kernel/drivers/staging/lustre/lustre/llite/rw26.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2011, 2012, Intel Corporation.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * lustre/lustre/llite/rw26.c
  37  *
  38  * Lustre Lite I/O page cache routines for the 2.5/2.6 kernel version
  39  */
  40
  41 #include <linux/kernel.h>
  42 #include <linux/mm.h>
  43 #include <linux/string.h>
  44 #include <linux/stat.h>
  45 #include <linux/errno.h>
  46 #include <linux/unistd.h>
  47 #include <linux/uaccess.h>
  48
  49 #include <linux/migrate.h>
  50 #include <linux/fs.h>
  51 #include <linux/buffer_head.h>
  52 #include <linux/mpage.h>
  53 #include <linux/writeback.h>
  54 #include <linux/pagemap.h>
  55
  56 #define DEBUG_SUBSYSTEM S_LLITE
  57
  58 #include "../include/lustre_lite.h"
  59 #include "llite_internal.h"
  60 #include "../include/linux/lustre_compat25.h"
  61
  62 /**
  63  * Implements Linux VM address_space::invalidatepage() method. This method is
  64  * called when the page is truncate from a file, either as a result of
  65  * explicit truncate, or when inode is removed from memory (as a result of
  66  * final iput(), umount, or memory pressure induced icache shrinking).
  67  *
  68  * [0, offset] bytes of the page remain valid (this is for a case of not-page
  69  * aligned truncate). Lustre leaves partially truncated page in the cache,
  70  * relying on struct inode::i_size to limit further accesses.
  71  */
  72 static void ll_invalidatepage(struct page *vmpage, unsigned int offset,
  73                               unsigned int length)
  74 {
  75         struct inode     *inode;
  76         struct lu_env    *env;
  77         struct cl_page   *page;
  78         struct cl_object *obj;
  79
  80         int refcheck;
  81
  82         LASSERT(PageLocked(vmpage));
  83         LASSERT(!PageWriteback(vmpage));
  84
  85         /*
  86          * It is safe to not check anything in invalidatepage/releasepage
  87          * below because they are run with page locked and all our io is
  88          * happening with locked page too
  89          */
  90         if (offset == 0 && length == PAGE_CACHE_SIZE) {
  91                 env = cl_env_get(&refcheck);
  92                 if (!IS_ERR(env)) {
  93                         inode = vmpage->mapping->host;
  94                         obj = ll_i2info(inode)->lli_clob;
  95                         if (obj != NULL) {
  96                                 page = cl_vmpage_page(vmpage, obj);
  97                                 if (page != NULL) {
  98                                         lu_ref_add(&page->cp_reference,
  99                                                    "delete", vmpage);
 100                                         cl_page_delete(env, page);
 101                                         lu_ref_del(&page->cp_reference,
 102                                                    "delete", vmpage);
 103                                         cl_page_put(env, page);
 104                                 }
 105                         } else
 106                                 LASSERT(vmpage->private == 0);
 107                         cl_env_put(env, &refcheck);
 108                 }
 109         }
 110 }
 111
 112 #ifdef HAVE_RELEASEPAGE_WITH_INT
 113 #define RELEASEPAGE_ARG_TYPE int
 114 #else
 115 #define RELEASEPAGE_ARG_TYPE gfp_t
 116 #endif
 117 static int ll_releasepage(struct page *vmpage, RELEASEPAGE_ARG_TYPE gfp_mask)
 118 {
 119         struct cl_env_nest nest;
 120         struct lu_env     *env;
 121         struct cl_object  *obj;
 122         struct cl_page    *page;
 123         struct address_space *mapping;
 124         int result;
 125
 126         LASSERT(PageLocked(vmpage));
 127         if (PageWriteback(vmpage) || PageDirty(vmpage))
 128                 return 0;
 129
 130         mapping = vmpage->mapping;
 131         if (mapping == NULL)
 132                 return 1;
 133
 134         obj = ll_i2info(mapping->host)->lli_clob;
 135         if (obj == NULL)
 136                 return 1;
 137
 138         /* 1 for page allocator, 1 for cl_page and 1 for page cache */
 139         if (page_count(vmpage) > 3)
 140                 return 0;
 141
 142         /* TODO: determine what gfp should be used by @gfp_mask. */
 143         env = cl_env_nested_get(&nest);
 144         if (IS_ERR(env))
 145                 /* If we can't allocate an env we won't call cl_page_put()
 146                  * later on which further means it's impossible to drop
 147                  * page refcount by cl_page, so ask kernel to not free
 148                  * this page. */
 149                 return 0;
 150
 151         page = cl_vmpage_page(vmpage, obj);
 152         result = page == NULL;
 153         if (page != NULL) {
 154                 if (!cl_page_in_use(page)) {
 155                         result = 1;
 156                         cl_page_delete(env, page);
 157                 }
 158                 cl_page_put(env, page);
 159         }
 160         cl_env_nested_put(&nest, env);
 161         return result;
 162 }
 163
 164 static int ll_set_page_dirty(struct page *vmpage)
 165 {
 166 #if 0
 167         struct cl_page    *page = vvp_vmpage_page_transient(vmpage);
 168         struct vvp_object *obj  = cl_inode2vvp(vmpage->mapping->host);
 169         struct vvp_page   *cpg;
 170
 171         /*
 172          * XXX should page method be called here?
