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/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 #include <linux/mm.h>
  61 #include <linux/pagemap.h>
  62 #include <linux/smp_lock.h>
  63
  64 #define DEBUG_SUBSYSTEM S_LLITE
  65
  66 //#include <lustre_mdc.h>
  67 #include <lustre_lite.h>
  68 #include "llite_internal.h"
  69 #include <linux/lustre_compat25.h>
  70
  71 /**
  72  * Implements Linux VM address_space::invalidatepage() method. This method is
  73  * called when the page is truncate from a file, either as a result of
  74  * explicit truncate, or when inode is removed from memory (as a result of
  75  * final iput(), umount, or memory pressure induced icache shrinking).
  76  *
  77  * [0, offset] bytes of the page remain valid (this is for a case of not-page
  78  * aligned truncate). Lustre leaves partially truncated page in the cache,
  79  * relying on struct inode::i_size to limit further accesses.
  80  */
  81 static int cl_invalidatepage(struct page *vmpage, unsigned long offset)
  82 {
  83         struct inode     *inode;
  84         struct lu_env    *env;
  85         struct cl_page   *page;
  86         struct cl_object *obj;
  87
  88         int result;
  89         int refcheck;
  90
  91         LASSERT(PageLocked(vmpage));
  92         LASSERT(!PageWriteback(vmpage));
  93
  94         /*
  95          * It is safe to not check anything in invalidatepage/releasepage
  96          * below because they are run with page locked and all our io is
  97          * happening with locked page too
  98          */
  99         result = 0;
 100         if (offset == 0) {
 101                 env = cl_env_get(&refcheck);
 102                 if (!IS_ERR(env)) {
 103                         inode = vmpage->mapping->host;
 104                         obj = ll_i2info(inode)->lli_clob;
 105                         if (obj != NULL) {
 106                                 page = cl_vmpage_page(vmpage, obj);
 107                                 if (page != NULL) {
 108                                         lu_ref_add(&page->cp_reference,
 109                                                    "delete", vmpage);
 110                                         cl_page_delete(env, page);
 111                                         result = 1;
 112                                         lu_ref_del(&page->cp_reference,
 113                                                    "delete", vmpage);
 114                                         cl_page_put(env, page);
 115                                 }
 116                         } else
 117                                 LASSERT(vmpage->private == 0);
 118                         cl_env_put(env, &refcheck);
 119                 }
 120         }
 121         return result;
 122 }
 123
 124 #ifdef HAVE_INVALIDATEPAGE_RETURN_INT
 125 static int ll_invalidatepage(struct page *page, unsigned long offset)
 126 {
 127         return cl_invalidatepage(page, offset);
 128 }
 129 #else /* !HAVE_INVALIDATEPAGE_RETURN_INT */
 130 static void ll_invalidatepage(struct page *page, unsigned long offset)
 131 {
 132         cl_invalidatepage(page, offset);
 133 }
 134 #endif
 135
 136 #ifdef HAVE_RELEASEPAGE_WITH_INT
 137 #define RELEASEPAGE_ARG_TYPE int
 138 #else
 139 #define RELEASEPAGE_ARG_TYPE gfp_t
 140 #endif
 141 static int ll_releasepage(struct page *page, RELEASEPAGE_ARG_TYPE gfp_mask)
 142 {
 143         void *cookie;
 144
 145         cookie = cl_env_reenter();
 146         ll_invalidatepage(page, 0);
 147         cl_env_reexit(cookie);
 148         return 1;
 149 }
 150
 151 static int ll_set_page_dirty(struct page *vmpage)
 152 {
 153 #if 0
 154         struct cl_page    *page = vvp_vmpage_page_transient(vmpage);
 155         struct vvp_object *obj  = cl_inode2vvp(vmpage->mapping->host);
 156         struct vvp_page   *cpg;
 157
 158         /*
 159          * XXX should page method be called here?
 160          */
 161         LASSERT(&obj->co_cl == page->cp_obj);
 162         cpg = cl2vvp_page(cl_page_at(page, &vvp_device_type));
 163         /*
 164          * XXX cannot do much here, because page is possibly not locked:
 165          * sys_munmap()->...
 166          *     ->unmap_page_range()->zap_pte_range()->set_page_dirty().
