The Design and Implementation of the FreeBSD Operating System, Second Edition
Now available: The Design and Implementation of the FreeBSD Operating System (Second Edition)


[ source navigation ] [ diff markup ] [ identifier search ] [ freetext search ] [ file search ] [ list types ] [ track identifier ]

FreeBSD/Linux Kernel Cross Reference
sys/ofed/drivers/infiniband/core/ib_umem_odp.c

Version: -  FREEBSD  -  FREEBSD-13-STABLE  -  FREEBSD-13-0  -  FREEBSD-12-STABLE  -  FREEBSD-12-0  -  FREEBSD-11-STABLE  -  FREEBSD-11-0  -  FREEBSD-10-STABLE  -  FREEBSD-10-0  -  FREEBSD-9-STABLE  -  FREEBSD-9-0  -  FREEBSD-8-STABLE  -  FREEBSD-8-0  -  FREEBSD-7-STABLE  -  FREEBSD-7-0  -  FREEBSD-6-STABLE  -  FREEBSD-6-0  -  FREEBSD-5-STABLE  -  FREEBSD-5-0  -  FREEBSD-4-STABLE  -  FREEBSD-3-STABLE  -  FREEBSD22  -  l41  -  OPENBSD  -  linux-2.6  -  MK84  -  PLAN9  -  xnu-8792 
SearchContext: -  none  -  3  -  10 

    1 /*-
    2  * SPDX-License-Identifier: BSD-2-Clause OR GPL-2.0
    3  *
    4  * Copyright (c) 2014 Mellanox Technologies. All rights reserved.
    5  *
    6  * This software is available to you under a choice of one of two
    7  * licenses.  You may choose to be licensed under the terms of the GNU
    8  * General Public License (GPL) Version 2, available from the file
    9  * COPYING in the main directory of this source tree, or the
   10  * OpenIB.org BSD license below:
   11  *
   12  *     Redistribution and use in source and binary forms, with or
   13  *     without modification, are permitted provided that the following
   14  *     conditions are met:
   15  *
   16  *      - Redistributions of source code must retain the above
   17  *        copyright notice, this list of conditions and the following
   18  *        disclaimer.
   19  *
   20  *      - Redistributions in binary form must reproduce the above
   21  *        copyright notice, this list of conditions and the following
   22  *        disclaimer in the documentation and/or other materials
   23  *        provided with the distribution.
   24  *
   25  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
   26  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
   27  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
   28  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
   29  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
   30  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
   31  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
   32  * SOFTWARE.
   33  */
   34 
   35 #include <sys/cdefs.h>
   36 __FBSDID("$FreeBSD$");
   37 
   38 #include <linux/types.h>
   39 #include <linux/sched.h>
   40 #include <linux/slab.h>
   41 #include <linux/vmalloc.h>
   42 
   43 #include <rdma/ib_verbs.h>
   44 #include <rdma/ib_umem.h>
   45 #include <rdma/ib_umem_odp.h>
   46 
   47 static void ib_umem_notifier_start_account(struct ib_umem *item)
   48 {
   49         mutex_lock(&item->odp_data->umem_mutex);
   50 
   51         /* Only update private counters for this umem if it has them.
   52          * Otherwise skip it. All page faults will be delayed for this umem. */
   53         if (item->odp_data->mn_counters_active) {
   54                 int notifiers_count = item->odp_data->notifiers_count++;
   55 
   56                 if (notifiers_count == 0)
   57                         /* Initialize the completion object for waiting on
   58                          * notifiers. Since notifier_count is zero, no one
   59                          * should be waiting right now. */
   60                         reinit_completion(&item->odp_data->notifier_completion);
   61         }
   62         mutex_unlock(&item->odp_data->umem_mutex);
   63 }
   64 
   65 static void ib_umem_notifier_end_account(struct ib_umem *item)
   66 {
   67         mutex_lock(&item->odp_data->umem_mutex);
   68 
   69         /* Only update private counters for this umem if it has them.
   70          * Otherwise skip it. All page faults will be delayed for this umem. */
   71         if (item->odp_data->mn_counters_active) {
   72                 /*
   73                  * This sequence increase will notify the QP page fault that
   74                  * the page that is going to be mapped in the spte could have
   75                  * been freed.
