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_sem must have been held by caller. */
  61         LASSERT(!down_write_trylock(&mm->mmap_sem));
  62
  63         for(vma = find_vma(mm, addr);
  64             vma != NULL && vma->vm_start < (addr + count); vma = vma->vm_next) {
  65                 if (vma->vm_ops && vma->vm_ops == &ll_file_vm_ops &&
  66                     vma->vm_flags & VM_SHARED) {
  67                         ret = vma;
  68                         break;
  69                 }
  70         }
  71         RETURN(ret);
  72 }
  73
  74 /**
  75  * API independent part for page fault initialization.
  76  * \param env - corespondent lu_env to processing
  77  * \param vma - virtual memory area addressed to page fault
  78  * \param index - page index corespondent to fault.
  79  *
  80  * \return error codes from cl_io_init.
  81  */
  82 static struct cl_io *
  83 ll_fault_io_init(struct lu_env *env, struct vm_area_struct *vma, pgoff_t index)
  84 {
  85         struct file            *file = vma->vm_file;
  86         struct inode           *inode = file_inode(file);
  87         struct cl_io           *io;
  88         struct cl_fault_io     *fio;
  89         int                     rc;
  90         ENTRY;
  91
  92         if (ll_file_nolock(file))
  93                 RETURN(ERR_PTR(-EOPNOTSUPP));
  94
  95 restart:
  96         io = vvp_env_thread_io(env);
  97         io->ci_obj = ll_i2info(inode)->lli_clob;
  98         LASSERT(io->ci_obj != NULL);
  99
 100         fio = &io->u.ci_fault;
 101         fio->ft_index = index;
 102         fio->ft_executable = vma->vm_flags & VM_EXEC;
 103
 104         CDEBUG(D_MMAP,
 105                DFID": vma=%p start=%#lx end=%#lx vm_flags=%#lx idx=%lu\n",
 106                PFID(&ll_i2info(inode)->lli_fid), vma, vma->vm_start,
 107                vma->vm_end, vma->vm_flags, fio->ft_index);
 108
 109         if (vma->vm_flags & VM_SEQ_READ)
 110                 io->ci_seq_read = 1;
 111         else if (vma->vm_flags & VM_RAND_READ)
 112                 io->ci_rand_read = 1;
 113
 114         rc = cl_io_init(env, io, CIT_FAULT, io->ci_obj);
 115         if (rc == 0) {
 116                 struct vvp_io *vio = vvp_env_io(env);
 117                 struct ll_file_data *fd = file->private_data;
 118
 119                 LASSERT(vio->vui_cl.cis_io == io);
 120
 121                 /* mmap lock must be MANDATORY it has to cache
 122                  * pages. */
 123                 io->ci_lockreq = CILR_MANDATORY;
 124                 vio->vui_fd = fd;
 125         } else {
 126                 LASSERT(rc < 0);
 127                 cl_io_fini(env, io);
 128                 if (io->ci_need_restart)
 129                         goto restart;
 130
 131                 io = ERR_PTR(rc);
 132         }
 133
 134         RETURN(io);
 135 }
 136
 137 /* Sharing code of page_mkwrite method for rhel5 and rhel6 */
 138 static int ll_page_mkwrite0(struct vm_area_struct *vma, struct page *vmpage,
 139                             bool *retry)
 140 {
 141         struct lu_env           *env;
 142         struct cl_io            *io;
 143         struct vvp_io           *vio;
 144         int                      result;
 145         __u16                    refcheck;
 146         sigset_t old, new;
 147         struct inode             *inode = NULL;
 148         struct ll_inode_info     *lli;
 149         ENTRY;
 150
 151         LASSERT(vmpage != NULL);
 152         env = cl_env_get(&refcheck);
 153         if (IS_ERR(env))
 154                 RETURN(PTR_ERR(env));
 155
 156         io = ll_fault_io_init(env, vma, vmpage->index);
