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/config.h>
  13 #include <linux/slab.h>
  14 #include <linux/file.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 #define NFS_PARANOIA 1
  23
  24 static kmem_cache_t *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, SLAB_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, and avoids
  54  * a possible deadlock when we reach the hard limit on the number
  55  * of dirty pages.
  56  * User should ensure it is safe to sleep in this function.
  57  */
  58 struct nfs_page *
  59 nfs_create_request(struct nfs_open_context *ctx, struct inode *inode,
  60                    struct page *page,
  61                    unsigned int offset, unsigned int count)
  62 {
  63         struct nfs_server *server = NFS_SERVER(inode);
  64         struct nfs_page         *req;
  65
  66         /* Deal with hard limits.  */
  67         for (;;) {
  68                 /* try to allocate the request struct */
  69                 req = nfs_page_alloc();
  70                 if (req != NULL)
  71                         break;
  72
  73                 /* Try to free up at least one request in order to stay
  74                  * below the hard limit
  75                  */
  76                 if (signalled() && (server->flags & NFS_MOUNT_INTR))
  77                         return ERR_PTR(-ERESTARTSYS);
  78                 yield();
  79         }
  80
  81         /* Initialize the request struct. Initially, we assume a
  82          * long write-back delay. This will be adjusted in
  83          * update_nfs_request below if the region is not locked. */
  84         req->wb_page    = page;
  85         atomic_set(&req->wb_complete, 0);
  86         req->wb_index   = page->index;
  87         page_cache_get(page);
  88         req->wb_offset  = offset;
  89         req->wb_pgbase  = offset;
  90         req->wb_bytes   = count;
  91         atomic_set(&req->wb_count, 1);
  92         req->wb_context = get_nfs_open_context(ctx);
  93
  94         return req;
  95 }
  96
  97 /**
  98  * nfs_unlock_request - Unlock request and wake up sleepers.
  99  * @req:
 100  */
 101 void nfs_unlock_request(struct nfs_page *req)
 102 {
 103         if (!NFS_WBACK_BUSY(req)) {
 104                 printk(KERN_ERR "NFS: Invalid unlock attempted\n");
 105                 BUG();
 106         }
 107         smp_mb__before_clear_bit();
 108         clear_bit(PG_BUSY, &req->wb_flags);
 109         smp_mb__after_clear_bit();
 110         wake_up_bit(&req->wb_flags, PG_BUSY);
 111         nfs_release_request(req);
 112 }
 113
 114 /**
 115  * nfs_set_page_writeback_locked - Lock a request for writeback
 116  * @req:
 117  */
 118 int nfs_set_page_writeback_locked(struct nfs_page *req)
 119 {
 120         struct nfs_inode *nfsi = NFS_I(req->wb_context->dentry->d_inode);
 121
 122         if (!nfs_lock_request(req))
 123                 return 0;
 124         radix_tree_tag_set(&nfsi->nfs_page_tree, req->wb_index, NFS_PAGE_TAG_WRITEBACK);
 125         return 1;
 126 }
 127
 128 /**
 129  * nfs_clear_page_writeback - Unlock request and wake up sleepers
 130  */
 131 void nfs_clear_page_writeback(struct nfs_page *req)
 132 {
 133         struct nfs_inode *nfsi = NFS_I(req->wb_context->dentry->d_inode);
 134
 135         spin_lock(&nfsi->req_lock);
 136         radix_tree_tag_clear(&nfsi->nfs_page_tree, req->wb_index, NFS_PAGE_TAG_WRITEBACK);
 137         spin_unlock(&nfsi->req_lock);
 138         nfs_unlock_request(req);
 139 }
 140
 141 /**
 142  * nfs_clear_request - Free up all resources allocated to the request
 143  * @req:
 144  *
 145  * Release page resources associated with a write request after it
 146  * has completed.
 147  */
 148 void nfs_clear_request(struct nfs_page *req)
 149 {
 150         if (req->wb_page) {
 151                 page_cache_release(req->wb_page);
 152                 req->wb_page = NULL;
 153         }
 154 }
 155
 156
 157 /**
 158  * nfs_release_request - Release the count on an NFS read/write request
 159  * @req: request to release
 160  *
 161  * Note: Should never be called with the spinlock held!
 162  */
 163 void
 164 nfs_release_request(struct nfs_page *req)
 165 {
 166         if (!atomic_dec_and_test(&req->wb_count))
 167                 return;
 168
 169 #ifdef NFS_PARANOIA
 170         BUG_ON (!list_empty(&req->wb_list));
 171         BUG_ON (NFS_WBACK_BUSY(req));
 172 #endif
 173
 174         /* Release struct file or cached credential */
 175         nfs_clear_request(req);
 176         put_nfs_open_context(req->wb_context);
 177         nfs_page_free(req);
 178 }
 179
 180 /**
 181  * nfs_list_add_request - Insert a request into a sorted list
 182  * @req: request
 183  * @head: head of list into which to insert the request.
 184  *
 185  * Note that the wb_list is sorted by page index in order to facilitate
 186  * coalescing of requests.
 187  * We use an insertion sort that is optimized for the case of appended
 188  * writes.
