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 <asm/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/stat.h>
  55 #include <asm/uaccess.h>
  56 #include <linux/mm.h>
  57 #include <linux/pagemap.h>
  58
  59 #define DEBUG_SUBSYSTEM S_LLITE
  60
  61 #include <lustre_lite.h>
  62 #include "llite_internal.h"
  63 #include <linux/lustre_compat25.h>
  64
  65 /**
  66  * Implements Linux VM address_space::invalidatepage() method. This method is
  67  * called when the page is truncate from a file, either as a result of
  68  * explicit truncate, or when inode is removed from memory (as a result of
  69  * final iput(), umount, or memory pressure induced icache shrinking).
  70  *
  71  * [0, offset] bytes of the page remain valid (this is for a case of not-page
  72  * aligned truncate). Lustre leaves partially truncated page in the cache,
  73  * relying on struct inode::i_size to limit further accesses.
  74  */
  75 static void ll_invalidatepage(struct page *vmpage, unsigned long offset)
  76 {
  77         struct inode     *inode;
  78         struct lu_env    *env;
  79         struct cl_page   *page;
  80         struct cl_object *obj;
  81
  82         int refcheck;
  83
  84         LASSERT(PageLocked(vmpage));
  85         LASSERT(!PageWriteback(vmpage));
  86
  87         /*
  88          * It is safe to not check anything in invalidatepage/releasepage
  89          * below because they are run with page locked and all our io is
  90          * happening with locked page too
  91          */
  92         if (offset == 0) {
  93                 env = cl_env_get(&refcheck);
  94                 if (!IS_ERR(env)) {
  95                         inode = vmpage->mapping->host;
  96                         obj = ll_i2info(inode)->lli_clob;
  97                         if (obj != NULL) {
  98                                 page = cl_vmpage_page(vmpage, obj);
  99                                 if (page != NULL) {
 100                                         lu_ref_add(&page->cp_reference,
 101                                                    "delete", vmpage);
 102                                         cl_page_delete(env, page);
 103                                         lu_ref_del(&page->cp_reference,
 104                                                    "delete", vmpage);
 105                                         cl_page_put(env, page);
 106                                 }
 107                         } else
 108                                 LASSERT(vmpage->private == 0);
 109                         cl_env_put(env, &refcheck);
 110                 }
 111         }
 112 }
 113
 114 #ifdef HAVE_RELEASEPAGE_WITH_INT
 115 #define RELEASEPAGE_ARG_TYPE int
 116 #else
 117 #define RELEASEPAGE_ARG_TYPE gfp_t
 118 #endif
 119 static int ll_releasepage(struct page *vmpage, RELEASEPAGE_ARG_TYPE gfp_mask)
 120 {
 121         struct cl_env_nest nest;
 122         struct lu_env     *env;
 123         struct cl_object  *obj;
 124         struct cl_page    *page;
 125         struct address_space *mapping;
 126         int result;
 127
 128         LASSERT(PageLocked(vmpage));
 129         if (PageWriteback(vmpage) || PageDirty(vmpage))
 130                 return 0;
 131
 132         mapping = vmpage->mapping;
 133         if (mapping == NULL)
 134                 return 1;
 135
 136         obj = ll_i2info(mapping->host)->lli_clob;
 137         if (obj == NULL)
 138                 return 1;
 139
 140         /* 1 for page allocator, 1 for cl_page and 1 for page cache */
 141         if (page_count(vmpage) > 3)
 142                 return 0;
 143
 144         /* TODO: determine what gfp should be used by @gfp_mask. */
 145         env = cl_env_nested_get(&nest);
 146         if (IS_ERR(env))
 147                 /* If we can't allocate an env we won't call cl_page_put()
 148                  * later on which further means it's impossible to drop
 149                  * page refcount by cl_page, so ask kernel to not free
 150                  * this page. */
 151                 return 0;
 152
 153         page = cl_vmpage_page(vmpage, obj);
 154         result = page == NULL;
 155         if (page != NULL) {
 156                 if (!cl_page_in_use(page)) {
 157                         result = 1;
 158                         cl_page_delete(env, page);
 159                 }
 160                 cl_page_put(env, page);
 161         }
 162         cl_env_nested_put(&nest, env);
 163         return result;
 164 }
 165
 166 static int ll_set_page_dirty(struct page *vmpage)
 167 {
 168 #if 0
 169         struct cl_page    *page = vvp_vmpage_page_transient(vmpage);
 170         struct vvp_object *obj  = cl_inode2vvp(vmpage->mapping->host);
 171         struct vvp_page   *cpg;
 172
 173         /*
 174          * XXX should page method be called here?
