lustre/llite/llite_mmap.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.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2004, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2011, 2016, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  */
  32
  33 #include <linux/errno.h>
  34 #include <linux/delay.h>
  35 #include <linux/kernel.h>
  36 #include <linux/mm.h>
  37
  38 #define DEBUG_SUBSYSTEM S_LLITE
  39
  40 #include "llite_internal.h"
  41 #include <lustre_compat.h>
  42
  43 static const struct vm_operations_struct ll_file_vm_ops;
  44
  45 void policy_from_vma(union ldlm_policy_data *policy, struct vm_area_struct *vma,
  46                      unsigned long addr, size_t count)
  47 {
  48         policy->l_extent.start = ((addr - vma->vm_start) & PAGE_MASK) +
  49                                  (vma->vm_pgoff << PAGE_SHIFT);
  50         policy->l_extent.end = (policy->l_extent.start + count - 1) |
  51                                ~PAGE_MASK;
  52 }
  53
  54 struct vm_area_struct *our_vma(struct mm_struct *mm, unsigned long addr,
  55                                size_t count)
  56 {
  57         struct vm_area_struct *vma, *ret = NULL;
  58         ENTRY;
  59
  60         /* mmap_lock must have been held by caller. */
  61         LASSERT(!mmap_write_trylock(mm));
  62
  63         for (vma = find_vma(mm, addr);
  64              vma != NULL && vma->vm_start < (addr + count);
  65              vma = vma->vm_next) {
  66                 if (vma->vm_ops && vma->vm_ops == &ll_file_vm_ops &&
  67                     vma->vm_flags & VM_SHARED) {
  68                         ret = vma;
  69                         break;
  70                 }
  71         }
  72         RETURN(ret);
  73 }
  74
  75 /**
  76  * API independent part for page fault initialization.
  77  * \param env - corespondent lu_env to processing
  78  * \param vma - virtual memory area addressed to page fault
  79  * \param index - page index corespondent to fault.
  80  *
  81  * \return error codes from cl_io_init.
  82  */
  83 static struct cl_io *
  84 ll_fault_io_init(struct lu_env *env, struct vm_area_struct *vma, pgoff_t index)
  85 {
  86         struct file            *file = vma->vm_file;
  87         struct inode           *inode = file_inode(file);
  88         struct cl_io           *io;
  89         struct cl_fault_io     *fio;
  90         int                     rc;
  91         ENTRY;
  92
  93         if (ll_file_nolock(file))
  94                 RETURN(ERR_PTR(-EOPNOTSUPP));
  95
  96 restart:
  97         io = vvp_env_thread_io(env);
  98         io->ci_obj = ll_i2info(inode)->lli_clob;
  99         LASSERT(io->ci_obj != NULL);
 100
 101         fio = &io->u.ci_fault;
 102         fio->ft_index = index;
 103         fio->ft_executable = vma->vm_flags & VM_EXEC;
 104
 105         CDEBUG(D_MMAP,
 106                DFID": vma=%p start=%#lx end=%#lx vm_flags=%#lx idx=%lu\n",
 107                PFID(&ll_i2info(inode)->lli_fid), vma, vma->vm_start,
 108                vma->vm_end, vma->vm_flags, fio->ft_index);
 109
 110         if (vma->vm_flags & VM_SEQ_READ)
 111                 io->ci_seq_read = 1;
 112         else if (vma->vm_flags & VM_RAND_READ)
 113                 io->ci_rand_read = 1;
 114
 115         rc = cl_io_init(env, io, CIT_FAULT, io->ci_obj);
 116         if (rc == 0) {
 117                 struct vvp_io *vio = vvp_env_io(env);
 118                 struct ll_file_data *fd = file->private_data;
 119
 120                 LASSERT(vio->vui_cl.cis_io == io);
 121
 122                 /* mmap lock must be MANDATORY it has to cache
 123                  * pages. */
 124                 io->ci_lockreq = CILR_MANDATORY;
 125                 vio->vui_fd = fd;
 126         } else {
 127                 LASSERT(rc < 0);
 128                 cl_io_fini(env, io);
 129                 if (io->ci_need_restart)
 130                         goto restart;
 131
 132                 io = ERR_PTR(rc);
 133         }
 134
 135         RETURN(io);
 136 }
 137
 138 /* Sharing code of page_mkwrite method for rhel5 and rhel6 */
 139 static int ll_page_mkwrite0(struct vm_area_struct *vma, struct page *vmpage,
 140                             bool *retry)
