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  */
  31
  32 #include <linux/errno.h>
  33 #include <linux/delay.h>
  34 #include <linux/kernel.h>
  35 #include <linux/mm.h>
  36 #include <linux/file.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  * \param mkwrite - whether it is mmap write.
  81  *
  82  * \return error codes from cl_io_init.
  83  */
  84 static struct cl_io *
  85 ll_fault_io_init(struct lu_env *env, struct vm_area_struct *vma,
  86                 pgoff_t index, bool mkwrite)
  87 {
  88         struct file            *file = vma->vm_file;
  89         struct inode           *inode = file_inode(file);
  90         struct cl_io           *io;
  91         struct cl_fault_io     *fio;
  92         int                     rc;
  93         ENTRY;
  94
  95         if (ll_file_nolock(file))
  96                 RETURN(ERR_PTR(-EOPNOTSUPP));
  97
  98 restart:
  99         io = vvp_env_thread_io(env);
 100         io->ci_obj = ll_i2info(inode)->lli_clob;
 101         LASSERT(io->ci_obj != NULL);
 102
 103         fio = &io->u.ci_fault;
 104         fio->ft_index = index;
 105         fio->ft_executable = vma->vm_flags & VM_EXEC;
 106
 107         if (mkwrite) {
 108                 fio->ft_mkwrite = 1;
 109                 fio->ft_writable = 1;
 110         }
 111
 112         CDEBUG(D_MMAP,
 113                DFID": vma=%p start=%#lx end=%#lx vm_flags=%#lx idx=%lu\n",
 114                PFID(&ll_i2info(inode)->lli_fid), vma, vma->vm_start,
 115                vma->vm_end, vma->vm_flags, fio->ft_index);
 116
 117         if (vma->vm_flags & VM_SEQ_READ)
 118                 io->ci_seq_read = 1;
 119         else if (vma->vm_flags & VM_RAND_READ)
 120                 io->ci_rand_read = 1;
 121
 122         rc = cl_io_init(env, io, CIT_FAULT, io->ci_obj);
 123         if (rc == 0) {
 124                 struct vvp_io *vio = vvp_env_io(env);
 125                 struct ll_file_data *fd = file->private_data;
 126
 127                 LASSERT(vio->vui_cl.cis_io == io);
 128
 129                 /* mmap lock must be MANDATORY it has to cache
 130                  * pages. */
 131                 io->ci_lockreq = CILR_MANDATORY;
 132                 vio->vui_fd = fd;
 133         } else {
 134                 cl_io_fini(env, io);
 135                 if (io->ci_need_restart)
 136                         goto restart;
 137
 138                 io = ERR_PTR(rc);
 139         }
 140
 141         RETURN(io);
 142 }
 143
 144 /* Sharing code of page_mkwrite method for rhel5 and rhel6 */
 145 static int ll_page_mkwrite0(struct vm_area_struct *vma, struct page *vmpage,
 146                             bool *retry)
 147 {
 148         struct lu_env           *env;
 149         struct cl_io            *io;
 150         struct vvp_io           *vio;
 151         int                      result;
 152         __u16                    refcheck;
 153         sigset_t old, new;
 154         struct inode             *inode = NULL;
 155         struct ll_inode_info     *lli;
 156         ENTRY;
 157
 158         LASSERT(vmpage != NULL);
 159         env = cl_env_get(&refcheck);
 160         if (IS_ERR(env))
 161                 RETURN(PTR_ERR(env));
 162
 163         io = ll_fault_io_init(env, vma, vmpage->index, true);
 164         if (IS_ERR(io))
 165                 GOTO(out, result = PTR_ERR(io));
 166
 167         result = io->ci_result;
 168         if (result < 0)
 169                 GOTO(out_io, result);
 170
 171         vio = vvp_env_io(env);
 172         vio->u.fault.ft_vma    = vma;
 173         vio->u.fault.ft_vmpage = vmpage;
 174
 175         siginitsetinv(&new, sigmask(SIGKILL) | sigmask(SIGTERM));
 176         sigprocmask(SIG_BLOCK, &new, &old);
 177
 178         inode = vvp_object_inode(io->ci_obj);
 179         lli = ll_i2info(inode);
 180
 181         result = cl_io_loop(env, io);
 182
 183         sigprocmask(SIG_SETMASK, &old, NULL);
 184
 185         if (result == 0) {
 186                 lock_page(vmpage);
 187                 if (vmpage->mapping == NULL) {
 188                         unlock_page(vmpage);
 189
 190                         /* page was truncated and lock was cancelled, return
 191                          * ENODATA so that VM_FAULT_NOPAGE will be returned
 192                          * to handle_mm_fault(). */
 193                         if (result == 0)
 194                                 result = -ENODATA;
 195                 } else if (!PageDirty(vmpage)) {
 196                         /* race, the page has been cleaned by ptlrpcd after
 197                          * it was unlocked, it has to be added into dirty
 198                          * cache again otherwise this soon-to-dirty page won't
 199                          * consume any grants, even worse if this page is being
 200                          * transferred because it will break RPC checksum.
