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> |
6dd01bed | 7 | #include <linux/module.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 | |
9eb912d1 BG |
17 | DEFINE_PER_CPU_SHARED_ALIGNED(struct tlb_state, cpu_tlbstate) |
18 | = { &init_mm, 0, }; | |
19 | ||
c048fdfe GC |
20 | /* |
21 | * Smarter SMP flushing macros. | |
22 | * c/o Linus Torvalds. | |
23 | * | |
24 | * These mean you can really definitely utterly forget about | |
25 | * writing to user space from interrupts. (Its not allowed anyway). | |
26 | * | |
27 | * Optimizations Manfred Spraul <manfred@colorfullife.com> | |
28 | * | |
29 | * More scalable flush, from Andi Kleen | |
30 | * | |
52aec330 | 31 | * Implement flush IPI by CALL_FUNCTION_VECTOR, Alex Shi |
c048fdfe GC |
32 | */ |
33 | ||
52aec330 AS |
34 | struct flush_tlb_info { |
35 | struct mm_struct *flush_mm; | |
36 | unsigned long flush_start; | |
37 | unsigned long flush_end; | |
38 | }; | |
93296720 | 39 | |
c048fdfe GC |
40 | /* |
41 | * We cannot call mmdrop() because we are in interrupt context, | |
42 | * instead update mm->cpu_vm_mask. | |
43 | */ | |
44 | void leave_mm(int cpu) | |
45 | { | |
02171b4a | 46 | struct mm_struct *active_mm = this_cpu_read(cpu_tlbstate.active_mm); |
c6ae41e7 | 47 | if (this_cpu_read(cpu_tlbstate.state) == TLBSTATE_OK) |
c048fdfe | 48 | BUG(); |
a6fca40f SS |
49 | if (cpumask_test_cpu(cpu, mm_cpumask(active_mm))) { |
50 | cpumask_clear_cpu(cpu, mm_cpumask(active_mm)); | |
51 | load_cr3(swapper_pg_dir); | |
7c7f1547 DH |
52 | /* |
53 | * This gets called in the idle path where RCU | |
54 | * functions differently. Tracing normally | |
55 | * uses RCU, so we have to call the tracepoint | |
56 | * specially here. | |
57 | */ | |
58 | trace_tlb_flush_rcuidle(TLB_FLUSH_ON_TASK_SWITCH, TLB_FLUSH_ALL); | |
a6fca40f | 59 | } |
c048fdfe GC |
60 | } |
61 | EXPORT_SYMBOL_GPL(leave_mm); | |
62 | ||
63 | /* | |
c048fdfe GC |
64 | * The flush IPI assumes that a thread switch happens in this order: |
65 | * [cpu0: the cpu that switches] | |
66 | * 1) switch_mm() either 1a) or 1b) | |
67 | * 1a) thread switch to a different mm | |
52aec330 AS |
68 | * 1a1) set cpu_tlbstate to TLBSTATE_OK |
69 | * Now the tlb flush NMI handler flush_tlb_func won't call leave_mm | |
70 | * if cpu0 was in lazy tlb mode. | |
71 | * 1a2) update cpu active_mm | |
c048fdfe | 72 | * Now cpu0 accepts tlb flushes for the new mm. |
52aec330 | 73 | * 1a3) cpu_set(cpu, new_mm->cpu_vm_mask); |
c048fdfe GC |
74 | * Now the other cpus will send tlb flush ipis. |
75 | * 1a4) change cr3. | |
52aec330 AS |
76 | * 1a5) cpu_clear(cpu, old_mm->cpu_vm_mask); |
77 | * Stop ipi delivery for the old mm. This is not synchronized with | |
78 | * the other cpus, but flush_tlb_func ignore flush ipis for the wrong | |
79 | * mm, and in the worst case we perform a superfluous tlb flush. | |
c048fdfe | 80 | * 1b) thread switch without mm change |
52aec330 AS |
81 | * cpu active_mm is correct, cpu0 already handles flush ipis. |
82 | * 1b1) set cpu_tlbstate to TLBSTATE_OK | |
c048fdfe GC |
83 | * 1b2) test_and_set the cpu bit in cpu_vm_mask. |
84 | * Atomically set the bit [other cpus will start sending flush ipis], | |
85 | * and test the bit. | |