   76                  */
   77                 ++item->odp_data->notifiers_seq;
   78                 if (--item->odp_data->notifiers_count == 0)
   79                         complete_all(&item->odp_data->notifier_completion);
   80         }
   81         mutex_unlock(&item->odp_data->umem_mutex);
   82 }
   83 
   84 /* Account for a new mmu notifier in an ib_ucontext. */
   85 static void ib_ucontext_notifier_start_account(struct ib_ucontext *context)
   86 {
   87         atomic_inc(&context->notifier_count);
   88 }
   89 
   90 /* Account for a terminating mmu notifier in an ib_ucontext.
   91  *
   92  * Must be called with the ib_ucontext->umem_rwsem semaphore unlocked, since
   93  * the function takes the semaphore itself. */
   94 static void ib_ucontext_notifier_end_account(struct ib_ucontext *context)
   95 {
   96         int zero_notifiers = atomic_dec_and_test(&context->notifier_count);
   97 
   98         if (zero_notifiers &&
   99             !list_empty(&context->no_private_counters)) {
  100                 /* No currently running mmu notifiers. Now is the chance to
  101                  * add private accounting to all previously added umems. */
  102                 struct ib_umem_odp *odp_data, *next;
  103 
  104                 /* Prevent concurrent mmu notifiers from working on the
  105                  * no_private_counters list. */
  106                 down_write(&context->umem_rwsem);
  107 
  108                 /* Read the notifier_count again, with the umem_rwsem
  109                  * semaphore taken for write. */
  110                 if (!atomic_read(&context->notifier_count)) {
  111                         list_for_each_entry_safe(odp_data, next,
  112                                                  &context->no_private_counters,
  113                                                  no_private_counters) {
  114                                 mutex_lock(&odp_data->umem_mutex);
  115                                 odp_data->mn_counters_active = true;
  116                                 list_del(&odp_data->no_private_counters);
  117                                 complete_all(&odp_data->notifier_completion);
  118                                 mutex_unlock(&odp_data->umem_mutex);
  119                         }
  120                 }
  121 
  122                 up_write(&context->umem_rwsem);
  123         }
  124 }
  125 
  126 static int ib_umem_notifier_release_trampoline(struct ib_umem *item, u64 start,
  127                                                u64 end, void *cookie) {
  128         /*
  129          * Increase the number of notifiers running, to
  130          * prevent any further fault handling on this MR.
  131          */
  132         ib_umem_notifier_start_account(item);
  133         item->odp_data->dying = 1;
  134         /* Make sure that the fact the umem is dying is out before we release
  135          * all pending page faults. */
  136         smp_wmb();
  137         complete_all(&item->odp_data->notifier_completion);
  138         item->context->invalidate_range(item, ib_umem_start(item),
  139                                         ib_umem_end(item));
  140         return 0;
  141 }
  142 
  143 static void ib_umem_notifier_release(struct mmu_notifier *mn,
  144                                      struct mm_struct *mm)
  145 {
  146         struct ib_ucontext *context = container_of(mn, struct ib_ucontext, mn);
  147 
  148         if (!context->invalidate_range)
  149                 return;
  150 
  151         ib_ucontext_notifier_start_account(context);
  152         down_read(&context->umem_rwsem);
  153         rbt_ib_umem_for_each_in_range(&context->umem_tree, 0,
  154                                       ULLONG_MAX,
  155                                       ib_umem_notifier_release_trampoline,
  156                                       NULL);
  157         up_read(&context->umem_rwsem);
  158 }
  159 
  160 static int invalidate_page_trampoline(struct ib_umem *item, u64 start,
  161                                       u64 end, void *cookie)
  162 {
  163         ib_umem_notifier_start_account(item);
  164         item->context->invalidate_range(item, start, start + PAGE_SIZE);
  165         ib_umem_notifier_end_account(item);
  166         return 0;
  167 }
  168 
  169 static void ib_umem_notifier_invalidate_page(struct mmu_notifier *mn,
  170                                              struct mm_struct *mm,
  171                                              unsigned long address)
  172 {
  173         struct ib_ucontext *context = container_of(mn, struct ib_ucontext, mn);
  174 
  175         if (!context->invalidate_range)
  176                 return;
  177 
  178         ib_ucontext_notifier_start_account(context);
  179         down_read(&context->umem_rwsem);