 175          */
 176         LASSERT(&obj->co_cl == page->cp_obj);
 177         cpg = cl2vvp_page(cl_page_at(page, &vvp_device_type));
 178         /*
 179          * XXX cannot do much here, because page is possibly not locked:
 180          * sys_munmap()->...
 181          *     ->unmap_page_range()->zap_pte_range()->set_page_dirty().
 182          */
 183         vvp_write_pending(obj, cpg);
 184 #endif
 185         RETURN(__set_page_dirty_nobuffers(vmpage));
 186 }
 187
 188 #define MAX_DIRECTIO_SIZE 2*1024*1024*1024UL
 189
 190 static inline int ll_get_user_pages(int rw, unsigned long user_addr,
 191                                     size_t size, struct page ***pages,
 192                                     int *max_pages)
 193 {
 194         int result = -ENOMEM;
 195
 196         /* set an arbitrary limit to prevent arithmetic overflow */
 197         if (size > MAX_DIRECTIO_SIZE) {
 198                 *pages = NULL;
 199                 return -EFBIG;
 200         }
 201
 202         *max_pages = (user_addr + size + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
 203         *max_pages -= user_addr >> PAGE_CACHE_SHIFT;
 204
 205         OBD_ALLOC_LARGE(*pages, *max_pages * sizeof(**pages));
 206         if (*pages) {
 207                 down_read(&current->mm->mmap_sem);
 208                 result = get_user_pages(current, current->mm, user_addr,
 209                                         *max_pages, (rw == READ), 0, *pages,
 210                                         NULL);
 211                 up_read(&current->mm->mmap_sem);
 212                 if (unlikely(result <= 0))
 213                         OBD_FREE_LARGE(*pages, *max_pages * sizeof(**pages));
 214         }
 215
 216         return result;
 217 }
 218
 219 /*  ll_free_user_pages - tear down page struct array
 220  *  @pages: array of page struct pointers underlying target buffer */
 221 static void ll_free_user_pages(struct page **pages, int npages, int do_dirty)
 222 {
 223         int i;
 224
 225         for (i = 0; i < npages; i++) {
 226                 if (pages[i] == NULL)
 227                         break;
 228                 if (do_dirty)
 229                         set_page_dirty_lock(pages[i]);
 230                 page_cache_release(pages[i]);
 231         }
 232
 233         OBD_FREE_LARGE(pages, npages * sizeof(*pages));
 234 }
 235
 236 ssize_t ll_direct_rw_pages(const struct lu_env *env, struct cl_io *io,
 237                            int rw, struct inode *inode,
 238                            struct ll_dio_pages *pv)
 239 {
 240         struct cl_page    *clp;
 241         struct cl_2queue  *queue;
 242         struct cl_object  *obj = io->ci_obj;
 243         int i;
 244         ssize_t rc = 0;
 245         loff_t file_offset  = pv->ldp_start_offset;
 246         long size          = pv->ldp_size;
 247         int page_count      = pv->ldp_nr;
 248         struct page **pages = pv->ldp_pages;
 249         long page_size      = cl_page_size(obj);
 250         bool do_io;
 251         int  io_pages       = 0;
 252         ENTRY;
 253
 254         queue = &io->ci_queue;
 255         cl_2queue_init(queue);
 256         for (i = 0; i < page_count; i++) {
 257                 if (pv->ldp_offsets)
 258                     file_offset = pv->ldp_offsets[i];
 259
 260                 LASSERT(!(file_offset & (page_size - 1)));
 261                 clp = cl_page_find(env, obj, cl_index(obj, file_offset),
 262                                    pv->ldp_pages[i], CPT_TRANSIENT);
 263                 if (IS_ERR(clp)) {
 264                         rc = PTR_ERR(clp);
 265                         break;
 266                 }
 267
 268                 rc = cl_page_own(env, io, clp);
 269                 if (rc) {
 270                         LASSERT(clp->cp_state == CPS_FREEING);
 271                         cl_page_put(env, clp);
 272                         break;
 273                 }
 274
 275                 do_io = true;
 276
 277                 /* check the page type: if the page is a host page, then do
 278                  * write directly */
 279                 if (clp->cp_type == CPT_CACHEABLE) {
 280                         struct page *vmpage = cl_page_vmpage(env, clp);
