fs/nfs/pagelist.c

   1 /*
   2  * linux/fs/nfs/pagelist.c
   3  *
   4  * A set of helper functions for managing NFS read and write requests.
   5  * The main purpose of these routines is to provide support for the
   6  * coalescing of several requests into a single RPC call.
   7  *
   8  * Copyright 2000, 2001 (c) Trond Myklebust <trond.myklebust@fys.uio.no>
   9  *
  10  */
  11
  12 #include <linux/slab.h>
  13 #include <linux/file.h>
  14 #include <linux/sched.h>
  15 #include <linux/sunrpc/clnt.h>
  16 #include <linux/nfs.h>
  17 #include <linux/nfs3.h>
  18 #include <linux/nfs4.h>
  19 #include <linux/nfs_page.h>
  20 #include <linux/nfs_fs.h>
  21 #include <linux/nfs_mount.h>
  22 #include <linux/export.h>
  23
  24 #include "internal.h"
  25 #include "pnfs.h"
  26
  27 static struct kmem_cache *nfs_page_cachep;
  28
  29 bool nfs_pgarray_set(struct nfs_page_array *p, unsigned int pagecount)
  30 {
  31         p->npages = pagecount;
  32         if (pagecount <= ARRAY_SIZE(p->page_array))
  33                 p->pagevec = p->page_array;
  34         else {
  35                 p->pagevec = kcalloc(pagecount, sizeof(struct page *), GFP_KERNEL);
  36                 if (!p->pagevec)
  37                         p->npages = 0;
  38         }
  39         return p->pagevec != NULL;
  40 }
  41
  42 void nfs_pgheader_init(struct nfs_pageio_descriptor *desc,
  43                        struct nfs_pgio_header *hdr,
  44                        void (*release)(struct nfs_pgio_header *hdr))
  45 {
  46         hdr->req = nfs_list_entry(desc->pg_list.next);
  47         hdr->inode = desc->pg_inode;
  48         hdr->cred = hdr->req->wb_context->cred;
  49         hdr->io_start = req_offset(hdr->req);
  50         hdr->good_bytes = desc->pg_count;
  51         hdr->dreq = desc->pg_dreq;
  52         hdr->release = release;
  53         hdr->completion_ops = desc->pg_completion_ops;
  54         if (hdr->completion_ops->init_hdr)
  55                 hdr->completion_ops->init_hdr(hdr);
  56 }
  57
  58 void nfs_set_pgio_error(struct nfs_pgio_header *hdr, int error, loff_t pos)
  59 {
  60         spin_lock(&hdr->lock);
  61         if (pos < hdr->io_start + hdr->good_bytes) {
  62                 set_bit(NFS_IOHDR_ERROR, &hdr->flags);
  63                 clear_bit(NFS_IOHDR_EOF, &hdr->flags);
  64                 hdr->good_bytes = pos - hdr->io_start;
  65                 hdr->error = error;
  66         }
  67         spin_unlock(&hdr->lock);
  68 }
  69
  70 static inline struct nfs_page *
  71 nfs_page_alloc(void)
  72 {
  73         struct nfs_page *p = kmem_cache_zalloc(nfs_page_cachep, GFP_KERNEL);
  74         if (p)
  75                 INIT_LIST_HEAD(&p->wb_list);
  76         return p;
  77 }
  78
  79 static inline void
  80 nfs_page_free(struct nfs_page *p)
  81 {
  82         kmem_cache_free(nfs_page_cachep, p);
  83 }
  84
  85 /**
  86  * nfs_create_request - Create an NFS read/write request.
  87  * @ctx: open context to use
  88  * @inode: inode to which the request is attached
  89  * @page: page to write
  90  * @offset: starting offset within the page for the write
  91  * @count: number of bytes to read/write
  92  *
  93  * The page must be locked by the caller. This makes sure we never
  94  * create two different requests for the same page.
  95  * User should ensure it is safe to sleep in this function.
