hw/hfi1/mmu_rb.c

0001 // SPDX-License-Identifier: GPL-2.0 or BSD-3-Clause
0002 /*
0003  * Copyright(c) 2020 Cornelis Networks, Inc.
0004  * Copyright(c) 2016 - 2017 Intel Corporation.
0005  */
0006
0007 #include <linux/list.h>
0008 #include <linux/rculist.h>
0009 #include <linux/mmu_notifier.h>
0010 #include <linux/interval_tree_generic.h>
0011 #include <linux/sched/mm.h>
0012
0013 #include "mmu_rb.h"
0014 #include "trace.h"
0015
0016 static unsigned long mmu_node_start(struct mmu_rb_node *);
0017 static unsigned long mmu_node_last(struct mmu_rb_node *);
0018 static int mmu_notifier_range_start(struct mmu_notifier *,
0019         const struct mmu_notifier_range *);
0020 static struct mmu_rb_node *__mmu_rb_search(struct mmu_rb_handler *,
0021                        unsigned long, unsigned long);
0022 static void do_remove(struct mmu_rb_handler *handler,
0023               struct list_head *del_list);
0024 static void handle_remove(struct work_struct *work);
0025
0026 static const struct mmu_notifier_ops mn_opts = {
0027     .invalidate_range_start = mmu_notifier_range_start,
0028 };
0029
0030 INTERVAL_TREE_DEFINE(struct mmu_rb_node, node, unsigned long, __last,
0031              mmu_node_start, mmu_node_last, static, __mmu_int_rb);
0032
0033 static unsigned long mmu_node_start(struct mmu_rb_node *node)
0034 {
0035     return node->addr & PAGE_MASK;
0036 }
0037
0038 static unsigned long mmu_node_last(struct mmu_rb_node *node)
0039 {
0040     return PAGE_ALIGN(node->addr + node->len) - 1;
0041 }
0042
0043 int hfi1_mmu_rb_register(void *ops_arg,
0044              struct mmu_rb_ops *ops,
0045              struct workqueue_struct *wq,
0046              struct mmu_rb_handler **handler)
0047 {
0048     struct mmu_rb_handler *h;
0049     int ret;
0050
0051     h = kzalloc(sizeof(*h), GFP_KERNEL);
0052     if (!h)
0053         return -ENOMEM;
0054
0055     h->root = RB_ROOT_CACHED;
0056     h->ops = ops;
0057     h->ops_arg = ops_arg;
0058     INIT_HLIST_NODE(&h->mn.hlist);
0059     spin_lock_init(&h->lock);
0060     h->mn.ops = &mn_opts;
0061     INIT_WORK(&h->del_work, handle_remove);
0062     INIT_LIST_HEAD(&h->del_list);
0063     INIT_LIST_HEAD(&h->lru_list);
0064     h->wq = wq;
0065
0066     ret = mmu_notifier_register(&h->mn, current->mm);
0067     if (ret) {
0068         kfree(h);
0069         return ret;
0070     }
0071
0072     *handler = h;
0073     return 0;
0074 }
0075
0076 void hfi1_mmu_rb_unregister(struct mmu_rb_handler *handler)
0077 {
0078     struct mmu_rb_node *rbnode;
0079     struct rb_node *node;
0080     unsigned long flags;
0081     struct list_head del_list;
0082
0083     /* Prevent freeing of mm until we are completely finished. */
0084     mmgrab(handler->mn.mm);
0085
0086     /* Unregister first so we don't get any more notifications. */
0087     mmu_notifier_unregister(&handler->mn, handler->mn.mm);
0088
0089     /*
0090      * Make sure the wq delete handler is finished running.  It will not
0091      * be triggered once the mmu notifiers are unregistered above.