 157         if (IS_ERR(io))
 158                 GOTO(out, result = PTR_ERR(io));
 159
 160         result = io->ci_result;
 161         if (result < 0)
 162                 GOTO(out_io, result);
 163
 164         io->u.ci_fault.ft_mkwrite = 1;
 165         io->u.ci_fault.ft_writable = 1;
 166
 167         vio = vvp_env_io(env);
 168         vio->u.fault.ft_vma    = vma;
 169         vio->u.fault.ft_vmpage = vmpage;
 170
 171         siginitsetinv(&new, sigmask(SIGKILL) | sigmask(SIGTERM));
 172         sigprocmask(SIG_BLOCK, &new, &old);
 173
 174         inode = vvp_object_inode(io->ci_obj);
 175         lli = ll_i2info(inode);
 176
 177         result = cl_io_loop(env, io);
 178
 179         sigprocmask(SIG_SETMASK, &old, NULL);
 180
 181         if (result == 0) {
 182                 lock_page(vmpage);
 183                 if (vmpage->mapping == NULL) {
 184                         unlock_page(vmpage);
 185
 186                         /* page was truncated and lock was cancelled, return
 187                          * ENODATA so that VM_FAULT_NOPAGE will be returned
 188                          * to handle_mm_fault(). */
 189                         if (result == 0)
 190                                 result = -ENODATA;
 191                 } else if (!PageDirty(vmpage)) {
 192                         /* race, the page has been cleaned by ptlrpcd after
 193                          * it was unlocked, it has to be added into dirty
 194                          * cache again otherwise this soon-to-dirty page won't
 195                          * consume any grants, even worse if this page is being
 196                          * transferred because it will break RPC checksum.
 197                          */
 198                         unlock_page(vmpage);
 199
 200                         CDEBUG(D_MMAP, "Race on page_mkwrite %p/%lu, page has "
 201                                "been written out, retry.\n",
 202                                vmpage, vmpage->index);
 203
 204                         *retry = true;
 205                         result = -EAGAIN;
 206                 }
 207
 208                 if (result == 0)
 209                         ll_file_set_flag(lli, LLIF_DATA_MODIFIED);
 210         }
 211         EXIT;
 212
 213 out_io:
 214         cl_io_fini(env, io);
 215 out:
 216         cl_env_put(env, &refcheck);
 217         CDEBUG(D_MMAP, "%s mkwrite with %d\n", current->comm, result);
 218         LASSERT(ergo(result == 0, PageLocked(vmpage)));
 219
 220         /* if page has been unmapped, presumably due to lock reclaim for
 221          * concurrent usage, add some delay before retrying to prevent
 222          * entering live-lock situation with competitors
 223          */
 224         if (result == -ENODATA && inode != NULL) {
 225                 CDEBUG(D_MMAP, "delaying new page-fault for inode %p to "
 226                                "prevent live-lock\n", inode);
 227                 msleep(10);
 228         }
 229
 230         return result;
 231 }
 232
 233 static inline int to_fault_error(int result)
 234 {
 235         switch(result) {
 236         case 0:
 237                 result = VM_FAULT_LOCKED;
 238                 break;
 239         case -ENOMEM:
 240                 result = VM_FAULT_OOM;
 241                 break;
 242         default:
 243                 result = VM_FAULT_SIGBUS;
 244                 break;
 245         }
 246         return result;
 247 }
 248
 249 /**
 250  * Lustre implementation of a vm_operations_struct::fault() method, called by
 251  * VM to server page fault (both in kernel and user space).