 167          */
 168         vvp_write_pending(obj, cpg);
 169 #endif
 170         RETURN(__set_page_dirty_nobuffers(vmpage));
 171 }
 172
 173 #define MAX_DIRECTIO_SIZE 2*1024*1024*1024UL
 174
 175 static inline int ll_get_user_pages(int rw, unsigned long user_addr,
 176                                     size_t size, struct page ***pages,
 177                                     int *max_pages)
 178 {
 179         int result = -ENOMEM;
 180
 181         /* set an arbitrary limit to prevent arithmetic overflow */
 182         if (size > MAX_DIRECTIO_SIZE) {
 183                 *pages = NULL;
 184                 return -EFBIG;
 185         }
 186
 187         *max_pages = (user_addr + size + CFS_PAGE_SIZE - 1) >> CFS_PAGE_SHIFT;
 188         *max_pages -= user_addr >> CFS_PAGE_SHIFT;
 189
 190         OBD_ALLOC_WAIT(*pages, *max_pages * sizeof(**pages));
 191         if (*pages) {
 192                 down_read(&current->mm->mmap_sem);
 193                 result = get_user_pages(current, current->mm, user_addr,
 194                                         *max_pages, (rw == READ), 0, *pages,
 195                                         NULL);
 196                 up_read(&current->mm->mmap_sem);
 197                 if (unlikely(result <= 0))
 198                         OBD_FREE(*pages, *max_pages * sizeof(**pages));
 199         }
 200
 201         return result;
 202 }
 203
 204 /*  ll_free_user_pages - tear down page struct array
 205  *  @pages: array of page struct pointers underlying target buffer */
 206 static void ll_free_user_pages(struct page **pages, int npages, int do_dirty)
 207 {
 208         int i;
 209
 210         for (i = 0; i < npages; i++) {
 211                 if (pages[i] == NULL)
 212                         break;
 213                 if (do_dirty)
 214                         set_page_dirty_lock(pages[i]);
 215                 page_cache_release(pages[i]);
 216         }
 217
 218         OBD_FREE(pages, npages * sizeof(*pages));
 219 }
 220
 221 ssize_t ll_direct_rw_pages(const struct lu_env *env, struct cl_io *io,
 222                            int rw, struct inode *inode,
 223                            struct ll_dio_pages *pv)
 224 {
 225         struct cl_page    *clp;
 226         struct ccc_page   *clup;
 227         struct cl_2queue  *queue;
 228         struct cl_object  *obj = io->ci_obj;
 229         struct cl_sync_io *anchor = &ccc_env_info(env)->cti_sync_io;
 230         int i;
 231         ssize_t rc = 0;
 232         loff_t file_offset  = pv->ldp_start_offset;
 233         size_t size         = pv->ldp_size;
 234         int page_count      = pv->ldp_nr;
 235         struct page **pages = pv->ldp_pages;
 236         size_t page_size    = cl_page_size(obj);
 237         ENTRY;
 238
 239         cl_sync_io_init(anchor, page_count);
 240
 241         queue = &io->ci_queue;
 242         cl_2queue_init(queue);
 243         for (i = 0; i < page_count; i++) {
 244                 if (pv->ldp_offsets)
 245                     file_offset = pv->ldp_offsets[i];
 246                 LASSERT(!(file_offset & (page_size - 1)));
 247                 clp = cl_page_find(env, obj, cl_index(obj, file_offset),
 248                                    pv->ldp_pages[i], CPT_TRANSIENT);
 249                 if (IS_ERR(clp)) {
 250                         rc = PTR_ERR(clp);
 251                         break;
 252                 }
 253
 254                 /* check the page type: if the page is a host page, then do
 255                  * write directly */
 256                 /*
 257                  * Very rare case that the host pages can be found for
 258                  * directIO case, since linux kernel truncated all covered
 259                  * pages before getting here. So, to make the OST happy(to
 260                  * write a contiguous region), all pages are issued
 261                  * here. -jay */
 262                 if (clp->cp_type == CPT_CACHEABLE) {
 263                         cfs_page_t *vmpage = cl_page_vmpage(env, clp);
 264                         cfs_page_t *src_page;
 265                         cfs_page_t *dst_page;
 266                         void       *src;
 267                         void       *dst;
 268
 269                         src_page = (rw == WRITE) ? pages[i] : vmpage;
 270                         dst_page = (rw == WRITE) ? vmpage : pages[i];
 271
 272                         src = kmap_atomic(src_page, KM_USER0);
 273                         dst = kmap_atomic(dst_page, KM_USER1);
 274                         memcpy(dst, (const void *)src, min(page_size, size));
 275                         kunmap_atomic(dst, KM_USER1);
 276                         kunmap_atomic(src, KM_USER0);
 277
 278                         /* make sure page will be added to the transfer by
 279                          * cl_io_submit()->...->vvp_page_prep_write(). */
 280                         if (rw == WRITE)
 281                                 set_page_dirty(vmpage);
 282                         /*
 283                          * If direct-io read finds up-to-date page in the
 284                          * cache, just copy it to the user space. Page will be
 285                          * filtered out by vvp_page_prep_read(). This
 286                          * preserves an invariant, that page is read at most
 287                          * once, see cl_page_flags::CPF_READ_COMPLETED.
