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/nfs3.h>
  17 #include <linux/nfs4.h>
  18 #include <linux/nfs_page.h>
  19 #include <linux/nfs_fs.h>
  20 #include <linux/nfs_mount.h>
  21
  22 #include "internal.h"
  23
  24 static struct kmem_cache *nfs_page_cachep;
  25
  26 static inline struct nfs_page *
  27 nfs_page_alloc(void)
  28 {
  29         struct nfs_page *p;
  30         p = kmem_cache_alloc(nfs_page_cachep, GFP_KERNEL);
  31         if (p) {
  32                 memset(p, 0, sizeof(*p));
  33                 INIT_LIST_HEAD(&p->wb_list);
  34         }
  35         return p;
  36 }
  37
  38 static inline void
  39 nfs_page_free(struct nfs_page *p)
  40 {
  41         kmem_cache_free(nfs_page_cachep, p);
  42 }
  43
  44 /**
  45  * nfs_create_request - Create an NFS read/write request.
  46  * @file: file descriptor to use
  47  * @inode: inode to which the request is attached
  48  * @page: page to write
  49  * @offset: starting offset within the page for the write
  50  * @count: number of bytes to read/write
  51  *
  52  * The page must be locked by the caller. This makes sure we never
  53  * create two different requests for the same page.
  54  * User should ensure it is safe to sleep in this function.
  55  */
  56 struct nfs_page *
  57 nfs_create_request(struct nfs_open_context *ctx, struct inode *inode,
  58                    struct page *page,
  59                    unsigned int offset, unsigned int count)
  60 {
  61         struct nfs_page         *req;
  62
  63         for (;;) {
  64                 /* try to allocate the request struct */
  65                 req = nfs_page_alloc();
  66                 if (req != NULL)
  67                         break;
  68
  69                 if (fatal_signal_pending(current))
  70                         return ERR_PTR(-ERESTARTSYS);
  71                 yield();
  72         }
  73
  74         /* Initialize the request struct. Initially, we assume a
  75          * long write-back delay. This will be adjusted in
  76          * update_nfs_request below if the region is not locked. */
  77         req->wb_page    = page;
  78         atomic_set(&req->wb_complete, 0);
  79         req->wb_index   = page->index;
  80         page_cache_get(page);
  81         BUG_ON(PagePrivate(page));
  82         BUG_ON(!PageLocked(page));
  83         BUG_ON(page->mapping->host != inode);
  84         req->wb_offset  = offset;
  85         req->wb_pgbase  = offset;
  86         req->wb_bytes   = count;
  87         req->wb_context = get_nfs_open_context(ctx);
  88         kref_init(&req->wb_kref);
  89         return req;
  90 }
  91
  92 /**
  93  * nfs_unlock_request - Unlock request and wake up sleepers.
  94  * @req:
  95  */
  96 void nfs_unlock_request(struct nfs_page *req)
  97 {
  98         if (!NFS_WBACK_BUSY(req)) {
  99                 printk(KERN_ERR "NFS: Invalid unlock attempted\n");
 100                 BUG();
 101         }
 102         smp_mb__before_clear_bit();
 103         clear_bit(PG_BUSY, &req->wb_flags);
 104         smp_mb__after_clear_bit();
 105         wake_up_bit(&req->wb_flags, PG_BUSY);
 106         nfs_release_request(req);
 107 }
 108
 109 /**
 110  * nfs_set_page_tag_locked - Tag a request as locked
 111  * @req:
 112  */
 113 int nfs_set_page_tag_locked(struct nfs_page *req)
 114 {
 115         if (!nfs_lock_request_dontget(req))
 116                 return 0;
 117         if (req->wb_page != NULL)
 118                 radix_tree_tag_set(&NFS_I(req->wb_context->path.dentry->d_inode)->nfs_page_tree, req->wb_index, NFS_PAGE_TAG_LOCKED);
 119         return 1;
 120 }
 121
 122 /**
 123  * nfs_clear_page_tag_locked - Clear request tag and wake up sleepers
 124  */
 125 void nfs_clear_page_tag_locked(struct nfs_page *req)
 126 {
 127         if (req->wb_page != NULL) {
 128                 struct inode *inode = req->wb_context->path.dentry->d_inode;
 129                 struct nfs_inode *nfsi = NFS_I(inode);
 130
 131                 spin_lock(&inode->i_lock);
 132                 radix_tree_tag_clear(&nfsi->nfs_page_tree, req->wb_index, NFS_PAGE_TAG_LOCKED);
 133                 nfs_unlock_request(req);
 134                 spin_unlock(&inode->i_lock);
 135         } else
 136                 nfs_unlock_request(req);
 137 }
 138
 139 /**
 140  * nfs_clear_request - Free up all resources allocated to the request
 141  * @req:
 142  *
 143  * Release page and open context resources associated with a read/write
 144  * request after it has completed.