 252  *
 253  * \param vma - is virtiual area struct related to page fault
 254  * \param vmf - structure which describe type and address where hit fault
 255  *
 256  * \return allocated and filled _locked_ page for address
 257  * \retval VM_FAULT_ERROR on general error
 258  * \retval NOPAGE_OOM not have memory for allocate new page
 259  */
 260 static vm_fault_t ll_fault0(struct vm_area_struct *vma, struct vm_fault *vmf)
 261 {
 262         struct lu_env           *env;
 263         struct cl_io            *io;
 264         struct vvp_io           *vio = NULL;
 265         struct page             *vmpage;
 266         int                      result = 0;
 267         int                      fault_ret = 0;
 268         __u16                    refcheck;
 269         ENTRY;
 270
 271         env = cl_env_get(&refcheck);
 272         if (IS_ERR(env))
 273                 RETURN(PTR_ERR(env));
 274
 275         if (ll_sbi_has_fast_read(ll_i2sbi(file_inode(vma->vm_file)))) {
 276                 /* do fast fault */
 277                 bool has_retry = vmf->flags & FAULT_FLAG_RETRY_NOWAIT;
 278
 279                 /* To avoid loops, instruct downstream to not drop mmap_sem */
 280                 vmf->flags |= FAULT_FLAG_RETRY_NOWAIT;
 281                 ll_cl_add(vma->vm_file, env, NULL, LCC_MMAP);
 282                 fault_ret = ll_filemap_fault(vma, vmf);
 283                 ll_cl_remove(vma->vm_file, env);
 284                 if (!has_retry)
 285                         vmf->flags &= ~FAULT_FLAG_RETRY_NOWAIT;
 286
 287                 /* - If there is no error, then the page was found in cache and
 288                  *   uptodate;
 289                  * - If VM_FAULT_RETRY is set, the page existed but failed to
 290                  *   lock. We will try slow path to avoid loops.
 291                  * - Otherwise, it should try normal fault under DLM lock. */
 292                 if (!(fault_ret & VM_FAULT_RETRY) &&
 293                     !(fault_ret & VM_FAULT_ERROR))
 294                         GOTO(out, result = 0);
 295
 296                 fault_ret = 0;
 297         }
 298
 299         io = ll_fault_io_init(env, vma, vmf->pgoff);
 300         if (IS_ERR(io))
 301                 GOTO(out, result = PTR_ERR(io));
 302
 303         result = io->ci_result;
 304         if (result == 0) {
 305                 vio = vvp_env_io(env);
 306                 vio->u.fault.ft_vma       = vma;
 307                 vio->u.fault.ft_vmpage    = NULL;
 308                 vio->u.fault.ft_vmf = vmf;
 309                 vio->u.fault.ft_flags = 0;
 310                 vio->u.fault.ft_flags_valid = 0;
 311
 312                 /* May call ll_readpage() */
 313                 ll_cl_add(vma->vm_file, env, io, LCC_MMAP);
 314
 315                 result = cl_io_loop(env, io);
 316
 317                 ll_cl_remove(vma->vm_file, env);
 318
 319                 /* ft_flags are only valid if we reached
 320                  * the call to filemap_fault */
 321                 if (vio->u.fault.ft_flags_valid)
 322                         fault_ret = vio->u.fault.ft_flags;
 323
 324                 vmpage = vio->u.fault.ft_vmpage;
 325                 if (result != 0 && vmpage != NULL) {
 326                         put_page(vmpage);
 327                         vmf->page = NULL;
 328                 }
 329         }
 330         cl_io_fini(env, io);
 331
 332 out:
 333         cl_env_put(env, &refcheck);
 334         if (result != 0 && !(fault_ret & VM_FAULT_RETRY))
 335                 fault_ret |= to_fault_error(result);
 336
 337         CDEBUG(D_MMAP, "%s fault %d/%d\n", current->comm, fault_ret, result);
 338         RETURN(fault_ret);
 339 }
 340
 341 #ifdef HAVE_VM_OPS_USE_VM_FAULT_ONLY
 342 static vm_fault_t ll_fault(struct vm_fault *vmf)
 343 {
 344         struct vm_area_struct *vma = vmf->vma;
 345 #else
 346 static vm_fault_t ll_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
 347 {
 348 #endif
 349         int count = 0;
 350         bool printed = false;
 351         bool cached;
 352         vm_fault_t result;
 353         ktime_t kstart = ktime_get();
 354         sigset_t old, new;
 355
 356         result = pcc_fault(vma, vmf, &cached);
 357         if (cached)
 358                 goto out;
 359
 360         CDEBUG(D_MMAP, DFID": vma=%p start=%#lx end=%#lx vm_flags=%#lx\n",
 361                PFID(&ll_i2info(file_inode(vma->vm_file))->lli_fid),
 362                vma, vma->vm_start, vma->vm_end, vma->vm_flags);
 363
 364         /* Only SIGKILL and SIGTERM is allowed for fault/nopage/mkwrite
 365          * so that it can be killed by admin but not cause segfault by
 366          * other signals.