86 | * 1b3) if the bit was 0: leave_mm was called, flush the tlb. | |
87 | * 2) switch %%esp, ie current | |
88 | * | |
89 | * The interrupt must handle 2 special cases: | |
90 | * - cr3 is changed before %%esp, ie. it cannot use current->{active_,}mm. | |
91 | * - the cpu performs speculative tlb reads, i.e. even if the cpu only | |
92 | * runs in kernel space, the cpu could load tlb entries for user space | |
93 | * pages. | |
94 | * | |
52aec330 | 95 | * The good news is that cpu_tlbstate is local to each cpu, no |
c048fdfe GC |
96 | * write/read ordering problems. |
97 | */ | |
98 | ||
99 | /* | |
52aec330 | 100 | * TLB flush funcation: |
c048fdfe GC |
101 | * 1) Flush the tlb entries if the cpu uses the mm that's being flushed. |
102 | * 2) Leave the mm if we are in the lazy tlb mode. | |
02cf94c3 | 103 | */ |
52aec330 | 104 | static void flush_tlb_func(void *info) |
c048fdfe | 105 | { |
52aec330 | 106 | struct flush_tlb_info *f = info; |
c048fdfe | 107 | |
fd0f5869 TS |
108 | inc_irq_stat(irq_tlb_count); |
109 | ||
52aec330 AS |
110 | if (f->flush_mm != this_cpu_read(cpu_tlbstate.active_mm)) |
111 | return; | |
a23421f1 DH |
112 | if (!f->flush_end) |
113 | f->flush_end = f->flush_start + PAGE_SIZE; | |
c048fdfe | 114 | |
ec659934 | 115 | count_vm_tlb_event(NR_TLB_REMOTE_FLUSH_RECEIVED); |
52aec330 | 116 | if (this_cpu_read(cpu_tlbstate.state) == TLBSTATE_OK) { |
d17d8f9d | 117 | if (f->flush_end == TLB_FLUSH_ALL) { |
52aec330 | 118 | local_flush_tlb(); |
d17d8f9d DH |
119 | trace_tlb_flush(TLB_REMOTE_SHOOTDOWN, TLB_FLUSH_ALL); |
120 | } else { | |
52aec330 | 121 | unsigned long addr; |
d17d8f9d DH |
122 | unsigned long nr_pages = |
123 | f->flush_end - f->flush_start / PAGE_SIZE; | |
52aec330 AS |
124 | addr = f->flush_start; |
125 | while (addr < f->flush_end) { | |
126 | __flush_tlb_single(addr); | |
127 | addr += PAGE_SIZE; | |
e7b52ffd | 128 | } |
d17d8f9d | 129 | trace_tlb_flush(TLB_REMOTE_SHOOTDOWN, nr_pages); |
52aec330 AS |
130 | } |
131 | } else | |
132 | leave_mm(smp_processor_id()); | |
c048fdfe | 133 | |
c048fdfe GC |
134 | } |
135 | ||
4595f962 | 136 | void native_flush_tlb_others(const struct cpumask *cpumask, |
e7b52ffd AS |
137 | struct mm_struct *mm, unsigned long start, |
138 | unsigned long end) | |
4595f962 | 139 | { |
52aec330 AS |
140 | struct flush_tlb_info info; |
141 | info.flush_mm = mm; | |
142 | info.flush_start = start; | |
143 | info.flush_end = end; | |
144 | ||
ec659934 | 145 | count_vm_tlb_event(NR_TLB_REMOTE_FLUSH); |
4595f962 | 146 | if (is_uv_system()) { |
bdbcdd48 | 147 | unsigned int cpu; |
0e21990a | 148 | |
25542c64 | 149 | cpu = smp_processor_id(); |
e7b52ffd | 150 | cpumask = uv_flush_tlb_others(cpumask, mm, start, end, cpu); |
bdbcdd48 | 151 | if (cpumask) |
52aec330 AS |
152 | smp_call_function_many(cpumask, flush_tlb_func, |
153 | &info, 1); | |
0e21990a | 154 | return; |
4595f962 | 155 | } |
52aec330 | 156 | smp_call_function_many(cpumask, flush_tlb_func, &info, 1); |
c048fdfe | 157 | } |
c048fdfe GC |
158 | |
159 | void flush_tlb_current_task(void) | |
160 | { | |
161 | struct mm_struct *mm = current->mm; | |
c048fdfe GC |
162 | |
163 | preempt_disable(); | |
c048fdfe | 164 | |