 173          */
 174         LASSERT(&obj->co_cl == page->cp_obj);
 175         cpg = cl2vvp_page(cl_page_at(page, &vvp_device_type));
 176         /*
 177          * XXX cannot do much here, because page is possibly not locked:
 178          * sys_munmap()->...
 179          *     ->unmap_page_range()->zap_pte_range()->set_page_dirty().
 180          */
 181         vvp_write_pending(obj, cpg);
 182 #endif
 183         return __set_page_dirty_nobuffers(vmpage);
 184 }
 185
 186 #define MAX_DIRECTIO_SIZE (2*1024*1024*1024UL)
 187
 188 static inline int ll_get_user_pages(int rw, unsigned long user_addr,
 189                                     size_t size, struct page ***pages,
 190                                     int *max_pages)
 191 {
 192         int result = -ENOMEM;
 193
 194         /* set an arbitrary limit to prevent arithmetic overflow */
 195         if (size > MAX_DIRECTIO_SIZE) {
 196                 *pages = NULL;
 197                 return -EFBIG;
 198         }
 199
 200         *max_pages = (user_addr + size + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
 201         *max_pages -= user_addr >> PAGE_CACHE_SHIFT;
 202
 203         OBD_ALLOC_LARGE(*pages, *max_pages * sizeof(**pages));
 204         if (*pages) {
 205                 result = get_user_pages_fast(user_addr, *max_pages,
 206                                              (rw == READ), *pages);
 207                 if (unlikely(result <= 0))
 208                         OBD_FREE_LARGE(*pages, *max_pages * sizeof(**pages));
 209         }
 210
 211         return result;
 212 }
 213
 214 /*  ll_free_user_pages - tear down page struct array
 215  *  @pages: array of page struct pointers underlying target buffer */
 216 static void ll_free_user_pages(struct page **pages, int npages, int do_dirty)
 217 {
 218         int i;
 219
 220         for (i = 0; i < npages; i++) {
 221                 if (do_dirty)
 222                         set_page_dirty_lock(pages[i]);
 223                 page_cache_release(pages[i]);
 224         }
 225         kvfree(pages);
 226 }
 227
 228 ssize_t ll_direct_rw_pages(const struct lu_env *env, struct cl_io *io,
 229                            int rw, struct inode *inode,
 230                            struct ll_dio_pages *pv)
 231 {
 232         struct cl_page    *clp;
 233         struct cl_2queue  *queue;
 234         struct cl_object  *obj = io->ci_obj;
 235         int i;
 236         ssize_t rc = 0;
 237         loff_t file_offset  = pv->ldp_start_offset;
 238         long size          = pv->ldp_size;
 239         int page_count      = pv->ldp_nr;
 240         struct page **pages = pv->ldp_pages;
 241         long page_size      = cl_page_size(obj);
 242         bool do_io;
 243         int  io_pages       = 0;
 244
 245         queue = &io->ci_queue;
 246         cl_2queue_init(queue);
 247         for (i = 0; i < page_count; i++) {
 248                 if (pv->ldp_offsets)
 249                     file_offset = pv->ldp_offsets[i];
 250
 251                 LASSERT(!(file_offset & (page_size - 1)));
 252                 clp = cl_page_find(env, obj, cl_index(obj, file_offset),
 253                                    pv->ldp_pages[i], CPT_TRANSIENT);
 254                 if (IS_ERR(clp)) {
 255                         rc = PTR_ERR(clp);
 256                         break;
 257                 }
 258
 259                 rc = cl_page_own(env, io, clp);
 260                 if (rc) {
 261                         LASSERT(clp->cp_state == CPS_FREEING);
 262                         cl_page_put(env, clp);
 263                         break;
 264                 }
 265
 266                 do_io = true;
 267
 268                 /* check the page type: if the page is a host page, then do
 269                  * write directly */
