1 #include <linux/gfp.h>
2 #include <linux/highmem.h>
3 #include <linux/kernel.h>
4 #include <linux/mmdebug.h>
5 #include <linux/mm_types.h>
6 #include <linux/pagemap.h>
7 #include <linux/rcupdate.h>
8 #include <linux/smp.h>
9 #include <linux/swap.h>
10
11 #include <asm/pgalloc.h>
12 #include <asm/tlb.h>
13
14 #ifndef CONFIG_MMU_GATHER_NO_GATHER
15
tlb_next_batch(struct mmu_gather * tlb)16 static bool tlb_next_batch(struct mmu_gather *tlb)
17 {
18 struct mmu_gather_batch *batch;
19
20 batch = tlb->active;
21 if (batch->next) {
22 tlb->active = batch->next;
23 return true;
24 }
25
26 if (tlb->batch_count == MAX_GATHER_BATCH_COUNT)
27 return false;
28
29 batch = (void *)__get_free_pages(GFP_NOWAIT | __GFP_NOWARN, 0);
30 if (!batch)
31 return false;
32
33 tlb->batch_count++;
34 batch->next = NULL;
35 batch->nr = 0;
36 batch->max = MAX_GATHER_BATCH;
37
38 tlb->active->next = batch;
39 tlb->active = batch;
40
41 return true;
42 }
43
tlb_batch_pages_flush(struct mmu_gather * tlb)44 static void tlb_batch_pages_flush(struct mmu_gather *tlb)
45 {
46 struct mmu_gather_batch *batch;
47
48 for (batch = &tlb->local; batch && batch->nr; batch = batch->next) {
49 free_pages_and_swap_cache(batch->pages, batch->nr);
50 batch->nr = 0;
51 }
52 tlb->active = &tlb->local;
53 }
54
tlb_batch_list_free(struct mmu_gather * tlb)55 static void tlb_batch_list_free(struct mmu_gather *tlb)
56 {
57 struct mmu_gather_batch *batch, *next;
58
59 for (batch = tlb->local.next; batch; batch = next) {
60 next = batch->next;
61 free_pages((unsigned long)batch, 0);
62 }
63 tlb->local.next = NULL;
64 }
65
__tlb_remove_page_size(struct mmu_gather * tlb,struct page * page,int page_size)66 bool __tlb_remove_page_size(struct mmu_gather *tlb, struct page *page, int page_size)
67 {
68 struct mmu_gather_batch *batch;
69
70 VM_BUG_ON(!tlb->end);
71
72 #ifdef CONFIG_MMU_GATHER_PAGE_SIZE
73 VM_WARN_ON(tlb->page_size != page_size);
74 #endif
75
76 batch = tlb->active;
77 /*
78 * Add the page and check if we are full. If so
79 * force a flush.
80 */
81 batch->pages[batch->nr++] = page;
82 if (batch->nr == batch->max) {
83 if (!tlb_next_batch(tlb))
84 return true;
85 batch = tlb->active;
86 }
87 VM_BUG_ON_PAGE(batch->nr > batch->max, page);
88
89 return false;
90 }
91
92 #endif /* MMU_GATHER_NO_GATHER */
93
94 #ifdef CONFIG_MMU_GATHER_TABLE_FREE
95
__tlb_remove_table_free(struct mmu_table_batch * batch)96 static void __tlb_remove_table_free(struct mmu_table_batch *batch)
97 {
98 int i;
99
100 for (i = 0; i < batch->nr; i++)
101 __tlb_remove_table(batch->tables[i]);
102
103 free_page((unsigned long)batch);
104 }
105
106 #ifdef CONFIG_MMU_GATHER_RCU_TABLE_FREE
107
108 /*
109 * Semi RCU freeing of the page directories.
110 *
111 * This is needed by some architectures to implement software pagetable walkers.
112 *
113 * gup_fast() and other software pagetable walkers do a lockless page-table
114 * walk and therefore needs some synchronization with the freeing of the page
115 * directories. The chosen means to accomplish that is by disabling IRQs over
116 * the walk.
117 *
118 * Architectures that use IPIs to flush TLBs will then automagically DTRT,
119 * since we unlink the page, flush TLBs, free the page. Since the disabling of
120 * IRQs delays the completion of the TLB flush we can never observe an already
121 * freed page.
122 *
123 * Architectures that do not have this (PPC) need to delay the freeing by some
124 * other means, this is that means.
125 *
126 * What we do is batch the freed directory pages (tables) and RCU free them.