 145  */
 146 void nfs_clear_request(struct nfs_page *req)
 147 {
 148         struct page *page = req->wb_page;
 149         struct nfs_open_context *ctx = req->wb_context;
 150
 151         if (page != NULL) {
 152                 page_cache_release(page);
 153                 req->wb_page = NULL;
 154         }
 155         if (ctx != NULL) {
 156                 put_nfs_open_context(ctx);
 157                 req->wb_context = NULL;
 158         }
 159 }
 160
 161
 162 /**
 163  * nfs_release_request - Release the count on an NFS read/write request
 164  * @req: request to release
 165  *
 166  * Note: Should never be called with the spinlock held!
 167  */
 168 static void nfs_free_request(struct kref *kref)
 169 {
 170         struct nfs_page *req = container_of(kref, struct nfs_page, wb_kref);
 171
 172         /* Release struct file and open context */
 173         nfs_clear_request(req);
 174         nfs_page_free(req);
 175 }
 176
 177 void nfs_release_request(struct nfs_page *req)
 178 {
 179         kref_put(&req->wb_kref, nfs_free_request);
 180 }
 181
 182 static int nfs_wait_bit_uninterruptible(void *word)
 183 {
 184         io_schedule();
 185         return 0;
 186 }
 187
 188 /**
 189  * nfs_wait_on_request - Wait for a request to complete.
 190  * @req: request to wait upon.
 191  *
 192  * Interruptible by fatal signals only.
 193  * The user is responsible for holding a count on the request.
 194  */
 195 int
 196 nfs_wait_on_request(struct nfs_page *req)
 197 {
 198         return wait_on_bit(&req->wb_flags, PG_BUSY,
 199                         nfs_wait_bit_uninterruptible,
 200                         TASK_UNINTERRUPTIBLE);
 201 }
 202
 203 /**
 204  * nfs_pageio_init - initialise a page io descriptor
 205  * @desc: pointer to descriptor
 206  * @inode: pointer to inode
 207  * @doio: pointer to io function
 208  * @bsize: io block size
 209  * @io_flags: extra parameters for the io function
 210  */
 211 void nfs_pageio_init(struct nfs_pageio_descriptor *desc,
 212                      struct inode *inode,
 213                      int (*doio)(struct inode *, struct list_head *, unsigned int, size_t, int),
 214                      size_t bsize,
 215                      int io_flags)
 216 {
 217         INIT_LIST_HEAD(&desc->pg_list);
 218         desc->pg_bytes_written = 0;
 219         desc->pg_count = 0;
 220         desc->pg_bsize = bsize;
 221         desc->pg_base = 0;
 222         desc->pg_inode = inode;
 223         desc->pg_doio = doio;
 224         desc->pg_ioflags = io_flags;
 225         desc->pg_error = 0;
 226 }
 227
 228 /**
 229  * nfs_can_coalesce_requests - test two requests for compatibility
 230  * @prev: pointer to nfs_page
 231  * @req: pointer to nfs_page
 232  *
 233  * The nfs_page structures 'prev' and 'req' are compared to ensure that the
 234  * page data area they describe is contiguous, and that their RPC
 235  * credentials, NFSv4 open state, and lockowners are the same.