 141 {
 142         struct lu_env           *env;
 143         struct cl_io            *io;
 144         struct vvp_io           *vio;
 145         int                      result;
 146         __u16                    refcheck;
 147         sigset_t old, new;
 148         struct inode             *inode = NULL;
 149         struct ll_inode_info     *lli;
 150         ENTRY;
 151
 152         LASSERT(vmpage != NULL);
 153         env = cl_env_get(&refcheck);
 154         if (IS_ERR(env))
 155                 RETURN(PTR_ERR(env));
 156
 157         io = ll_fault_io_init(env, vma, vmpage->index);
 158         if (IS_ERR(io))
 159                 GOTO(out, result = PTR_ERR(io));
 160
 161         result = io->ci_result;
 162         if (result < 0)
 163                 GOTO(out_io, result);
 164
 165         io->u.ci_fault.ft_mkwrite = 1;
 166         io->u.ci_fault.ft_writable = 1;
 167
 168         vio = vvp_env_io(env);
 169         vio->u.fault.ft_vma    = vma;
 170         vio->u.fault.ft_vmpage = vmpage;
 171
 172         siginitsetinv(&new, sigmask(SIGKILL) | sigmask(SIGTERM));
 173         sigprocmask(SIG_BLOCK, &new, &old);
 174
 175         inode = vvp_object_inode(io->ci_obj);
 176         lli = ll_i2info(inode);
 177
 178         result = cl_io_loop(env, io);
 179
 180         sigprocmask(SIG_SETMASK, &old, NULL);
 181
 182         if (result == 0) {
 183                 lock_page(vmpage);
 184                 if (vmpage->mapping == NULL) {
 185                         unlock_page(vmpage);
 186
 187                         /* page was truncated and lock was cancelled, return
 188                          * ENODATA so that VM_FAULT_NOPAGE will be returned
 189                          * to handle_mm_fault(). */
 190                         if (result == 0)
 191                                 result = -ENODATA;
 192                 } else if (!PageDirty(vmpage)) {
 193                         /* race, the page has been cleaned by ptlrpcd after
 194                          * it was unlocked, it has to be added into dirty
 195                          * cache again otherwise this soon-to-dirty page won't
 196                          * consume any grants, even worse if this page is being
 197                          * transferred because it will break RPC checksum.
 198                          */
 199                         unlock_page(vmpage);
 200
 201                         CDEBUG(D_MMAP, "Race on page_mkwrite %p/%lu, page has "
 202                                "been written out, retry.\n",
 203                                vmpage, vmpage->index);
 204
 205                         *retry = true;
 206                         result = -EAGAIN;
 207                 }
 208
 209                 if (result == 0)
 210                         set_bit(LLIF_DATA_MODIFIED, &lli->lli_flags);
 211         }
 212         EXIT;
 213
 214 out_io:
 215         cl_io_fini(env, io);
 216 out:
 217         cl_env_put(env, &refcheck);
 218         CDEBUG(D_MMAP, "%s mkwrite with %d\n", current->comm, result);
 219         LASSERT(ergo(result == 0, PageLocked(vmpage)));
 220
 221         /* if page has been unmapped, presumably due to lock reclaim for
 222          * concurrent usage, add some delay before retrying to prevent
 223          * entering live-lock situation with competitors
 224          */
 225         if (result == -ENODATA && inode != NULL) {
 226                 CDEBUG(D_MMAP, "delaying new page-fault for inode %p to "
 227                                "prevent live-lock\n", inode);
 228                 msleep(10);
 229         }
 230
 231         return result;
 232 }
 233
 234 static inline int to_fault_error(int result)
 235 {
 236         switch(result) {
 237         case 0:
 238                 result = VM_FAULT_LOCKED;
 239                 break;
 240         case -ENOMEM:
 241                 result = VM_FAULT_OOM;
 242                 break;
 243         default:
 244                 result = VM_FAULT_SIGBUS;
 245                 break;
 246         }
 247         return result;
 248 }
 249
 250 /**
 251  * Lustre implementation of a vm_operations_struct::fault() method, called by
 252  * VM to server page fault (both in kernel and user space).
