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_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) {
  96                                 page = cl_vmpage_page(vmpage, obj);
  97                                 if (page) {
  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 static int ll_releasepage(struct page *vmpage, gfp_t gfp_mask)
 113 {
 114         struct cl_env_nest nest;
 115         struct lu_env     *env;
 116         struct cl_object  *obj;
 117         struct cl_page    *page;
 118         struct address_space *mapping;
 119         int result;
 120
 121         LASSERT(PageLocked(vmpage));
 122         if (PageWriteback(vmpage) || PageDirty(vmpage))
 123                 return 0;
 124
 125         mapping = vmpage->mapping;
 126         if (!mapping)
 127                 return 1;
 128
 129         obj = ll_i2info(mapping->host)->lli_clob;
 130         if (!obj)
 131                 return 1;
 132
 133         /* 1 for page allocator, 1 for cl_page and 1 for page cache */
 134         if (page_count(vmpage) > 3)
 135                 return 0;
 136
 137         /* TODO: determine what gfp should be used by @gfp_mask. */
 138         env = cl_env_nested_get(&nest);
 139         if (IS_ERR(env))
 140                 /* If we can't allocate an env we won't call cl_page_put()
 141                  * later on which further means it's impossible to drop
 142                  * page refcount by cl_page, so ask kernel to not free
 143                  * this page.
 144                  */
 145                 return 0;
 146
 147         page = cl_vmpage_page(vmpage, obj);
 148         result = !page;
 149         if (page) {
 150                 if (!cl_page_in_use(page)) {
 151                         result = 1;
 152                         cl_page_delete(env, page);
 153                 }
 154                 cl_page_put(env, page);
 155         }
 156         cl_env_nested_put(&nest, env);
 157         return result;
 158 }
 159
 160 static int ll_set_page_dirty(struct page *vmpage)
 161 {
 162 #if 0
 163         struct cl_page    *page = vvp_vmpage_page_transient(vmpage);
 164         struct vvp_object *obj  = cl_inode2vvp(vmpage->mapping->host);
 165         struct vvp_page   *cpg;
 166
 167         /*
 168          * XXX should page method be called here?
 169          */
 170         LASSERT(&obj->co_cl == page->cp_obj);
 171         cpg = cl2vvp_page(cl_page_at(page, &vvp_device_type));
 172         /*
 173          * XXX cannot do much here, because page is possibly not locked:
 174          * sys_munmap()->...
 175          *     ->unmap_page_range()->zap_pte_range()->set_page_dirty().
 176          */
 177         vvp_write_pending(obj, cpg);
 178 #endif
 179         return __set_page_dirty_nobuffers(vmpage);
 180 }
 181
 182 #define MAX_DIRECTIO_SIZE (2*1024*1024*1024UL)
 183
 184 static inline int ll_get_user_pages(int rw, unsigned long user_addr,
 185                                     size_t size, struct page ***pages,
 186                                     int *max_pages)
 187 {
 188         int result = -ENOMEM;
 189
 190         /* set an arbitrary limit to prevent arithmetic overflow */
 191         if (size > MAX_DIRECTIO_SIZE) {
 192                 *pages = NULL;
 193                 return -EFBIG;
 194         }
 195
 196         *max_pages = (user_addr + size + PAGE_SIZE - 1) >> PAGE_SHIFT;
 197         *max_pages -= user_addr >> PAGE_SHIFT;
 198
 199         *pages = libcfs_kvzalloc(*max_pages * sizeof(**pages), GFP_NOFS);
 200         if (*pages) {
 201                 result = get_user_pages_fast(user_addr, *max_pages,
 202                                              (rw == READ), *pages);
 203                 if (unlikely(result <= 0))
 204                         kvfree(*pages);
 205         }
 206
 207         return result;
 208 }
 209
 210 /*  ll_free_user_pages - tear down page struct array
 211  *  @pages: array of page struct pointers underlying target buffer
 212  */
 213 static void ll_free_user_pages(struct page **pages, int npages, int do_dirty)
