore and updating the i_size.  The NFS server will set
1048  * the new i_size and this client must read the updated size
1049  * back into its cache.  We let the server do generic write
1050  * parameter checking and report problems.
1051  *
1052  * We eliminate local atime updates, see direct read above.
1053  *
1054  * We avoid unnecessary page cache invalidations for normal cached
1055  * readers of this file.
1056  *
1057  * Note that O_APPEND is not supported for NFS direct writes, as there
1058  * is no atomic O_APPEND write facility in the NFS protocol.
1059  */
1060 ssize_t nfs_file_direct_write(struct kiocb *iocb, struct iov_iter *iter,
1061                               loff_t pos)
1062 {
1063         ssize_t retval = -EINVAL;
1064         struct file *file = iocb->ki_filp;
1065         struct address_space *mapping = file->f_mapping;
1066         size_t count;
1067
1068         count = iov_iter_count(iter);
1069         nfs_add_stats(mapping->host, NFSIOS_DIRECTWRITTENBYTES, count);
1070
1071         dfprintk(FILE, "NFS: direct write(%s/%s, %zd@%Ld)\n",
1072                 file->f_path.dentry->d_parent->d_name.name,
1073                 file->f_path.dentry->d_name.name,
1074                 count, (long long) pos);
1075
1076         retval = generic_write_checks(file, &pos, &count, 0);
1077         if (retval)
1078                 goto out;
1079
1080         retval = -EINVAL;
1081         if ((ssize_t) count < 0)
1082                 goto out;
1083         retval = 0;
1084         if (!count)
1085                 goto out;
1086
1087         retval = nfs_sync_mapping(mapping);
1088         if (retval)
1089                 goto out;
1090
1091         task_io_account_write(count);
1092
1093         retval = nfs_direct_write(iocb, iter, pos);
1094         if (retval > 0) {
1095                 struct inode *inode = mapping->host;
1096
1097                 iocb->ki_pos = pos + retval;
1098                 spin_lock(&inode->i_lock);
1099                 if (i_size_read(inode) < iocb->ki_pos)
1100                         i_size_write(inode, iocb->ki_pos);
1101                 spin_unlock(&inode->i_lock);
1102         }
1103 out:
1104         return retval;
1105 }
1106
1107 /**
1108  * nfs_init_directcache - create a slab cache for nfs_direct_req structures
1109  *
1110  */
1111 int __init nfs_init_directcache(void)
1112 {
1113         nfs_direct_cachep = kmem_cache_create("nfs_direct_cache",
1114                                                 sizeof(struct nfs_direct_req),
1115                                                 0, (SLAB_RECLAIM_ACCOUNT|
1116                                                         SLAB_MEM_SPREAD),
1117                                                 NULL);
1118         if (nfs_direct_cachep == NULL)
1119                 return -ENOMEM;
1120
1121         return 0;
1122 }
1123
1124 /**
1125  * nfs_destroy_directcache - destroy the slab cache for nfs_direct_req structures
1126  *
1127  */
1128 void nfs_destroy_directcache(void)
1129 {
1130         kmem_cache_destroy(nfs_direct_cachep);
1131 }