  96  */
  97 struct nfs_page *
  98 nfs_create_request(struct nfs_open_context *ctx, struct inode *inode,
  99                    struct page *page,
 100                    unsigned int offset, unsigned int count)
 101 {
 102         struct nfs_page         *req;
 103
 104         /* try to allocate the request struct */
 105         req = nfs_page_alloc();
 106         if (req == NULL)
 107                 return ERR_PTR(-ENOMEM);
 108
 109         /* get lock context early so we can deal with alloc failures */
 110         req->wb_lock_context = nfs_get_lock_context(ctx);
 111         if (req->wb_lock_context == NULL) {
 112                 nfs_page_free(req);
 113                 return ERR_PTR(-ENOMEM);
 114         }
 115
 116         /* Initialize the request struct. Initially, we assume a
 117          * long write-back delay. This will be adjusted in
 118          * update_nfs_request below if the region is not locked. */
 119         req->wb_page    = page;
 120         req->wb_index   = page->index;
 121         page_cache_get(page);
 122         req->wb_offset  = offset;
 123         req->wb_pgbase  = offset;
 124         req->wb_bytes   = count;
 125         req->wb_context = get_nfs_open_context(ctx);
 126         kref_init(&req->wb_kref);
 127         return req;
 128 }
 129
 130 /**
 131  * nfs_unlock_request - Unlock request and wake up sleepers.
 132  * @req:
 133  */
 134 void nfs_unlock_request(struct nfs_page *req)
 135 {
 136         if (!NFS_WBACK_BUSY(req)) {
 137                 printk(KERN_ERR "NFS: Invalid unlock attempted\n");
 138                 BUG();
 139         }
 140         smp_mb__before_clear_bit();
 141         clear_bit(PG_BUSY, &req->wb_flags);
 142         smp_mb__after_clear_bit();
 143         wake_up_bit(&req->wb_flags, PG_BUSY);
 144 }
 145
 146 /**
 147  * nfs_unlock_and_release_request - Unlock request and release the nfs_page
 148  * @req:
 149  */
 150 void nfs_unlock_and_release_request(struct nfs_page *req)
 151 {
 152         nfs_unlock_request(req);
 153         nfs_release_request(req);
 154 }
 155
 156 /*
 157  * nfs_clear_request - Free up all resources allocated to the request
 158  * @req:
 159  *
 160  * Release page and open context resources associated with a read/write
 161  * request after it has completed.
 162  */
 163 static void nfs_clear_request(struct nfs_page *req)
 164 {
 165         struct page *page = req->wb_page;
 166         struct nfs_open_context *ctx = req->wb_context;
 167         struct nfs_lock_context *l_ctx = req->wb_lock_context;
 168
 169         if (page != NULL) {
 170                 page_cache_release(page);
 171                 req->wb_page = NULL;
 172         }
 173         if (l_ctx != NULL) {
 174                 nfs_put_lock_context(l_ctx);
 175                 req->wb_lock_context = NULL;
 176         }
 177         if (ctx != NULL) {
 178                 put_nfs_open_context(ctx);
 179                 req->wb_context = NULL;
 180         }
 181 }
 182
 183
 184 /**
 185  * nfs_release_request - Release the count on an NFS read/write request
 186  * @req: request to release
 187  *
 188  * Note: Should never be called with the spinlock held!
 189  */
 190 static void nfs_free_request(struct kref *kref)
 191 {
 192         struct nfs_page *req = container_of(kref, struct nfs_page, wb_kref);
 193
 194         /* Release struct file and open context */
 195         nfs_clear_request(req);
 196         nfs_page_free(req);
 197 }
 198
 199 void nfs_release_request(struct nfs_page *req)
 200 {
 201         kref_put(&req->wb_kref, nfs_free_request);
 202 }
 203
 204 static int nfs_wait_bit_uninterruptible(void *word)
 205 {
 206         io_schedule();
 207         return 0;
 208 }
 209
 210 /**
 211  * nfs_wait_on_request - Wait for a request to complete.
 212  * @req: request to wait upon.
 213  *
 214  * Interruptible by fatal signals only.
 215  * The user is responsible for holding a count on the request.
 216  */
 217 int
 218 nfs_wait_on_request(struct nfs_page *req)
 219 {
 220         return wait_on_bit(&req->wb_flags, PG_BUSY,
 221                         nfs_wait_bit_uninterruptible,
 222                         TASK_UNINTERRUPTIBLE);
 223 }
 224
 225 bool nfs_generic_pg_test(struct nfs_pageio_descriptor *desc, struct nfs_page *prev, struct nfs_page *req)
 226 {
 227         /*
 228          * FIXME: ideally we should be able to coalesce all requests
 229          * that are not block boundary aligned, but currently this
 230          * is problematic for the case of bsize < PAGE_CACHE_SIZE,
 231          * since nfs_flush_multi and nfs_pagein_multi assume you
 232          * can have only one struct nfs_page.
