]>
Commit | Line | Data |
---|---|---|
c048fdfe GC |
1 | #include <linux/init.h> |
2 | ||
3 | #include <linux/mm.h> | |
c048fdfe GC |
4 | #include <linux/spinlock.h> |
5 | #include <linux/smp.h> | |
c048fdfe | 6 | #include <linux/interrupt.h> |
4b599fed | 7 | #include <linux/export.h> |
93296720 | 8 | #include <linux/cpu.h> |
c048fdfe | 9 | |
c048fdfe | 10 | #include <asm/tlbflush.h> |
c048fdfe | 11 | #include <asm/mmu_context.h> |
350f8f56 | 12 | #include <asm/cache.h> |
6dd01bed | 13 | #include <asm/apic.h> |
bdbcdd48 | 14 | #include <asm/uv/uv.h> |
3df3212f | 15 | #include <linux/debugfs.h> |
5af5573e | 16 | |
c048fdfe | 17 | /* |
ce4a4e56 | 18 | * TLB flushing, formerly SMP-only |
c048fdfe GC |
19 | * c/o Linus Torvalds. |
20 | * | |
21 | * These mean you can really definitely utterly forget about | |
22 | * writing to user space from interrupts. (Its not allowed anyway). | |
23 | * | |
24 | * Optimizations Manfred Spraul <manfred@colorfullife.com> | |
25 | * | |
26 | * More scalable flush, from Andi Kleen | |
27 | * | |
52aec330 | 28 | * Implement flush IPI by CALL_FUNCTION_VECTOR, Alex Shi |
c048fdfe GC |
29 | */ |
30 | ||
f39681ed AL |
31 | atomic64_t last_mm_ctx_id = ATOMIC64_INIT(1); |
32 | ||
c048fdfe GC |
33 | void leave_mm(int cpu) |
34 | { | |
3d28ebce AL |
35 | struct mm_struct *loaded_mm = this_cpu_read(cpu_tlbstate.loaded_mm); |
36 | ||
37 | /* | |
38 | * It's plausible that we're in lazy TLB mode while our mm is init_mm. | |
39 | * If so, our callers still expect us to flush the TLB, but there | |
40 | * aren't any user TLB entries in init_mm to worry about. | |
41 | * | |
42 | * This needs to happen before any other sanity checks due to | |
43 | * intel_idle's shenanigans. | |
44 | */ | |
45 | if (loaded_mm == &init_mm) | |
46 | return; | |
47 | ||
c6ae41e7 | 48 | if (this_cpu_read(cpu_tlbstate.state) == TLBSTATE_OK) |
c048fdfe | 49 | BUG(); |
3d28ebce AL |
50 | |
51 | switch_mm(NULL, &init_mm, NULL); | |
c048fdfe GC |
52 | } |
53 | EXPORT_SYMBOL_GPL(leave_mm); | |
54 | ||
69c0319a AL |
55 | void switch_mm(struct mm_struct *prev, struct mm_struct *next, |
56 | struct task_struct *tsk) | |
078194f8 AL |
57 | { |
58 | unsigned long flags; | |
59 | ||
60 | local_irq_save(flags); | |
61 | switch_mm_irqs_off(prev, next, tsk); | |
62 | local_irq_restore(flags); | |
63 | } | |
64 | ||
65 | void switch_mm_irqs_off(struct mm_struct *prev, struct mm_struct *next, | |
66 | struct task_struct *tsk) | |
69c0319a AL |
67 | { |
68 | unsigned cpu = smp_processor_id(); | |
3d28ebce | 69 | struct mm_struct *real_prev = this_cpu_read(cpu_tlbstate.loaded_mm); |
69c0319a | 70 | |
3d28ebce AL |
71 | /* |
72 | * NB: The scheduler will call us with prev == next when | |
73 | * switching from lazy TLB mode to normal mode if active_mm | |
74 | * isn't changing. When this happens, there is no guarantee | |
75 | * that CR3 (and hence cpu_tlbstate.loaded_mm) matches next. | |
76 | * | |
77 | * NB: leave_mm() calls us with prev == NULL and tsk == NULL. | |
78 | */ | |
