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 = kmem_cache_zalloc(nfs_page_cachep, GFP_KERNEL);
  30         if (p)
  31                 INIT_LIST_HEAD(&p->wb_list);
  32         return p;
  33 }
  34
  35 static inline void
  36 nfs_page_free(struct nfs_page *p)
  37 {
  38         kmem_cache_free(nfs_page_cachep, p);
  39 }
  40
  41 /**
  42  * nfs_create_request - Create an NFS read/write request.
  43  * @file: file descriptor to use
  44  * @inode: inode to which the request is attached
  45  * @page: page to write
  46  * @offset: starting offset within the page for the write
  47  * @count: number of bytes to read/write
  48  *
  49  * The page must be locked by the caller. This makes sure we never
  50  * create two different requests for the same page.
  51  * User should ensure it is safe to sleep in this function.
  52  */
  53 struct nfs_page *
  54 nfs_create_request(struct nfs_open_context *ctx, struct inode *inode,
  55                    struct page *page,
  56                    unsigned int offset, unsigned int count)
  57 {
  58         struct nfs_page         *req;
  59
  60         /* try to allocate the request struct */
  61         req = nfs_page_alloc();
  62         if (req == NULL)
  63                 return ERR_PTR(-ENOMEM);
  64
  65         /* get lock context early so we can deal with alloc failures */
  66         req->wb_lock_context = nfs_get_lock_context(ctx);
  67         if (req->wb_lock_context == NULL) {
  68                 nfs_page_free(req);
  69                 return ERR_PTR(-ENOMEM);
  70         }
  71
  72         /* Initialize the request struct. Initially, we assume a
  73          * long write-back delay. This will be adjusted in
  74          * update_nfs_request below if the region is not locked. */
  75         req->wb_page    = page;
  76         atomic_set(&req->wb_complete, 0);
  77         req->wb_index   = page->index;
  78         page_cache_get(page);
  79         BUG_ON(PagePrivate(page));
  80         BUG_ON(!PageLocked(page));
  81         BUG_ON(page->mapping->host != inode);
  82         req->wb_offset  = offset;
  83         req->wb_pgbase  = offset;
  84         req->wb_bytes   = count;
  85         req->wb_context = get_nfs_open_context(ctx);
  86         kref_init(&req->wb_kref);
  87         return req;
  88 }
  89
  90 /**
  91  * nfs_unlock_request - Unlock request and wake up sleepers.
  92  * @req:
  93  */
  94 void nfs_unlock_request(struct nfs_page *req)
  95 {
  96         if (!NFS_WBACK_BUSY(req)) {
  97                 printk(KERN_ERR "NFS: Invalid unlock attempted\n");
  98                 BUG();
  99         }
 100         smp_mb__before_clear_bit();
 101         clear_bit(PG_BUSY, &req->wb_flags);
 102         smp_mb__after_clear_bit();
 103         wake_up_bit(&req->wb_flags, PG_BUSY);
 104         nfs_release_request(req);
 105 }
 106
 107 /**
 108  * nfs_set_page_tag_locked - Tag a request as locked
 109  * @req:
 110  */
 111 int nfs_set_page_tag_locked(struct nfs_page *req)
 112 {
 113         if (!nfs_lock_request_dontget(req))
 114                 return 0;
 115         if (test_bit(PG_MAPPED, &req->wb_flags))
 116                 radix_tree_tag_set(&NFS_I(req->wb_context->path.dentry->d_inode)->nfs_page_tree, req->wb_index, NFS_PAGE_TAG_LOCKED);
 117         return 1;
 118 }
 119
 120 /**
 121  * nfs_clear_page_tag_locked - Clear request tag and wake up sleepers
 122  */
 123 void nfs_clear_page_tag_locked(struct nfs_page *req)
 124 {
 125         if (test_bit(PG_MAPPED, &req->wb_flags)) {
 126                 struct inode *inode = req->wb_context->path.dentry->d_inode;
 127                 struct nfs_inode *nfsi = NFS_I(inode);
 128
 129                 spin_lock(&inode->i_lock);
 130                 radix_tree_tag_clear(&nfsi->nfs_page_tree, req->wb_index, NFS_PAGE_TAG_LOCKED);
 131                 nfs_unlock_request(req);
 132                 spin_unlock(&inode->i_lock);
 133         } else
 134                 nfs_unlock_request(req);
 135 }
 136
 137 /**
 138  * nfs_clear_request - Free up all resources allocated to the request
 139  * @req:
 140  *
 141  * Release page and open context resources associated with a read/write
 142  * request after it has completed.