 233          */
 234         if (desc->pg_bsize < PAGE_SIZE)
 235                 return 0;
 236
 237         return desc->pg_count + req->wb_bytes <= desc->pg_bsize;
 238 }
 239 EXPORT_SYMBOL_GPL(nfs_generic_pg_test);
 240
 241 /**
 242  * nfs_pageio_init - initialise a page io descriptor
 243  * @desc: pointer to descriptor
 244  * @inode: pointer to inode
 245  * @doio: pointer to io function
 246  * @bsize: io block size
 247  * @io_flags: extra parameters for the io function
 248  */
 249 void nfs_pageio_init(struct nfs_pageio_descriptor *desc,
 250                      struct inode *inode,
 251                      const struct nfs_pageio_ops *pg_ops,
 252                      const struct nfs_pgio_completion_ops *compl_ops,
 253                      size_t bsize,
 254                      int io_flags)
 255 {
 256         INIT_LIST_HEAD(&desc->pg_list);
 257         desc->pg_bytes_written = 0;
 258         desc->pg_count = 0;
 259         desc->pg_bsize = bsize;
 260         desc->pg_base = 0;
 261         desc->pg_moreio = 0;
 262         desc->pg_recoalesce = 0;
 263         desc->pg_inode = inode;
 264         desc->pg_ops = pg_ops;
 265         desc->pg_completion_ops = compl_ops;
 266         desc->pg_ioflags = io_flags;
 267         desc->pg_error = 0;
 268         desc->pg_lseg = NULL;
 269         desc->pg_dreq = NULL;
 270 }
 271
 272 /**
 273  * nfs_can_coalesce_requests - test two requests for compatibility
 274  * @prev: pointer to nfs_page
 275  * @req: pointer to nfs_page
 276  *
 277  * The nfs_page structures 'prev' and 'req' are compared to ensure that the
 278  * page data area they describe is contiguous, and that their RPC
 279  * credentials, NFSv4 open state, and lockowners are the same.
 280  *
 281  * Return 'true' if this is the case, else return 'false'.
 282  */
 283 static bool nfs_can_coalesce_requests(struct nfs_page *prev,
 284                                       struct nfs_page *req,
 285                                       struct nfs_pageio_descriptor *pgio)
 286 {
 287         if (req->wb_context->cred != prev->wb_context->cred)
 288                 return false;
 289         if (req->wb_lock_context->lockowner != prev->wb_lock_context->lockowner)
 290                 return false;
 291         if (req->wb_context->state != prev->wb_context->state)
 292                 return false;
 293         if (req->wb_pgbase != 0)
 294                 return false;
 295         if (prev->wb_pgbase + prev->wb_bytes != PAGE_CACHE_SIZE)
 296                 return false;
 297         if (req_offset(req) != req_offset(prev) + prev->wb_bytes)
 298                 return false;
 299         return pgio->pg_ops->pg_test(pgio, prev, req);
 300 }
 301
 302 /**
 303  * nfs_pageio_do_add_request - Attempt to coalesce a request into a page list.
 304  * @desc: destination io descriptor
 305  * @req: request
 306  *
 307  * Returns true if the request 'req' was successfully coalesced into the
 308  * existing list of pages 'desc'.
 309  */
 310 static int nfs_pageio_do_add_request(struct nfs_pageio_descriptor *desc,
 311                                      struct nfs_page *req)
 312 {
 313         if (desc->pg_count != 0) {
 314                 struct nfs_page *prev;
 315
 316                 prev = nfs_list_entry(desc->pg_list.prev);
 317                 if (!nfs_can_coalesce_requests(prev, req, desc))
 318                         return 0;
 319         } else {
 320                 if (desc->pg_ops->pg_init)
 321                         desc->pg_ops->pg_init(desc, req);
 322                 desc->pg_base = req->wb_pgbase;
 323         }
 324         nfs_list_remove_request(req);
 325         nfs_list_add_request(req, &desc->pg_list);
 326         desc->pg_count += req->wb_bytes;
 327         return 1;
 328 }
 329
 330 /*
 331  * Helper for nfs_pageio_add_request and nfs_pageio_complete
 332  */
 333 static void nfs_pageio_doio(struct nfs_pageio_descriptor *desc)
 334 {
 335         if (!list_empty(&desc->pg_list)) {
 336                 int error = desc->pg_ops->pg_doio(desc);
 337                 if (error < 0)
 338                         desc->pg_error = error;
 339                 else
 340                         desc->pg_bytes_written += desc->pg_count;
 341         }
 342         if (list_empty(&desc->pg_list)) {
 343                 desc->pg_count = 0;
 344                 desc->pg_base = 0;
 345         }
 346 }
 347
 348 /**
 349  * nfs_pageio_add_request - Attempt to coalesce a request into a page list.
