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>
9 #include <linux/swap.h>
11 #include <asm/pgalloc.h>
14 #ifndef CONFIG_MMU_GATHER_NO_GATHER
16 static bool tlb_next_batch(struct mmu_gather *tlb)
18 struct mmu_gather_batch *batch;
22 tlb->active = batch->next;
26 if (tlb->batch_count == MAX_GATHER_BATCH_COUNT)
29 batch = (void *)__get_free_pages(GFP_NOWAIT | __GFP_NOWARN, 0);
36 batch->max = MAX_GATHER_BATCH;
38 tlb->active->next = batch;
44 static void tlb_batch_pages_flush(struct mmu_gather *tlb)
46 struct mmu_gather_batch *batch;
48 for (batch = &tlb->local; batch && batch->nr; batch = batch->next) {
49 free_pages_and_swap_cache(batch->pages, batch->nr);
52 tlb->active = &tlb->local;
55 static void tlb_batch_list_free(struct mmu_gather *tlb)
57 struct mmu_gather_batch *batch, *next;
59 for (batch = tlb->local.next; batch; batch = next) {
61 free_pages((unsigned long)batch, 0);
63 tlb->local.next = NULL;
66 bool __tlb_remove_page_size(struct mmu_gather *tlb, struct page *page, int page_size)
68 struct mmu_gather_batch *batch;
72 #ifdef CONFIG_MMU_GATHER_PAGE_SIZE
73 VM_WARN_ON(tlb->page_size != page_size);
78 * Add the page and check if we are full. If so
81 batch->pages[batch->nr++] = page;
82 if (batch->nr == batch->max) {
83 if (!tlb_next_batch(tlb))
87 VM_BUG_ON_PAGE(batch->nr > batch->max, page);
92 #endif /* MMU_GATHER_NO_GATHER */
94 #ifdef CONFIG_MMU_GATHER_TABLE_FREE
96 static void __tlb_remove_table_free(struct mmu_table_batch *batch)
100 for (i = 0; i < batch->nr; i++)
101 __tlb_remove_table(batch->tables[i]);
103 free_page((unsigned long)batch);
106 #ifdef CONFIG_MMU_GATHER_RCU_TABLE_FREE
109 * Semi RCU freeing of the page directories.
111 * This is needed by some architectures to implement software pagetable walkers.
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
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
123 * Architectures that do not have this (PPC) need to delay the freeing by some
124 * other means, this is that means.
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.
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().
137 static void tlb_remove_table_smp_sync(void *arg)
139 /* Simply deliver the interrupt */
142 static void tlb_remove_table_sync_one(void)
145 * This isn't an RCU grace period and hence the page-tables cannot be
146 * assumed to be actually RCU-freed.
148 * It is however sufficient for software page-table walkers that rely on
151 smp_call_function(tlb_remove_table_smp_sync, NULL, 1);
154 static void tlb_remove_table_rcu(struct rcu_head *head)
156 __tlb_remove_table_free(container_of(head, struct mmu_table_batch, rcu));
159 static void tlb_remove_table_free(struct mmu_table_batch *batch)
161 call_rcu(&batch->rcu, tlb_remove_table_rcu);
164 #else /* !CONFIG_MMU_GATHER_RCU_TABLE_FREE */
166 static void tlb_remove_table_sync_one(void) { }
168 static void tlb_remove_table_free(struct mmu_table_batch *batch)
170 __tlb_remove_table_free(batch);
173 #endif /* CONFIG_MMU_GATHER_RCU_TABLE_FREE */
176 * If we want tlb_remove_table() to imply TLB invalidates.
178 static inline void tlb_table_invalidate(struct mmu_gather *tlb)
180 if (tlb_needs_table_invalidate()) {
182 * Invalidate page-table caches used by hardware walkers. Then
183 * we still need to RCU-sched wait while freeing the pages
184 * because software walkers can still be in-flight.