 143  */
 144 void nfs_clear_request(struct nfs_page *req)
 145 {
 146         struct page *page = req->wb_page;
 147         struct nfs_open_context *ctx = req->wb_context;
 148         struct nfs_lock_context *l_ctx = req->wb_lock_context;
 149
 150         if (page != NULL) {
 151                 page_cache_release(page);
 152                 req->wb_page = NULL;
 153         }
 154         if (l_ctx != NULL) {
 155                 nfs_put_lock_context(l_ctx);
 156                 req->wb_lock_context = NULL;
 157         }
 158         if (ctx != NULL) {
 159                 put_nfs_open_context(ctx);
 160                 req->wb_context = NULL;
 161         }
 162 }
 163
 164
 165 /**
 166  * nfs_release_request - Release the count on an NFS read/write request
 167  * @req: request to release
 168  *
 169  * Note: Should never be called with the spinlock held!
 170  */
 171 static void nfs_free_request(struct kref *kref)
 172 {
 173         struct nfs_page *req = container_of(kref, struct nfs_page, wb_kref);
 174
 175         /* Release struct file and open context */
 176         nfs_clear_request(req);
 177         nfs_page_free(req);
 178 }
 179
 180 void nfs_release_request(struct nfs_page *req)
 181 {
 182         kref_put(&req->wb_kref, nfs_free_request);
 183 }
 184
 185 static int nfs_wait_bit_uninterruptible(void *word)
 186 {
 187         io_schedule();
 188         return 0;
 189 }
 190
 191 /**
 192  * nfs_wait_on_request - Wait for a request to complete.
 193  * @req: request to wait upon.
 194  *
 195  * Interruptible by fatal signals only.
 196  * The user is responsible for holding a count on the request.
 197  */
 198 int
 199 nfs_wait_on_request(struct nfs_page *req)
 200 {
 201         return wait_on_bit(&req->wb_flags, PG_BUSY,
 202                         nfs_wait_bit_uninterruptible,
 203                         TASK_UNINTERRUPTIBLE);
 204 }
 205
 206 /**
 207  * nfs_pageio_init - initialise a page io descriptor
 208  * @desc: pointer to descriptor
 209  * @inode: pointer to inode
 210  * @doio: pointer to io function
 211  * @bsize: io block size
 212  * @io_flags: extra parameters for the io function
 213  */
 214 void nfs_pageio_init(struct nfs_pageio_descriptor *desc,
 215                      struct inode *inode,
 216                      int (*doio)(struct inode *, struct list_head *, unsigned int, size_t, int),
 217                      size_t bsize,
 218                      int io_flags)
 219 {
 220         INIT_LIST_HEAD(&desc->pg_list);
 221         desc->pg_bytes_written = 0;
 222         desc->pg_count = 0;
 223         desc->pg_bsize = bsize;
 224         desc->pg_base = 0;
 225         desc->pg_inode = inode;
 226         desc->pg_doio = doio;
 227         desc->pg_ioflags = io_flags;
 228         desc->pg_error = 0;
 229 }
 230
 231 /**
 232  * nfs_can_coalesce_requests - test two requests for compatibility
 233  * @prev: pointer to nfs_page
 234  * @req: pointer to nfs_page
 235  *
 236  * The nfs_page structures 'prev' and 'req' are compared to ensure that the
 237  * page data area they describe is contiguous, and that their RPC
 238  * credentials, NFSv4 open state, and lockowners are the same.
 239  *
 240  * Return 'true' if this is the case, else return 'false'.