 201                          */
 202                         unlock_page(vmpage);
 203
 204                         CDEBUG(D_MMAP, "Race on page_mkwrite %p/%lu, page has "
 205                                "been written out, retry.\n",
 206                                vmpage, vmpage->index);
 207
 208                         *retry = true;
 209                         result = -EAGAIN;
 210                 }
 211
 212                 if (result == 0)
 213                         set_bit(LLIF_DATA_MODIFIED, &lli->lli_flags);
 214         }
 215         EXIT;
 216
 217 out_io:
 218         cl_io_fini(env, io);
 219 out:
 220         cl_env_put(env, &refcheck);
 221         CDEBUG(D_MMAP, "%s mkwrite with %d\n", current->comm, result);
 222         LASSERT(ergo(result == 0, PageLocked(vmpage)));
 223
 224         /* if page has been unmapped, presumably due to lock reclaim for
 225          * concurrent usage, add some delay before retrying to prevent
 226          * entering live-lock situation with competitors
 227          */
 228         if (result == -ENODATA && inode != NULL) {
 229                 CDEBUG(D_MMAP, "delaying new page-fault for inode %p to "
 230                                "prevent live-lock\n", inode);
 231                 msleep(10);
 232         }
 233
 234         return result;
 235 }
 236
 237 static inline int to_fault_error(int result)
 238 {
 239         switch(result) {
 240         case 0:
 241                 result = VM_FAULT_LOCKED;
 242                 break;
 243         case -ENOMEM:
 244                 result = VM_FAULT_OOM;
 245                 break;
 246         default:
 247                 result = VM_FAULT_SIGBUS;
 248                 break;
 249         }
 250         return result;
 251 }
 252
 253 /**
 254  * Lustre implementation of a vm_operations_struct::fault() method, called by
 255  * VM to server page fault (both in kernel and user space).
 256  *
 257  * \param vma - is virtiual area struct related to page fault
 258  * \param vmf - structure which describe type and address where hit fault
 259  *
 260  * \return allocated and filled _locked_ page for address
 261  * \retval VM_FAULT_ERROR on general error
 262  * \retval NOPAGE_OOM not have memory for allocate new page
 263  */
 264 static vm_fault_t ll_fault0(struct vm_area_struct *vma, struct vm_fault *vmf)
 265 {
 266         struct inode *inode = file_inode(vma->vm_file);
 267         struct lu_env           *env;
 268         struct cl_io            *io;
 269         struct vvp_io           *vio = NULL;
 270         struct page             *vmpage;
 271         int                      result = 0;
 272         int                      fault_ret = 0;
 273         __u16                    refcheck;
 274         ENTRY;
 275
 276         env = cl_env_get(&refcheck);
 277         if (IS_ERR(env))
 278                 RETURN(PTR_ERR(env));
 279
 280         if (ll_sbi_has_fast_read(ll_i2sbi(inode))) {
 281                 /* do fast fault */
 282                 bool allow_retry = vmf->flags & FAULT_FLAG_ALLOW_RETRY;
 283                 bool has_retry = vmf->flags & FAULT_FLAG_RETRY_NOWAIT;
 284
 285                 /* To avoid loops, instruct downstream to not drop mmap_sem */
 286                 /**
 287                  * only need FAULT_FLAG_ALLOW_RETRY prior to Linux 5.1
 288                  * (6b4c9f4469819), where FAULT_FLAG_RETRY_NOWAIT is enough
 289                  * to not drop mmap_sem when failed to lock the page.