 281                         struct page *src_page;
 282                         struct page *dst_page;
 283                         void       *src;
 284                         void       *dst;
 285
 286                         src_page = (rw == WRITE) ? pages[i] : vmpage;
 287                         dst_page = (rw == WRITE) ? vmpage : pages[i];
 288
 289                         src = ll_kmap_atomic(src_page, KM_USER0);
 290                         dst = ll_kmap_atomic(dst_page, KM_USER1);
 291                         memcpy(dst, src, min(page_size, size));
 292                         ll_kunmap_atomic(dst, KM_USER1);
 293                         ll_kunmap_atomic(src, KM_USER0);
 294
 295                         /* make sure page will be added to the transfer by
 296                          * cl_io_submit()->...->vvp_page_prep_write(). */
 297                         if (rw == WRITE)
 298                                 set_page_dirty(vmpage);
 299
 300                         if (rw == READ) {
 301                                 /* do not issue the page for read, since it
 302                                  * may reread a ra page which has NOT uptodate
 303                                  * bit set. */
 304                                 cl_page_disown(env, io, clp);
 305                                 do_io = false;
 306                         }
 307                 }
 308
 309                 if (likely(do_io)) {
 310                         cl_2queue_add(queue, clp);
 311
 312                         /*
 313                          * Set page clip to tell transfer formation engine
 314                          * that page has to be sent even if it is beyond KMS.
 315                          */
 316                         cl_page_clip(env, clp, 0, min(size, page_size));
 317
 318                         ++io_pages;
 319                 }
 320
 321                 /* drop the reference count for cl_page_find */
 322                 cl_page_put(env, clp);
 323                 size -= page_size;
 324                 file_offset += page_size;
 325         }
 326
 327         if (rc == 0 && io_pages) {
 328                 rc = cl_io_submit_sync(env, io,
 329                                        rw == READ ? CRT_READ : CRT_WRITE,
 330                                        queue, 0);
 331         }
 332         if (rc == 0)
 333                 rc = pv->ldp_size;
 334
 335         cl_2queue_discard(env, io, queue);
 336         cl_2queue_disown(env, io, queue);
 337         cl_2queue_fini(env, queue);
 338         RETURN(rc);
 339 }
 340 EXPORT_SYMBOL(ll_direct_rw_pages);
 341
 342 static ssize_t ll_direct_IO_26_seg(const struct lu_env *env, struct cl_io *io,
 343                                    int rw, struct inode *inode,
 344                                    struct address_space *mapping,
 345                                    size_t size, loff_t file_offset,
 346                                    struct page **pages, int page_count)
 347 {
 348     struct ll_dio_pages pvec = { .ldp_pages     = pages,
 349                                  .ldp_nr           = page_count,
 350                                  .ldp_size       = size,
 351                                  .ldp_offsets      = NULL,
 352                                  .ldp_start_offset = file_offset
 353                                };
 354
 355     return ll_direct_rw_pages(env, io, rw, inode, &pvec);
 356 }
 357
 358 #ifdef KMALLOC_MAX_SIZE
 359 #define MAX_MALLOC KMALLOC_MAX_SIZE
 360 #else
 361 #define MAX_MALLOC (128 * 1024)
 362 #endif
 363
 364 /* This is the maximum size of a single O_DIRECT request, based on the
 365  * kmalloc limit.  We need to fit all of the brw_page structs, each one
 366  * representing PAGE_SIZE worth of user data, into a single buffer, and
 367  * then truncate this to be a full-sized RPC.  For 4kB PAGE_SIZE this is
 368  * up to 22MB for 128kB kmalloc and up to 682MB for 4MB kmalloc. */
 369 #define MAX_DIO_SIZE ((MAX_MALLOC / sizeof(struct brw_page) * PAGE_CACHE_SIZE) & \
 370                       ~(DT_MAX_BRW_SIZE - 1))
 371 static ssize_t ll_direct_IO_26(int rw, struct kiocb *iocb,
 372                                const struct iovec *iov, loff_t file_offset,
