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         /* try to allocate the request struct */
  64         req = nfs_page_alloc();
  65         if (req == NULL)
  66                 return ERR_PTR(-ENOMEM);
  67
  68         /* get lock context early so we can deal with alloc failures */
  69         req->wb_lock_context = nfs_get_lock_context(ctx);
  70         if (req->wb_lock_context == NULL) {
  71                 nfs_page_free(req);
  72                 return ERR_PTR(-ENOMEM);
  73         }
  74
  75         /* Initialize the request struct. Initially, we assume a
  76          * long write-back delay. This will be adjusted in
  77          * update_nfs_request below if the region is not locked. */
  78         req->wb_page    = page;
  79         atomic_set(&req->wb_complete, 0);
  80         req->wb_index   = page->index;
  81         page_cache_get(page);
  82         BUG_ON(PagePrivate(page));
  83         BUG_ON(!PageLocked(page));
  84         BUG_ON(page->mapping->host != inode);
  85         req->wb_offset  = offset;
  86         req->wb_pgbase  = offset;
  87         req->wb_bytes   = count;
  88         req->wb_context = get_nfs_open_context(ctx);
  89         kref_init(&req->wb_kref);
  90         return req;
  91 }
  92
  93 /**
  94  * nfs_unlock_request - Unlock request and wake up sleepers.
  95  * @req:
  96  */
  97 void nfs_unlock_request(struct nfs_page *req)
  98 {
  99         if (!NFS_WBACK_BUSY(req)) {
 100                 printk(KERN_ERR "NFS: Invalid unlock attempted\n");
 101                 BUG();
 102         }
 103         smp_mb__before_clear_bit();
 104         clear_bit(PG_BUSY, &req->wb_flags);
 105         smp_mb__after_clear_bit();
 106         wake_up_bit(&req->wb_flags, PG_BUSY);
 107         nfs_release_request(req);
 108 }
 109
 110 /**
 111  * nfs_set_page_tag_locked - Tag a request as locked
 112  * @req:
 113  */
 114 int nfs_set_page_tag_locked(struct nfs_page *req)
 115 {
 116         if (!nfs_lock_request_dontget(req))
 117                 return 0;
 118         if (req->wb_page != NULL)
 119                 radix_tree_tag_set(&NFS_I(req->wb_context->path.dentry->d_inode)->nfs_page_tree, req->wb_index, NFS_PAGE_TAG_LOCKED);
 120         return 1;
 121 }
 122
 123 /**
 124  * nfs_clear_page_tag_locked - Clear request tag and wake up sleepers
 125  */
 126 void nfs_clear_page_tag_locked(struct nfs_page *req)
 127 {
 128         if (req->wb_page != NULL) {
 129                 struct inode *inode = req->wb_context->path.dentry->d_inode;
 130                 struct nfs_inode *nfsi = NFS_I(inode);
 131
 132                 spin_lock(&inode->i_lock);
 133                 radix_tree_tag_clear(&nfsi->nfs_page_tree, req->wb_index, NFS_PAGE_TAG_LOCKED);
 134                 nfs_unlock_request(req);
 135                 spin_unlock(&inode->i_lock);
 136         } else
 137                 nfs_unlock_request(req);
 138 }
 139
 140 /**
 141  * nfs_clear_request - Free up all resources allocated to the request
 142  * @req:
 143  *
 144  * Release page and open context resources associated with a read/write
 145  * request after it has completed.
 146  */
 147 void nfs_clear_request(struct nfs_page *req)
 148 {
 149         struct page *page = req->wb_page;
 150         struct nfs_open_context *ctx = req->wb_context;
 151         struct nfs_lock_context *l_ctx = req->wb_lock_context;
 152
 153         if (page != NULL) {
 154                 page_cache_release(page);
 155                 req->wb_page = NULL;
 156         }
 157         if (l_ctx != NULL) {
 158                 nfs_put_lock_context(l_ctx);
 159                 req->wb_lock_context = NULL;
 160         }
 161         if (ctx != NULL) {
 162                 put_nfs_open_context(ctx);
 163                 req->wb_context = NULL;
 164         }
 165 }
 166
 167
 168 /**
 169  * nfs_release_request - Release the count on an NFS read/write request
 170  * @req: request to release
 171  *
 172  * Note: Should never be called with the spinlock held!
 173  */
 174 static void nfs_free_request(struct kref *kref)
 175 {
 176         struct nfs_page *req = container_of(kref, struct nfs_page, wb_kref);
 177
 178         /* Release struct file and open context */
 179         nfs_clear_request(req);
 180         nfs_page_free(req);
 181 }
 182
 183 void nfs_release_request(struct nfs_page *req)
 184 {
 185         kref_put(&req->wb_kref, nfs_free_request);
 186 }
 187
 188 static int nfs_wait_bit_uninterruptible(void *word)
 189 {
 190         io_schedule();
 191         return 0;
 192 }
 193
 194 /**
 195  * nfs_wait_on_request - Wait for a request to complete.