 290                  */
 291                 vmf->flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT;
 292                 ll_cl_add(inode, env, NULL, LCC_MMAP);
 293                 fault_ret = ll_filemap_fault(vma, vmf);
 294                 ll_cl_remove(inode, env);
 295                 if (!has_retry)
 296                         vmf->flags &= ~FAULT_FLAG_RETRY_NOWAIT;
 297                 if (!allow_retry)
 298                         vmf->flags &= ~FAULT_FLAG_ALLOW_RETRY;
 299
 300                 /* - If there is no error, then the page was found in cache and
 301                  *   uptodate;
 302                  * - If VM_FAULT_RETRY is set, the page existed but failed to
 303                  *   lock. We will try slow path to avoid loops.
 304                  * - Otherwise, it should try normal fault under DLM lock. */
 305                 if (!(fault_ret & VM_FAULT_RETRY) &&
 306                     !(fault_ret & VM_FAULT_ERROR))
 307                         GOTO(out, result = 0);
 308
 309                 fault_ret = 0;
 310         }
 311
 312         io = ll_fault_io_init(env, vma, vmf->pgoff, false);
 313         if (IS_ERR(io))
 314                 GOTO(out, result = PTR_ERR(io));
 315
 316         result = io->ci_result;
 317         if (result == 0) {
 318                 vio = vvp_env_io(env);
 319                 vio->u.fault.ft_vma       = vma;
 320                 vio->u.fault.ft_vmpage    = NULL;
 321                 vio->u.fault.ft_vmf = vmf;
 322                 vio->u.fault.ft_flags = 0;
 323                 vio->u.fault.ft_flags_valid = 0;
 324
 325                 /* May call ll_readpage() */
 326                 ll_cl_add(inode, env, io, LCC_MMAP);
 327
 328                 result = cl_io_loop(env, io);
 329
 330                 ll_cl_remove(inode, env);
 331
 332                 /* ft_flags are only valid if we reached
 333                  * the call to filemap_fault */
 334                 if (vio->u.fault.ft_flags_valid)
 335                         fault_ret = vio->u.fault.ft_flags;
 336
 337                 vmpage = vio->u.fault.ft_vmpage;
 338                 if (result != 0 && vmpage != NULL) {
 339                         put_page(vmpage);
 340                         vmf->page = NULL;
 341                 }
 342         }
 343         cl_io_fini(env, io);
 344
 345 out:
 346         cl_env_put(env, &refcheck);
 347         if (result != 0 && !(fault_ret & VM_FAULT_RETRY))
 348                 fault_ret |= to_fault_error(result);
 349
 350         CDEBUG(D_MMAP, "%s fault %d/%d\n", current->comm, fault_ret, result);
 351         RETURN(fault_ret);
 352 }
 353
 354 #ifdef HAVE_VM_OPS_USE_VM_FAULT_ONLY
 355 static vm_fault_t ll_fault(struct vm_fault *vmf)
 356 {
 357         struct vm_area_struct *vma = vmf->vma;
 358 #else
 359 static vm_fault_t ll_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
 360 {
 361 #endif
 362         int count = 0;
 363         bool printed = false;
 364         bool cached;
 365         vm_fault_t result;
 366         ktime_t kstart = ktime_get();
 367         sigset_t old, new;
 368
 369         result = pcc_fault(vma, vmf, &cached);
 370         if (cached)
 371                 goto out;
 372
 373         CDEBUG(D_MMAP, DFID": vma=%p start=%#lx end=%#lx vm_flags=%#lx\n",
 374                PFID(&ll_i2info(file_inode(vma->vm_file))->lli_fid),
 375                vma, vma->vm_start, vma->vm_end, vma->vm_flags);
 376
 377         /* Only SIGKILL and SIGTERM is allowed for fault/nopage/mkwrite
 378          * so that it can be killed by admin but not cause segfault by
 379          * other signals.