  180         rbt_ib_umem_for_each_in_range(&context->umem_tree, address,
  181                                       address + PAGE_SIZE,
  182                                       invalidate_page_trampoline, NULL);
  183         up_read(&context->umem_rwsem);
  184         ib_ucontext_notifier_end_account(context);
  185 }
  186 
  187 static int invalidate_range_start_trampoline(struct ib_umem *item, u64 start,
  188                                              u64 end, void *cookie)
  189 {
  190         ib_umem_notifier_start_account(item);
  191         item->context->invalidate_range(item, start, end);
  192         return 0;
  193 }
  194 
  195 static void ib_umem_notifier_invalidate_range_start(struct mmu_notifier *mn,
  196                                                     struct mm_struct *mm,
  197                                                     unsigned long start,
  198                                                     unsigned long end)
  199 {
  200         struct ib_ucontext *context = container_of(mn, struct ib_ucontext, mn);
  201 
  202         if (!context->invalidate_range)
  203                 return;
  204 
  205         ib_ucontext_notifier_start_account(context);
  206         down_read(&context->umem_rwsem);
  207         rbt_ib_umem_for_each_in_range(&context->umem_tree, start,
  208                                       end,
  209                                       invalidate_range_start_trampoline, NULL);
  210         up_read(&context->umem_rwsem);
  211 }
  212 
  213 static int invalidate_range_end_trampoline(struct ib_umem *item, u64 start,
  214                                            u64 end, void *cookie)
  215 {
  216         ib_umem_notifier_end_account(item);
  217         return 0;
  218 }
  219 
  220 static void ib_umem_notifier_invalidate_range_end(struct mmu_notifier *mn,
  221                                                   struct mm_struct *mm,
  222                                                   unsigned long start,
  223                                                   unsigned long end)
  224 {
  225         struct ib_ucontext *context = container_of(mn, struct ib_ucontext, mn);
  226 
  227         if (!context->invalidate_range)
  228                 return;
  229 
  230         down_read(&context->umem_rwsem);
  231         rbt_ib_umem_for_each_in_range(&context->umem_tree, start,
  232                                       end,
  233                                       invalidate_range_end_trampoline, NULL);
  234         up_read(&context->umem_rwsem);
  235         ib_ucontext_notifier_end_account(context);
  236 }
  237 
  238 static const struct mmu_notifier_ops ib_umem_notifiers = {
  239         .release                    = ib_umem_notifier_release,
  240         .invalidate_page            = ib_umem_notifier_invalidate_page,
  241         .invalidate_range_start     = ib_umem_notifier_invalidate_range_start,
  242         .invalidate_range_end       = ib_umem_notifier_invalidate_range_end,
  243 };
  244 
  245 int ib_umem_odp_get(struct ib_ucontext *context, struct ib_umem *umem)
  246 {
  247         int ret_val;
  248         pid_t our_pid;
  249         struct mm_struct *mm = get_task_mm(current);
  250 
  251         if (!mm)
  252                 return -EINVAL;
  253 
  254         /* Prevent creating ODP MRs in child processes */
  255         rcu_read_lock();
  256         our_pid = get_pid(task_pid_group_leader(current));
  257         rcu_read_unlock();
  258         put_pid(our_pid);
  259         if (context->tgid != our_pid) {
  260                 ret_val = -EINVAL;
  261                 goto out_mm;
  262         }
  263 
  264         umem->odp_data = kzalloc(sizeof(*umem->odp_data), GFP_KERNEL);
  265         if (!umem->odp_data) {
  266                 ret_val = -ENOMEM;
  267                 goto out_mm;
  268         }
  269         umem->odp_data->umem = umem;
  270 
  271         mutex_init(&umem->odp_data->umem_mutex);
  272 
  273         init_completion(&umem->odp_data->notifier_completion);
  274 
  275         umem->odp_data->page_list = vzalloc(ib_umem_num_pages(umem) *
  276                                             sizeof(*umem->odp_data->page_list));
  277         if (!umem->odp_data->page_list) {
  278                 ret_val = -ENOMEM;
  279                 goto out_odp_data;
  280         }
  281 
  282         umem->odp_data->dma_list = vzalloc(ib_umem_num_pages(umem) *
  283                                           sizeof(*umem->odp_data->dma_list));
  284         if (!umem->odp_data->dma_list) {
  285                 ret_val = -ENOMEM;
  286                 goto out_page_list;
  287         }
  288 
  289         /*
  290          * When using MMU notifiers, we will get a
  291          * notification before the "current" task (and MM) is
  292          * destroyed. We use the umem_rwsem semaphore to synchronize.