 253  *
 254  * \param vma - is virtiual area struct related to page fault
 255  * \param vmf - structure which describe type and address where hit fault
 256  *
 257  * \return allocated and filled _locked_ page for address
 258  * \retval VM_FAULT_ERROR on general error
 259  * \retval NOPAGE_OOM not have memory for allocate new page
 260  */
 261 static vm_fault_t ll_fault0(struct vm_area_struct *vma, struct vm_fault *vmf)
 262 {
 263         struct lu_env           *env;
 264         struct cl_io            *io;
 265         struct vvp_io           *vio = NULL;
 266         struct page             *vmpage;
 267         int                      result = 0;
 268         int                      fault_ret = 0;
 269         __u16                    refcheck;
 270         ENTRY;
 271
 272         env = cl_env_get(&refcheck);
 273         if (IS_ERR(env))
 274                 RETURN(PTR_ERR(env));
 275
 276         if (ll_sbi_has_fast_read(ll_i2sbi(file_inode(vma->vm_file)))) {
 277                 /* do fast fault */
 278                 bool has_retry = vmf->flags & FAULT_FLAG_RETRY_NOWAIT;
 279
 280                 /* To avoid loops, instruct downstream to not drop mmap_sem */
 281                 vmf->flags |= FAULT_FLAG_RETRY_NOWAIT;
 282                 ll_cl_add(vma->vm_file, env, NULL, LCC_MMAP);
 283                 fault_ret = ll_filemap_fault(vma, vmf);
 284                 ll_cl_remove(vma->vm_file, env);
 285                 if (!has_retry)
 286                         vmf->flags &= ~FAULT_FLAG_RETRY_NOWAIT;
 287
 288                 /* - If there is no error, then the page was found in cache and
 289                  *   uptodate;
 290                  * - If VM_FAULT_RETRY is set, the page existed but failed to
 291                  *   lock. We will try slow path to avoid loops.
 292                  * - Otherwise, it should try normal fault under DLM lock. */
 293                 if (!(fault_ret & VM_FAULT_RETRY) &&
 294                     !(fault_ret & VM_FAULT_ERROR))
 295                         GOTO(out, result = 0);
 296
 297                 fault_ret = 0;
 298         }
 299
 300         io = ll_fault_io_init(env, vma, vmf->pgoff);
 301         if (IS_ERR(io))
 302                 GOTO(out, result = PTR_ERR(io));
 303
 304         result = io->ci_result;
 305         if (result == 0) {
 306                 vio = vvp_env_io(env);
 307                 vio->u.fault.ft_vma       = vma;
 308                 vio->u.fault.ft_vmpage    = NULL;
 309                 vio->u.fault.ft_vmf = vmf;
 310                 vio->u.fault.ft_flags = 0;
 311                 vio->u.fault.ft_flags_valid = 0;
 312
 313                 /* May call ll_readpage() */
 314                 ll_cl_add(vma->vm_file, env, io, LCC_MMAP);
 315
 316                 result = cl_io_loop(env, io);
 317
 318                 ll_cl_remove(vma->vm_file, env);
 319
 320                 /* ft_flags are only valid if we reached
 321                  * the call to filemap_fault */
 322                 if (vio->u.fault.ft_flags_valid)
 323                         fault_ret = vio->u.fault.ft_flags;
 324
 325                 vmpage = vio->u.fault.ft_vmpage;
 326                 if (result != 0 && vmpage != NULL) {
 327                         put_page(vmpage);
 328                         vmf->page = NULL;
 329                 }
 330         }
 331         cl_io_fini(env, io);
 332
 333 out:
 334         cl_env_put(env, &refcheck);
 335         if (result != 0 && !(fault_ret & VM_FAULT_RETRY))
 336                 fault_ret |= to_fault_error(result);
 337
 338         CDEBUG(D_MMAP, "%s fault %d/%d\n", current->comm, fault_ret, result);
 339         RETURN(fault_ret);
 340 }
 341
 342 #ifdef HAVE_VM_OPS_USE_VM_FAULT_ONLY
 343 static vm_fault_t ll_fault(struct vm_fault *vmf)
 344 {
 345         struct vm_area_struct *vma = vmf->vma;
 346 #else
 347 static vm_fault_t ll_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
 348 {
 349 #endif
 350         int count = 0;
 351         bool printed = false;
 352         bool cached;
 353         vm_fault_t result;
 354         ktime_t kstart = ktime_get();
 355         sigset_t old, new;
 356
 357         result = pcc_fault(vma, vmf, &cached);
 358         if (cached)
 359                 goto out;
 360
 361         CDEBUG(D_MMAP, DFID": vma=%p start=%#lx end=%#lx vm_flags=%#lx\n",
 362                PFID(&ll_i2info(file_inode(vma->vm_file))->lli_fid),
 363                vma, vma->vm_start, vma->vm_end, vma->vm_flags);
 364
 365         /* Only SIGKILL and SIGTERM is allowed for fault/nopage/mkwrite
 366          * so that it can be killed by admin but not cause segfault by
 367          * other signals.