 373                                unsigned long nr_segs)
 374 {
 375         struct lu_env *env;
 376         struct cl_io *io;
 377         struct file *file = iocb->ki_filp;
 378         struct inode *inode = file->f_mapping->host;
 379         struct ccc_object *obj = cl_inode2ccc(inode);
 380         long count = iov_length(iov, nr_segs);
 381         long tot_bytes = 0, result = 0;
 382         struct ll_inode_info *lli = ll_i2info(inode);
 383         unsigned long seg = 0;
 384         long size = MAX_DIO_SIZE;
 385         int refcheck;
 386         ENTRY;
 387
 388         if (!lli->lli_has_smd)
 389                 RETURN(-EBADF);
 390
 391         /* FIXME: io smaller than PAGE_SIZE is broken on ia64 ??? */
 392         if ((file_offset & ~CFS_PAGE_MASK) || (count & ~CFS_PAGE_MASK))
 393                 RETURN(-EINVAL);
 394
 395         CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p), size=%lu (max %lu), "
 396                "offset=%lld=%llx, pages %lu (max %lu)\n",
 397                inode->i_ino, inode->i_generation, inode, count, MAX_DIO_SIZE,
 398                file_offset, file_offset, count >> PAGE_CACHE_SHIFT,
 399                MAX_DIO_SIZE >> PAGE_CACHE_SHIFT);
 400
 401         /* Check that all user buffers are aligned as well */
 402         for (seg = 0; seg < nr_segs; seg++) {
 403                 if (((unsigned long)iov[seg].iov_base & ~CFS_PAGE_MASK) ||
 404                     (iov[seg].iov_len & ~CFS_PAGE_MASK))
 405                         RETURN(-EINVAL);
 406         }
 407
 408         env = cl_env_get(&refcheck);
 409         LASSERT(!IS_ERR(env));
 410         io = ccc_env_io(env)->cui_cl.cis_io;
 411         LASSERT(io != NULL);
 412
 413         /* 0. Need locking between buffered and direct access. and race with
 414          *    size changing by concurrent truncates and writes.
 415          * 1. Need inode mutex to operate transient pages.
 416          */
 417         if (rw == READ)
 418                 mutex_lock(&inode->i_mutex);
 419
 420         LASSERT(obj->cob_transient_pages == 0);
 421         for (seg = 0; seg < nr_segs; seg++) {
 422                 long iov_left = iov[seg].iov_len;
 423                 unsigned long user_addr = (unsigned long)iov[seg].iov_base;
 424
 425                 if (rw == READ) {
 426                         if (file_offset >= i_size_read(inode))
 427                                 break;
 428                         if (file_offset + iov_left > i_size_read(inode))
 429                                 iov_left = i_size_read(inode) - file_offset;
 430                 }
 431
 432                 while (iov_left > 0) {
 433                         struct page **pages;
 434                         int page_count, max_pages = 0;
 435                         long bytes;
 436
 437                         bytes = min(size, iov_left);
 438                         page_count = ll_get_user_pages(rw, user_addr, bytes,
 439                                                        &pages, &max_pages);
 440                         if (likely(page_count > 0)) {
 441                                 if (unlikely(page_count <  max_pages))
 442                                         bytes = page_count << PAGE_CACHE_SHIFT;
 443                                 result = ll_direct_IO_26_seg(env, io, rw, inode,
 444                                                              file->f_mapping,
 445                                                              bytes, file_offset,
 446                                                              pages, page_count);
 447                                 ll_free_user_pages(pages, max_pages, rw==READ);
 448                         } else if (page_count == 0) {
 449                                 GOTO(out, result = -EFAULT);
 450                         } else {
 451                                 result = page_count;
 452                         }
 453                         if (unlikely(result <= 0)) {
 454                                 /* If we can't allocate a large enough buffer
 455                                  * for the request, shrink it to a smaller
 456                                  * PAGE_SIZE multiple and try again.