 350  * @desc: destination io descriptor
 351  * @req: request
 352  *
 353  * Returns true if the request 'req' was successfully coalesced into the
 354  * existing list of pages 'desc'.
 355  */
 356 static int __nfs_pageio_add_request(struct nfs_pageio_descriptor *desc,
 357                            struct nfs_page *req)
 358 {
 359         while (!nfs_pageio_do_add_request(desc, req)) {
 360                 desc->pg_moreio = 1;
 361                 nfs_pageio_doio(desc);
 362                 if (desc->pg_error < 0)
 363                         return 0;
 364                 desc->pg_moreio = 0;
 365                 if (desc->pg_recoalesce)
 366                         return 0;
 367         }
 368         return 1;
 369 }
 370
 371 static int nfs_do_recoalesce(struct nfs_pageio_descriptor *desc)
 372 {
 373         LIST_HEAD(head);
 374
 375         do {
 376                 list_splice_init(&desc->pg_list, &head);
 377                 desc->pg_bytes_written -= desc->pg_count;
 378                 desc->pg_count = 0;
 379                 desc->pg_base = 0;
 380                 desc->pg_recoalesce = 0;
 381
 382                 while (!list_empty(&head)) {
 383                         struct nfs_page *req;
 384
 385                         req = list_first_entry(&head, struct nfs_page, wb_list);
 386                         nfs_list_remove_request(req);
 387                         if (__nfs_pageio_add_request(desc, req))
 388                                 continue;
 389                         if (desc->pg_error < 0)
 390                                 return 0;
 391                         break;
 392                 }
 393         } while (desc->pg_recoalesce);
 394         return 1;
 395 }
 396
 397 int nfs_pageio_add_request(struct nfs_pageio_descriptor *desc,
 398                 struct nfs_page *req)
 399 {
 400         int ret;
 401
 402         do {
 403                 ret = __nfs_pageio_add_request(desc, req);
 404                 if (ret)
 405                         break;
 406                 if (desc->pg_error < 0)
 407                         break;
 408                 ret = nfs_do_recoalesce(desc);
 409         } while (ret);
 410         return ret;
 411 }
 412
 413 /**
 414  * nfs_pageio_complete - Complete I/O on an nfs_pageio_descriptor
 415  * @desc: pointer to io descriptor
 416  */
 417 void nfs_pageio_complete(struct nfs_pageio_descriptor *desc)
 418 {
 419         for (;;) {
 420                 nfs_pageio_doio(desc);
 421                 if (!desc->pg_recoalesce)
 422                         break;
 423                 if (!nfs_do_recoalesce(desc))
 424                         break;
 425         }
 426 }
 427
 428 /**
 429  * nfs_pageio_cond_complete - Conditional I/O completion
 430  * @desc: pointer to io descriptor
 431  * @index: page index
 432  *
 433  * It is important to ensure that processes don't try to take locks
 434  * on non-contiguous ranges of pages as that might deadlock. This
 435  * function should be called before attempting to wait on a locked
 436  * nfs_page. It will complete the I/O if the page index 'index'
 437  * is not contiguous with the existing list of pages in 'desc'.
 438  */
 439 void nfs_pageio_cond_complete(struct nfs_pageio_descriptor *desc, pgoff_t index)
 440 {
 441         if (!list_empty(&desc->pg_list)) {
 442                 struct nfs_page *prev = nfs_list_entry(desc->pg_list.prev);
 443                 if (index != prev->wb_index + 1)
 444                         nfs_pageio_complete(desc);
 445         }
 446 }
 447
 448 int __init nfs_init_nfspagecache(void)
 449 {
 450         nfs_page_cachep = kmem_cache_create("nfs_page",
 451                                             sizeof(struct nfs_page),
 452                                             0, SLAB_HWCACHE_ALIGN,
 453                                             NULL);
 454         if (nfs_page_cachep == NULL)
 455                 return -ENOMEM;
 456
 457         return 0;
 458 }
 459
 460 void nfs_destroy_nfspagecache(void)
 461 {
 462         kmem_cache_destroy(nfs_page_cachep);
 463 }
 464