1 /* 2 * linux/mm/mmu_notifier.c 3 * 4 * Copyright (C) 2008 Qumranet, Inc. 5 * Copyright (C) 2008 SGI 6 * Christoph Lameter <clameter@sgi.com> 7 * 8 * This work is licensed under the terms of the GNU GPL, version 2. See 9 * the COPYING file in the top-level directory. 10 */ 11 12 #include <linux/rculist.h> 13 #include <linux/mmu_notifier.h> 14 #include <linux/module.h> 15 #include <linux/mm.h> 16 #include <linux/err.h> 17 #include <linux/rcupdate.h> 18 #include <linux/sched.h> 19 20 /* 21 * This function can't run concurrently against mmu_notifier_register 22 * because mm->mm_users > 0 during mmu_notifier_register and exit_mmap 23 * runs with mm_users == 0. Other tasks may still invoke mmu notifiers 24 * in parallel despite there being no task using this mm any more, 25 * through the vmas outside of the exit_mmap context, such as with 26 * vmtruncate. This serializes against mmu_notifier_unregister with 27 * the mmu_notifier_mm->lock in addition to RCU and it serializes 28 * against the other mmu notifiers with RCU. struct mmu_notifier_mm 29 * can't go away from under us as exit_mmap holds an mm_count pin 30 * itself. 31 */ 32 void __mmu_notifier_release(struct mm_struct *mm) 33 { 34 struct mmu_notifier *mn; 35 36 spin_lock(&mm->mmu_notifier_mm->lock); 37 while (unlikely(!hlist_empty(&mm->mmu_notifier_mm->list))) { 38 mn = hlist_entry(mm->mmu_notifier_mm->list.first, 39 struct mmu_notifier, 40 hlist); 41 /* 42 * We arrived before mmu_notifier_unregister so 43 * mmu_notifier_unregister will do nothing other than 44 * to wait ->release to finish and 45 * mmu_notifier_unregister to return. 46 */ 47 hlist_del_init_rcu(&mn->hlist); 48 /* 49 * RCU here will block mmu_notifier_unregister until 50 * ->release returns. 51 */ 52 rcu_read_lock(); 53 spin_unlock(&mm->mmu_notifier_mm->lock); 54 /* 55 * if ->release runs before mmu_notifier_unregister it 56 * must be handled as it's the only way for the driver 57 * to flush all existing sptes and stop the driver 58 * from establishing any more sptes before all the 59 * pages in the mm are freed. 60 */ 61 if (mn->ops->release) 62 mn->ops->release(mn, mm); 63 rcu_read_unlock(); 64 spin_lock(&mm->mmu_notifier_mm->lock); 65 } 66 spin_unlock(&mm->mmu_notifier_mm->lock); 67 68 /* 69 * synchronize_rcu here prevents mmu_notifier_release to 70 * return to exit_mmap (which would proceed freeing all pages 71 * in the mm) until the ->release method returns, if it was 72 * invoked by mmu_notifier_unregister. 73 * 74 * The mmu_notifier_mm can't go away from under us because one 75 * mm_count is hold by exit_mmap. 76 */ 77 synchronize_rcu(); 78 } 79 80 /* 81 * If no young bitflag is supported by the hardware, ->clear_flush_young can 82 * unmap the address and return 1 or 0 depending if the mapping previously 83 * existed or not. 84 */ 85 int __mmu_notifier_clear_flush_young(struct mm_struct *mm, 86 unsigned long address) 87 { 88 struct mmu_notifier *mn; 89 struct hlist_node *n; 90 int young = 0; 91 92 rcu_read_lock(); 93 hlist_for_each_entry_rcu(mn, n, &mm->mmu_notifier_mm->list, hlist) { 94 if (mn->ops->clear_flush_young) 95 young |= mn->ops->clear_flush_young(mn, mm, address); 96 } 97 rcu_read_unlock(); 98 99 return young; 100 } 101 102 void __mmu_notifier_invalidate_page(struct mm_struct *mm, 103 unsigned long address) 104 { 105 struct mmu_notifier *mn; 106 struct hlist_node *n; 107 108 rcu_read_lock(); 109 hlist_for_each_entry_rcu(mn, n, &mm->mmu_notifier_mm->list, hlist) { 110 if (mn->ops->invalidate_page) 111 mn->ops->invalidate_page(mn, mm, address); 112 } 113 rcu_read_unlock(); 114 } 115 116 void __mmu_notifier_invalidate_range_start(struct mm_struct *mm, 117 unsigned long start, unsigned long end) 118 { 119 struct mmu_notifier *mn; 120 struct hlist_node *n; 121 122 rcu_read_lock(); 123 hlist_for_each_entry_rcu(mn, n, &mm->mmu_notifier_mm->list, hlist) { 124 if (mn->ops->invalidate_range_start) 125 mn->ops->invalidate_range_start(mn, mm, start, end); 126 } 127 rcu_read_unlock(); 128 } 129 130 void __mmu_notifier_invalidate_range_end(struct mm_struct *mm, 131 unsigned long start, unsigned long end) 132 { 133 struct mmu_notifier *mn; 134 struct hlist_node *n; 135 136 rcu_read_lock(); 137 hlist_for_each_entry_rcu(mn, n, &mm->mmu_notifier_mm->list, hlist) { 138 if (mn->ops->invalidate_range_end) 139 mn->ops->invalidate_range_end(mn, mm, start, end); 140 } 141 rcu_read_unlock(); 142 } 143 144 static int do_mmu_notifier_register(struct mmu_notifier *mn, 145 struct mm_struct *mm, 146 int take_mmap_sem) 147 { 148 struct mmu_notifier_mm *mmu_notifier_mm; 149 int ret; 150 151 BUG_ON(atomic_read(&mm->mm_users) <= 0); 152 153 ret = -ENOMEM; 154 mmu_notifier_mm = kmalloc(sizeof(struct mmu_notifier_mm), GFP_KERNEL); 155 if (unlikely(!mmu_notifier_mm)) 156 goto out; 157 158 if (take_mmap_sem) 159 down_write(&mm->mmap_sem); 160 ret = mm_take_all_locks(mm); 161 if (unlikely(ret)) 162 goto out_cleanup; 163 164 if (!mm_has_notifiers(mm)) { 165 INIT_HLIST_HEAD(&mmu_notifier_mm->list); 166 spin_lock_init(&mmu_notifier_mm->lock); 167 mm->mmu_notifier_mm = mmu_notifier_mm; 168 mmu_notifier_mm = NULL; 169 } 170 atomic_inc(&mm->mm_count); 171 172 /* 173 * Serialize the update against mmu_notifier_unregister. A 174 * side note: mmu_notifier_release can't run concurrently with 175 * us because we hold the mm_users pin (either implicitly as 176 * current->mm or explicitly with get_task_mm() or similar). 177 * We can't race against any other mmu notifier method either 178 * thanks to mm_take_all_locks(). 179 */ 180 spin_lock(&mm->mmu_notifier_mm->lock); 181 hlist_add_head(&mn->hlist, &mm->mmu_notifier_mm->list); 182 spin_unlock(&mm->mmu_notifier_mm->lock); 183 184 mm_drop_all_locks(mm); 185 out_cleanup: 186 if (take_mmap_sem) 187 up_write(&mm->mmap_sem); 188 /* kfree() does nothing if mmu_notifier_mm is NULL */ 189 kfree(mmu_notifier_mm); 190 out: 191 BUG_ON(atomic_read(&mm->mm_users) <= 0); 192 return ret; 193 } 194 195 /* 196 * Must not hold mmap_sem nor any other VM related lock when calling 197 * this registration function. Must also ensure mm_users can't go down 198 * to zero while this runs to avoid races with mmu_notifier_release, 199 * so mm has to be current->mm or the mm should be pinned safely such 200 * as with get_task_mm(). If the mm is not current->mm, the mm_users 201 * pin should be released by calling mmput after mmu_notifier_register 202 * returns. mmu_notifier_unregister must be always called to 203 * unregister the notifier. mm_count is automatically pinned to allow 204 * mmu_notifier_unregister to safely run at any time later, before or 205 * after exit_mmap. ->release will always be called before exit_mmap 206 * frees the pages. 207 */ 208 int mmu_notifier_register(struct mmu_notifier *mn, struct mm_struct *mm) 209 { 210 return do_mmu_notifier_register(mn, mm, 1); 211 } 212 EXPORT_SYMBOL_GPL(mmu_notifier_register); 213 214 /* 215 * Same as mmu_notifier_register but here the caller must hold the 216 * mmap_sem in write mode. 217 */ 218 int __mmu_notifier_register(struct mmu_notifier *mn, struct mm_struct *mm) 219 { 220 return do_mmu_notifier_register(mn, mm, 0); 221 } 222 EXPORT_SYMBOL_GPL(__mmu_notifier_register); 223 224 /* this is called after the last mmu_notifier_unregister() returned */ 225 void __mmu_notifier_mm_destroy(struct mm_struct *mm) 226 { 227 BUG_ON(!hlist_empty(&mm->mmu_notifier_mm->list)); 228 kfree(mm->mmu_notifier_mm); 229 mm->mmu_notifier_mm = LIST_POISON1; /* debug */ 230 } 231 232 /* 233 * This releases the mm_count pin automatically and frees the mm 234 * structure if it was the last user of it. It serializes against 235 * running mmu notifiers with RCU and against mmu_notifier_unregister 236 * with the unregister lock + RCU. All sptes must be dropped before 237 * calling mmu_notifier_unregister. ->release or any other notifier 238 * method may be invoked concurrently with mmu_notifier_unregister, 239 * and only after mmu_notifier_unregister returned we're guaranteed 240 * that ->release or any other method can't run anymore. 241 */ 242 void mmu_notifier_unregister(struct mmu_notifier *mn, struct mm_struct *mm) 243 { 244 BUG_ON(atomic_read(&mm->mm_count) <= 0); 245 246 spin_lock(&mm->mmu_notifier_mm->lock); 247 if (!hlist_unhashed(&mn->hlist)) { 248 hlist_del_rcu(&mn->hlist); 249 250 /* 251 * RCU here will force exit_mmap to wait ->release to finish 252 * before freeing the pages. 253 */ 254 rcu_read_lock(); 255 spin_unlock(&mm->mmu_notifier_mm->lock); 256 /* 257 * exit_mmap will block in mmu_notifier_release to 258 * guarantee ->release is called before freeing the 259 * pages. 260 */ 261 if (mn->ops->release) 262 mn->ops->release(mn, mm); 263 rcu_read_unlock(); 264 } else 265 spin_unlock(&mm->mmu_notifier_mm->lock); 266 267 /* 268 * Wait any running method to finish, of course including 269 * ->release if it was run by mmu_notifier_relase instead of us. 270 */ 271 synchronize_rcu(); 272 273 BUG_ON(atomic_read(&mm->mm_count) <= 0); 274 275 mmdrop(mm); 276 } 277 EXPORT_SYMBOL_GPL(mmu_notifier_unregister); 278