lustre/llite/rw26.c

   1 /*
   2  * -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   3  * vim:expandtab:shiftwidth=8:tabstop=8:
   4  *
   5  * GPL HEADER START
   6  *
   7  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   8  *
   9  * This program is free software; you can redistribute it and/or modify
  10  * it under the terms of the GNU General Public License version 2 only,
  11  * as published by the Free Software Foundation.
  12  *
  13  * This program is distributed in the hope that it will be useful, but
  14  * WITHOUT ANY WARRANTY; without even the implied warranty of
  15  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  16  * General Public License version 2 for more details (a copy is included
  17  * in the LICENSE file that accompanied this code).
  18  *
  19  * You should have received a copy of the GNU General Public License
  20  * version 2 along with this program; If not, see [sun.com URL with a
  21  * copy of GPLv2].
  22  *
  23  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  24  * CA 95054 USA or visit www.sun.com if you need additional information or
  25  * have any questions.
  26  *
  27  * GPL HEADER END
  28  */
  29 /*
  30  * Copyright  2008 Sun Microsystems, Inc. All rights reserved
  31  * Use is subject to license terms.
  32  */
  33 /*
  34  * This file is part of Lustre, http://www.lustre.org/
  35  * Lustre is a trademark of Sun Microsystems, Inc.
  36  *
  37  * lustre/lustre/llite/rw26.c
  38  *
  39  * Lustre Lite I/O page cache routines for the 2.5/2.6 kernel version
  40  */
  41
  42 #ifndef AUTOCONF_INCLUDED
  43 #include <linux/config.h>
  44 #endif
  45 #include <linux/kernel.h>
  46 #include <linux/mm.h>
  47 #include <linux/string.h>
  48 #include <linux/stat.h>
  49 #include <linux/errno.h>
  50 #include <linux/smp_lock.h>
  51 #include <linux/unistd.h>
  52 #include <linux/version.h>
  53 #include <asm/system.h>
  54 #include <asm/uaccess.h>
  55
  56 #include <linux/fs.h>
  57 #include <linux/buffer_head.h>
  58 #include <linux/mpage.h>
  59 #include <linux/writeback.h>
  60 #include <linux/stat.h>
  61 #include <asm/uaccess.h>
  62 #include <linux/mm.h>
  63 #include <linux/pagemap.h>
  64 #include <linux/smp_lock.h>
  65
  66 #define DEBUG_SUBSYSTEM S_LLITE
  67
  68 //#include <lustre_mdc.h>
  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 == 0 && PagePrivate(page))
  94                 ll_removepage(page);
  95 }
  96 #endif
  97 static int ll_releasepage(struct page *page, gfp_t gfp_mask)
  98 {
  99         if (PagePrivate(page))
 100                 ll_removepage(page);
 101         return 1;
 102 }
 103
 104 static int ll_set_page_dirty(struct page *page)
 105 {
 106         struct ll_async_page *llap;
 107         ENTRY;
 108
 109         llap = llap_from_page(page, LLAP_ORIGIN_UNKNOWN);
 110         if (IS_ERR(llap))
 111                 RETURN(PTR_ERR(llap));
 112
 113         llap_write_pending(page->mapping->host, llap);
 114         RETURN(__set_page_dirty_nobuffers(page));
 115 }
 116
 117 #define MAX_DIRECTIO_SIZE 2*1024*1024*1024UL
 118
 119 static inline int ll_get_user_pages(int rw, unsigned long user_addr,
 120                                     size_t size, struct page ***pages)
 121 {
 122         int result = -ENOMEM;
 123         int page_count;
 124
 125         /* set an arbitrary limit to prevent arithmetic overflow */
 126         if (size > MAX_DIRECTIO_SIZE) {
 127                 *pages = NULL;
 128                 return -EFBIG;
 129         }
 130
 131         page_count = (user_addr + size + CFS_PAGE_SIZE - 1) >> CFS_PAGE_SHIFT;
 132         page_count -= user_addr >> CFS_PAGE_SHIFT;
 133
 134         OBD_ALLOC_WAIT(*pages, page_count * sizeof(**pages));
 135         if (*pages) {
 136                 down_read(&current->mm->mmap_sem);
 137                 result = get_user_pages(current, current->mm, user_addr,
 138                                         page_count, (rw == READ), 0, *pages,
 139                                         NULL);
 140                 up_read(&current->mm->mmap_sem);
 141                 if (result < 0)
 142                         OBD_FREE(*pages, page_count * sizeof(**pages));
 143         }
 144
 145         return result;
 146 }
 147
 148 /*  ll_free_user_pages - tear down page struct array
 149  *  @pages: array of page struct pointers underlying target buffer */
 150 static void ll_free_user_pages(struct page **pages, int npages, int do_dirty)
 151 {
 152         int i;
 153
 154         for (i = 0; i < npages; i++) {
 155                 if (do_dirty)
 156                         set_page_dirty_lock(pages[i]);
 157                 page_cache_release(pages[i]);
 158         }
 159
 160         OBD_FREE(pages, npages * sizeof(*pages));
 161 }
 162