186 tlb_flush_mmu_tlbonly(tlb);
190 static void tlb_remove_table_one(void *table)
192 tlb_remove_table_sync_one();
193 __tlb_remove_table(table);
196 static void tlb_table_flush(struct mmu_gather *tlb)
198 struct mmu_table_batch **batch = &tlb->batch;
201 tlb_table_invalidate(tlb);
202 tlb_remove_table_free(*batch);
207 void tlb_remove_table(struct mmu_gather *tlb, void *table)
209 struct mmu_table_batch **batch = &tlb->batch;
211 if (*batch == NULL) {
212 *batch = (struct mmu_table_batch *)__get_free_page(GFP_NOWAIT | __GFP_NOWARN);
213 if (*batch == NULL) {
214 tlb_table_invalidate(tlb);
215 tlb_remove_table_one(table);
221 (*batch)->tables[(*batch)->nr++] = table;
222 if ((*batch)->nr == MAX_TABLE_BATCH)
223 tlb_table_flush(tlb);
226 static inline void tlb_table_init(struct mmu_gather *tlb)
231 #else /* !CONFIG_MMU_GATHER_TABLE_FREE */
233 static inline void tlb_table_flush(struct mmu_gather *tlb) { }
234 static inline void tlb_table_init(struct mmu_gather *tlb) { }
236 #endif /* CONFIG_MMU_GATHER_TABLE_FREE */
238 static void tlb_flush_mmu_free(struct mmu_gather *tlb)
240 tlb_table_flush(tlb);
241 #ifndef CONFIG_MMU_GATHER_NO_GATHER
242 tlb_batch_pages_flush(tlb);
246 void tlb_flush_mmu(struct mmu_gather *tlb)
248 tlb_flush_mmu_tlbonly(tlb);
249 tlb_flush_mmu_free(tlb);
253 * tlb_gather_mmu - initialize an mmu_gather structure for page-table tear-down
254 * @tlb: the mmu_gather structure to initialize
255 * @mm: the mm_struct of the target address space
256 * @fullmm: @mm is without users and we're going to destroy the full address
257 * space (exit/execve)
259 * Called to initialize an (on-stack) mmu_gather structure for page-table
260 * tear-down from @mm.
262 static void __tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm,
266 tlb->fullmm = fullmm;
268 #ifndef CONFIG_MMU_GATHER_NO_GATHER
269 tlb->need_flush_all = 0;
270 tlb->local.next = NULL;
272 tlb->local.max = ARRAY_SIZE(tlb->__pages);
273 tlb->active = &tlb->local;
274 tlb->batch_count = 0;
278 #ifdef CONFIG_MMU_GATHER_PAGE_SIZE
282 __tlb_reset_range(tlb);
283 inc_tlb_flush_pending(tlb->mm);
286 void tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm)
288 __tlb_gather_mmu(tlb, mm, false);
291 void tlb_gather_mmu_fullmm(struct mmu_gather *tlb, struct mm_struct *mm)
293 __tlb_gather_mmu(tlb, mm, true);
297 * tlb_finish_mmu - finish an mmu_gather structure
298 * @tlb: the mmu_gather structure to finish
300 * Called at the end of the shootdown operation to free up any resources that
303 void tlb_finish_mmu(struct mmu_gather *tlb)
306 * If there are parallel threads are doing PTE changes on same range
307 * under non-exclusive lock (e.g., mmap_lock read-side) but defer TLB
308 * flush by batching, one thread may end up seeing inconsistent PTEs
309 * and result in having stale TLB entries. So flush TLB forcefully
310 * if we detect parallel PTE batching threads.
312 * However, some syscalls, e.g. munmap(), may free page tables, this
313 * needs force flush everything in the given range. Otherwise this
314 * may result in having stale TLB entries for some architectures,
315 * e.g. aarch64, that could specify flush what level TLB.
317 if (mm_tlb_flush_nested(tlb->mm)) {
319 * The aarch64 yields better performance with fullmm by
320 * avoiding multiple CPUs spamming TLBI messages at the
323 * On x86 non-fullmm doesn't yield significant difference
327 __tlb_reset_range(tlb);
328 tlb->freed_tables = 1;
333 #ifndef CONFIG_MMU_GATHER_NO_GATHER
334 tlb_batch_list_free(tlb);
336 dec_tlb_flush_pending(tlb->mm);