 457                                  * We should always be able to kmalloc for a
 458                                  * page worth of page pointers = 4MB on i386. */
 459                                 if (result == -ENOMEM &&
 460                                     size > (PAGE_CACHE_SIZE / sizeof(*pages)) *
 461                                            PAGE_CACHE_SIZE) {
 462                                         size = ((((size / 2) - 1) |
 463                                                  ~CFS_PAGE_MASK) + 1) &
 464                                                 CFS_PAGE_MASK;
 465                                         CDEBUG(D_VFSTRACE,"DIO size now %lu\n",
 466                                                size);
 467                                         continue;
 468                                 }
 469
 470                                 GOTO(out, result);
 471                         }
 472
 473                         tot_bytes += result;
 474                         file_offset += result;
 475                         iov_left -= result;
 476                         user_addr += result;
 477                 }
 478         }
 479 out:
 480         LASSERT(obj->cob_transient_pages == 0);
 481         if (rw == READ)
 482                 mutex_unlock(&inode->i_mutex);
 483
 484         if (tot_bytes > 0) {
 485                 if (rw == WRITE) {
 486                         struct lov_stripe_md *lsm;
 487
 488                         lsm = ccc_inode_lsm_get(inode);
 489                         LASSERT(lsm != NULL);
 490                         lov_stripe_lock(lsm);
 491                         obd_adjust_kms(ll_i2dtexp(inode), lsm, file_offset, 0);
 492                         lov_stripe_unlock(lsm);
 493                         ccc_inode_lsm_put(inode, lsm);
 494                 }
 495         }
 496
 497         cl_env_put(env, &refcheck);
 498         RETURN(tot_bytes ? : result);
 499 }
 500
 501 static int ll_write_begin(struct file *file, struct address_space *mapping,
 502                          loff_t pos, unsigned len, unsigned flags,
 503                          struct page **pagep, void **fsdata)
 504 {
 505         pgoff_t index = pos >> PAGE_CACHE_SHIFT;
 506         struct page *page;
 507         int rc;
 508         unsigned from = pos & (PAGE_CACHE_SIZE - 1);
 509         ENTRY;
 510
 511         page = grab_cache_page_write_begin(mapping, index, flags);
 512         if (!page)
 513                 RETURN(-ENOMEM);
 514
 515         *pagep = page;
 516
 517         rc = ll_prepare_write(file, page, from, from + len);
 518         if (rc) {
 519                 unlock_page(page);
 520                 page_cache_release(page);
 521         }
 522         RETURN(rc);
 523 }
 524
 525 static int ll_write_end(struct file *file, struct address_space *mapping,
 526                         loff_t pos, unsigned len, unsigned copied,
 527                         struct page *page, void *fsdata)
 528 {
 529         unsigned from = pos & (PAGE_CACHE_SIZE - 1);
 530         int rc;
 531
 532         rc = ll_commit_write(file, page, from, from + copied);
 533         unlock_page(page);
 534         page_cache_release(page);
 535
 536         return rc ?: copied;
 537 }
 538
 539 #ifdef CONFIG_MIGRATION
 540 int ll_migratepage(struct address_space *mapping,
 541                 struct page *newpage, struct page *page
 542                 , enum migrate_mode mode
 543                 )
 544 {
 545         /* Always fail page migration until we have a proper implementation */
 546         return -EIO;
 547 }
 548 #endif
 549
 550 #ifndef MS_HAS_NEW_AOPS
 551 struct address_space_operations ll_aops = {
 552         .readpage       = ll_readpage,
 553 //      .readpages      = ll_readpages,
 554         .direct_IO      = ll_direct_IO_26,
 555         .writepage      = ll_writepage,
 556         .writepages     = ll_writepages,
 557         .set_page_dirty = ll_set_page_dirty,
 558         .write_begin    = ll_write_begin,
 559         .write_end      = ll_write_end,
 560         .invalidatepage = ll_invalidatepage,
 561         .releasepage    = (void *)ll_releasepage,
 562 #ifdef CONFIG_MIGRATION
 563         .migratepage    = ll_migratepage,
 564 #endif
 565         .bmap      = NULL
 566 };
 567 #else
 568 struct address_space_operations_ext ll_aops = {
 569         .orig_aops.readpage       = ll_readpage,
 570 //      .orig_aops.readpages      = ll_readpages,
 571         .orig_aops.direct_IO      = ll_direct_IO_26,
 572         .orig_aops.writepage      = ll_writepage,
 573         .orig_aops.writepages     = ll_writepages,
 574         .orig_aops.set_page_dirty = ll_set_page_dirty,
 575         .orig_aops.prepare_write  = ll_prepare_write,
 576         .orig_aops.commit_write   = ll_commit_write,
 577         .orig_aops.invalidatepage = ll_invalidatepage,
 578         .orig_aops.releasepage    = ll_releasepage,
 579 #ifdef CONFIG_MIGRATION
 580         .orig_aops.migratepage    = ll_migratepage,
 581 #endif
 582         .orig_aops.bmap    = NULL,
 583         .write_begin    = ll_write_begin,
 584         .write_end      = ll_write_end
 585 };
 586 #endif