0092      */
0093     flush_work(&handler->del_work);
0094
0095     INIT_LIST_HEAD(&del_list);
0096
0097     spin_lock_irqsave(&handler->lock, flags);
0098     while ((node = rb_first_cached(&handler->root))) {
0099         rbnode = rb_entry(node, struct mmu_rb_node, node);
0100         rb_erase_cached(node, &handler->root);
0101         /* move from LRU list to delete list */
0102         list_move(&rbnode->list, &del_list);
0103     }
0104     spin_unlock_irqrestore(&handler->lock, flags);
0105
0106     do_remove(handler, &del_list);
0107
0108     /* Now the mm may be freed. */
0109     mmdrop(handler->mn.mm);
0110
0111     kfree(handler);
0112 }
0113
0114 int hfi1_mmu_rb_insert(struct mmu_rb_handler *handler,
0115                struct mmu_rb_node *mnode)
0116 {
0117     struct mmu_rb_node *node;
0118     unsigned long flags;
0119     int ret = 0;
0120
0121     trace_hfi1_mmu_rb_insert(mnode->addr, mnode->len);
0122
0123     if (current->mm != handler->mn.mm)
0124         return -EPERM;
0125
0126     spin_lock_irqsave(&handler->lock, flags);
0127     node = __mmu_rb_search(handler, mnode->addr, mnode->len);
0128     if (node) {
0129         ret = -EINVAL;
0130         goto unlock;
0131     }
0132     __mmu_int_rb_insert(mnode, &handler->root);
0133     list_add(&mnode->list, &handler->lru_list);
0134
0135     ret = handler->ops->insert(handler->ops_arg, mnode);
0136     if (ret) {
0137         __mmu_int_rb_remove(mnode, &handler->root);
0138         list_del(&mnode->list); /* remove from LRU list */
0139     }
0140     mnode->handler = handler;
0141 unlock:
0142     spin_unlock_irqrestore(&handler->lock, flags);
0143     return ret;
0144 }
0145
0146 /* Caller must hold handler lock */
0147 static struct mmu_rb_node *__mmu_rb_search(struct mmu_rb_handler *handler,
0148                        unsigned long addr,
0149                        unsigned long len)
0150 {
0151     struct mmu_rb_node *node = NULL;
0152
0153     trace_hfi1_mmu_rb_search(addr, len);
0154     if (!handler->ops->filter) {
0155         node = __mmu_int_rb_iter_first(&handler->root, addr,
0156                            (addr + len) - 1);
0157     } else {
0158         for (node = __mmu_int_rb_iter_first(&handler->root, addr,
0159                             (addr + len) - 1);
0160              node;
0161              node = __mmu_int_rb_iter_next(node, addr,
0162                            (addr + len) - 1)) {
0163             if (handler->ops->filter(node, addr, len))
0164                 return node;
0165         }
0166     }
0167     return node;
0168 }
0169
0170 bool hfi1_mmu_rb_remove_unless_exact(struct mmu_rb_handler *handler,
0171                      unsigned long addr, unsigned long len,
0172                      struct mmu_rb_node **rb_node)
0173 {
0174     struct mmu_rb_node *node;
0175     unsigned long flags;
0176     bool ret = false;
0177
0178     if (current->mm != handler->mn.mm)
0179         return ret;
0180
0181     spin_lock_irqsave(&handler->lock, flags);
0182     node = __mmu_rb_search(handler, addr, len);
0183     if (node) {
0184         if (node->addr == addr && node->len == len)
0185             goto unlock;
0186         __mmu_int_rb_remove(node, &handler->root);
0187         list_del(&node->list); /* remove from LRU list */
0188         ret = true;
0189     }
0190 unlock:
0191     spin_unlock_irqrestore(&handler->lock, flags);
0192     *rb_node = node;
0193     return ret;
0194 }
0195
0196 void hfi1_mmu_rb_evict(struct mmu_rb_handler *handler, void *evict_arg)
0197 {
0198     struct mmu_rb_node *rbnode, *ptr;
0199     struct list_head del_list;
0200     unsigned long flags;
0201     bool stop = false;
0202
0203     if (current->mm != handler->mn.mm)
0204         return;
0205
0206     INIT_LIST_HEAD(&del_list);
0207
0208     spin_lock_irqsave(&handler->lock, flags);
0209     list_for_each_entry_safe_reverse(rbnode, ptr, &handler->lru_list,
0210                      list) {
0211         if (handler->ops->evict(handler->ops_arg, rbnode, evict_arg,
0212                     &stop)) {
0213             __mmu_int_rb_remove(rbnode, &handler->root);
0214             /* move from LRU list to delete list */
0215             list_move(&rbnode->list, &del_list);
0216         }
0217         if (stop)
0218             break;
0219     }
0220     spin_unlock_irqrestore(&handler->lock, flags);
0221
0222     while (!list_empty(&del_list)) {
0223         rbnode = list_first_entry(&del_list, struct mmu_rb_node, list);
0224         list_del(&rbnode->list);
0225         handler->ops->remove(handler->ops_arg, rbnode);
0226     }
0227 }
0228
0229 /*
0230  * It is up to the caller to ensure that this function does not race with the
0231  * mmu invalidate notifier which may be calling the users remove callback on
0232  * 'node'.