 241  */
 242 static int nfs_can_coalesce_requests(struct nfs_page *prev,
 243                                      struct nfs_page *req)
 244 {
 245         if (req->wb_context->cred != prev->wb_context->cred)
 246                 return 0;
 247         if (req->wb_lock_context->lockowner != prev->wb_lock_context->lockowner)
 248                 return 0;
 249         if (req->wb_context->state != prev->wb_context->state)
 250                 return 0;
 251         if (req->wb_index != (prev->wb_index + 1))
 252                 return 0;
 253         if (req->wb_pgbase != 0)
 254                 return 0;
 255         if (prev->wb_pgbase + prev->wb_bytes != PAGE_CACHE_SIZE)
 256                 return 0;
 257         return 1;
 258 }
 259
 260 /**
 261  * nfs_pageio_do_add_request - Attempt to coalesce a request into a page list.
 262  * @desc: destination io descriptor
 263  * @req: request
 264  *
 265  * Returns true if the request 'req' was successfully coalesced into the
 266  * existing list of pages 'desc'.
 267  */
 268 static int nfs_pageio_do_add_request(struct nfs_pageio_descriptor *desc,
 269                                      struct nfs_page *req)
 270 {
 271         size_t newlen = req->wb_bytes;
 272
 273         if (desc->pg_count != 0) {
 274                 struct nfs_page *prev;
 275
 276                 /*
 277                  * FIXME: ideally we should be able to coalesce all requests
 278                  * that are not block boundary aligned, but currently this
 279                  * is problematic for the case of bsize < PAGE_CACHE_SIZE,
 280                  * since nfs_flush_multi and nfs_pagein_multi assume you
 281                  * can have only one struct nfs_page.
 282                  */
 283                 if (desc->pg_bsize < PAGE_SIZE)
 284                         return 0;
 285                 newlen += desc->pg_count;
 286                 if (newlen > desc->pg_bsize)
 287                         return 0;
 288                 prev = nfs_list_entry(desc->pg_list.prev);
 289                 if (!nfs_can_coalesce_requests(prev, req))
 290                         return 0;
 291         } else
 292                 desc->pg_base = req->wb_pgbase;
 293         nfs_list_remove_request(req);
 294         nfs_list_add_request(req, &desc->pg_list);
 295         desc->pg_count = newlen;
 296         return 1;
 297 }
 298
 299 /*
 300  * Helper for nfs_pageio_add_request and nfs_pageio_complete
 301  */
 302 static void nfs_pageio_doio(struct nfs_pageio_descriptor *desc)
 303 {
 304         if (!list_empty(&desc->pg_list)) {
 305                 int error = desc->pg_doio(desc->pg_inode,
 306                                           &desc->pg_list,
 307                                           nfs_page_array_len(desc->pg_base,
 308                                                              desc->pg_count),
 309                                           desc->pg_count,
 310                                           desc->pg_ioflags);
 311                 if (error < 0)
 312                         desc->pg_error = error;
 313                 else
 314                         desc->pg_bytes_written += desc->pg_count;
 315         }
 316         if (list_empty(&desc->pg_list)) {
 317                 desc->pg_count = 0;
 318                 desc->pg_base = 0;
 319         }
 320 }
 321
 322 /**
 323  * nfs_pageio_add_request - Attempt to coalesce a request into a page list.
 324  * @desc: destination io descriptor
 325  * @req: request
 326  *
 327  * Returns true if the request 'req' was successfully coalesced into the
 328  * existing list of pages 'desc'.