127 * We use the sched RCU variant, as that guarantees that IRQ/preempt disabling
128 * holds off grace periods.
129 *
130 * However, in order to batch these pages we need to allocate storage, this
131 * allocation is deep inside the MM code and can thus easily fail on memory
132 * pressure. To guarantee progress we fall back to single table freeing, see
133 * the implementation of tlb_remove_table_one().
134 *
135 */
136
tlb_remove_table_smp_sync(void * arg)137 static void tlb_remove_table_smp_sync(void *arg)
138 {
139 /* Simply deliver the interrupt */
140 }
141
tlb_remove_table_sync_one(void)142 void tlb_remove_table_sync_one(void)
143 {
144 /*
145 * This isn't an RCU grace period and hence the page-tables cannot be
146 * assumed to be actually RCU-freed.
147 *
148 * It is however sufficient for software page-table walkers that rely on
149 * IRQ disabling.
150 */
151 smp_call_function(tlb_remove_table_smp_sync, NULL, 1);
152 }
153
tlb_remove_table_rcu(struct rcu_head * head)154 static void tlb_remove_table_rcu(struct rcu_head *head)
155 {
156 __tlb_remove_table_free(container_of(head, struct mmu_table_batch, rcu));
157 }
158
tlb_remove_table_free(struct mmu_table_batch * batch)159 static void tlb_remove_table_free(struct mmu_table_batch *batch)
160 {
161 call_rcu(&batch->rcu, tlb_remove_table_rcu);
162 }
163
164 #else /* !CONFIG_MMU_GATHER_RCU_TABLE_FREE */
165
tlb_remove_table_free(struct mmu_table_batch * batch)166 static void tlb_remove_table_free(struct mmu_table_batch *batch)
167 {
168 __tlb_remove_table_free(batch);
169 }
170
171 #endif /* CONFIG_MMU_GATHER_RCU_TABLE_FREE */
172
173 /*
174 * If we want tlb_remove_table() to imply TLB invalidates.
175 */
tlb_table_invalidate(struct mmu_gather * tlb)176 static inline void tlb_table_invalidate(struct mmu_gather *tlb)
177 {
178 if (tlb_needs_table_invalidate()) {
179 /*
180 * Invalidate page-table caches used by hardware walkers. Then
181 * we still need to RCU-sched wait while freeing the pages
182 * because software walkers can still be in-flight.
183 */
184 tlb_flush_mmu_tlbonly(tlb);
185 }
186 }
187
tlb_remove_table_one(void * table)188 static void tlb_remove_table_one(void *table)
189 {
190 tlb_remove_table_sync_one();
191 __tlb_remove_table(table);
192 }
193
tlb_table_flush(struct mmu_gather * tlb)194 static void tlb_table_flush(struct mmu_gather *tlb)
195 {
196 struct mmu_table_batch **batch = &tlb->batch;
197
198 if (*batch) {
199 tlb_table_invalidate(tlb);
200 tlb_remove_table_free(*batch);
201 *batch = NULL;
202 }
203 }
204
tlb_remove_table(struct mmu_gather * tlb,void * table)205 void tlb_remove_table(struct mmu_gather *tlb, void *table)
206 {
207 struct mmu_table_batch **batch = &tlb->batch;
208
209 if (*batch == NULL) {
210 *batch = (struct mmu_table_batch *)__get_free_page(GFP_NOWAIT | __GFP_NOWARN);
211 if (*batch == NULL) {
212 tlb_table_invalidate(tlb);
213 tlb_remove_table_one(table);
214 return;
215 }
216 (*batch)->nr = 0;
217 }
218
219 (*batch)->tables[(*batch)->nr++] = table;
220 if ((*batch)->nr == MAX_TABLE_BATCH)
221 tlb_table_flush(tlb);
222 }
223
tlb_table_init(struct mmu_gather * tlb)224 static inline void tlb_table_init(struct mmu_gather *tlb)
225 {
226 tlb->batch = NULL;
227 }
228
229 #else /* !CONFIG_MMU_GATHER_TABLE_FREE */
230
tlb_table_flush(struct mmu_gather * tlb)231 static inline void tlb_table_flush(struct mmu_gather *tlb) { }
tlb_table_init(struct mmu_gather * tlb)232 static inline void tlb_table_init(struct mmu_gather *tlb) { }
233
234 #endif /* CONFIG_MMU_GATHER_TABLE_FREE */
235
