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