ec659934 | 165 | count_vm_tlb_event(NR_TLB_LOCAL_FLUSH_ALL); |
c048fdfe | 166 | local_flush_tlb(); |
d17d8f9d | 167 | trace_tlb_flush(TLB_LOCAL_SHOOTDOWN, TLB_FLUSH_ALL); |
78f1c4d6 | 168 | if (cpumask_any_but(mm_cpumask(mm), smp_processor_id()) < nr_cpu_ids) |
e7b52ffd | 169 | flush_tlb_others(mm_cpumask(mm), mm, 0UL, TLB_FLUSH_ALL); |
c048fdfe GC |
170 | preempt_enable(); |
171 | } | |
172 | ||
a5102476 DH |
173 | /* |
174 | * See Documentation/x86/tlb.txt for details. We choose 33 | |
175 | * because it is large enough to cover the vast majority (at | |
176 | * least 95%) of allocations, and is small enough that we are | |
177 | * confident it will not cause too much overhead. Each single | |
178 | * flush is about 100 ns, so this caps the maximum overhead at | |
179 | * _about_ 3,000 ns. | |
180 | * | |
181 | * This is in units of pages. | |
182 | */ | |
86426851 | 183 | static unsigned long tlb_single_page_flush_ceiling __read_mostly = 33; |
e9f4e0a9 | 184 | |
611ae8e3 AS |
185 | void flush_tlb_mm_range(struct mm_struct *mm, unsigned long start, |
186 | unsigned long end, unsigned long vmflag) | |
187 | { | |
188 | unsigned long addr; | |
9dfa6dee DH |
189 | /* do a global flush by default */ |
190 | unsigned long base_pages_to_flush = TLB_FLUSH_ALL; | |
e7b52ffd AS |
191 | |
192 | preempt_disable(); | |
611ae8e3 | 193 | if (current->active_mm != mm) |
4995ab9c | 194 | goto out; |
e7b52ffd | 195 | |
611ae8e3 AS |
196 | if (!current->mm) { |
197 | leave_mm(smp_processor_id()); | |
4995ab9c | 198 | goto out; |
611ae8e3 | 199 | } |
c048fdfe | 200 | |
9dfa6dee DH |
201 | if ((end != TLB_FLUSH_ALL) && !(vmflag & VM_HUGETLB)) |
202 | base_pages_to_flush = (end - start) >> PAGE_SHIFT; | |
e7b52ffd | 203 | |
9dfa6dee DH |
204 | if (base_pages_to_flush > tlb_single_page_flush_ceiling) { |
205 | base_pages_to_flush = TLB_FLUSH_ALL; | |
ec659934 | 206 | count_vm_tlb_event(NR_TLB_LOCAL_FLUSH_ALL); |
611ae8e3 | 207 | local_flush_tlb(); |
9824cf97 | 208 | } else { |
611ae8e3 | 209 | /* flush range by one by one 'invlpg' */ |
9824cf97 | 210 | for (addr = start; addr < end; addr += PAGE_SIZE) { |
ec659934 | 211 | count_vm_tlb_event(NR_TLB_LOCAL_FLUSH_ONE); |
611ae8e3 | 212 | __flush_tlb_single(addr); |
9824cf97 | 213 | } |
e7b52ffd | 214 | } |
d17d8f9d | 215 | trace_tlb_flush(TLB_LOCAL_MM_SHOOTDOWN, base_pages_to_flush); |
4995ab9c | 216 | out: |
9dfa6dee | 217 | if (base_pages_to_flush == TLB_FLUSH_ALL) { |
4995ab9c DH |
218 | start = 0UL; |
219 | end = TLB_FLUSH_ALL; | |
220 | } | |
e7b52ffd | 221 | if (cpumask_any_but(mm_cpumask(mm), smp_processor_id()) < nr_cpu_ids) |
4995ab9c | 222 | flush_tlb_others(mm_cpumask(mm), mm, start, end); |
c048fdfe GC |
223 | preempt_enable(); |
224 | } | |
225 | ||
e7b52ffd | 226 | void flush_tlb_page(struct vm_area_struct *vma, unsigned long start) |
c048fdfe GC |
227 | { |
228 | struct mm_struct *mm = vma->vm_mm; | |
c048fdfe GC |
229 | |
230 | preempt_disable(); | |
c048fdfe GC |
231 | |
232 | if (current->active_mm == mm) { | |
233 | if (current->mm) | |
e7b52ffd | 234 | __flush_tlb_one(start); |
c048fdfe GC |
235 | else |
236 | leave_mm(smp_processor_id()); | |
237 | } | |
238 | ||