0233  */
0234 void hfi1_mmu_rb_remove(struct mmu_rb_handler *handler,
0235             struct mmu_rb_node *node)
0236 {
0237     unsigned long flags;
0238
0239     if (current->mm != handler->mn.mm)
0240         return;
0241
0242     /* Validity of handler and node pointers has been checked by caller. */
0243     trace_hfi1_mmu_rb_remove(node->addr, node->len);
0244     spin_lock_irqsave(&handler->lock, flags);
0245     __mmu_int_rb_remove(node, &handler->root);
0246     list_del(&node->list); /* remove from LRU list */
0247     spin_unlock_irqrestore(&handler->lock, flags);
0248
0249     handler->ops->remove(handler->ops_arg, node);
0250 }
0251
0252 static int mmu_notifier_range_start(struct mmu_notifier *mn,
0253         const struct mmu_notifier_range *range)
0254 {
0255     struct mmu_rb_handler *handler =
0256         container_of(mn, struct mmu_rb_handler, mn);
0257     struct rb_root_cached *root = &handler->root;
0258     struct mmu_rb_node *node, *ptr = NULL;
0259     unsigned long flags;
0260     bool added = false;
0261
0262     spin_lock_irqsave(&handler->lock, flags);
0263     for (node = __mmu_int_rb_iter_first(root, range->start, range->end-1);
0264          node; node = ptr) {
0265         /* Guard against node removal. */
0266         ptr = __mmu_int_rb_iter_next(node, range->start,
0267                          range->end - 1);
0268         trace_hfi1_mmu_mem_invalidate(node->addr, node->len);
0269         if (handler->ops->invalidate(handler->ops_arg, node)) {
0270             __mmu_int_rb_remove(node, root);
0271             /* move from LRU list to delete list */
0272             list_move(&node->list, &handler->del_list);
0273             added = true;
0274         }
0275     }
0276     spin_unlock_irqrestore(&handler->lock, flags);
0277
0278     if (added)
0279         queue_work(handler->wq, &handler->del_work);
0280
0281     return 0;
0282 }
0283
0284 /*
0285  * Call the remove function for the given handler and the list.  This
0286  * is expected to be called with a delete list extracted from handler.
0287  * The caller should not be holding the handler lock.
0288  */
0289 static void do_remove(struct mmu_rb_handler *handler,
0290               struct list_head *del_list)
0291 {
0292     struct mmu_rb_node *node;
0293
0294     while (!list_empty(del_list)) {
0295         node = list_first_entry(del_list, struct mmu_rb_node, list);
0296         list_del(&node->list);
0297         handler->ops->remove(handler->ops_arg, node);
0298     }
0299 }
0300
0301 /*
0302  * Work queue function to remove all nodes that have been queued up to
0303  * be removed.  The key feature is that mm->mmap_lock is not being held
0304  * and the remove callback can sleep while taking it, if needed.
0305  */
0306 static void handle_remove(struct work_struct *work)
0307 {
0308     struct mmu_rb_handler *handler = container_of(work,
0309                         struct mmu_rb_handler,
0310                         del_work);
0311     struct list_head del_list;
0312     unsigned long flags;
0313
0314     /* remove anything that is queued to get removed */
0315     spin_lock_irqsave(&handler->lock, flags);
0316     list_replace_init(&handler->del_list, &del_list);
0317     spin_unlock_irqrestore(&handler->lock, flags);
0318
0319     do_remove(handler, &del_list);
0320 }