 270                 if (clp->cp_type == CPT_CACHEABLE) {
 271                         struct page *vmpage = cl_page_vmpage(env, clp);
 272                         struct page *src_page;
 273                         struct page *dst_page;
 274                         void       *src;
 275                         void       *dst;
 276
 277                         src_page = (rw == WRITE) ? pages[i] : vmpage;
 278                         dst_page = (rw == WRITE) ? vmpage : pages[i];
 279
 280                         src = kmap_atomic(src_page);
 281                         dst = kmap_atomic(dst_page);
 282                         memcpy(dst, src, min(page_size, size));
 283                         kunmap_atomic(dst);
 284                         kunmap_atomic(src);
 285
 286                         /* make sure page will be added to the transfer by
 287                          * cl_io_submit()->...->vvp_page_prep_write(). */
 288                         if (rw == WRITE)
 289                                 set_page_dirty(vmpage);
 290
 291                         if (rw == READ) {
 292                                 /* do not issue the page for read, since it
 293                                  * may reread a ra page which has NOT uptodate
 294                                  * bit set. */
 295                                 cl_page_disown(env, io, clp);
 296                                 do_io = false;
 297                         }
 298                 }
 299
 300                 if (likely(do_io)) {
 301                         cl_2queue_add(queue, clp);
 302
 303                         /*
 304                          * Set page clip to tell transfer formation engine
 305                          * that page has to be sent even if it is beyond KMS.
 306                          */
 307                         cl_page_clip(env, clp, 0, min(size, page_size));
 308
 309                         ++io_pages;
 310                 }
 311
 312                 /* drop the reference count for cl_page_find */
 313                 cl_page_put(env, clp);
 314                 size -= page_size;
 315                 file_offset += page_size;
 316         }
 317
 318         if (rc == 0 && io_pages) {
 319                 rc = cl_io_submit_sync(env, io,
 320                                        rw == READ ? CRT_READ : CRT_WRITE,
 321                                        queue, 0);
 322         }
 323         if (rc == 0)
 324                 rc = pv->ldp_size;
 325
 326         cl_2queue_discard(env, io, queue);
 327         cl_2queue_disown(env, io, queue);
 328         cl_2queue_fini(env, queue);
 329         return rc;
 330 }
 331 EXPORT_SYMBOL(ll_direct_rw_pages);
 332
 333 static ssize_t ll_direct_IO_26_seg(const struct lu_env *env, struct cl_io *io,
 334                                    int rw, struct inode *inode,
 335                                    struct address_space *mapping,
 336                                    size_t size, loff_t file_offset,
 337                                    struct page **pages, int page_count)
 338 {
 339     struct ll_dio_pages pvec = { .ldp_pages     = pages,
 340                                  .ldp_nr           = page_count,
 341                                  .ldp_size       = size,
 342                                  .ldp_offsets      = NULL,
 343                                  .ldp_start_offset = file_offset
 344                                };
 345
 346     return ll_direct_rw_pages(env, io, rw, inode, &pvec);
 347 }
 348
 349 #ifdef KMALLOC_MAX_SIZE
 350 #define MAX_MALLOC KMALLOC_MAX_SIZE
 351 #else
 352 #define MAX_MALLOC (128 * 1024)
 353 #endif
 354
 355 /* This is the maximum size of a single O_DIRECT request, based on the