 236  *
 237  * Return 'true' if this is the case, else return 'false'.
 238  */
 239 static int nfs_can_coalesce_requests(struct nfs_page *prev,
 240                                      struct nfs_page *req)
 241 {
 242         if (req->wb_context->cred != prev->wb_context->cred)
 243                 return 0;
 244         if (req->wb_context->lockowner != prev->wb_context->lockowner)
 245                 return 0;
 246         if (req->wb_context->state != prev->wb_context->state)
 247                 return 0;
 248         if (req->wb_index != (prev->wb_index + 1))
 249                 return 0;
 250         if (req->wb_pgbase != 0)
 251                 return 0;
 252         if (prev->wb_pgbase + prev->wb_bytes != PAGE_CACHE_SIZE)
 253                 return 0;
 254         return 1;
 255 }
 256
 257 /**
 258  * nfs_pageio_do_add_request - Attempt to coalesce a request into a page list.
 259  * @desc: destination io descriptor
 260  * @req: request
 261  *
 262  * Returns true if the request 'req' was successfully coalesced into the
 263  * existing list of pages 'desc'.
 264  */
 265 static int nfs_pageio_do_add_request(struct nfs_pageio_descriptor *desc,
 266                                      struct nfs_page *req)
 267 {
 268         size_t newlen = req->wb_bytes;
 269
 270         if (desc->pg_count != 0) {
 271                 struct nfs_page *prev;
 272
 273                 /*
 274                  * FIXME: ideally we should be able to coalesce all requests
 275                  * that are not block boundary aligned, but currently this
 276                  * is problematic for the case of bsize < PAGE_CACHE_SIZE,
 277                  * since nfs_flush_multi and nfs_pagein_multi assume you
 278                  * can have only one struct nfs_page.
 279                  */
 280                 if (desc->pg_bsize < PAGE_SIZE)
 281                         return 0;
 282                 newlen += desc->pg_count;
 283                 if (newlen > desc->pg_bsize)
 284                         return 0;
 285                 prev = nfs_list_entry(desc->pg_list.prev);
 286                 if (!nfs_can_coalesce_requests(prev, req))
 287                         return 0;
 288         } else
 289                 desc->pg_base = req->wb_pgbase;
 290         nfs_list_remove_request(req);
 291         nfs_list_add_request(req, &desc->pg_list);
 292         desc->pg_count = newlen;
 293         return 1;
 294 }
 295
 296 /*
 297  * Helper for nfs_pageio_add_request and nfs_pageio_complete
 298  */
 299 static void nfs_pageio_doio(struct nfs_pageio_descriptor *desc)
 300 {
 301         if (!list_empty(&desc->pg_list)) {
 302                 int error = desc->pg_doio(desc->pg_inode,
 303                                           &desc->pg_list,
 304                                           nfs_page_array_len(desc->pg_base,
 305                                                              desc->pg_count),
 306                                           desc->pg_count,
 307                                           desc->pg_ioflags);
 308                 if (error < 0)
 309                         desc->pg_error = error;
 310                 else
 311                         desc->pg_bytes_written += desc->pg_count;
 312         }
 313         if (list_empty(&desc->pg_list)) {
 314                 desc->pg_count = 0;
 315                 desc->pg_base = 0;
 316         }
 317 }
 318
 319 /**
 320  * nfs_pageio_add_request - Attempt to coalesce a request into a page list.
 321  * @desc: destination io descriptor
 322  * @req: request
 323  *
 324  * Returns true if the request 'req' was successfully coalesced into the
 325  * existing list of pages 'desc'.