 214 {
 215         int i;
 216
 217         for (i = 0; i < npages; i++) {
 218                 if (do_dirty)
 219                         set_page_dirty_lock(pages[i]);
 220                 put_page(pages[i]);
 221         }
 222         kvfree(pages);
 223 }
 224
 225 ssize_t ll_direct_rw_pages(const struct lu_env *env, struct cl_io *io,
 226                            int rw, struct inode *inode,
 227                            struct ll_dio_pages *pv)
 228 {
 229         struct cl_page    *clp;
 230         struct cl_2queue  *queue;
 231         struct cl_object  *obj = io->ci_obj;
 232         int i;
 233         ssize_t rc = 0;
 234         loff_t file_offset  = pv->ldp_start_offset;
 235         long size          = pv->ldp_size;
 236         int page_count      = pv->ldp_nr;
 237         struct page **pages = pv->ldp_pages;
 238         long page_size      = cl_page_size(obj);
 239         bool do_io;
 240         int  io_pages       = 0;
 241
 242         queue = &io->ci_queue;
 243         cl_2queue_init(queue);
 244         for (i = 0; i < page_count; i++) {
 245                 if (pv->ldp_offsets)
 246                         file_offset = pv->ldp_offsets[i];
 247
 248                 LASSERT(!(file_offset & (page_size - 1)));
 249                 clp = cl_page_find(env, obj, cl_index(obj, file_offset),
 250                                    pv->ldp_pages[i], CPT_TRANSIENT);
 251                 if (IS_ERR(clp)) {
 252                         rc = PTR_ERR(clp);
 253                         break;
 254                 }
 255
 256                 rc = cl_page_own(env, io, clp);
 257                 if (rc) {
 258                         LASSERT(clp->cp_state == CPS_FREEING);
 259                         cl_page_put(env, clp);
 260                         break;
 261                 }
 262
 263                 do_io = true;
 264
 265                 /* check the page type: if the page is a host page, then do
 266                  * write directly
 267                  */
 268                 if (clp->cp_type == CPT_CACHEABLE) {
 269                         struct page *vmpage = cl_page_vmpage(env, clp);
 270                         struct page *src_page;
 271                         struct page *dst_page;
 272                         void       *src;
 273                         void       *dst;
 274
 275                         src_page = (rw == WRITE) ? pages[i] : vmpage;
 276                         dst_page = (rw == WRITE) ? vmpage : pages[i];
 277
 278                         src = kmap_atomic(src_page);
 279                         dst = kmap_atomic(dst_page);
 280                         memcpy(dst, src, min(page_size, size));
 281                         kunmap_atomic(dst);
 282                         kunmap_atomic(src);
 283
 284                         /* make sure page will be added to the transfer by
 285                          * cl_io_submit()->...->vvp_page_prep_write().
 286                          */
 287                         if (rw == WRITE)
 288                                 set_page_dirty(vmpage);
 289
 290                         if (rw == READ) {
 291                                 /* do not issue the page for read, since it
 292                                  * may reread a ra page which has NOT uptodate
 293                                  * bit set.
 294                                  */
 295                                 cl_page_disown(env, io, clp);
 296                                 do_io = false;
 297                         }
 298                 }
 299
 300                 if (likely(do_io)) {
 301                         /*
 302                          * Add a page to the incoming page list of 2-queue.
 303                          */
 304                         cl_page_list_add(&queue->c2_qin, clp);
 305
 306                         /*
 307                          * Set page clip to tell transfer formation engine
 308                          * that page has to be sent even if it is beyond KMS.