 356  * kmalloc limit.  We need to fit all of the brw_page structs, each one
 357  * representing PAGE_SIZE worth of user data, into a single buffer, and
 358  * then truncate this to be a full-sized RPC.  For 4kB PAGE_SIZE this is
 359  * up to 22MB for 128kB kmalloc and up to 682MB for 4MB kmalloc. */
 360 #define MAX_DIO_SIZE ((MAX_MALLOC / sizeof(struct brw_page) * PAGE_CACHE_SIZE) & \
 361                       ~(DT_MAX_BRW_SIZE - 1))
 362 static ssize_t ll_direct_IO_26(struct kiocb *iocb, struct iov_iter *iter,
 363                                loff_t file_offset)
 364 {
 365         struct lu_env *env;
 366         struct cl_io *io;
 367         struct file *file = iocb->ki_filp;
 368         struct inode *inode = file->f_mapping->host;
 369         struct ccc_object *obj = cl_inode2ccc(inode);
 370         ssize_t count = iov_iter_count(iter);
 371         ssize_t tot_bytes = 0, result = 0;
 372         struct ll_inode_info *lli = ll_i2info(inode);
 373         long size = MAX_DIO_SIZE;
 374         int refcheck;
 375
 376         if (!lli->lli_has_smd)
 377                 return -EBADF;
 378
 379         /* FIXME: io smaller than PAGE_SIZE is broken on ia64 ??? */
 380         if ((file_offset & ~CFS_PAGE_MASK) || (count & ~CFS_PAGE_MASK))
 381                 return -EINVAL;
 382
 383         CDEBUG(D_VFSTRACE,
 384                "VFS Op:inode=%lu/%u(%p), size=%zd (max %lu), offset=%lld=%llx, pages %zd (max %lu)\n",
 385                inode->i_ino, inode->i_generation, inode, count, MAX_DIO_SIZE,
 386                file_offset, file_offset, count >> PAGE_CACHE_SHIFT,
 387                MAX_DIO_SIZE >> PAGE_CACHE_SHIFT);
 388
 389         /* Check that all user buffers are aligned as well */
 390         if (iov_iter_alignment(iter) & ~CFS_PAGE_MASK)
 391                 return -EINVAL;
 392
 393         env = cl_env_get(&refcheck);
 394         LASSERT(!IS_ERR(env));
 395         io = ccc_env_io(env)->cui_cl.cis_io;
 396         LASSERT(io != NULL);
 397
 398         /* 0. Need locking between buffered and direct access. and race with
 399          *    size changing by concurrent truncates and writes.
 400          * 1. Need inode mutex to operate transient pages.
 401          */
 402         if (iov_iter_rw(iter) == READ)
 403                 mutex_lock(&inode->i_mutex);
 404
 405         LASSERT(obj->cob_transient_pages == 0);
 406         while (iov_iter_count(iter)) {
 407                 struct page **pages;
 408                 size_t offs;
 409
 410                 count = min_t(size_t, iov_iter_count(iter), size);
 411                 if (iov_iter_rw(iter) == READ) {
 412                         if (file_offset >= i_size_read(inode))
 413                                 break;
 414                         if (file_offset + count > i_size_read(inode))
 415                                 count = i_size_read(inode) - file_offset;
 416                 }
 417
 418                 result = iov_iter_get_pages_alloc(iter, &pages, count, &offs);
 419                 if (likely(result > 0)) {
 420                         int n = DIV_ROUND_UP(result + offs, PAGE_SIZE);
 421                         result = ll_direct_IO_26_seg(env, io, iov_iter_rw(iter),
 422                                                      inode, file->f_mapping,
 423                                                      result, file_offset, pages,
 424                                                      n);
 425                         ll_free_user_pages(pages, n, iov_iter_rw(iter) == READ);
 426                 }
 427                 if (unlikely(result <= 0)) {
 428                         /* If we can't allocate a large enough buffer
 429                          * for the request, shrink it to a smaller
 430                          * PAGE_SIZE multiple and try again.