tlb_flush_mmu_free(struct mmu_gather * tlb)236 static void tlb_flush_mmu_free(struct mmu_gather *tlb)
237 {
238 tlb_table_flush(tlb);
239 #ifndef CONFIG_MMU_GATHER_NO_GATHER
240 tlb_batch_pages_flush(tlb);
241 #endif
242 }
243
tlb_flush_mmu(struct mmu_gather * tlb)244 void tlb_flush_mmu(struct mmu_gather *tlb)
245 {
246 tlb_flush_mmu_tlbonly(tlb);
247 tlb_flush_mmu_free(tlb);
248 }
249
__tlb_gather_mmu(struct mmu_gather * tlb,struct mm_struct * mm,bool fullmm)250 static void __tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm,
251 bool fullmm)
252 {
253 tlb->mm = mm;
254 tlb->fullmm = fullmm;
255
256 #ifndef CONFIG_MMU_GATHER_NO_GATHER
257 tlb->need_flush_all = 0;
258 tlb->local.next = NULL;
259 tlb->local.nr = 0;
260 tlb->local.max = ARRAY_SIZE(tlb->__pages);
261 tlb->active = &tlb->local;
262 tlb->batch_count = 0;
263 #endif
264
265 tlb_table_init(tlb);
266 #ifdef CONFIG_MMU_GATHER_PAGE_SIZE
267 tlb->page_size = 0;
268 #endif
269
270 __tlb_reset_range(tlb);
271 inc_tlb_flush_pending(tlb->mm);
272 }
273
274 /**
275 * tlb_gather_mmu - initialize an mmu_gather structure for page-table tear-down
276 * @tlb: the mmu_gather structure to initialize
277 * @mm: the mm_struct of the target address space
278 *
279 * Called to initialize an (on-stack) mmu_gather structure for page-table
280 * tear-down from @mm.
281 */
tlb_gather_mmu(struct mmu_gather * tlb,struct mm_struct * mm)282 void tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm)
283 {
284 __tlb_gather_mmu(tlb, mm, false);
285 }
286
287 /**
288 * tlb_gather_mmu_fullmm - initialize an mmu_gather structure for page-table tear-down
289 * @tlb: the mmu_gather structure to initialize
290 * @mm: the mm_struct of the target address space
291 *
292 * In this case, @mm is without users and we're going to destroy the
293 * full address space (exit/execve).
294 *
295 * Called to initialize an (on-stack) mmu_gather structure for page-table
296 * tear-down from @mm.
297 */
tlb_gather_mmu_fullmm(struct mmu_gather * tlb,struct mm_struct * mm)298 void tlb_gather_mmu_fullmm(struct mmu_gather *tlb, struct mm_struct *mm)
299 {
300 __tlb_gather_mmu(tlb, mm, true);
301 }
302
303 /**
304 * tlb_finish_mmu - finish an mmu_gather structure
305 * @tlb: the mmu_gather structure to finish
306 *
307 * Called at the end of the shootdown operation to free up any resources that
308 * were required.
309 */
tlb_finish_mmu(struct mmu_gather * tlb)310 void tlb_finish_mmu(struct mmu_gather *tlb)
311 {
312 /*
313 * If there are parallel threads are doing PTE changes on same range
314 * under non-exclusive lock (e.g., mmap_lock read-side) but defer TLB
315 * flush by batching, one thread may end up seeing inconsistent PTEs
316 * and result in having stale TLB entries. So flush TLB forcefully
317 * if we detect parallel PTE batching threads.
318 *
319 * However, some syscalls, e.g. munmap(), may free page tables, this
320 * needs force flush everything in the given range. Otherwise this
321 * may result in having stale TLB entries for some architectures,
322 * e.g. aarch64, that could specify flush what level TLB.
323 */
324 if (mm_tlb_flush_nested(tlb->mm)) {
325 /*
326 * The aarch64 yields better performance with fullmm by
327 * avoiding multiple CPUs spamming TLBI messages at the
328 * same time.
329 *
330 * On x86 non-fullmm doesn't yield significant difference
331 * against fullmm.
332 */
333 tlb->fullmm = 1;
334 __tlb_reset_range(tlb);
335 tlb->freed_tables = 1;
336 }
337
338 tlb_flush_mmu(tlb);
339
340 #ifndef CONFIG_MMU_GATHER_NO_GATHER
341 tlb_batch_list_free(tlb);
342 #endif
343 dec_tlb_flush_pending(tlb->mm);
344 }
345