 309                          */
 310                         cl_page_clip(env, clp, 0, min(size, page_size));
 311
 312                         ++io_pages;
 313                 }
 314
 315                 /* drop the reference count for cl_page_find */
 316                 cl_page_put(env, clp);
 317                 size -= page_size;
 318                 file_offset += page_size;
 319         }
 320
 321         if (rc == 0 && io_pages) {
 322                 rc = cl_io_submit_sync(env, io,
 323                                        rw == READ ? CRT_READ : CRT_WRITE,
 324                                        queue, 0);
 325         }
 326         if (rc == 0)
 327                 rc = pv->ldp_size;
 328
 329         cl_2queue_discard(env, io, queue);
 330         cl_2queue_disown(env, io, queue);
 331         cl_2queue_fini(env, queue);
 332         return rc;
 333 }
 334 EXPORT_SYMBOL(ll_direct_rw_pages);
 335
 336 static ssize_t ll_direct_IO_26_seg(const struct lu_env *env, struct cl_io *io,
 337                                    int rw, struct inode *inode,
 338                                    struct address_space *mapping,
 339                                    size_t size, loff_t file_offset,
 340                                    struct page **pages, int page_count)
 341 {
 342         struct ll_dio_pages pvec = {
 343                 .ldp_pages      = pages,
 344                 .ldp_nr         = page_count,
 345                 .ldp_size       = size,
 346                 .ldp_offsets    = NULL,
 347                 .ldp_start_offset = file_offset
 348         };
 349
 350         return ll_direct_rw_pages(env, io, rw, inode, &pvec);
 351 }
 352
 353 /* This is the maximum size of a single O_DIRECT request, based on the
 354  * kmalloc limit.  We need to fit all of the brw_page structs, each one
 355  * representing PAGE_SIZE worth of user data, into a single buffer, and
 356  * then truncate this to be a full-sized RPC.  For 4kB PAGE_SIZE this is
 357  * up to 22MB for 128kB kmalloc and up to 682MB for 4MB kmalloc.
 358  */
 359 #define MAX_DIO_SIZE ((KMALLOC_MAX_SIZE / sizeof(struct brw_page) *       \
 360                        PAGE_SIZE) & ~(DT_MAX_BRW_SIZE - 1))
 361 static ssize_t ll_direct_IO_26(struct kiocb *iocb, struct iov_iter *iter)
 362 {
 363         struct lu_env *env;
 364         struct cl_io *io;
 365         struct file *file = iocb->ki_filp;
 366         struct inode *inode = file->f_mapping->host;
 367         struct ccc_object *obj = cl_inode2ccc(inode);
 368         loff_t file_offset = iocb->ki_pos;
 369         ssize_t count = iov_iter_count(iter);
 370         ssize_t tot_bytes = 0, result = 0;
 371         struct ll_inode_info *lli = ll_i2info(inode);
 372         long size = MAX_DIO_SIZE;
 373         int refcheck;
 374
 375         if (!lli->lli_has_smd)
 376                 return -EBADF;
 377
 378         /* FIXME: io smaller than PAGE_SIZE is broken on ia64 ??? */
 379         if ((file_offset & ~CFS_PAGE_MASK) || (count & ~CFS_PAGE_MASK))
 380                 return -EINVAL;
 381
 382         CDEBUG(D_VFSTRACE,
 383                "VFS Op:inode=%lu/%u(%p), size=%zd (max %lu), offset=%lld=%llx, pages %zd (max %lu)\n",
 384                inode->i_ino, inode->i_generation, inode, count, MAX_DIO_SIZE,
 385                file_offset, file_offset, count >> PAGE_SHIFT,
 386                MAX_DIO_SIZE >> PAGE_SHIFT);
 387
 388         /* Check that all user buffers are aligned as well */
 389         if (iov_iter_alignment(iter) & ~CFS_PAGE_MASK)
 390                 return -EINVAL;
 391
 392         env = cl_env_get(&refcheck);
 393         LASSERT(!IS_ERR(env));
 394         io = ccc_env_io(env)->cui_cl.cis_io;
 395         LASSERT(io);
 396
 397         /* 0. Need locking between buffered and direct access. and race with
 398          *    size changing by concurrent truncates and writes.
 399          * 1. Need inode mutex to operate transient pages.
