lustre/llite/rw26.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * Lustre Lite I/O page cache routines for the 2.5/2.6 kernel version
   5  *
   6  *  Copyright (c) 2001-2003 Cluster File Systems, Inc.
   7  *
   8  *   This file is part of Lustre, http://www.lustre.org.
   9  *
  10  *   Lustre is free software; you can redistribute it and/or
  11  *   modify it under the terms of version 2 of the GNU General Public
  12  *   License as published by the Free Software Foundation.
  13  *
  14  *   Lustre is distributed in the hope that it will be useful,
  15  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  17  *   GNU General Public License for more details.
  18  *
  19  *   You should have received a copy of the GNU General Public License
  20  *   along with Lustre; if not, write to the Free Software
  21  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  22  */
  23 #ifdef HAVE_KERNEL_CONFIG_H
  24 #include <linux/config.h>
  25 #endif
  26 #include <linux/kernel.h>
  27 #include <linux/mm.h>
  28 #include <linux/string.h>
  29 #include <linux/stat.h>
  30 #include <linux/errno.h>
  31 #include <linux/smp_lock.h>
  32 #include <linux/unistd.h>
  33 #include <linux/version.h>
  34 #include <asm/system.h>
  35 #include <asm/uaccess.h>
  36
  37 #include <linux/fs.h>
  38 #include <linux/buffer_head.h>
  39 #include <linux/mpage.h>
  40 #include <linux/writeback.h>
  41 #include <linux/stat.h>
  42 #include <asm/uaccess.h>
  43 #include <asm/segment.h>
  44 #include <linux/mm.h>
  45 #include <linux/pagemap.h>
  46 #include <linux/smp_lock.h>
  47
  48 #define DEBUG_SUBSYSTEM S_LLITE
  49
  50 #include <lustre_lite.h>
  51 #include "llite_internal.h"
  52 #include <linux/lustre_compat25.h>
  53
  54 static int ll_writepage_26(struct page *page, struct writeback_control *wbc)
  55 {
  56         return ll_writepage(page);
  57 }
  58
  59 /* It is safe to not check anything in invalidatepage/releasepage below
  60    because they are run with page locked and all our io is happening with
  61    locked page too */
  62 #ifdef HAVE_INVALIDATEPAGE_RETURN_INT
  63 static int ll_invalidatepage(struct page *page, unsigned long offset)
  64 {
  65         if (offset)
  66                 return 0;
  67         if (PagePrivate(page))
  68                 ll_removepage(page);
  69         return 1;
  70 }
  71 #else
  72 static void ll_invalidatepage(struct page *page, unsigned long offset)
  73 {
  74         if (offset)
  75                 return;
  76         if (PagePrivate(page))
  77                 ll_removepage(page);
  78 }
  79 #endif
  80
  81 static int ll_releasepage(struct page *page, gfp_t gfp_mask)
  82 {
  83         if (PagePrivate(page))
  84                 ll_removepage(page);
  85         return 1;
  86 }
  87
  88 #define MAX_DIRECTIO_SIZE 2*1024*1024*1024UL
  89
  90 static inline int ll_get_user_pages(int rw, unsigned long user_addr,
  91                                     size_t size, struct page ***pages)
  92 {
  93         int result = -ENOMEM;
  94         int page_count;
  95
  96         /* set an arbitrary limit to prevent arithmetic overflow */
  97         if (size > MAX_DIRECTIO_SIZE) {
  98                 *pages = NULL;
  99                 return -EFBIG;
 100         }
 101
 102         page_count = ((user_addr + size + CFS_PAGE_SIZE - 1) >> CFS_PAGE_SHIFT)-
 103                       (user_addr >> CFS_PAGE_SHIFT);
 104
 105         OBD_ALLOC_GFP(*pages, page_count * sizeof(**pages), GFP_KERNEL);
 106         if (*pages) {
 107                 down_read(&current->mm->mmap_sem);
 108                 result = get_user_pages(current, current->mm, user_addr,
 109                                         page_count, (rw == READ), 0, *pages,
 110                                         NULL);
 111                 up_read(&current->mm->mmap_sem);
 112                 if (result < 0)
 113                         OBD_FREE(*pages, page_count * sizeof(**pages));
 114         }
 115
 116         return result;
 117 }
 118
 119 /*  ll_free_user_pages - tear down page struct array
 120  *  @pages: array of page struct pointers underlying target buffer */
 121 static void ll_free_user_pages(struct page **pages, int npages, int do_dirty)