78f1c4d6 | 239 | if (cpumask_any_but(mm_cpumask(mm), smp_processor_id()) < nr_cpu_ids) |
e7b52ffd | 240 | flush_tlb_others(mm_cpumask(mm), mm, start, 0UL); |
c048fdfe GC |
241 | |
242 | preempt_enable(); | |
243 | } | |
244 | ||
245 | static void do_flush_tlb_all(void *info) | |
246 | { | |
ec659934 | 247 | count_vm_tlb_event(NR_TLB_REMOTE_FLUSH_RECEIVED); |
c048fdfe | 248 | __flush_tlb_all(); |
c6ae41e7 | 249 | if (this_cpu_read(cpu_tlbstate.state) == TLBSTATE_LAZY) |
3f8afb77 | 250 | leave_mm(smp_processor_id()); |
c048fdfe GC |
251 | } |
252 | ||
253 | void flush_tlb_all(void) | |
254 | { | |
ec659934 | 255 | count_vm_tlb_event(NR_TLB_REMOTE_FLUSH); |
15c8b6c1 | 256 | on_each_cpu(do_flush_tlb_all, NULL, 1); |
c048fdfe | 257 | } |
3df3212f | 258 | |
effee4b9 AS |
259 | static void do_kernel_range_flush(void *info) |
260 | { | |
261 | struct flush_tlb_info *f = info; | |
262 | unsigned long addr; | |
263 | ||
264 | /* flush range by one by one 'invlpg' */ | |
6df46865 | 265 | for (addr = f->flush_start; addr < f->flush_end; addr += PAGE_SIZE) |
effee4b9 AS |
266 | __flush_tlb_single(addr); |
267 | } | |
268 | ||
269 | void flush_tlb_kernel_range(unsigned long start, unsigned long end) | |
270 | { | |
effee4b9 AS |
271 | |
272 | /* Balance as user space task's flush, a bit conservative */ | |
e9f4e0a9 DH |
273 | if (end == TLB_FLUSH_ALL || |
274 | (end - start) > tlb_single_page_flush_ceiling * PAGE_SIZE) { | |
effee4b9 | 275 | on_each_cpu(do_flush_tlb_all, NULL, 1); |
e9f4e0a9 DH |
276 | } else { |
277 | struct flush_tlb_info info; | |
effee4b9 AS |
278 | info.flush_start = start; |
279 | info.flush_end = end; | |
280 | on_each_cpu(do_kernel_range_flush, &info, 1); | |
281 | } | |
282 | } | |
2d040a1c DH |
283 | |
284 | static ssize_t tlbflush_read_file(struct file *file, char __user *user_buf, | |
285 | size_t count, loff_t *ppos) | |
286 | { | |
287 | char buf[32]; | |
288 | unsigned int len; | |
289 | ||
290 | len = sprintf(buf, "%ld\n", tlb_single_page_flush_ceiling); | |
291 | return simple_read_from_buffer(user_buf, count, ppos, buf, len); | |
292 | } | |
293 | ||
294 | static ssize_t tlbflush_write_file(struct file *file, | |
295 | const char __user *user_buf, size_t count, loff_t *ppos) | |
296 | { | |
297 | char buf[32]; | |
298 | ssize_t len; | |
299 | int ceiling; | |
300 | ||
301 | len = min(count, sizeof(buf) - 1); | |
302 | if (copy_from_user(buf, user_buf, len)) | |
303 | return -EFAULT; | |
304 | ||
305 | buf[len] = '\0'; | |
306 | if (kstrtoint(buf, 0, &ceiling)) | |
307 | return -EINVAL; | |
308 | ||
309 | if (ceiling < 0) | |
310 | return -EINVAL; | |
311 | ||
312 | tlb_single_page_flush_ceiling = ceiling; | |
313 | return count; | |
314 | } | |
315 | ||
316 | static const struct file_operations fops_tlbflush = { | |
317 | .read = tlbflush_read_file, | |
318 | .write = tlbflush_write_file, | |
319 | .llseek = default_llseek, | |
320 | }; | |
321 | ||
322 | static int __init create_tlb_single_page_flush_ceiling(void) | |
323 | { | |
324 | debugfs_create_file("tlb_single_page_flush_ceiling", S_IRUSR | S_IWUSR, | |
325 | arch_debugfs_dir, NULL, &fops_tlbflush); | |
326 | return 0; | |
327 | } | |
328 | late_initcall(create_tlb_single_page_flush_ceiling); |