 368          */
 369         siginitsetinv(&new, sigmask(SIGKILL) | sigmask(SIGTERM));
 370         sigprocmask(SIG_BLOCK, &new, &old);
 371
 372         /* make sure offset is not a negative number */
 373         if (vmf->pgoff > (MAX_LFS_FILESIZE >> PAGE_SHIFT))
 374                 return VM_FAULT_SIGBUS;
 375
 376 restart:
 377         result = ll_fault0(vma, vmf);
 378         if (vmf->page &&
 379             !(result & (VM_FAULT_RETRY | VM_FAULT_ERROR | VM_FAULT_LOCKED))) {
 380                 struct page *vmpage = vmf->page;
 381
 382                 /* check if this page has been truncated */
 383                 lock_page(vmpage);
 384                 if (unlikely(vmpage->mapping == NULL)) { /* unlucky */
 385                         unlock_page(vmpage);
 386                         put_page(vmpage);
 387                         vmf->page = NULL;
 388
 389                         if (!printed && ++count > 16) {
 390                                 CWARN("the page is under heavy contention, maybe your app(%s) needs revising :-)\n",
 391                                       current->comm);
 392                                 printed = true;
 393                         }
 394
 395                         goto restart;
 396                 }
 397
 398                 result |= VM_FAULT_LOCKED;
 399         }
 400         sigprocmask(SIG_SETMASK, &old, NULL);
 401
 402 out:
 403         if (vmf->page && result == VM_FAULT_LOCKED) {
 404                 ll_rw_stats_tally(ll_i2sbi(file_inode(vma->vm_file)),
 405                                   current->pid, vma->vm_file->private_data,
 406                                   cl_offset(NULL, vmf->page->index), PAGE_SIZE,
 407                                   READ);
 408                 ll_stats_ops_tally(ll_i2sbi(file_inode(vma->vm_file)),
 409                                    LPROC_LL_FAULT,
 410                                    ktime_us_delta(ktime_get(), kstart));
 411         }
 412
 413         return result;
 414 }
 415
 416 #ifdef HAVE_VM_OPS_USE_VM_FAULT_ONLY
 417 static vm_fault_t ll_page_mkwrite(struct vm_fault *vmf)
 418 {
 419         struct vm_area_struct *vma = vmf->vma;
 420 #else
 421 static vm_fault_t ll_page_mkwrite(struct vm_area_struct *vma,
 422                                   struct vm_fault *vmf)
 423 {
 424 #endif
 425         int count = 0;
 426         bool printed = false;
 427         bool retry;
 428         bool cached;
 429         ktime_t kstart = ktime_get();
 430         vm_fault_t result;
 431
 432         result = pcc_page_mkwrite(vma, vmf, &cached);
 433         if (cached)
 434                 goto out;
 435
 436         file_update_time(vma->vm_file);
 437         do {
 438                 retry = false;
 439                 result = ll_page_mkwrite0(vma, vmf->page, &retry);
 440
 441                 if (!printed && ++count > 16) {
 442                         const struct dentry *de = file_dentry(vma->vm_file);
 443
 444                         CWARN("app(%s): the page %lu of file "DFID" is under heavy contention\n",
 445                               current->comm, vmf->pgoff,
 446                               PFID(ll_inode2fid(de->d_inode)));
 447                         printed = true;
 448                 }
 449         } while (retry);
 450
 451         switch (result) {
 452         case 0:
 453                 LASSERT(PageLocked(vmf->page));
 454                 result = VM_FAULT_LOCKED;
 455                 break;
 456         case -ENODATA:
 457         case -EFAULT:
