lustre/llite/rw26.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * GPL HEADER START
   5  *
   6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   7  *
   8  * This program is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License version 2 only,
  10  * as published by the Free Software Foundation.
  11  *
  12  * This program is distributed in the hope that it will be useful, but
  13  * WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * General Public License version 2 for more details (a copy is included
  16  * in the LICENSE file that accompanied this code).
  17  *
  18  * You should have received a copy of the GNU General Public License
  19  * version 2 along with this program; If not, see
  20  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  21  *
  22  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  23  * CA 95054 USA or visit www.sun.com if you need additional information or
  24  * have any questions.
  25  *
  26  * GPL HEADER END
  27  */
  28 /*
  29  * Copyright  2008 Sun Microsystems, Inc. All rights reserved
  30  * Use is subject to license terms.
  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/llite/rw26.c
  37  *
  38  * Lustre Lite I/O page cache routines for the 2.5/2.6 kernel version
  39  */
  40
  41 #ifndef AUTOCONF_INCLUDED
  42 #include <linux/config.h>
  43 #endif
  44 #include <linux/kernel.h>
  45 #include <linux/mm.h>
  46 #include <linux/string.h>
  47 #include <linux/stat.h>
  48 #include <linux/errno.h>
  49 #include <linux/smp_lock.h>
  50 #include <linux/unistd.h>
  51 #include <linux/version.h>
  52 #include <asm/system.h>
  53 #include <asm/uaccess.h>
  54
  55 #include <linux/fs.h>
  56 #include <linux/buffer_head.h>
  57 #include <linux/writeback.h>
  58 #include <linux/stat.h>
  59 #include <asm/uaccess.h>
  60 #ifdef HAVE_SEGMENT_H
  61 # include <asm/segment.h>
  62 #endif
  63 #include <linux/mm.h>
  64 #include <linux/pagemap.h>
  65 #include <linux/smp_lock.h>
  66
  67 #define DEBUG_SUBSYSTEM S_LLITE
  68
  69 #include <lustre_lite.h>
  70 #include "llite_internal.h"
  71 #include <linux/lustre_compat25.h>
  72
  73 static int ll_writepage_26(struct page *page, struct writeback_control *wbc)
  74 {
  75         return ll_writepage(page);
  76 }
  77
  78 /* It is safe to not check anything in invalidatepage/releasepage below
  79    because they are run with page locked and all our io is happening with
  80    locked page too */
  81 #ifdef HAVE_INVALIDATEPAGE_RETURN_INT
  82 static int ll_invalidatepage(struct page *page, unsigned long offset)
  83 {
  84         if (offset)
  85                 return 0;
  86         if (PagePrivate(page))
  87                 ll_removepage(page);
  88         return 1;
  89 }
  90 #else
  91 static void ll_invalidatepage(struct page *page, unsigned long offset)
  92 {
  93         if (offset)
  94                 return;
  95         if (PagePrivate(page))
  96                 ll_removepage(page);
  97 }
  98 #endif
  99
 100 #ifdef HAVE_RELEASEPAGE_WITH_GFP
 101 #define RELEASEPAGE_ARG_TYPE gfp_t
 102 #else
 103 #define RELEASEPAGE_ARG_TYPE int
 104 #endif
 105 static int ll_releasepage(struct page *page, RELEASEPAGE_ARG_TYPE gfp_mask)
 106 {
 107         if (PagePrivate(page))
 108                 ll_removepage(page);
 109         return 1;
 110 }
 111
 112 #define MAX_DIRECTIO_SIZE 2*1024*1024*1024UL
 113
 114 static inline int ll_get_user_pages(int rw, unsigned long user_addr,
 115                                     size_t size, struct page ***pages)
 116 {
 117         int result = -ENOMEM;
 118         int page_count;
 119
 120         /* set an arbitrary limit to prevent arithmetic overflow */
 121         if (size > MAX_DIRECTIO_SIZE) {
 122                 *pages = NULL;
 123                 return -EFBIG;
 124         }
 125
 126         page_count = ((user_addr + size + CFS_PAGE_SIZE - 1) >> CFS_PAGE_SHIFT)-
 127                       (user_addr >> CFS_PAGE_SHIFT);
 128
 129         OBD_ALLOC_WAIT(*pages, page_count * sizeof(**pages));
 130         if (*pages) {
 131                 down_read(&current->mm->mmap_sem);
 132                 result = get_user_pages(current, current->mm, user_addr,
 133                                         page_count, (rw == READ), 0, *pages,
 134                                         NULL);
 135                 up_read(&current->mm->mmap_sem);
 136                 if (result < 0)
 137                         OBD_FREE(*pages, page_count * sizeof(**pages));
 138         }
 139
 140         return result;
 141 }
 142
 143 /*  ll_free_user_pages - tear down page struct array
 144  *  @pages: array of page struct pointers underlying target buffer */
 145 static void ll_free_user_pages(struct page **pages, int npages, int do_dirty)
 146 {
 147         int i;
 148