 329  */
 330 int nfs_pageio_add_request(struct nfs_pageio_descriptor *desc,
 331                            struct nfs_page *req)
 332 {
 333         while (!nfs_pageio_do_add_request(desc, req)) {
 334                 nfs_pageio_doio(desc);
 335                 if (desc->pg_error < 0)
 336                         return 0;
 337         }
 338         return 1;
 339 }
 340
 341 /**
 342  * nfs_pageio_complete - Complete I/O on an nfs_pageio_descriptor
 343  * @desc: pointer to io descriptor
 344  */
 345 void nfs_pageio_complete(struct nfs_pageio_descriptor *desc)
 346 {
 347         nfs_pageio_doio(desc);
 348 }
 349
 350 /**
 351  * nfs_pageio_cond_complete - Conditional I/O completion
 352  * @desc: pointer to io descriptor
 353  * @index: page index
 354  *
 355  * It is important to ensure that processes don't try to take locks
 356  * on non-contiguous ranges of pages as that might deadlock. This
 357  * function should be called before attempting to wait on a locked
 358  * nfs_page. It will complete the I/O if the page index 'index'
 359  * is not contiguous with the existing list of pages in 'desc'.
 360  */
 361 void nfs_pageio_cond_complete(struct nfs_pageio_descriptor *desc, pgoff_t index)
 362 {
 363         if (!list_empty(&desc->pg_list)) {
 364                 struct nfs_page *prev = nfs_list_entry(desc->pg_list.prev);
 365                 if (index != prev->wb_index + 1)
 366                         nfs_pageio_doio(desc);
 367         }
 368 }
 369
 370 #define NFS_SCAN_MAXENTRIES 16
 371 /**
 372  * nfs_scan_list - Scan a list for matching requests
 373  * @nfsi: NFS inode
 374  * @dst: Destination list
 375  * @idx_start: lower bound of page->index to scan
 376  * @npages: idx_start + npages sets the upper bound to scan.
 377  * @tag: tag to scan for
 378  *
 379  * Moves elements from one of the inode request lists.
 380  * If the number of requests is set to 0, the entire address_space
 381  * starting at index idx_start, is scanned.
 382  * The requests are *not* checked to ensure that they form a contiguous set.
 383  * You must be holding the inode's i_lock when calling this function
 384  */
 385 int nfs_scan_list(struct nfs_inode *nfsi,
 386                 struct list_head *dst, pgoff_t idx_start,
 387                 unsigned int npages, int tag)
 388 {
 389         struct nfs_page *pgvec[NFS_SCAN_MAXENTRIES];
 390         struct nfs_page *req;
 391         pgoff_t idx_end;
 392         int found, i;
 393         int res;
 394
 395         res = 0;
 396         if (npages == 0)
 397                 idx_end = ~0;
 398         else
 399                 idx_end = idx_start + npages - 1;
 400
 401         for (;;) {
 402                 found = radix_tree_gang_lookup_tag(&nfsi->nfs_page_tree,
 403                                 (void **)&pgvec[0], idx_start,
 404                                 NFS_SCAN_MAXENTRIES, tag);
 405                 if (found <= 0)
 406                         break;
 407                 for (i = 0; i < found; i++) {
 408                         req = pgvec[i];
 409                         if (req->wb_index > idx_end)
 410                                 goto out;
 411                         idx_start = req->wb_index + 1;
 412                         if (nfs_set_page_tag_locked(req)) {
 413                                 kref_get(&req->wb_kref);
 414                                 nfs_list_remove_request(req);
 415                                 radix_tree_tag_clear(&nfsi->nfs_page_tree,
 416                                                 req->wb_index, tag);
 417                                 nfs_list_add_request(req, dst);
 418                                 res++;
 419                                 if (res == INT_MAX)
 420                                         goto out;
 421                         }
 422                 }
 423                 /* for latency reduction */
 424                 cond_resched_lock(&nfsi->vfs_inode.i_lock);
 425         }
 426 out:
 427         return res;
 428 }
 429
 430 int __init nfs_init_nfspagecache(void)
 431 {
 432         nfs_page_cachep = kmem_cache_create("nfs_page",
 433                                             sizeof(struct nfs_page),
 434                                             0, SLAB_HWCACHE_ALIGN,
 435                                             NULL);
 436         if (nfs_page_cachep == NULL)
 437                 return -ENOMEM;
 438
 439         return 0;
 440 }
 441
 442 void nfs_destroy_nfspagecache(void)
 443 {
 444         kmem_cache_destroy(nfs_page_cachep);
 445 }
 446