 122 {
 123         int i;
 124
 125         for (i = 0; i < npages; i++) {
 126                 if (do_dirty)
 127                         set_page_dirty_lock(pages[i]);
 128                 page_cache_release(pages[i]);
 129         }
 130
 131         OBD_FREE(pages, npages * sizeof(*pages));
 132 }
 133
 134 static ssize_t ll_direct_IO_26_seg(int rw, struct inode *inode,
 135                                    struct address_space *mapping,
 136                                    struct lov_stripe_md *lsm,
 137                                    size_t size, loff_t file_offset,
 138                                    struct page **pages, int page_count)
 139 {
 140         struct brw_page *pga;
 141         struct obdo oa;
 142         int i, rc = 0;
 143         size_t length;
 144         ENTRY;
 145
 146         OBD_ALLOC(pga, sizeof(*pga) * page_count);
 147         if (!pga) {
 148                 CDEBUG(D_VFSTRACE, "sizeof(*pga) = %u page_count = %u\n",
 149                       (int)sizeof(*pga), page_count);
 150                 RETURN(-ENOMEM);
 151         }
 152
 153         for (i = 0, length = size; length > 0;
 154              length -=pga[i].count, file_offset +=pga[i].count,i++) {/*i last!*/
 155                 pga[i].pg = pages[i];
 156                 pga[i].off = file_offset;
 157                 /* To the end of the page, or the length, whatever is less */
 158                 pga[i].count = min_t(int, CFS_PAGE_SIZE -(file_offset & ~CFS_PAGE_MASK),
 159                                      length);
 160                 pga[i].flag = 0;
 161                 if (rw == READ)
 162                         POISON_PAGE(pages[i], 0x0d);
 163         }
 164
 165         ll_inode_fill_obdo(inode, rw, &oa);
 166
 167         rc = obd_brw_rqset(rw == WRITE ? OBD_BRW_WRITE : OBD_BRW_READ,
 168                            ll_i2obdexp(inode), &oa, lsm, page_count, pga, NULL);
 169         if (rc == 0) {
 170                 rc = size;
 171                 if (rw == WRITE) {
 172                         lov_stripe_lock(lsm);
 173                         obd_adjust_kms(ll_i2obdexp(inode), lsm, file_offset, 0);
 174                         lov_stripe_unlock(lsm);
 175                 }
 176         }
 177
 178         OBD_FREE(pga, sizeof(*pga) * page_count);
 179         RETURN(rc);
 180 }
 181
 182 /* This is the maximum size of a single O_DIRECT request, based on a 128kB
 183  * kmalloc limit.  We need to fit all of the brw_page structs, each one
 184  * representing PAGE_SIZE worth of user data, into a single buffer, and
 185  * then truncate this to be a full-sized RPC.  This is 22MB for 4kB pages. */
 186 #define MAX_DIO_SIZE ((128 * 1024 / sizeof(struct brw_page) * CFS_PAGE_SIZE) & \
 187                       ~(PTLRPC_MAX_BRW_SIZE - 1))
 188 static ssize_t ll_direct_IO_26(int rw, struct kiocb *iocb,
 189                                const struct iovec *iov, loff_t file_offset,
 190                                unsigned long nr_segs)
 191 {
 192         struct file *file = iocb->ki_filp;
 193         struct inode *inode = file->f_mapping->host;
 194         ssize_t count = iov_length(iov, nr_segs), tot_bytes = 0;
 195         struct ll_inode_info *lli = ll_i2info(inode);
 196         unsigned long seg;
 197         size_t size = MAX_DIO_SIZE;
 198         ENTRY;
 199
 200         if (!lli->lli_smd || !lli->lli_smd->lsm_object_id)
 201                 RETURN(-EBADF);
 202
 203         /* FIXME: io smaller than CFS_PAGE_SIZE is broken on ia64 ??? */
 204         if ((file_offset & (~CFS_PAGE_MASK)) || (count & ~CFS_PAGE_MASK))
 205                 RETURN(-EINVAL);
 206
 207         CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p), size="LPSZ" (max "LPSZ
 208                "), offset=%lld=%llx, pages "LPSZ" (max "LPSZ")\n",
 209                inode->i_ino, inode->i_generation, inode, count, MAX_DIO_SIZE,
 210                file_offset, file_offset, count >> CFS_PAGE_SHIFT,
 211                MAX_DIO_SIZE >> CFS_PAGE_SHIFT);
 212
 213         if (rw == WRITE)
 214                 lprocfs_counter_add(ll_i2sbi(inode)->ll_stats,
 215                                     LPROC_LL_DIRECT_WRITE, count);