 458                 result = VM_FAULT_NOPAGE;
 459                 break;
 460         case -ENOMEM:
 461                 result = VM_FAULT_OOM;
 462                 break;
 463         case -EAGAIN:
 464                 result = VM_FAULT_RETRY;
 465                 break;
 466         default:
 467                 result = VM_FAULT_SIGBUS;
 468                 break;
 469         }
 470
 471 out:
 472         if (result == VM_FAULT_LOCKED) {
 473                 ll_rw_stats_tally(ll_i2sbi(file_inode(vma->vm_file)),
 474                                   current->pid, vma->vm_file->private_data,
 475                                   cl_offset(NULL, vmf->page->index), PAGE_SIZE,
 476                                   WRITE);
 477                 ll_stats_ops_tally(ll_i2sbi(file_inode(vma->vm_file)),
 478                                    LPROC_LL_MKWRITE,
 479                                    ktime_us_delta(ktime_get(), kstart));
 480         }
 481
 482         return result;
 483 }
 484
 485 /**
 486  *  To avoid cancel the locks covering mmapped region for lock cache pressure,
 487  *  we track the mapped vma count in vvp_object::vob_mmap_cnt.
 488  */
 489 static void ll_vm_open(struct vm_area_struct * vma)
 490 {
 491         struct inode *inode    = file_inode(vma->vm_file);
 492         struct vvp_object *vob = cl_inode2vvp(inode);
 493
 494         ENTRY;
 495         LASSERT(atomic_read(&vob->vob_mmap_cnt) >= 0);
 496         atomic_inc(&vob->vob_mmap_cnt);
 497         pcc_vm_open(vma);
 498         EXIT;
 499 }
 500
 501 /**
 502  * Dual to ll_vm_open().
 503  */
 504 static void ll_vm_close(struct vm_area_struct *vma)
 505 {
 506         struct inode      *inode = file_inode(vma->vm_file);
 507         struct vvp_object *vob   = cl_inode2vvp(inode);
 508
 509         ENTRY;
 510         atomic_dec(&vob->vob_mmap_cnt);
 511         LASSERT(atomic_read(&vob->vob_mmap_cnt) >= 0);
 512         pcc_vm_close(vma);
 513         EXIT;
 514 }
 515
 516 static const struct vm_operations_struct ll_file_vm_ops = {
 517         .fault                  = ll_fault,
 518         .page_mkwrite           = ll_page_mkwrite,
 519         .open                   = ll_vm_open,
 520         .close                  = ll_vm_close,
 521 };
 522
 523 int ll_file_mmap(struct file *file, struct vm_area_struct * vma)
 524 {
 525         struct inode *inode = file_inode(file);
 526         ktime_t kstart = ktime_get();
 527         bool cached;
 528         int rc;
 529
 530         ENTRY;
 531         CDEBUG(D_VFSTRACE | D_MMAP,
 532                "VFS_Op: fid="DFID" vma=%p start=%#lx end=%#lx vm_flags=%#lx\n",
 533                PFID(&ll_i2info(inode)->lli_fid),
 534                vma, vma->vm_start, vma->vm_end, vma->vm_flags);
 535
 536         if (ll_file_nolock(file))
 537                 RETURN(-EOPNOTSUPP);
 538
 539         rc = pcc_file_mmap(file, vma, &cached);
 540         if (cached && rc != 0)
 541                 RETURN(rc);
 542
 543         rc = generic_file_mmap(file, vma);
 544         if (rc == 0) {
 545                 vma->vm_ops = &ll_file_vm_ops;
 546                 vma->vm_ops->open(vma);
 547                 /* update the inode's size and mtime */
 548                 if (!cached)
 549                         rc = ll_glimpse_size(inode);
 550         }
 551
 552         if (!rc)
 553                 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_MMAP,
 554                                    ktime_us_delta(ktime_get(), kstart));
 555
 556         RETURN(rc);
 557 }