 149         for (i = 0; i < npages; i++) {
 150                 if (do_dirty)
 151                         set_page_dirty_lock(pages[i]);
 152                 page_cache_release(pages[i]);
 153         }
 154
 155         OBD_FREE(pages, npages * sizeof(*pages));
 156 }
 157
 158 static ssize_t ll_direct_IO_26_seg(int rw, struct inode *inode,
 159                                    struct address_space *mapping,
 160                                    struct obd_info *oinfo,
 161                                    struct ptlrpc_request_set *set,
 162                                    size_t size, loff_t file_offset,
 163                                    struct page **pages, int page_count)
 164 {
 165         struct brw_page *pga;
 166         int i, rc = 0;
 167         size_t length;
 168         ENTRY;
 169
 170         OBD_ALLOC(pga, sizeof(*pga) * page_count);
 171         if (!pga) {
 172                 CDEBUG(D_VFSTRACE, "sizeof(*pga) = %u page_count = %u\n",
 173                       (int)sizeof(*pga), page_count);
 174                 RETURN(-ENOMEM);
 175         }
 176
 177         for (i = 0, length = size; length > 0;
 178              length -=pga[i].count, file_offset +=pga[i].count,i++) {/*i last!*/
 179                 pga[i].pg = pages[i];
 180                 pga[i].off = file_offset;
 181                 /* To the end of the page, or the length, whatever is less */
 182                 pga[i].count = min_t(int, CFS_PAGE_SIZE -(file_offset & ~CFS_PAGE_MASK),
 183                                      length);
 184                 pga[i].flag = OBD_BRW_SYNC;
 185                 if (rw == READ)
 186                         POISON_PAGE(pages[i], 0x0d);
 187         }
 188
 189         rc = obd_brw_async(rw == WRITE ? OBD_BRW_WRITE : OBD_BRW_READ,
 190                            ll_i2obdexp(inode), oinfo, page_count,
 191                            pga, NULL, set);
 192         if (rc == 0)
 193                 rc = size;
 194
 195         OBD_FREE(pga, sizeof(*pga) * page_count);
 196         RETURN(rc);
 197 }
 198
 199 /* This is the maximum size of a single O_DIRECT request, based on a 128kB
 200  * kmalloc limit.  We need to fit all of the brw_page structs, each one
 201  * representing PAGE_SIZE worth of user data, into a single buffer, and
 202  * then truncate this to be a full-sized RPC.  This is 22MB for 4kB pages. */
 203 #define MAX_DIO_SIZE ((128 * 1024 / sizeof(struct brw_page) * CFS_PAGE_SIZE) & \
 204                       ~(PTLRPC_MAX_BRW_SIZE - 1))
 205 static ssize_t ll_direct_IO_26(int rw, struct kiocb *iocb,
 206                                const struct iovec *iov, loff_t file_offset,
 207                                unsigned long nr_segs)
 208 {
 209         struct file *file = iocb->ki_filp;
 210         struct inode *inode = file->f_mapping->host;
 211         ssize_t count = iov_length(iov, nr_segs), tot_bytes = 0;
 212         struct ll_inode_info *lli = ll_i2info(inode);
 213         struct lov_stripe_md *lsm = lli->lli_smd;
 214         struct ptlrpc_request_set *set;
 215         struct obd_info oinfo;
 216         struct obdo oa;
 217         unsigned long seg;
 218         size_t size = MAX_DIO_SIZE;
 219         ENTRY;
 220
 221         if (!lli->lli_smd || !lli->lli_smd->lsm_object_id)
 222                 RETURN(-EBADF);
 223
 224         /* FIXME: io smaller than CFS_PAGE_SIZE is broken on ia64 ??? */
 225         if ((file_offset & (~CFS_PAGE_MASK)) || (count & ~CFS_PAGE_MASK))
 226                 RETURN(-EINVAL);
 227
 228         CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p), size="LPSZ" (max %lu), "
 229                "offset=%lld=%llx, pages "LPSZ" (max %lu)\n",
 230                inode->i_ino, inode->i_generation, inode, count, MAX_DIO_SIZE,
 231                file_offset, file_offset, count >> CFS_PAGE_SHIFT,
 232                MAX_DIO_SIZE >> CFS_PAGE_SHIFT);
 233
 234         if (rw == WRITE)
 235                 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_DIRECT_WRITE, count);
 236         else
 237                 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_DIRECT_READ, count);
 238
 239         /* Check that all user buffers are aligned as well */
 240         for (seg = 0; seg < nr_segs; seg++) {
 241                 if (((unsigned long)iov[seg].iov_base & ~CFS_PAGE_MASK) ||
 242                     (iov[seg].iov_len & ~CFS_PAGE_MASK))
 243                         RETURN(-EINVAL);
 244         }
 245
 246         set = ptlrpc_prep_set();
 247         if (set == NULL)
 248                 RETURN(-ENOMEM);
 249
 250         ll_inode_fill_obdo(inode, rw == WRITE ? OBD_BRW_WRITE : OBD_BRW_READ, &oa);
 251         oinfo.oi_oa = &oa;
 252         oinfo.oi_md = lsm;
 253
 254         /* need locking between buffered and direct access. and race with