 189  */
 190 void
 191 nfs_list_add_request(struct nfs_page *req, struct list_head *head)
 192 {
 193         struct list_head *pos;
 194
 195 #ifdef NFS_PARANOIA
 196         if (!list_empty(&req->wb_list)) {
 197                 printk(KERN_ERR "NFS: Add to list failed!\n");
 198                 BUG();
 199         }
 200 #endif
 201         list_for_each_prev(pos, head) {
 202                 struct nfs_page *p = nfs_list_entry(pos);
 203                 if (p->wb_index < req->wb_index)
 204                         break;
 205         }
 206         list_add(&req->wb_list, pos);
 207         req->wb_list_head = head;
 208 }
 209
 210 static int nfs_wait_bit_interruptible(void *word)
 211 {
 212         int ret = 0;
 213
 214         if (signal_pending(current))
 215                 ret = -ERESTARTSYS;
 216         else
 217                 schedule();
 218         return ret;
 219 }
 220
 221 /**
 222  * nfs_wait_on_request - Wait for a request to complete.
 223  * @req: request to wait upon.
 224  *
 225  * Interruptible by signals only if mounted with intr flag.
 226  * The user is responsible for holding a count on the request.
 227  */
 228 int
 229 nfs_wait_on_request(struct nfs_page *req)
 230 {
 231         struct rpc_clnt *clnt = NFS_CLIENT(req->wb_context->dentry->d_inode);
 232         sigset_t oldmask;
 233         int ret = 0;
 234
 235         if (!test_bit(PG_BUSY, &req->wb_flags))
 236                 goto out;
 237         /*
 238          * Note: the call to rpc_clnt_sigmask() suffices to ensure that we
 239          *       are not interrupted if intr flag is not set
 240          */
 241         rpc_clnt_sigmask(clnt, &oldmask);
 242         ret = out_of_line_wait_on_bit(&req->wb_flags, PG_BUSY,
 243                         nfs_wait_bit_interruptible, TASK_INTERRUPTIBLE);
 244         rpc_clnt_sigunmask(clnt, &oldmask);
 245 out:
 246         return ret;
 247 }
 248
 249 /**
 250  * nfs_coalesce_requests - Split coalesced requests out from a list.
 251  * @head: source list
 252  * @dst: destination list
 253  * @nmax: maximum number of requests to coalesce
 254  *
 255  * Moves a maximum of 'nmax' elements from one list to another.
 256  * The elements are checked to ensure that they form a contiguous set
 257  * of pages, and that the RPC credentials are the same.
 258  */
 259 int
 260 nfs_coalesce_requests(struct list_head *head, struct list_head *dst,
 261                       unsigned int nmax)
 262 {
 263         struct nfs_page         *req = NULL;
 264         unsigned int            npages = 0;
 265
 266         while (!list_empty(head)) {
 267                 struct nfs_page *prev = req;
 268
 269                 req = nfs_list_entry(head->next);
 270                 if (prev) {
 271                         if (req->wb_context->cred != prev->wb_context->cred)
 272                                 break;
 273                         if (req->wb_context->lockowner != prev->wb_context->lockowner)
 274                                 break;
 275                         if (req->wb_context->state != prev->wb_context->state)
 276                                 break;
 277                         if (req->wb_index != (prev->wb_index + 1))
 278                                 break;
 279
 280                         if (req->wb_pgbase != 0)
 281                                 break;
 282                 }
 283                 nfs_list_remove_request(req);
 284                 nfs_list_add_request(req, dst);
 285                 npages++;
 286                 if (req->wb_pgbase + req->wb_bytes != PAGE_CACHE_SIZE)
 287                         break;
 288                 if (npages >= nmax)
 289                         break;
 290         }
 291         return npages;
 292 }
 293
 294 /**
 295  * nfs_scan_list - Scan a list for matching requests
 296  * @head: One of the NFS inode request lists
 297  * @dst: Destination list
 298  * @idx_start: lower bound of page->index to scan
 299  * @npages: idx_start + npages sets the upper bound to scan.
 300  *
 301  * Moves elements from one of the inode request lists.
 302  * If the number of requests is set to 0, the entire address_space
 303  * starting at index idx_start, is scanned.
 304  * The requests are *not* checked to ensure that they form a contiguous set.
 305  * You must be holding the inode's req_lock when calling this function
 306  */
 307 int
 308 nfs_scan_list(struct list_head *head, struct list_head *dst,
 309               unsigned long idx_start, unsigned int npages)
 310 {
 311         struct list_head        *pos, *tmp;
 312         struct nfs_page         *req;
 313         unsigned long           idx_end;
 314         int                     res;
 315
 316         res = 0;
 317         if (npages == 0)
 318                 idx_end = ~0;
 319         else
 320                 idx_end = idx_start + npages - 1;
 321
 322         list_for_each_safe(pos, tmp, head) {
 323
 324                 req = nfs_list_entry(pos);
 325
 326                 if (req->wb_index < idx_start)
 327                         continue;
 328                 if (req->wb_index > idx_end)
 329                         break;
 330
 331                 if (!nfs_set_page_writeback_locked(req))
 332                         continue;
 333                 nfs_list_remove_request(req);
 334                 nfs_list_add_request(req, dst);
 335                 res++;
 336         }
 337         return res;
 338 }
 339
 340 int nfs_init_nfspagecache(void)
 341 {
 342         nfs_page_cachep = kmem_cache_create("nfs_page",
 343                                             sizeof(struct nfs_page),
 344                                             0, SLAB_HWCACHE_ALIGN,
 345                                             NULL, NULL);
 346         if (nfs_page_cachep == NULL)
 347                 return -ENOMEM;
 348
 349         return 0;
 350 }
 351
 352 void nfs_destroy_nfspagecache(void)
 353 {
 354         if (kmem_cache_destroy(nfs_page_cachep))
 355                 printk(KERN_INFO "nfs_page: not all structures were freed\n");
 356 }
 357