e37e43a4 | 79 | |
3d28ebce | 80 | this_cpu_write(cpu_tlbstate.state, TLBSTATE_OK); |
e37e43a4 | 81 | |
3d28ebce AL |
82 | if (real_prev == next) { |
83 | /* | |
84 | * There's nothing to do: we always keep the per-mm control | |
85 | * regs in sync with cpu_tlbstate.loaded_mm. Just | |
86 | * sanity-check mm_cpumask. | |
87 | */ | |
88 | if (WARN_ON_ONCE(!cpumask_test_cpu(cpu, mm_cpumask(next)))) | |
89 | cpumask_set_cpu(cpu, mm_cpumask(next)); | |
90 | return; | |
91 | } | |
69c0319a | 92 | |
3d28ebce | 93 | if (IS_ENABLED(CONFIG_VMAP_STACK)) { |
69c0319a | 94 | /* |
3d28ebce AL |
95 | * If our current stack is in vmalloc space and isn't |
96 | * mapped in the new pgd, we'll double-fault. Forcibly | |
97 | * map it. | |
69c0319a | 98 | */ |
3d28ebce | 99 | unsigned int stack_pgd_index = pgd_index(current_stack_pointer()); |
69c0319a | 100 | |
3d28ebce | 101 | pgd_t *pgd = next->pgd + stack_pgd_index; |
69c0319a | 102 | |
3d28ebce AL |
103 | if (unlikely(pgd_none(*pgd))) |
104 | set_pgd(pgd, init_mm.pgd[stack_pgd_index]); | |
105 | } | |
69c0319a | 106 | |
3d28ebce AL |
107 | this_cpu_write(cpu_tlbstate.loaded_mm, next); |
108 | ||
109 | WARN_ON_ONCE(cpumask_test_cpu(cpu, mm_cpumask(next))); | |
110 | cpumask_set_cpu(cpu, mm_cpumask(next)); | |
111 | ||
112 | /* | |
113 | * Re-load page tables. | |
114 | * | |
115 | * This logic has an ordering constraint: | |
116 | * | |
117 | * CPU 0: Write to a PTE for 'next' | |
118 | * CPU 0: load bit 1 in mm_cpumask. if nonzero, send IPI. | |
119 | * CPU 1: set bit 1 in next's mm_cpumask | |
120 | * CPU 1: load from the PTE that CPU 0 writes (implicit) | |
121 | * | |
122 | * We need to prevent an outcome in which CPU 1 observes | |
123 | * the new PTE value and CPU 0 observes bit 1 clear in | |
124 | * mm_cpumask. (If that occurs, then the IPI will never | |
125 | * be sent, and CPU 0's TLB will contain a stale entry.) | |
126 | * | |
127 | * The bad outcome can occur if either CPU's load is | |
128 | * reordered before that CPU's store, so both CPUs must | |
129 | * execute full barriers to prevent this from happening. | |
130 | * | |
131 | * Thus, switch_mm needs a full barrier between the | |
132 | * store to mm_cpumask and any operation that could load | |
133 | * from next->pgd. TLB fills are special and can happen | |
134 | * due to instruction fetches or for no reason at all, | |
135 | * and neither LOCK nor MFENCE orders them. | |
136 | * Fortunately, load_cr3() is serializing and gives the | |
137 | * ordering guarantee we need. | |
138 | */ | |
139 | load_cr3(next->pgd); | |
140 | ||
141 | /* | |
142 | * This gets called via leave_mm() in the idle path where RCU | |
143 | * functions differently. Tracing normally uses RCU, so we have to | |
144 | * call the tracepoint specially here. | |
145 | */ | |
146 | trace_tlb_flush_rcuidle(TLB_FLUSH_ON_TASK_SWITCH, TLB_FLUSH_ALL); | |
147 | ||
148 | /* Stop flush ipis for the previous mm */ | |
149 | WARN_ON_ONCE(!cpumask_test_cpu(cpu, mm_cpumask(real_prev)) && | |