 400          */
 401         if (iov_iter_rw(iter) == READ)
 402                 inode_lock(inode);
 403
 404         LASSERT(obj->cob_transient_pages == 0);
 405         while (iov_iter_count(iter)) {
 406                 struct page **pages;
 407                 size_t offs;
 408
 409                 count = min_t(size_t, iov_iter_count(iter), size);
 410                 if (iov_iter_rw(iter) == READ) {
 411                         if (file_offset >= i_size_read(inode))
 412                                 break;
 413                         if (file_offset + count > i_size_read(inode))
 414                                 count = i_size_read(inode) - file_offset;
 415                 }
 416
 417                 result = iov_iter_get_pages_alloc(iter, &pages, count, &offs);
 418                 if (likely(result > 0)) {
 419                         int n = DIV_ROUND_UP(result + offs, PAGE_SIZE);
 420
 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                          */
 434                         if (result == -ENOMEM &&
 435                             size > (PAGE_SIZE / sizeof(*pages)) *
 436                             PAGE_SIZE) {
 437                                 size = ((((size / 2) - 1) |
 438                                          ~CFS_PAGE_MASK) + 1) &
 439                                         CFS_PAGE_MASK;
 440                                 CDEBUG(D_VFSTRACE, "DIO size now %lu\n",
 441                                        size);
 442                                 continue;
 443                         }
 444
 445                         goto out;
 446                 }
 447                 iov_iter_advance(iter, result);
 448                 tot_bytes += result;
 449                 file_offset += result;
 450         }
 451 out:
 452         LASSERT(obj->cob_transient_pages == 0);
 453         if (iov_iter_rw(iter) == READ)
 454                 inode_unlock(inode);
 455
 456         if (tot_bytes > 0) {
 457                 if (iov_iter_rw(iter) == WRITE) {
 458                         struct lov_stripe_md *lsm;
 459
 460                         lsm = ccc_inode_lsm_get(inode);
 461                         LASSERT(lsm);
 462                         lov_stripe_lock(lsm);
 463                         obd_adjust_kms(ll_i2dtexp(inode), lsm, file_offset, 0);
 464                         lov_stripe_unlock(lsm);
 465                         ccc_inode_lsm_put(inode, lsm);
 466                 }
 467         }
 468
 469         cl_env_put(env, &refcheck);
 470         return tot_bytes ? : result;
 471 }
 472
 473 static int ll_write_begin(struct file *file, struct address_space *mapping,
 474                           loff_t pos, unsigned len, unsigned flags,
 475                           struct page **pagep, void **fsdata)
 476 {
 477         pgoff_t index = pos >> PAGE_SHIFT;
 478         struct page *page;
 479         int rc;
 480         unsigned from = pos & (PAGE_SIZE - 1);
 481
 482         page = grab_cache_page_write_begin(mapping, index, flags);
 483         if (!page)
 484                 return -ENOMEM;
 485
 486         *pagep = page;
 487
 488         rc = ll_prepare_write(file, page, from, from + len);
 489         if (rc) {
 490                 unlock_page(page);
 491                 put_page(page);
 492         }
 493         return rc;
 494 }
 495
 496 static int ll_write_end(struct file *file, struct address_space *mapping,
 497                         loff_t pos, unsigned len, unsigned copied,
 498                         struct page *page, void *fsdata)
 499 {
 500         unsigned from = pos & (PAGE_SIZE - 1);
 501         int rc;
 502
 503         rc = ll_commit_write(file, page, from, from + copied);
 504         unlock_page(page);
 505         put_page(page);
 506
 507         return rc ?: copied;
 508 }
 509
 510 #ifdef CONFIG_MIGRATION
 511 static int ll_migratepage(struct address_space *mapping,
 512                           struct page *newpage, struct page *page,
 513                           enum migrate_mode mode
 514                 )
 515 {
 516         /* Always fail page migration until we have a proper implementation */
 517         return -EIO;
 518 }
 519 #endif
 520
 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 };