 326  */
 327 int nfs_pageio_add_request(struct nfs_pageio_descriptor *desc,
 328                            struct nfs_page *req)
 329 {
 330         while (!nfs_pageio_do_add_request(desc, req)) {
 331                 nfs_pageio_doio(desc);
 332                 if (desc->pg_error < 0)
 333                         return 0;
 334         }
 335         return 1;
 336 }
 337
 338 /**
 339  * nfs_pageio_complete - Complete I/O on an nfs_pageio_descriptor
 340  * @desc: pointer to io descriptor
 341  */
 342 void nfs_pageio_complete(struct nfs_pageio_descriptor *desc)
 343 {
 344         nfs_pageio_doio(desc);
 345 }
 346
 347 /**
 348  * nfs_pageio_cond_complete - Conditional I/O completion
 349  * @desc: pointer to io descriptor
 350  * @index: page index
 351  *
 352  * It is important to ensure that processes don't try to take locks
 353  * on non-contiguous ranges of pages as that might deadlock. This
 354  * function should be called before attempting to wait on a locked
 355  * nfs_page. It will complete the I/O if the page index 'index'
 356  * is not contiguous with the existing list of pages in 'desc'.
 357  */
 358 void nfs_pageio_cond_complete(struct nfs_pageio_descriptor *desc, pgoff_t index)
 359 {
 360         if (!list_empty(&desc->pg_list)) {
 361                 struct nfs_page *prev = nfs_list_entry(desc->pg_list.prev);
 362                 if (index != prev->wb_index + 1)
 363                         nfs_pageio_doio(desc);
 364         }
 365 }
 366
 367 #define NFS_SCAN_MAXENTRIES 16
 368 /**
 369  * nfs_scan_list - Scan a list for matching requests
 370  * @nfsi: NFS inode
 371  * @dst: Destination list
 372  * @idx_start: lower bound of page->index to scan
 373  * @npages: idx_start + npages sets the upper bound to scan.
 374  * @tag: tag to scan for
 375  *
 376  * Moves elements from one of the inode request lists.
 377  * If the number of requests is set to 0, the entire address_space
 378  * starting at index idx_start, is scanned.
 379  * The requests are *not* checked to ensure that they form a contiguous set.
 380  * You must be holding the inode's i_lock when calling this function
 381  */
 382 int nfs_scan_list(struct nfs_inode *nfsi,
 383                 struct list_head *dst, pgoff_t idx_start,
 384                 unsigned int npages, int tag)
 385 {
 386         struct nfs_page *pgvec[NFS_SCAN_MAXENTRIES];
 387         struct nfs_page *req;
 388         pgoff_t idx_end;
 389         int found, i;
 390         int res;
 391
 392         res = 0;
 393         if (npages == 0)
 394                 idx_end = ~0;
 395         else
 396                 idx_end = idx_start + npages - 1;
 397
 398         for (;;) {
 399                 found = radix_tree_gang_lookup_tag(&nfsi->nfs_page_tree,
 400                                 (void **)&pgvec[0], idx_start,
 401                                 NFS_SCAN_MAXENTRIES, tag);
 402                 if (found <= 0)
 403                         break;
 404                 for (i = 0; i < found; i++) {
 405                         req = pgvec[i];
 406                         if (req->wb_index > idx_end)
 407                                 goto out;
 408                         idx_start = req->wb_index + 1;
 409                         if (nfs_set_page_tag_locked(req)) {
 410                                 kref_get(&req->wb_kref);
 411                                 nfs_list_remove_request(req);
 412                                 radix_tree_tag_clear(&nfsi->nfs_page_tree,
 413                                                 req->wb_index, tag);
 414                                 nfs_list_add_request(req, dst);
 415                                 res++;
 416                                 if (res == INT_MAX)
 417                                         goto out;
 418                         }
 419                 }
 420                 /* for latency reduction */
 421                 cond_resched_lock(&nfsi->vfs_inode.i_lock);
 422         }
 423 out:
 424         return res;
 425 }
 426
 427 int __init nfs_init_nfspagecache(void)
 428 {
 429         nfs_page_cachep = kmem_cache_create("nfs_page",
 430                                             sizeof(struct nfs_page),
 431                                             0, SLAB_HWCACHE_ALIGN,
 432                                             NULL);
 433         if (nfs_page_cachep == NULL)
 434                 return -ENOMEM;
 435
 436         return 0;
 437 }
 438
 439 void nfs_destroy_nfspagecache(void)
 440 {
 441         kmem_cache_destroy(nfs_page_cachep);
 442 }
 443