 367          */
 368         siginitsetinv(&new, sigmask(SIGKILL) | sigmask(SIGTERM));
 369         sigprocmask(SIG_BLOCK, &new, &old);
 370
 371         /* make sure offset is not a negative number */
 372         if (vmf->pgoff > (MAX_LFS_FILESIZE >> PAGE_SHIFT))
 373                 return VM_FAULT_SIGBUS;
 374
 375 restart:
 376         result = ll_fault0(vma, vmf);
 377         if (vmf->page &&
 378             !(result & (VM_FAULT_RETRY | VM_FAULT_ERROR | VM_FAULT_LOCKED))) {
 379                 struct page *vmpage = vmf->page;
 380
 381                 /* check if this page has been truncated */
 382                 lock_page(vmpage);
 383                 if (unlikely(vmpage->mapping == NULL)) { /* unlucky */
 384                         unlock_page(vmpage);
 385                         put_page(vmpage);
 386                         vmf->page = NULL;
 387
 388                         if (!printed && ++count > 16) {
 389                                 CWARN("the page is under heavy contention, maybe your app(%s) needs revising :-)\n",
 390                                       current->comm);
 391                                 printed = true;
 392                         }
 393
 394                         goto restart;
 395                 }
 396
 397                 result |= VM_FAULT_LOCKED;
 398         }
 399         sigprocmask(SIG_SETMASK, &old, NULL);
 400
 401 out:
 402         if (vmf->page && result == VM_FAULT_LOCKED) {
 403                 ll_rw_stats_tally(ll_i2sbi(file_inode(vma->vm_file)),
 404                                   current->pid, vma->vm_file->private_data,
 405                                   cl_offset(NULL, vmf->page->index), PAGE_SIZE,
 406                                   READ);
 407                 ll_stats_ops_tally(ll_i2sbi(file_inode(vma->vm_file)),
 408                                    LPROC_LL_FAULT,
 409                                    ktime_us_delta(ktime_get(), kstart));
 410         }
 411
 412         return result;
 413 }
 414
 415 #ifdef HAVE_VM_OPS_USE_VM_FAULT_ONLY
 416 static vm_fault_t ll_page_mkwrite(struct vm_fault *vmf)
 417 {
 418         struct vm_area_struct *vma = vmf->vma;
 419 #else
 420 static vm_fault_t ll_page_mkwrite(struct vm_area_struct *vma,
 421                                   struct vm_fault *vmf)
 422 {
 423 #endif
 424         int count = 0;
 425         bool printed = false;
 426         bool retry;
 427         bool cached;
 428         ktime_t kstart = ktime_get();
 429         vm_fault_t result;
 430
 431         result = pcc_page_mkwrite(vma, vmf, &cached);
 432         if (cached)
 433                 goto out;
 434
 435         file_update_time(vma->vm_file);
 436         do {
 437                 retry = false;
 438                 result = ll_page_mkwrite0(vma, vmf->page, &retry);
 439
 440                 if (!printed && ++count > 16) {
 441                         const struct dentry *de = file_dentry(vma->vm_file);
 442
 443                         CWARN("app(%s): the page %lu of file "DFID" is under heavy contention\n",
 444                               current->comm, vmf->pgoff,
 445                               PFID(ll_inode2fid(de->d_inode)));
 446                         printed = true;
 447                 }
 448         } while (retry);
 449
 450         switch (result) {
 451         case 0:
 452                 LASSERT(PageLocked(vmf->page));
 453                 result = VM_FAULT_LOCKED;
 454                 break;
 455         case -ENODATA:
 456         case -EFAULT:
 457                 result = VM_FAULT_NOPAGE;
 458                 break;
 459         case -ENOMEM:
 460                 result = VM_FAULT_OOM;
 461                 break;
 462         case -EAGAIN:
 463                 result = VM_FAULT_RETRY;
 464                 break;
 465         default:
 466                 result = VM_FAULT_SIGBUS;
 467                 break;
 468         }
 469
 470 out:
 471         if (result == VM_FAULT_LOCKED) {
 472                 ll_rw_stats_tally(ll_i2sbi(file_inode(vma->vm_file)),
 473                                   current->pid, vma->vm_file->private_data,
 474                                   cl_offset(NULL, vmf->page->index), PAGE_SIZE,
 475                                   WRITE);
 476                 ll_stats_ops_tally(ll_i2sbi(file_inode(vma->vm_file)),
 477                                    LPROC_LL_MKWRITE,
 478                                    ktime_us_delta(ktime_get(), kstart));
 479         }
 480
 481         return result;
 482 }
 483
 484 /**
 485  *  To avoid cancel the locks covering mmapped region for lock cache pressure,
 486  *  we track the mapped vma count in vvp_object::vob_mmap_cnt.