 216         else
 217                 lprocfs_counter_add(ll_i2sbi(inode)->ll_stats,
 218                                     LPROC_LL_DIRECT_READ, count);
 219
 220         /* Check that all user buffers are aligned as well */
 221         for (seg = 0; seg < nr_segs; seg++) {
 222                 if (((unsigned long)iov[seg].iov_base & ~CFS_PAGE_MASK) ||
 223                     (iov[seg].iov_len & ~CFS_PAGE_MASK))
 224                         RETURN(-EINVAL);
 225         }
 226
 227         for (seg = 0; seg < nr_segs; seg++) {
 228                 size_t iov_left = iov[seg].iov_len;
 229                 unsigned long user_addr = (unsigned long)iov[seg].iov_base;
 230
 231                 while (iov_left > 0) {
 232                         struct page **pages;
 233                         int page_count;
 234                         ssize_t result;
 235
 236                         page_count = ll_get_user_pages(rw, user_addr,
 237                                                        min(size, iov_left),
 238                                                        &pages);
 239                         LASSERT(page_count != 0);
 240                         if (page_count > 0) {
 241                                 result = ll_direct_IO_26_seg(rw, inode,
 242                                                              file->f_mapping,
 243                                                              lli->lli_smd,
 244                                                              min(size,iov_left),
 245                                                              file_offset, pages,
 246                                                              page_count);
 247                                 ll_free_user_pages(pages, page_count, rw==READ);
 248                         } else {
 249                                 result = 0;
 250                         }
 251                         if (page_count < 0 || result <= 0) {
 252                                 /* If we can't allocate a large enough buffer
 253                                  * for the request, shrink it to a smaller
 254                                  * PAGE_SIZE multiple and try again.
 255                                  * We should always be able to kmalloc for a
 256                                  * page worth of page pointers = 4MB on i386. */
 257                                 if ((page_count == -ENOMEM||result == -ENOMEM)&&
 258                                     size > (CFS_PAGE_SIZE / sizeof(*pages)) *
 259                                            CFS_PAGE_SIZE) {
 260                                         size = ((((size / 2) - 1) |
 261                                                  ~CFS_PAGE_MASK) + 1) &
 262                                                 CFS_PAGE_MASK;
 263                                         CDEBUG(D_VFSTRACE, "DIO size now %u\n",
 264                                                (int)size);
 265                                         continue;
 266                                 }
 267                                 if (tot_bytes > 0)
 268                                         RETURN(tot_bytes);
 269                                 RETURN(page_count < 0 ? page_count : result);
 270                         }
 271
 272                         tot_bytes += result;
 273                         file_offset += result;
 274                         iov_left -= result;
 275                         user_addr += result;
 276                 }
 277         }
 278         RETURN(tot_bytes);
 279 }
 280
 281 struct address_space_operations ll_aops = {
 282         .readpage       = ll_readpage,
 283 //        .readpages      = ll_readpages,
 284         .direct_IO      = ll_direct_IO_26,
 285         .writepage      = ll_writepage_26,
 286         .writepages     = generic_writepages,
 287         .set_page_dirty = __set_page_dirty_nobuffers,
 288         .sync_page      = NULL,
 289         .prepare_write  = ll_prepare_write,
 290         .commit_write   = ll_commit_write,
 291         .invalidatepage = ll_invalidatepage,
 292         .releasepage    = ll_releasepage,
 293         .bmap           = NULL
 294 };