150 | real_prev != &init_mm); | |
151 | cpumask_clear_cpu(cpu, mm_cpumask(real_prev)); | |
152 | ||
73534258 | 153 | /* Load per-mm CR4 and LDTR state */ |
3d28ebce | 154 | load_mm_cr4(next); |
73534258 | 155 | switch_ldt(real_prev, next); |
69c0319a AL |
156 | } |
157 | ||
454bbad9 AL |
158 | static void flush_tlb_func_common(const struct flush_tlb_info *f, |
159 | bool local, enum tlb_flush_reason reason) | |
c048fdfe | 160 | { |
bc0d5a89 AL |
161 | /* This code cannot presently handle being reentered. */ |
162 | VM_WARN_ON(!irqs_disabled()); | |
163 | ||
b3b90e5a | 164 | if (this_cpu_read(cpu_tlbstate.state) != TLBSTATE_OK) { |
52aec330 | 165 | leave_mm(smp_processor_id()); |
b3b90e5a AL |
166 | return; |
167 | } | |
c048fdfe | 168 | |
a2055abe | 169 | if (f->end == TLB_FLUSH_ALL) { |
b3b90e5a | 170 | local_flush_tlb(); |
454bbad9 AL |
171 | if (local) |
172 | count_vm_tlb_event(NR_TLB_LOCAL_FLUSH_ALL); | |
173 | trace_tlb_flush(reason, TLB_FLUSH_ALL); | |
b3b90e5a AL |
174 | } else { |
175 | unsigned long addr; | |
be4ffc0d | 176 | unsigned long nr_pages = (f->end - f->start) >> PAGE_SHIFT; |
a2055abe AL |
177 | addr = f->start; |
178 | while (addr < f->end) { | |
b3b90e5a AL |
179 | __flush_tlb_single(addr); |
180 | addr += PAGE_SIZE; | |
181 | } | |
454bbad9 AL |
182 | if (local) |
183 | count_vm_tlb_events(NR_TLB_LOCAL_FLUSH_ONE, nr_pages); | |
184 | trace_tlb_flush(reason, nr_pages); | |
b3b90e5a | 185 | } |
c048fdfe GC |
186 | } |
187 | ||
454bbad9 AL |
188 | static void flush_tlb_func_local(void *info, enum tlb_flush_reason reason) |
189 | { | |
190 | const struct flush_tlb_info *f = info; | |
191 | ||
192 | flush_tlb_func_common(f, true, reason); | |
193 | } | |
194 | ||
195 | static void flush_tlb_func_remote(void *info) | |
196 | { | |
197 | const struct flush_tlb_info *f = info; | |
198 | ||
199 | inc_irq_stat(irq_tlb_count); | |
200 | ||
3d28ebce | 201 | if (f->mm && f->mm != this_cpu_read(cpu_tlbstate.loaded_mm)) |
454bbad9 AL |
202 | return; |
203 | ||
204 | count_vm_tlb_event(NR_TLB_REMOTE_FLUSH_RECEIVED); | |
205 | flush_tlb_func_common(f, false, TLB_REMOTE_SHOOTDOWN); | |
206 | } | |
207 | ||
4595f962 | 208 | void native_flush_tlb_others(const struct cpumask *cpumask, |
a2055abe | 209 | const struct flush_tlb_info *info) |
4595f962 | 210 | { |
ec659934 | 211 | count_vm_tlb_event(NR_TLB_REMOTE_FLUSH); |
a2055abe | 212 | if (info->end == TLB_FLUSH_ALL) |
18c98243 NA |
213 | trace_tlb_flush(TLB_REMOTE_SEND_IPI, TLB_FLUSH_ALL); |
214 | else | |
215 | trace_tlb_flush(TLB_REMOTE_SEND_IPI, | |
a2055abe | 216 | (info->end - info->start) >> PAGE_SHIFT); |
18c98243 | 217 | |
4595f962 | 218 | if (is_uv_system()) { |
bdbcdd48 | 219 | unsigned int cpu; |
0e21990a | 220 | |
25542c64 | 221 | cpu = smp_processor_id(); |
a2055abe | 222 | cpumask = uv_flush_tlb_others(cpumask, info); |
bdbcdd48 | 223 | if (cpumask) |
454bbad9 | 224 | smp_call_function_many(cpumask, flush_tlb_func_remote, |
a2055abe | 225 | (void *)info, 1); |
0e21990a | 226 | return; |