 380          */
 381         siginitsetinv(&new, sigmask(SIGKILL) | sigmask(SIGTERM));
 382         sigprocmask(SIG_BLOCK, &new, &old);
 383
 384         /* make sure offset is not a negative number */
 385         if (vmf->pgoff > (MAX_LFS_FILESIZE >> PAGE_SHIFT))
 386                 return VM_FAULT_SIGBUS;
 387
 388 restart:
 389         result = ll_fault0(vma, vmf);
 390         if (vmf->page &&
 391             !(result & (VM_FAULT_RETRY | VM_FAULT_ERROR | VM_FAULT_LOCKED))) {
 392                 struct page *vmpage = vmf->page;
 393
 394                 /* check if this page has been truncated */
 395                 lock_page(vmpage);
 396                 if (unlikely(vmpage->mapping == NULL)) { /* unlucky */
 397                         unlock_page(vmpage);
 398                         put_page(vmpage);
 399                         vmf->page = NULL;
 400
 401                         if (!printed && ++count > 16) {
 402                                 CWARN("the page is under heavy contention, maybe your app(%s) needs revising :-)\n",
 403                                       current->comm);
 404                                 printed = true;
 405                         }
 406
 407                         goto restart;
 408                 }
 409
 410                 result |= VM_FAULT_LOCKED;
 411         }
 412         sigprocmask(SIG_SETMASK, &old, NULL);
 413
 414 out:
 415         if (vmf->page && result == VM_FAULT_LOCKED) {
 416                 ll_rw_stats_tally(ll_i2sbi(file_inode(vma->vm_file)),
 417                                   current->pid, vma->vm_file->private_data,
 418                                   cl_offset(NULL, vmf->page->index), PAGE_SIZE,
 419                                   READ);
 420                 ll_stats_ops_tally(ll_i2sbi(file_inode(vma->vm_file)),
 421                                    LPROC_LL_FAULT,
 422                                    ktime_us_delta(ktime_get(), kstart));
 423         }
 424
 425         return result;
 426 }
 427
 428 #ifdef HAVE_VM_OPS_USE_VM_FAULT_ONLY
 429 static vm_fault_t ll_page_mkwrite(struct vm_fault *vmf)
 430 {
 431         struct vm_area_struct *vma = vmf->vma;
 432 #else
 433 static vm_fault_t ll_page_mkwrite(struct vm_area_struct *vma,
 434                                   struct vm_fault *vmf)
 435 {
 436 #endif
 437         int count = 0;
 438         bool printed = false;
 439         bool retry;
 440         bool cached;
 441         ktime_t kstart = ktime_get();
 442         vm_fault_t result;
 443
 444         result = pcc_page_mkwrite(vma, vmf, &cached);
 445         if (cached)
 446                 goto out;
 447
 448         file_update_time(vma->vm_file);
 449         do {
 450                 retry = false;
 451                 result = ll_page_mkwrite0(vma, vmf->page, &retry);
 452
 453                 if (!printed && ++count > 16) {
 454                         const struct dentry *de = file_dentry(vma->vm_file);
 455
 456                         CWARN("app(%s): the page %lu of file "DFID" is under heavy contention\n",
 457                               current->comm, vmf->pgoff,
 458                               PFID(ll_inode2fid(de->d_inode)));
 459                         printed = true;
 460                 }
 461         } while (retry);
 462
 463         switch (result) {
 464         case 0:
 465                 LASSERT(PageLocked(vmf->page));
 466                 result = VM_FAULT_LOCKED;
 467                 break;
 468         case -ENODATA:
 469         case -EFAULT:
 470                 result = VM_FAULT_NOPAGE;
 471                 break;
 472         case -ENOMEM:
 473                 result = VM_FAULT_OOM;
 474                 break;
 475         case -EAGAIN:
 476                 result = VM_FAULT_RETRY;
 477                 break;
 478         default:
 479                 result = VM_FAULT_SIGBUS;
 480                 break;
 481         }
 482
 483 out:
 484         if (result == VM_FAULT_LOCKED) {
 485                 ll_rw_stats_tally(ll_i2sbi(file_inode(vma->vm_file)),
 486                                   current->pid, vma->vm_file->private_data,
 487                                   cl_offset(NULL, vmf->page->index), PAGE_SIZE,
 488                                   WRITE);
 489                 ll_stats_ops_tally(ll_i2sbi(file_inode(vma->vm_file)),
 490                                    LPROC_LL_MKWRITE,
 491                                    ktime_us_delta(ktime_get(), kstart));
 492         }
 493
 494         return result;
 495 }
 496
 497 /**
 498  *  To avoid cancel the locks covering mmapped region for lock cache pressure,
 499  *  we track the mapped vma count in vvp_object::vob_mmap_cnt.
 500  */
 501 static void ll_vm_open(struct vm_area_struct * vma)
 502 {
 503         struct inode *inode    = file_inode(vma->vm_file);
 504         struct vvp_object *vob = cl_inode2vvp(inode);
 505
 506         ENTRY;
 507         LASSERT(atomic_read(&vob->vob_mmap_cnt) >= 0);
 508         atomic_inc(&vob->vob_mmap_cnt);
 509         pcc_vm_open(vma);
 510         EXIT;
 511 }
 512
 513 /**
 514  * Dual to ll_vm_open().
 515  */
 516 static void ll_vm_close(struct vm_area_struct *vma)
 517 {
 518         struct inode      *inode = file_inode(vma->vm_file);
 519         struct vvp_object *vob   = cl_inode2vvp(inode);
 520
 521         ENTRY;
 522         atomic_dec(&vob->vob_mmap_cnt);
 523         LASSERT(atomic_read(&vob->vob_mmap_cnt) >= 0);
 524         pcc_vm_close(vma);
 525         EXIT;
 526 }
 527
 528 static const struct vm_operations_struct ll_file_vm_ops = {
 529         .fault                  = ll_fault,
 530         .page_mkwrite           = ll_page_mkwrite,
 531         .open                   = ll_vm_open,
 532         .close                  = ll_vm_close,
 533 };
 534
 535 int ll_file_mmap(struct file *file, struct vm_area_struct * vma)
 536 {
 537         struct inode *inode = file_inode(file);
 538         ktime_t kstart = ktime_get();
 539         bool cached;
 540         int rc;
 541
 542         ENTRY;
 543         CDEBUG(D_VFSTRACE | D_MMAP,
 544                "VFS_Op: fid="DFID" vma=%p start=%#lx end=%#lx vm_flags=%#lx\n",
 545                PFID(&ll_i2info(inode)->lli_fid),
 546                vma, vma->vm_start, vma->vm_end, vma->vm_flags);
 547
 548         if (ll_file_nolock(file))
 549                 RETURN(-EOPNOTSUPP);
 550
 551         rc = pcc_file_mmap(file, vma, &cached);
 552         if (cached && rc != 0)
 553                 RETURN(rc);
 554
 555         rc = generic_file_mmap(file, vma);
 556         if (rc == 0) {
 557                 vma->vm_ops = &ll_file_vm_ops;
 558                 vma->vm_ops->open(vma);
 559                 /* update the inode's size and mtime */
 560                 if (!cached)
 561                         rc = ll_glimpse_size(inode);
 562         }
 563
 564         if (!rc)
 565                 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_MMAP,
 566                                    ktime_us_delta(ktime_get(), kstart));
 567
 568         RETURN(rc);
 569 }