 163 static ssize_t ll_direct_IO_26_seg(int rw, struct inode *inode,
 164                                    struct address_space *mapping,
 165                                    struct obd_info *oinfo,
 166                                    struct ptlrpc_request_set *set,
 167                                    size_t size, loff_t file_offset,
 168                                    struct page **pages, int page_count)
 169 {
 170         struct brw_page *pga;
 171         int i, rc = 0;
 172         size_t length;
 173         ENTRY;
 174
 175         OBD_ALLOC(pga, sizeof(*pga) * page_count);
 176         if (!pga) {
 177                 CDEBUG(D_VFSTRACE, "sizeof(*pga) = %u page_count = %u\n",
 178                        (int)sizeof(*pga), page_count);
 179                 RETURN(-ENOMEM);
 180         }
 181
 182         for (i = 0, length = size; length > 0;
 183              length -=pga[i].count, file_offset +=pga[i].count,i++) {/*i last!*/
 184                 pga[i].pg = pages[i];
 185                 pga[i].off = file_offset;
 186                 /* To the end of the page, or the length, whatever is less */
 187                 pga[i].count = min_t(int, CFS_PAGE_SIZE -
 188                                           (file_offset & ~CFS_PAGE_MASK),
 189                                      length);
 190                 pga[i].flag = 0;
 191                 if (rw == READ)
 192                         POISON_PAGE(pages[i], 0x0d);
 193         }
 194
 195         rc = obd_brw_async(rw == WRITE ? OBD_BRW_WRITE : OBD_BRW_READ,
 196                                 ll_i2dtexp(inode), oinfo, page_count,
 197                                 pga, NULL, set);
 198         if (rc == 0)
 199                 rc = size;
 200
 201         OBD_FREE(pga, sizeof(*pga) * page_count);
 202         RETURN(rc);
 203 }
 204
 205 /* This is the maximum size of a single O_DIRECT request, based on a 128kB
 206  * kmalloc limit.  We need to fit all of the brw_page structs, each one
 207  * representing PAGE_SIZE worth of user data, into a single buffer, and
 208  * then truncate this to be a full-sized RPC.  This is 22MB for 4kB pages. */
 209 #define MAX_DIO_SIZE ((128 * 1024 / sizeof(struct brw_page) * CFS_PAGE_SIZE) & \
 210                       ~(PTLRPC_MAX_BRW_SIZE - 1))
 211 static ssize_t ll_direct_IO_26(int rw, struct kiocb *iocb,
 212                                const struct iovec *iov, loff_t file_offset,
 213                                unsigned long nr_segs)
 214 {
 215         struct file *file = iocb->ki_filp;
 216         struct inode *inode = file->f_mapping->host;
 217         ssize_t count = iov_length(iov, nr_segs), tot_bytes = 0;
 218         struct ll_inode_info *lli = ll_i2info(inode);
 219         struct lov_stripe_md *lsm = lli->lli_smd;
 220         struct ptlrpc_request_set *set;
 221         struct obd_info oinfo;
 222         struct obdo oa;
 223         unsigned long seg = 0;
 224         size_t size = MAX_DIO_SIZE;
 225         int opc;
 226         ENTRY;
 227
 228         if (!lli->lli_smd || !lli->lli_smd->lsm_object_id)
 229                 RETURN(-EBADF);
 230
 231         /* FIXME: io smaller than PAGE_SIZE is broken on ia64 ??? */
 232         if ((file_offset & ~CFS_PAGE_MASK) || (count & ~CFS_PAGE_MASK))
 233                 RETURN(-EINVAL);
 234
 235         CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p), size="LPSZ" (max %lu), "
 236                "offset=%lld=%llx, pages "LPSZ" (max %lu)\n",
 237                inode->i_ino, inode->i_generation, inode, count, MAX_DIO_SIZE,
 238                file_offset, file_offset, count >> CFS_PAGE_SHIFT,
 239                MAX_DIO_SIZE >> CFS_PAGE_SHIFT);
 240
 241         if (rw == WRITE) {
 242                 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_DIRECT_WRITE, count);
 243                 opc = CAPA_OPC_OSS_WRITE;
 244                 llap_write_pending(inode, NULL);
 245         } else {
 246                 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_DIRECT_READ, count);
 247                 opc = CAPA_OPC_OSS_RW;
 248         }
 249
 250         /* Check that all user buffers are aligned as well */
 251         for (seg = 0; seg < nr_segs; seg++) {
 252                 if (((unsigned long)iov[seg].iov_base & ~CFS_PAGE_MASK) ||
 253                     (iov[seg].iov_len & ~CFS_PAGE_MASK))
 254                         RETURN(-EINVAL);
 255         }
 256
 257         set = ptlrpc_prep_set();
 258         if (set == NULL)
 259                 RETURN(-ENOMEM);
 260
 261         ll_inode_fill_obdo(inode, rw, &oa);
 262         oinfo.oi_oa = &oa;
 263         oinfo.oi_md = lsm;
 264         oinfo.oi_capa = ll_osscapa_get(inode, opc);
 265
 266         /* need locking between buffered and direct access. and race with