  293          */
  294         down_write(&context->umem_rwsem);
  295         context->odp_mrs_count++;
  296         if (likely(ib_umem_start(umem) != ib_umem_end(umem)))
  297                 rbt_ib_umem_insert(&umem->odp_data->interval_tree,
  298                                    &context->umem_tree);
  299         if (likely(!atomic_read(&context->notifier_count)) ||
  300             context->odp_mrs_count == 1)
  301                 umem->odp_data->mn_counters_active = true;
  302         else
  303                 list_add(&umem->odp_data->no_private_counters,
  304                          &context->no_private_counters);
  305         downgrade_write(&context->umem_rwsem);
  306 
  307         if (context->odp_mrs_count == 1) {
  308                 /*
  309                  * Note that at this point, no MMU notifier is running
  310                  * for this context!
  311                  */
  312                 atomic_set(&context->notifier_count, 0);
  313                 INIT_HLIST_NODE(&context->mn.hlist);
  314                 context->mn.ops = &ib_umem_notifiers;
  315                 /*
  316                  * Lock-dep detects a false positive for mmap_sem vs.
  317                  * umem_rwsem, due to not grasping downgrade_write correctly.
  318                  */
  319                 ret_val = mmu_notifier_register(&context->mn, mm);
  320                 if (ret_val) {
  321                         pr_err("Failed to register mmu_notifier %d\n", ret_val);
  322                         ret_val = -EBUSY;
  323                         goto out_mutex;
  324                 }
  325         }
  326 
  327         up_read(&context->umem_rwsem);
  328 
  329         /*
  330          * Note that doing an mmput can cause a notifier for the relevant mm.
  331          * If the notifier is called while we hold the umem_rwsem, this will
  332          * cause a deadlock. Therefore, we release the reference only after we
  333          * released the semaphore.
  334          */
  335         mmput(mm);
  336         return 0;
  337 
  338 out_mutex:
  339         up_read(&context->umem_rwsem);
  340         vfree(umem->odp_data->dma_list);
  341 out_page_list:
  342         vfree(umem->odp_data->page_list);
  343 out_odp_data:
  344         kfree(umem->odp_data);
  345 out_mm:
  346         mmput(mm);
  347         return ret_val;
  348 }
  349 
  350 void ib_umem_odp_release(struct ib_umem *umem)
  351 {
  352         struct ib_ucontext *context = umem->context;
  353 
  354         /*
  355          * Ensure that no more pages are mapped in the umem.
  356          *
  357          * It is the driver's responsibility to ensure, before calling us,
  358          * that the hardware will not attempt to access the MR any more.
  359          */
  360         ib_umem_odp_unmap_dma_pages(umem, ib_umem_start(umem),
  361                                     ib_umem_end(umem));
  362 
  363         down_write(&context->umem_rwsem);
  364         if (likely(ib_umem_start(umem) != ib_umem_end(umem)))
  365                 rbt_ib_umem_remove(&umem->odp_data->interval_tree,
  366                                    &context->umem_tree);
  367         context->odp_mrs_count--;
  368         if (!umem->odp_data->mn_counters_active) {
  369                 list_del(&umem->odp_data->no_private_counters);
  370                 complete_all(&umem->odp_data->notifier_completion);
  371         }
  372 
  373         /*
  374          * Downgrade the lock to a read lock. This ensures that the notifiers
  375          * (who lock the mutex for reading) will be able to finish, and we
  376          * will be able to enventually obtain the mmu notifiers SRCU. Note
  377          * that since we are doing it atomically, no other user could register
  378          * and unregister while we do the check.