 288                          */
 289                 }
 290
 291                 rc = cl_page_own(env, io, clp);
 292                 if (rc) {
 293                         LASSERT(clp->cp_state == CPS_FREEING);
 294                         cl_page_put(env, clp);
 295                         break;
 296                 }
 297
 298                 clup = cl2ccc_page(cl_page_at(clp, &vvp_device_type));
 299                 clup->cpg_sync_io = anchor;
 300                 cl_2queue_add(queue, clp);
 301
 302                 /* drop the reference count for cl_page_find, so that the page
 303                  * will be freed in cl_2queue_fini. */
 304                 cl_page_put(env, clp);
 305                 /*
 306                  * Set page clip to tell transfer formation engine that page
 307                  * has to be sent even if it is beyond KMS.
 308                  */
 309                 cl_page_clip(env, clp, 0, min(size, page_size));
 310                 size -= page_size;
 311                 file_offset += page_size;
 312         }
 313
 314         if (rc == 0) {
 315                 rc = cl_io_submit_rw(env, io, rw == READ ? CRT_READ : CRT_WRITE,
 316                                      queue, CRP_NORMAL);
 317                 if (rc == 0) {
 318                         /*
 319                          * If some pages weren't sent for any reason (e.g.,
 320                          * direct-io read found up-to-date pages in the
 321                          * cache), count them as completed to avoid infinite
 322                          * wait.
 323                          */
 324                         cl_page_list_for_each(clp, &queue->c2_qin)
 325                                 cl_sync_io_note(anchor, +1);
 326                         /* wait for the IO to be finished. */
 327                         rc = cl_sync_io_wait(env, io, &queue->c2_qout,
 328                                              anchor) ?: pv->ldp_size;
 329                 }
 330         }
 331
 332         cl_2queue_discard(env, io, queue);
 333         cl_2queue_disown(env, io, queue);
 334         cl_2queue_fini(env, queue);
 335         RETURN(rc);
 336 }
 337 EXPORT_SYMBOL(ll_direct_rw_pages);
 338
 339 static ssize_t ll_direct_IO_26_seg(const struct lu_env *env, struct cl_io *io,
 340                                    int rw, struct inode *inode,
 341                                    struct address_space *mapping,
 342                                    size_t size, loff_t file_offset,
 343                                    struct page **pages, int page_count)
 344 {
 345     struct ll_dio_pages pvec = { .ldp_pages        = pages,
 346                                  .ldp_nr           = page_count,
 347                                  .ldp_size         = size,
 348                                  .ldp_offsets      = NULL,
 349                                  .ldp_start_offset = file_offset
 350                                };
 351
 352     return ll_direct_rw_pages(env, io, rw, inode, &pvec);
 353 }
 354
 355 /* This is the maximum size of a single O_DIRECT request, based on a 128kB
 356  * kmalloc limit.  We need to fit all of the brw_page structs, each one
 357  * representing PAGE_SIZE worth of user data, into a single buffer, and
 358  * then truncate this to be a full-sized RPC.  This is 22MB for 4kB pages. */
 359 #define MAX_DIO_SIZE ((128 * 1024 / sizeof(struct brw_page) * CFS_PAGE_SIZE) & \
 360                       ~(PTLRPC_MAX_BRW_SIZE - 1))
 361 static ssize_t ll_direct_IO_26(int rw, struct kiocb *iocb,
 362                                const struct iovec *iov, loff_t file_offset,
 363                                unsigned long nr_segs)
 364 {
 365         struct lu_env *env;
 366         struct cl_io *io;
 367         struct file *file = iocb->ki_filp;
 368         struct inode *inode = file->f_mapping->host;
 369         struct ccc_object *obj = cl_inode2ccc(inode);
 370         ssize_t count = iov_length(iov, nr_segs);
 371         ssize_t tot_bytes = 0, result = 0;
 372         struct ll_inode_info *lli = ll_i2info(inode);
 373         struct lov_stripe_md *lsm = lli->lli_smd;
 374         unsigned long seg = 0;
 375         size_t size = MAX_DIO_SIZE;
 376         int refcheck;
 377         ENTRY;
 378
 379         if (!lli->lli_smd || !lli->lli_smd->lsm_object_id)
 380                 RETURN(-EBADF);
 381
 382         /* FIXME: io smaller than PAGE_SIZE is broken on ia64 ??? */
 383         if ((file_offset & ~CFS_PAGE_MASK) || (count & ~CFS_PAGE_MASK))
 384                 RETURN(-EINVAL);
 385
 386         CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p), size="LPSZ" (max %lu), "
 387                "offset=%lld=%llx, pages "LPSZ" (max %lu)\n",
 388                inode->i_ino, inode->i_generation, inode, count, MAX_DIO_SIZE,
 389                file_offset, file_offset, count >> CFS_PAGE_SHIFT,
 390                MAX_DIO_SIZE >> CFS_PAGE_SHIFT);
 391
 392         /* Check that all user buffers are aligned as well */
 393         for (seg = 0; seg < nr_segs; seg++) {
 394                 if (((unsigned long)iov[seg].iov_base & ~CFS_PAGE_MASK) ||
 395                     (iov[seg].iov_len & ~CFS_PAGE_MASK))
 396                         RETURN(-EINVAL);
 397         }
 398
 399         env = cl_env_get(&refcheck);
 400         LASSERT(!IS_ERR(env));
 401         io = ccc_env_io(env)->cui_cl.cis_io;
 402         LASSERT(io != NULL);
 403
 404         /* 0. Need locking between buffered and direct access. and race with
 405          *size changing by concurrent truncates and writes.