4595f962 | 227 | } |
454bbad9 | 228 | smp_call_function_many(cpumask, flush_tlb_func_remote, |
a2055abe | 229 | (void *)info, 1); |
c048fdfe | 230 | } |
c048fdfe | 231 | |
a5102476 DH |
232 | /* |
233 | * See Documentation/x86/tlb.txt for details. We choose 33 | |
234 | * because it is large enough to cover the vast majority (at | |
235 | * least 95%) of allocations, and is small enough that we are | |
236 | * confident it will not cause too much overhead. Each single | |
237 | * flush is about 100 ns, so this caps the maximum overhead at | |
238 | * _about_ 3,000 ns. | |
239 | * | |
240 | * This is in units of pages. | |
241 | */ | |
86426851 | 242 | static unsigned long tlb_single_page_flush_ceiling __read_mostly = 33; |
e9f4e0a9 | 243 | |
611ae8e3 AS |
244 | void flush_tlb_mm_range(struct mm_struct *mm, unsigned long start, |
245 | unsigned long end, unsigned long vmflag) | |
246 | { | |
454bbad9 | 247 | int cpu; |
ce27374f | 248 | |
454bbad9 AL |
249 | struct flush_tlb_info info = { |
250 | .mm = mm, | |
251 | }; | |
ce27374f | 252 | |
454bbad9 | 253 | cpu = get_cpu(); |
71b3c126 | 254 | |
f39681ed AL |
255 | /* This is also a barrier that synchronizes with switch_mm(). */ |
256 | inc_mm_tlb_gen(mm); | |
71b3c126 | 257 | |
454bbad9 AL |
258 | /* Should we flush just the requested range? */ |
259 | if ((end != TLB_FLUSH_ALL) && | |
260 | !(vmflag & VM_HUGETLB) && | |
261 | ((end - start) >> PAGE_SHIFT) <= tlb_single_page_flush_ceiling) { | |
262 | info.start = start; | |
263 | info.end = end; | |
9824cf97 | 264 | } else { |
a2055abe AL |
265 | info.start = 0UL; |
266 | info.end = TLB_FLUSH_ALL; | |
4995ab9c | 267 | } |
454bbad9 | 268 | |
bc0d5a89 AL |
269 | if (mm == this_cpu_read(cpu_tlbstate.loaded_mm)) { |
270 | VM_WARN_ON(irqs_disabled()); | |
271 | local_irq_disable(); | |
454bbad9 | 272 | flush_tlb_func_local(&info, TLB_LOCAL_MM_SHOOTDOWN); |
bc0d5a89 AL |
273 | local_irq_enable(); |
274 | } | |
275 | ||
454bbad9 | 276 | if (cpumask_any_but(mm_cpumask(mm), cpu) < nr_cpu_ids) |
a2055abe | 277 | flush_tlb_others(mm_cpumask(mm), &info); |
454bbad9 | 278 | put_cpu(); |
c048fdfe GC |
279 | } |
280 | ||
a2055abe | 281 | |
c048fdfe GC |
282 | static void do_flush_tlb_all(void *info) |
283 | { | |
ec659934 | 284 | count_vm_tlb_event(NR_TLB_REMOTE_FLUSH_RECEIVED); |
c048fdfe | 285 | __flush_tlb_all(); |
c6ae41e7 | 286 | if (this_cpu_read(cpu_tlbstate.state) == TLBSTATE_LAZY) |
3f8afb77 | 287 | leave_mm(smp_processor_id()); |
c048fdfe GC |
288 | } |
289 | ||
290 | void flush_tlb_all(void) | |
291 | { | |
ec659934 | 292 | count_vm_tlb_event(NR_TLB_REMOTE_FLUSH); |
15c8b6c1 | 293 | on_each_cpu(do_flush_tlb_all, NULL, 1); |
c048fdfe | 294 | } |
3df3212f | 295 | |
effee4b9 AS |
296 | static void do_kernel_range_flush(void *info) |
297 | { | |
298 | struct flush_tlb_info *f = info; | |
299 | unsigned long addr; | |
300 | ||
301 | /* flush range by one by one 'invlpg' */ | |
a2055abe | 302 | for (addr = f->start; addr < f->end; addr += PAGE_SIZE) |