  379          */
  380         downgrade_write(&context->umem_rwsem);
  381         if (!context->odp_mrs_count) {
  382                 struct task_struct *owning_process = NULL;
  383                 struct mm_struct *owning_mm        = NULL;
  384 
  385                 owning_process = get_pid_task(context->tgid,
  386                                               PIDTYPE_PID);
  387                 if (owning_process == NULL)
  388                         /*
  389                          * The process is already dead, notifier were removed
  390                          * already.
  391                          */
  392                         goto out;
  393 
  394                 owning_mm = get_task_mm(owning_process);
  395                 if (owning_mm == NULL)
  396                         /*
  397                          * The process' mm is already dead, notifier were
  398                          * removed already.
  399                          */
  400                         goto out_put_task;
  401                 mmu_notifier_unregister(&context->mn, owning_mm);
  402 
  403                 mmput(owning_mm);
  404 
  405 out_put_task:
  406                 put_task_struct(owning_process);
  407         }
  408 out:
  409         up_read(&context->umem_rwsem);
  410 
  411         vfree(umem->odp_data->dma_list);
  412         vfree(umem->odp_data->page_list);
  413         kfree(umem->odp_data);
  414         kfree(umem);
  415 }
  416 
  417 /*
  418  * Map for DMA and insert a single page into the on-demand paging page tables.
  419  *
  420  * @umem: the umem to insert the page to.
  421  * @page_index: index in the umem to add the page to.
  422  * @page: the page struct to map and add.
  423  * @access_mask: access permissions needed for this page.
  424  * @current_seq: sequence number for synchronization with invalidations.
  425  *               the sequence number is taken from
  426  *               umem->odp_data->notifiers_seq.
  427  *
  428  * The function returns -EFAULT if the DMA mapping operation fails. It returns
  429  * -EAGAIN if a concurrent invalidation prevents us from updating the page.
  430  *
  431  * The page is released via put_page even if the operation failed. For
  432  * on-demand pinning, the page is released whenever it isn't stored in the
  433  * umem.
  434  */
  435 static int ib_umem_odp_map_dma_single_page(
  436                 struct ib_umem *umem,
  437                 int page_index,
  438                 u64 base_virt_addr,
  439                 struct page *page,
  440                 u64 access_mask,
  441                 unsigned long current_seq)
  442 {
  443         struct ib_device *dev = umem->context->device;
  444         dma_addr_t dma_addr;
  445         int stored_page = 0;
  446         int remove_existing_mapping = 0;
  447         int ret = 0;
  448 
  449         /*
  450          * Note: we avoid writing if seq is different from the initial seq, to
  451          * handle case of a racing notifier. This check also allows us to bail
  452          * early if we have a notifier running in parallel with us.
  453          */
  454         if (ib_umem_mmu_notifier_retry(umem, current_seq)) {
  455                 ret = -EAGAIN;
  456                 goto out;
  457         }
  458         if (!(umem->odp_data->dma_list[page_index])) {
  459                 dma_addr = ib_dma_map_page(dev,
  460                                            page,
  461                                            0, PAGE_SIZE,
  462                                            DMA_BIDIRECTIONAL);
  463                 if (ib_dma_mapping_error(dev, dma_addr)) {
  464                         ret = -EFAULT;
  465                         goto out;
  466                 }
  467                 umem->odp_data->dma_list[page_index] = dma_addr | access_mask;
  468                 umem->odp_data->page_list[page_index] = page;
  469                 stored_page = 1;
  470         } else if (umem->odp_data->page_list[page_index] == page) {
  471                 umem->odp_data->dma_list[page_index] |= access_mask;
  472         } else {
  473                 pr_err("error: got different pages in IB device and from get_user_pages. IB device page: %p, gup page: %p\n",
  474                        umem->odp_data->page_list[page_index], page);
  475                 /* Better remove the mapping now, to prevent any further
  476                  * damage. */
  477                 remove_existing_mapping = 1;
  478         }
  479 
  480 out:
  481         /* On Demand Paging - avoid pinning the page */
  482         if (umem->context->invalidate_range || !stored_page)
  483                 put_page(page);
  484 
  485         if (remove_existing_mapping && umem->context->invalidate_range) {
  486                 invalidate_page_trampoline(
  487                         umem,
  488                         base_virt_addr + (page_index * PAGE_SIZE),
  489                         base_virt_addr + ((page_index+1)*PAGE_SIZE),
  490                         NULL);
  491                 ret = -EAGAIN;
  492         }
  493 
  494         return ret;
  495 }
  496 
  497 /**
  498  * ib_umem_odp_map_dma_pages - Pin and DMA map userspace memory in an ODP MR.