 196  * @req: request to wait upon.
 197  *
 198  * Interruptible by fatal signals only.
 199  * The user is responsible for holding a count on the request.
 200  */
 201 int
 202 nfs_wait_on_request(struct nfs_page *req)
 203 {
 204         return wait_on_bit(&req->wb_flags, PG_BUSY,
 205                         nfs_wait_bit_uninterruptible,
 206                         TASK_UNINTERRUPTIBLE);
 207 }
 208
 209 /**
 210  * nfs_pageio_init - initialise a page io descriptor
 211  * @desc: pointer to descriptor
 212  * @inode: pointer to inode
 213  * @doio: pointer to io function
 214  * @bsize: io block size
 215  * @io_flags: extra parameters for the io function
 216  */
 217 void nfs_pageio_init(struct nfs_pageio_descriptor *desc,
 218                      struct inode *inode,
 219                      int (*doio)(struct inode *, struct list_head *, unsigned int, size_t, int),
 220                      size_t bsize,
 221                      int io_flags)
 222 {
 223         INIT_LIST_HEAD(&desc->pg_list);
 224         desc->pg_bytes_written = 0;
 225         desc->pg_count = 0;
 226         desc->pg_bsize = bsize;
 227         desc->pg_base = 0;
 228         desc->pg_inode = inode;
 229         desc->pg_doio = doio;
 230         desc->pg_ioflags = io_flags;
 231         desc->pg_error = 0;
 232 }
 233
 234 /**
 235  * nfs_can_coalesce_requests - test two requests for compatibility
 236  * @prev: pointer to nfs_page
 237  * @req: pointer to nfs_page
 238  *
 239  * The nfs_page structures 'prev' and 'req' are compared to ensure that the
 240  * page data area they describe is contiguous, and that their RPC
 241  * credentials, NFSv4 open state, and lockowners are the same.
 242  *
 243  * Return 'true' if this is the case, else return 'false'.
 244  */
 245 static int nfs_can_coalesce_requests(struct nfs_page *prev,
 246                                      struct nfs_page *req)
 247 {
 248         if (req->wb_context->cred != prev->wb_context->cred)
 249                 return 0;
 250         if (req->wb_lock_context->lockowner != prev->wb_lock_context->lockowner)
 251                 return 0;
 252         if (req->wb_context->state != prev->wb_context->state)
 253                 return 0;
 254         if (req->wb_index != (prev->wb_index + 1))
 255                 return 0;
 256         if (req->wb_pgbase != 0)
 257                 return 0;
 258         if (prev->wb_pgbase + prev->wb_bytes != PAGE_CACHE_SIZE)
 259                 return 0;
 260         return 1;
 261 }
 262
 263 /**
 264  * nfs_pageio_do_add_request - Attempt to coalesce a request into a page list.
 265  * @desc: destination io descriptor
 266  * @req: request
 267  *
 268  * Returns true if the request 'req' was successfully coalesced into the
 269  * existing list of pages 'desc'.
 270  */
 271 static int nfs_pageio_do_add_request(struct nfs_pageio_descriptor *desc,
 272                                      struct nfs_page *req)
 273 {
 274         size_t newlen = req->wb_bytes;
 275
 276         if (desc->pg_count != 0) {
 277                 struct nfs_page *prev;
 278
 279                 /*
 280                  * FIXME: ideally we should be able to coalesce all requests
 281                  * that are not block boundary aligned, but currently this
 282                  * is problematic for the case of bsize < PAGE_CACHE_SIZE,
 283                  * since nfs_flush_multi and nfs_pagein_multi assume you
 284                  * can have only one struct nfs_page.
 285                  */
 286                 if (desc->pg_bsize < PAGE_SIZE)
 287                         return 0;
 288                 newlen += desc->pg_count;
 289                 if (newlen > desc->pg_bsize)
 290                         return 0;
 291                 prev = nfs_list_entry(desc->pg_list.prev);
 292                 if (!nfs_can_coalesce_requests(prev, req))
 293                         return 0;
 294         } else
 295                 desc->pg_base = req->wb_pgbase;
 296         nfs_list_remove_request(req);
 297         nfs_list_add_request(req, &desc->pg_list);
 298         desc->pg_count = newlen;
 299         return 1;
 300 }
 301
 302 /*
 303  * Helper for nfs_pageio_add_request and nfs_pageio_complete
 304  */
 305 static void nfs_pageio_doio(struct nfs_pageio_descriptor *desc)
 306 {
 307         if (!list_empty(&desc->pg_list)) {
 308                 int error = desc->pg_doio(desc->pg_inode,
 309                                           &desc->pg_list,
 310                                           nfs_page_array_len(desc->pg_base,
 311                                                              desc->pg_count),
 312                                           desc->pg_count,
 313                                           desc->pg_ioflags);
 314                 if (error < 0)
 315                         desc->pg_error = error;
 316                 else
 317                         desc->pg_bytes_written += desc->pg_count;
 318         }
 319         if (list_empty(&desc->pg_list)) {
 320                 desc->pg_count = 0;
 321                 desc->pg_base = 0;
 322         }
 323 }
 324
 325 /**
 326  * nfs_pageio_add_request - Attempt to coalesce a request into a page list.