 487  */
 488 static void ll_vm_open(struct vm_area_struct * vma)
 489 {
 490         struct inode *inode    = file_inode(vma->vm_file);
 491         struct vvp_object *vob = cl_inode2vvp(inode);
 492
 493         ENTRY;
 494         LASSERT(atomic_read(&vob->vob_mmap_cnt) >= 0);
 495         atomic_inc(&vob->vob_mmap_cnt);
 496         pcc_vm_open(vma);
 497         EXIT;
 498 }
 499
 500 /**
 501  * Dual to ll_vm_open().
 502  */
 503 static void ll_vm_close(struct vm_area_struct *vma)
 504 {
 505         struct inode      *inode = file_inode(vma->vm_file);
 506         struct vvp_object *vob   = cl_inode2vvp(inode);
 507
 508         ENTRY;
 509         atomic_dec(&vob->vob_mmap_cnt);
 510         LASSERT(atomic_read(&vob->vob_mmap_cnt) >= 0);
 511         pcc_vm_close(vma);
 512         EXIT;
 513 }
 514
 515 /* XXX put nice comment here.  talk about __free_pte -> dirty pages and
 516  * nopage's reference passing to the pte */
 517 int ll_teardown_mmaps(struct address_space *mapping, __u64 first, __u64 last)
 518 {
 519         int rc = -ENOENT;
 520         ENTRY;
 521
 522         LASSERTF(last > first, "last %llu first %llu\n", last, first);
 523         if (mapping_mapped(mapping)) {
 524                 rc = 0;
 525                 unmap_mapping_range(mapping, first + PAGE_SIZE - 1,
 526                                     last - first + 1, 0);
 527         }
 528
 529         RETURN(rc);
 530 }
 531
 532 static const struct vm_operations_struct ll_file_vm_ops = {
 533         .fault                  = ll_fault,
 534         .page_mkwrite           = ll_page_mkwrite,
 535         .open                   = ll_vm_open,
 536         .close                  = ll_vm_close,
 537 };
 538
 539 int ll_file_mmap(struct file *file, struct vm_area_struct * vma)
 540 {
 541         struct inode *inode = file_inode(file);
 542         ktime_t kstart = ktime_get();
 543         bool cached;
 544         int rc;
 545
 546         ENTRY;
 547         CDEBUG(D_VFSTRACE | D_MMAP,
 548                "VFS_Op: fid="DFID" vma=%p start=%#lx end=%#lx vm_flags=%#lx\n",
 549                PFID(&ll_i2info(inode)->lli_fid),
 550                vma, vma->vm_start, vma->vm_end, vma->vm_flags);
 551
 552         if (ll_file_nolock(file))
 553                 RETURN(-EOPNOTSUPP);
 554
 555         rc = pcc_file_mmap(file, vma, &cached);
 556         if (cached && rc != 0)
 557                 RETURN(rc);
 558
 559         rc = generic_file_mmap(file, vma);
 560         if (rc == 0) {
 561                 vma->vm_ops = &ll_file_vm_ops;
 562                 vma->vm_ops->open(vma);
 563                 /* update the inode's size and mtime */
 564                 if (!cached)
 565                         rc = ll_glimpse_size(inode);
 566         }
 567
 568         if (!rc)
 569                 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_MMAP,
 570                                    ktime_us_delta(ktime_get(), kstart));
 571
 572         RETURN(rc);
 573 }