effee4b9 AS |
303 | __flush_tlb_single(addr); |
304 | } | |
305 | ||
306 | void flush_tlb_kernel_range(unsigned long start, unsigned long end) | |
307 | { | |
effee4b9 AS |
308 | |
309 | /* Balance as user space task's flush, a bit conservative */ | |
e9f4e0a9 | 310 | if (end == TLB_FLUSH_ALL || |
be4ffc0d | 311 | (end - start) > tlb_single_page_flush_ceiling << PAGE_SHIFT) { |
effee4b9 | 312 | on_each_cpu(do_flush_tlb_all, NULL, 1); |
e9f4e0a9 DH |
313 | } else { |
314 | struct flush_tlb_info info; | |
a2055abe AL |
315 | info.start = start; |
316 | info.end = end; | |
effee4b9 AS |
317 | on_each_cpu(do_kernel_range_flush, &info, 1); |
318 | } | |
319 | } | |
2d040a1c | 320 | |
e73ad5ff AL |
321 | void arch_tlbbatch_flush(struct arch_tlbflush_unmap_batch *batch) |
322 | { | |
a2055abe AL |
323 | struct flush_tlb_info info = { |
324 | .mm = NULL, | |
325 | .start = 0UL, | |
326 | .end = TLB_FLUSH_ALL, | |
327 | }; | |
328 | ||
e73ad5ff AL |
329 | int cpu = get_cpu(); |
330 | ||
bc0d5a89 AL |
331 | if (cpumask_test_cpu(cpu, &batch->cpumask)) { |
332 | VM_WARN_ON(irqs_disabled()); | |
333 | local_irq_disable(); | |
3f79e4c7 | 334 | flush_tlb_func_local(&info, TLB_LOCAL_SHOOTDOWN); |
bc0d5a89 AL |
335 | local_irq_enable(); |
336 | } | |
337 | ||
e73ad5ff | 338 | if (cpumask_any_but(&batch->cpumask, cpu) < nr_cpu_ids) |
a2055abe | 339 | flush_tlb_others(&batch->cpumask, &info); |
e73ad5ff AL |
340 | cpumask_clear(&batch->cpumask); |
341 | ||
342 | put_cpu(); | |
343 | } | |
344 | ||
2d040a1c DH |
345 | static ssize_t tlbflush_read_file(struct file *file, char __user *user_buf, |
346 | size_t count, loff_t *ppos) | |
347 | { | |
348 | char buf[32]; | |
349 | unsigned int len; | |
350 | ||
351 | len = sprintf(buf, "%ld\n", tlb_single_page_flush_ceiling); | |
352 | return simple_read_from_buffer(user_buf, count, ppos, buf, len); | |
353 | } | |
354 | ||
355 | static ssize_t tlbflush_write_file(struct file *file, | |
356 | const char __user *user_buf, size_t count, loff_t *ppos) | |
357 | { | |
358 | char buf[32]; | |
359 | ssize_t len; | |
360 | int ceiling; | |
361 | ||
362 | len = min(count, sizeof(buf) - 1); | |
363 | if (copy_from_user(buf, user_buf, len)) | |
364 | return -EFAULT; | |
365 | ||
366 | buf[len] = '\0'; | |
367 | if (kstrtoint(buf, 0, &ceiling)) | |
368 | return -EINVAL; | |
369 | ||
370 | if (ceiling < 0) | |
371 | return -EINVAL; | |
372 | ||
373 | tlb_single_page_flush_ceiling = ceiling; | |
374 | return count; | |
375 | } | |
376 | ||
377 | static const struct file_operations fops_tlbflush = { | |
378 | .read = tlbflush_read_file, | |
379 | .write = tlbflush_write_file, | |
380 | .llseek = default_llseek, | |
381 | }; | |
382 | ||
383 | static int __init create_tlb_single_page_flush_ceiling(void) | |
384 | { | |
385 | debugfs_create_file("tlb_single_page_flush_ceiling", S_IRUSR | S_IWUSR, | |
386 | arch_debugfs_dir, NULL, &fops_tlbflush); | |
387 | return 0; | |
388 | } | |
389 | late_initcall(create_tlb_single_page_flush_ceiling); |