 431                          * We should always be able to kmalloc for a
 432                          * page worth of page pointers = 4MB on i386. */
 433                         if (result == -ENOMEM &&
 434                             size > (PAGE_CACHE_SIZE / sizeof(*pages)) *
 435                                    PAGE_CACHE_SIZE) {
 436                                 size = ((((size / 2) - 1) |
 437                                          ~CFS_PAGE_MASK) + 1) &
 438                                         CFS_PAGE_MASK;
 439                                 CDEBUG(D_VFSTRACE, "DIO size now %lu\n",
 440                                        size);
 441                                 continue;
 442                         }
 443
 444                         goto out;
 445                 }
 446                 iov_iter_advance(iter, result);
 447                 tot_bytes += result;
 448                 file_offset += result;
 449         }
 450 out:
 451         LASSERT(obj->cob_transient_pages == 0);
 452         if (iov_iter_rw(iter) == READ)
 453                 mutex_unlock(&inode->i_mutex);
 454
 455         if (tot_bytes > 0) {
 456                 if (iov_iter_rw(iter) == WRITE) {
 457                         struct lov_stripe_md *lsm;
 458
 459                         lsm = ccc_inode_lsm_get(inode);
 460                         LASSERT(lsm != NULL);
 461                         lov_stripe_lock(lsm);
 462                         obd_adjust_kms(ll_i2dtexp(inode), lsm, file_offset, 0);
 463                         lov_stripe_unlock(lsm);
 464                         ccc_inode_lsm_put(inode, lsm);
 465                 }
 466         }
 467
 468         cl_env_put(env, &refcheck);
 469         return tot_bytes ? : result;
 470 }
 471
 472 static int ll_write_begin(struct file *file, struct address_space *mapping,
 473                          loff_t pos, unsigned len, unsigned flags,
 474                          struct page **pagep, void **fsdata)
 475 {
 476         pgoff_t index = pos >> PAGE_CACHE_SHIFT;
 477         struct page *page;
 478         int rc;
 479         unsigned from = pos & (PAGE_CACHE_SIZE - 1);
 480
 481         page = grab_cache_page_write_begin(mapping, index, flags);
 482         if (!page)
 483                 return -ENOMEM;
 484
 485         *pagep = page;
 486
 487         rc = ll_prepare_write(file, page, from, from + len);
 488         if (rc) {
 489                 unlock_page(page);
 490                 page_cache_release(page);
 491         }
 492         return rc;
 493 }
 494
 495 static int ll_write_end(struct file *file, struct address_space *mapping,
 496                         loff_t pos, unsigned len, unsigned copied,
 497                         struct page *page, void *fsdata)
 498 {
 499         unsigned from = pos & (PAGE_CACHE_SIZE - 1);
 500         int rc;
 501
 502         rc = ll_commit_write(file, page, from, from + copied);
 503         unlock_page(page);
 504         page_cache_release(page);
 505
 506         return rc ?: copied;
 507 }
 508
 509 #ifdef CONFIG_MIGRATION
 510 static int ll_migratepage(struct address_space *mapping,
 511                          struct page *newpage, struct page *page,
 512                          enum migrate_mode mode
 513                 )
 514 {
 515         /* Always fail page migration until we have a proper implementation */
 516         return -EIO;
 517 }
 518 #endif
 519
 520 #ifndef MS_HAS_NEW_AOPS
 521 const struct address_space_operations ll_aops = {
 522         .readpage       = ll_readpage,
 523         .direct_IO      = ll_direct_IO_26,
 524         .writepage      = ll_writepage,
 525         .writepages     = ll_writepages,
 526         .set_page_dirty = ll_set_page_dirty,
 527         .write_begin    = ll_write_begin,
 528         .write_end      = ll_write_end,
 529         .invalidatepage = ll_invalidatepage,
 530         .releasepage    = (void *)ll_releasepage,
 531 #ifdef CONFIG_MIGRATION
 532         .migratepage    = ll_migratepage,
 533 #endif
 534 };
 535 #else
 536 const struct address_space_operations_ext ll_aops = {
 537         .orig_aops.readpage       = ll_readpage,
 538 /*      .orig_aops.readpages      = ll_readpages, */
 539         .orig_aops.direct_IO      = ll_direct_IO_26,
 540         .orig_aops.writepage      = ll_writepage,
 541         .orig_aops.writepages     = ll_writepages,
 542         .orig_aops.set_page_dirty = ll_set_page_dirty,
 543         .orig_aops.prepare_write  = ll_prepare_write,
 544         .orig_aops.commit_write   = ll_commit_write,
 545         .orig_aops.invalidatepage = ll_invalidatepage,
 546         .orig_aops.releasepage    = ll_releasepage,
 547 #ifdef CONFIG_MIGRATION
 548         .orig_aops.migratepage    = ll_migratepage,
 549 #endif
 550         .write_begin    = ll_write_begin,
 551         .write_end      = ll_write_end
 552 };
 553 #endif