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