 327  * @desc: destination io descriptor
 328  * @req: request
 329  *
 330  * Returns true if the request 'req' was successfully coalesced into the
 331  * existing list of pages 'desc'.
 332  */
 333 int nfs_pageio_add_request(struct nfs_pageio_descriptor *desc,
 334                            struct nfs_page *req)
 335 {
 336         while (!nfs_pageio_do_add_request(desc, req)) {
 337                 nfs_pageio_doio(desc);
 338                 if (desc->pg_error < 0)
 339                         return 0;
 340         }
 341         return 1;
 342 }
 343
 344 /**
 345  * nfs_pageio_complete - Complete I/O on an nfs_pageio_descriptor
 346  * @desc: pointer to io descriptor
 347  */
 348 void nfs_pageio_complete(struct nfs_pageio_descriptor *desc)
 349 {
 350         nfs_pageio_doio(desc);
 351 }
 352
 353 /**
 354  * nfs_pageio_cond_complete - Conditional I/O completion
 355  * @desc: pointer to io descriptor
 356  * @index: page index
 357  *
 358  * It is important to ensure that processes don't try to take locks
 359  * on non-contiguous ranges of pages as that might deadlock. This
 360  * function should be called before attempting to wait on a locked
 361  * nfs_page. It will complete the I/O if the page index 'index'
 362  * is not contiguous with the existing list of pages in 'desc'.
 363  */
 364 void nfs_pageio_cond_complete(struct nfs_pageio_descriptor *desc, pgoff_t index)
 365 {
 366         if (!list_empty(&desc->pg_list)) {
 367                 struct nfs_page *prev = nfs_list_entry(desc->pg_list.prev);
 368                 if (index != prev->wb_index + 1)
 369                         nfs_pageio_doio(desc);
 370         }
 371 }
 372
 373 #define NFS_SCAN_MAXENTRIES 16
 374 /**
 375  * nfs_scan_list - Scan a list for matching requests
 376  * @nfsi: NFS inode
 377  * @dst: Destination list
 378  * @idx_start: lower bound of page->index to scan
 379  * @npages: idx_start + npages sets the upper bound to scan.
 380  * @tag: tag to scan for
 381  *
 382  * Moves elements from one of the inode request lists.
 383  * If the number of requests is set to 0, the entire address_space
 384  * starting at index idx_start, is scanned.
 385  * The requests are *not* checked to ensure that they form a contiguous set.
 386  * You must be holding the inode's i_lock when calling this function
 387  */
 388 int nfs_scan_list(struct nfs_inode *nfsi,
 389                 struct list_head *dst, pgoff_t idx_start,
 390                 unsigned int npages, int tag)
 391 {
 392         struct nfs_page *pgvec[NFS_SCAN_MAXENTRIES];
 393         struct nfs_page *req;
 394         pgoff_t idx_end;
 395         int found, i;
 396         int res;
 397
 398         res = 0;
 399         if (npages == 0)
 400                 idx_end = ~0;
 401         else
 402                 idx_end = idx_start + npages - 1;
 403
 404         for (;;) {
 405                 found = radix_tree_gang_lookup_tag(&nfsi->nfs_page_tree,
 406                                 (void **)&pgvec[0], idx_start,
 407                                 NFS_SCAN_MAXENTRIES, tag);
 408                 if (found <= 0)
 409                         break;
 410                 for (i = 0; i < found; i++) {
 411                         req = pgvec[i];
 412                         if (req->wb_index > idx_end)
 413                                 goto out;
 414                         idx_start = req->wb_index + 1;
 415                         if (nfs_set_page_tag_locked(req)) {
 416                                 kref_get(&req->wb_kref);
 417                                 nfs_list_remove_request(req);
 418                                 radix_tree_tag_clear(&nfsi->nfs_page_tree,
 419                                                 req->wb_index, tag);
 420                                 nfs_list_add_request(req, dst);
 421                                 res++;
 422                                 if (res == INT_MAX)
 423                                         goto out;
 424                         }
 425                 }
 426                 /* for latency reduction */
 427                 cond_resched_lock(&nfsi->vfs_inode.i_lock);
 428         }
 429 out:
 430         return res;
 431 }
 432
 433 int __init nfs_init_nfspagecache(void)
 434 {
 435         nfs_page_cachep = kmem_cache_create("nfs_page",
 436                                             sizeof(struct nfs_page),
 437                                             0, SLAB_HWCACHE_ALIGN,
 438                                             NULL);
 439         if (nfs_page_cachep == NULL)
 440                 return -ENOMEM;
 441
 442         return 0;
 443 }
 444
 445 void nfs_destroy_nfspagecache(void)
 446 {
 447         kmem_cache_destroy(nfs_page_cachep);
 448 }
 449