  499  *
  500  * Pins the range of pages passed in the argument, and maps them to
  501  * DMA addresses. The DMA addresses of the mapped pages is updated in
  502  * umem->odp_data->dma_list.
  503  *
  504  * Returns the number of pages mapped in success, negative error code
  505  * for failure.
  506  * An -EAGAIN error code is returned when a concurrent mmu notifier prevents
  507  * the function from completing its task.
  508  *
  509  * @umem: the umem to map and pin
  510  * @user_virt: the address from which we need to map.
  511  * @bcnt: the minimal number of bytes to pin and map. The mapping might be
  512  *        bigger due to alignment, and may also be smaller in case of an error
  513  *        pinning or mapping a page. The actual pages mapped is returned in
  514  *        the return value.
  515  * @access_mask: bit mask of the requested access permissions for the given
  516  *               range.
  517  * @current_seq: the MMU notifiers sequance value for synchronization with
  518  *               invalidations. the sequance number is read from
  519  *               umem->odp_data->notifiers_seq before calling this function
  520  */
  521 int ib_umem_odp_map_dma_pages(struct ib_umem *umem, u64 user_virt, u64 bcnt,
  522                               u64 access_mask, unsigned long current_seq)
  523 {
  524         struct task_struct *owning_process  = NULL;
  525         struct mm_struct   *owning_mm       = NULL;
  526         struct page       **local_page_list = NULL;
  527         u64 off;
  528         int j, k, ret = 0, start_idx, npages = 0;
  529         u64 base_virt_addr;
  530         unsigned int flags = 0;
  531 
  532         if (access_mask == 0)
  533                 return -EINVAL;
  534 
  535         if (user_virt < ib_umem_start(umem) ||
  536             user_virt + bcnt > ib_umem_end(umem))
  537                 return -EFAULT;
  538 
  539         local_page_list = (struct page **)__get_free_page(GFP_KERNEL);
  540         if (!local_page_list)
  541                 return -ENOMEM;
  542 
  543         off = user_virt & (~PAGE_MASK);
  544         user_virt = user_virt & PAGE_MASK;
  545         base_virt_addr = user_virt;
  546         bcnt += off; /* Charge for the first page offset as well. */
  547 
  548         owning_process = get_pid_task(umem->context->tgid, PIDTYPE_PID);
  549         if (owning_process == NULL) {
  550                 ret = -EINVAL;
  551                 goto out_no_task;
  552         }
  553 
  554         owning_mm = get_task_mm(owning_process);
  555         if (owning_mm == NULL) {
  556                 ret = -EINVAL;
  557                 goto out_put_task;
  558         }
  559 
  560         if (access_mask & ODP_WRITE_ALLOWED_BIT)
  561                 flags |= FOLL_WRITE;
  562 
  563         start_idx = (user_virt - ib_umem_start(umem)) >> PAGE_SHIFT;
  564         k = start_idx;
  565 
  566         while (bcnt > 0) {
  567                 const size_t gup_num_pages =
  568                         min_t(size_t, ALIGN(bcnt, PAGE_SIZE) / PAGE_SIZE,
  569                               PAGE_SIZE / sizeof(struct page *));
  570 
  571                 down_read(&owning_mm->mmap_sem);
  572                 /*
  573                  * Note: this might result in redundent page getting. We can
  574                  * avoid this by checking dma_list to be 0 before calling
  575                  * get_user_pages. However, this make the code much more
  576                  * complex (and doesn't gain us much performance in most use
  577                  * cases).