 255          *size changing by concurrent truncates and writes. */
 256         if (rw == READ)
 257                 LOCK_INODE_MUTEX(inode);
 258
 259         for (seg = 0; seg < nr_segs; seg++) {
 260                 size_t iov_left = iov[seg].iov_len;
 261                 unsigned long user_addr = (unsigned long)iov[seg].iov_base;
 262
 263                 if (rw == READ) {
 264                         if (file_offset >= inode->i_size)
 265                                 break;
 266                         if (file_offset + iov_left > inode->i_size)
 267                                 iov_left = inode->i_size - file_offset;
 268                 }
 269
 270                 while (iov_left > 0) {
 271                         struct page **pages;
 272                         int page_count;
 273                         ssize_t result;
 274
 275                         page_count = ll_get_user_pages(rw, user_addr,
 276                                                        min(size, iov_left),
 277                                                        &pages);
 278                         LASSERT(page_count != 0);
 279                         if (page_count > 0) {
 280                                 result = ll_direct_IO_26_seg(rw, inode,
 281                                                              file->f_mapping,
 282                                                              &oinfo, set,
 283                                                              min(size,iov_left),
 284                                                              file_offset, pages,
 285                                                              page_count);
 286                                 ll_free_user_pages(pages, page_count, rw==READ);
 287                         } else {
 288                                 result = 0;
 289                         }
 290                         if (page_count < 0 || result <= 0) {
 291                                 /* If we can't allocate a large enough buffer
 292                                  * for the request, shrink it to a smaller
 293                                  * PAGE_SIZE multiple and try again.
 294                                  * We should always be able to kmalloc for a
 295                                  * page worth of page pointers = 4MB on i386. */
 296                                 if ((page_count == -ENOMEM||result == -ENOMEM)&&
 297                                     size > (CFS_PAGE_SIZE / sizeof(*pages)) *
 298                                            CFS_PAGE_SIZE) {
 299                                         size = ((((size / 2) - 1) |
 300                                                  ~CFS_PAGE_MASK) + 1) &
 301                                                 CFS_PAGE_MASK;
 302                                         CDEBUG(D_VFSTRACE, "DIO size now %u\n",
 303                                                (int)size);
 304                                         continue;
 305                                 }
 306                                 if (tot_bytes > 0)
 307                                         GOTO(wait_io, tot_bytes);
 308                                 GOTO(out, tot_bytes = page_count < 0 ? page_count : result);
 309                         }
 310
 311                         tot_bytes += result;
 312                         file_offset += result;
 313                         iov_left -= result;
 314                         user_addr += result;
 315                 }
 316         }
 317
 318         if (tot_bytes > 0) {
 319                 int rc;
 320         wait_io:
 321                 rc = ptlrpc_set_wait(set);
 322                 if (rc)
 323                         GOTO(out, tot_bytes = rc);
 324                 if (rw == WRITE) {
 325                         lov_stripe_lock(lsm);
 326                         obd_adjust_kms(ll_i2obdexp(inode), lsm, file_offset, 0);
 327                         lov_stripe_unlock(lsm);
 328                 }
 329         }
 330 out:
 331         if (rw == READ)
 332                 UNLOCK_INODE_MUTEX(inode);
 333
 334         ptlrpc_set_destroy(set);
 335         RETURN(tot_bytes);
 336 }
 337
 338 struct address_space_operations ll_aops = {
 339         .readpage       = ll_readpage,
 340 //        .readpages      = ll_readpages,
 341         .direct_IO      = ll_direct_IO_26,
 342         .writepage      = ll_writepage_26,
 343         .writepages     = generic_writepages,
 344         .set_page_dirty = __set_page_dirty_nobuffers,
 345         .sync_page      = NULL,
 346         .prepare_write  = ll_prepare_write,
 347         .commit_write   = ll_commit_write,
 348         .invalidatepage = ll_invalidatepage,
 349         .releasepage    = ll_releasepage,
 350         .bmap           = NULL
 351 };