 267          *size changing by concurrent truncates and writes. */
 268         if (rw == READ)
 269                 LOCK_INODE_MUTEX(inode);
 270
 271         for (seg = 0; seg < nr_segs; seg++) {
 272                 size_t iov_left = iov[seg].iov_len;
 273                 unsigned long user_addr = (unsigned long)iov[seg].iov_base;
 274
 275                 if (rw == READ) {
 276                         if (file_offset >= inode->i_size)
 277                                 break;
 278                         if (file_offset + iov_left > inode->i_size)
 279                                 iov_left = inode->i_size - file_offset;
 280                 }
 281
 282                 while (iov_left > 0) {
 283                         struct page **pages;
 284                         int page_count;
 285                         ssize_t result;
 286
 287                         page_count = ll_get_user_pages(rw, user_addr,
 288                                                        min(size, iov_left),
 289                                                        &pages);
 290                         LASSERT(page_count != 0);
 291                         if (page_count > 0) {
 292                                 result = ll_direct_IO_26_seg(rw, inode,
 293                                                              file->f_mapping,
 294                                                              &oinfo, set,
 295                                                              min(size,iov_left),
 296                                                              file_offset, pages,
 297                                                              page_count);
 298                                 ll_free_user_pages(pages, page_count, rw==READ);
 299                         } else {
 300                                 result = 0;
 301                         }
 302                         if (page_count < 0 || result <= 0) {
 303                                 /* If we can't allocate a large enough buffer
 304                                  * for the request, shrink it to a smaller
 305                                  * PAGE_SIZE multiple and try again.
 306                                  * We should always be able to kmalloc for a
 307                                  * page worth of page pointers = 4MB on i386. */
 308                                 if ((page_count == -ENOMEM||result == -ENOMEM)&&
 309                                     size > (CFS_PAGE_SIZE / sizeof(*pages)) *
 310                                            CFS_PAGE_SIZE) {
 311                                         size = ((((size / 2) - 1) |
 312                                                  ~CFS_PAGE_MASK) + 1) &
 313                                                 CFS_PAGE_MASK;
 314                                         CDEBUG(D_VFSTRACE, "DIO size now %u\n",
 315                                                (int)size);
 316                                         continue;
 317                                 }
 318
 319                                 if (tot_bytes <= 0)
 320                                         tot_bytes = page_count < 0 ? page_count : result;
 321                                 GOTO(out, tot_bytes);
 322                         }
 323
 324                         tot_bytes += result;
 325                         file_offset += result;
 326                         iov_left -= result;
 327                         user_addr += result;
 328                 }
 329         }
 330 out:
 331         if (rw == READ)
 332                 UNLOCK_INODE_MUTEX(inode);
 333
 334         if (tot_bytes > 0) {
 335                 int rc;
 336
 337                 rc = ptlrpc_set_wait(set);
 338                 if (rc) {
 339                         tot_bytes = rc;
 340                 } else if (rw == WRITE) {
 341                         lov_stripe_lock(lsm);
 342                         obd_adjust_kms(ll_i2dtexp(inode), lsm, file_offset, 0);
 343                         lov_stripe_unlock(lsm);
 344                 }
 345         }
 346
 347         capa_put(oinfo.oi_capa);
 348         ptlrpc_set_destroy(set);
 349         RETURN(tot_bytes);
 350 }
 351
 352 struct address_space_operations ll_aops = {
 353         .readpage       = ll_readpage,
 354 //        .readpages      = ll_readpages,
 355         .direct_IO      = ll_direct_IO_26,
 356         .writepage      = ll_writepage_26,
 357         .writepages     = generic_writepages,
 358         .set_page_dirty = ll_set_page_dirty,
 359         .sync_page      = NULL,
 360         .prepare_write  = ll_prepare_write,
 361         .commit_write   = ll_commit_write,
 362         .invalidatepage = ll_invalidatepage,
 363         .releasepage    = ll_releasepage,
 364         .bmap           = NULL
 365 };