  578                  */
  579                 npages = get_user_pages_remote(owning_process, owning_mm,
  580                                 user_virt, gup_num_pages,
  581                                 flags, local_page_list, NULL);
  582                 up_read(&owning_mm->mmap_sem);
  583 
  584                 if (npages < 0)
  585                         break;
  586 
  587                 bcnt -= min_t(size_t, npages << PAGE_SHIFT, bcnt);
  588                 user_virt += npages << PAGE_SHIFT;
  589                 mutex_lock(&umem->odp_data->umem_mutex);
  590                 for (j = 0; j < npages; ++j) {
  591                         ret = ib_umem_odp_map_dma_single_page(
  592                                 umem, k, base_virt_addr, local_page_list[j],
  593                                 access_mask, current_seq);
  594                         if (ret < 0)
  595                                 break;
  596                         k++;
  597                 }
  598                 mutex_unlock(&umem->odp_data->umem_mutex);
  599 
  600                 if (ret < 0) {
  601                         /* Release left over pages when handling errors. */
  602                         for (++j; j < npages; ++j)
  603                                 put_page(local_page_list[j]);
  604                         break;
  605                 }
  606         }
  607 
  608         if (ret >= 0) {
  609                 if (npages < 0 && k == start_idx)
  610                         ret = npages;
  611                 else
  612                         ret = k - start_idx;
  613         }
  614 
  615         mmput(owning_mm);
  616 out_put_task:
  617         put_task_struct(owning_process);
  618 out_no_task:
  619         free_page((unsigned long)local_page_list);
  620         return ret;
  621 }
  622 EXPORT_SYMBOL(ib_umem_odp_map_dma_pages);
  623 
  624 void ib_umem_odp_unmap_dma_pages(struct ib_umem *umem, u64 virt,
  625                                  u64 bound)
  626 {
  627         int idx;
  628         u64 addr;
  629         struct ib_device *dev = umem->context->device;
  630 
  631         virt  = max_t(u64, virt,  ib_umem_start(umem));
  632         bound = min_t(u64, bound, ib_umem_end(umem));
  633         /* Note that during the run of this function, the
  634          * notifiers_count of the MR is > 0, preventing any racing
  635          * faults from completion. We might be racing with other
  636          * invalidations, so we must make sure we free each page only
  637          * once. */
  638         mutex_lock(&umem->odp_data->umem_mutex);
  639         for (addr = virt; addr < bound; addr += (u64)umem->page_size) {
  640                 idx = (addr - ib_umem_start(umem)) / PAGE_SIZE;
  641                 if (umem->odp_data->page_list[idx]) {
  642                         struct page *page = umem->odp_data->page_list[idx];
  643                         dma_addr_t dma = umem->odp_data->dma_list[idx];
  644                         dma_addr_t dma_addr = dma & ODP_DMA_ADDR_MASK;
  645 
  646                         WARN_ON(!dma_addr);
  647 
  648                         ib_dma_unmap_page(dev, dma_addr, PAGE_SIZE,
  649                                           DMA_BIDIRECTIONAL);
  650                         if (dma & ODP_WRITE_ALLOWED_BIT) {
  651                                 struct page *head_page = compound_head(page);
  652                                 /*
  653                                  * set_page_dirty prefers being called with
  654                                  * the page lock. However, MMU notifiers are
  655                                  * called sometimes with and sometimes without
  656                                  * the lock. We rely on the umem_mutex instead
  657                                  * to prevent other mmu notifiers from
  658                                  * continuing and allowing the page mapping to
  659                                  * be removed.
  660                                  */
  661                                 set_page_dirty(head_page);
  662                         }
  663                         /* on demand pinning support */
  664                         if (!umem->context->invalidate_range)
  665                                 put_page(page);
  666                         umem->odp_data->page_list[idx] = NULL;
  667                         umem->odp_data->dma_list[idx] = 0;
  668                 }
  669         }
  670         mutex_unlock(&umem->odp_data->umem_mutex);
  671 }
  672 EXPORT_SYMBOL(ib_umem_odp_unmap_dma_pages);

Cache object: 31cda2acc2a3b763878fee6417b23342


[ source navigation ] [ diff markup ] [ identifier search ] [ freetext search ] [ file search ] [ list types ] [ track identifier ]


This page is part of the FreeBSD/Linux Linux Kernel Cross-Reference, and was automatically generated using a modified version of the LXR engine.