 406          * 1. Need inode sem to operate transient pages. */
 407         if (rw == READ)
 408                 LOCK_INODE_MUTEX(inode);
 409
 410         LASSERT(obj->cob_transient_pages == 0);
 411         for (seg = 0; seg < nr_segs; seg++) {
 412                 size_t iov_left = iov[seg].iov_len;
 413                 unsigned long user_addr = (unsigned long)iov[seg].iov_base;
 414
 415                 if (rw == READ) {
 416                         if (file_offset >= inode->i_size)
 417                                 break;
 418                         if (file_offset + iov_left > inode->i_size)
 419                                 iov_left = inode->i_size - file_offset;
 420                 }
 421
 422                 while (iov_left > 0) {
 423                         struct page **pages;
 424                         int page_count, max_pages = 0;
 425                         size_t bytes;
 426
 427                         bytes = min(size,iov_left);
 428                         page_count = ll_get_user_pages(rw, user_addr,
 429                                                        bytes,
 430                                                        &pages, &max_pages);
 431                         if (likely(page_count > 0)) {
 432                                 if (unlikely(page_count <  max_pages))
 433                                         bytes = page_count << CFS_PAGE_SHIFT;
 434                                 result = ll_direct_IO_26_seg(env, io, rw, inode,
 435                                                              file->f_mapping,
 436                                                              bytes,
 437                                                              file_offset, pages,
 438                                                              page_count);
 439                                 ll_free_user_pages(pages, max_pages, rw==READ);
 440                         } else if (page_count == 0) {
 441                                 GOTO(out, result = -EFAULT);
 442                         } else {
 443                                 result = page_count;
 444                         }
 445                         if (unlikely(result <= 0)) {
 446                                 /* If we can't allocate a large enough buffer
 447                                  * for the request, shrink it to a smaller
 448                                  * PAGE_SIZE multiple and try again.
 449                                  * We should always be able to kmalloc for a
 450                                  * page worth of page pointers = 4MB on i386. */
 451                                 if (result == -ENOMEM &&
 452                                     size > (CFS_PAGE_SIZE / sizeof(*pages)) *
 453                                            CFS_PAGE_SIZE) {
 454                                         size = ((((size / 2) - 1) |
 455                                                  ~CFS_PAGE_MASK) + 1) &
 456                                                 CFS_PAGE_MASK;
 457                                         CDEBUG(D_VFSTRACE, "DIO size now %u\n",
 458                                                (int)size);
 459                                         continue;
 460                                 }
 461
 462                                 GOTO(out, result);
 463                         }
 464
 465                         tot_bytes += result;
 466                         file_offset += result;
 467                         iov_left -= result;
 468                         user_addr += result;
 469                 }
 470         }
 471 out:
 472         LASSERT(obj->cob_transient_pages == 0);
 473         if (rw == READ)
 474                 UNLOCK_INODE_MUTEX(inode);
 475
 476         if (tot_bytes > 0) {
 477                 if (rw == WRITE) {
 478                         lov_stripe_lock(lsm);
 479                         obd_adjust_kms(ll_i2dtexp(inode), lsm, file_offset, 0);
 480                         lov_stripe_unlock(lsm);
 481                 }
 482         }
 483
 484         cl_env_put(env, &refcheck);
 485         RETURN(tot_bytes ? : result);
 486 }
 487
 488 struct address_space_operations ll_aops = {
 489         .readpage       = ll_readpage,
 490 //        .readpages      = ll_readpages,
 491         .direct_IO      = ll_direct_IO_26,
 492         .writepage      = ll_writepage,
 493         .writepages     = generic_writepages,
 494         .set_page_dirty = ll_set_page_dirty,
 495         .sync_page      = NULL,
 496         .prepare_write  = ll_prepare_write,
 497         .commit_write   = ll_commit_write,
 498         .invalidatepage = ll_invalidatepage,
 499         .releasepage    = (void *)ll_releasepage,
 500         .bmap           = NULL
 501 };