Upgrade to 4.4.50-rt62
[kvmfornfv.git] / kernel / arch / x86 / mm / tlb.c
1 #include <linux/init.h>
2
3 #include <linux/mm.h>
4 #include <linux/spinlock.h>
5 #include <linux/smp.h>
6 #include <linux/interrupt.h>
7 #include <linux/module.h>
8 #include <linux/cpu.h>
9
10 #include <asm/tlbflush.h>
11 #include <asm/mmu_context.h>
12 #include <asm/cache.h>
13 #include <asm/apic.h>
14 #include <asm/uv/uv.h>
15 #include <linux/debugfs.h>
16
17 /*
18  *      Smarter SMP flushing macros.
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  *
28  *      Implement flush IPI by CALL_FUNCTION_VECTOR, Alex Shi
29  */
30
31 struct flush_tlb_info {
32         struct mm_struct *flush_mm;
33         unsigned long flush_start;
34         unsigned long flush_end;
35 };
36
37 /*
38  * We cannot call mmdrop() because we are in interrupt context,
39  * instead update mm->cpu_vm_mask.
40  */
41 void leave_mm(int cpu)
42 {
43         struct mm_struct *active_mm = this_cpu_read(cpu_tlbstate.active_mm);
44         if (this_cpu_read(cpu_tlbstate.state) == TLBSTATE_OK)
45                 BUG();
46         if (cpumask_test_cpu(cpu, mm_cpumask(active_mm))) {
47                 cpumask_clear_cpu(cpu, mm_cpumask(active_mm));
48                 load_cr3(swapper_pg_dir);
49                 /*
50                  * This gets called in the idle path where RCU
51                  * functions differently.  Tracing normally
52                  * uses RCU, so we have to call the tracepoint
53                  * specially here.
54                  */
55                 trace_tlb_flush_rcuidle(TLB_FLUSH_ON_TASK_SWITCH, TLB_FLUSH_ALL);
56         }
57 }
58 EXPORT_SYMBOL_GPL(leave_mm);
59
60 /*
61  * The flush IPI assumes that a thread switch happens in this order:
62  * [cpu0: the cpu that switches]
63  * 1) switch_mm() either 1a) or 1b)
64  * 1a) thread switch to a different mm
65  * 1a1) set cpu_tlbstate to TLBSTATE_OK
66  *      Now the tlb flush NMI handler flush_tlb_func won't call leave_mm
67  *      if cpu0 was in lazy tlb mode.
68  * 1a2) update cpu active_mm
69  *      Now cpu0 accepts tlb flushes for the new mm.
70  * 1a3) cpu_set(cpu, new_mm->cpu_vm_mask);
71  *      Now the other cpus will send tlb flush ipis.
72  * 1a4) change cr3.
73  * 1a5) cpu_clear(cpu, old_mm->cpu_vm_mask);
74  *      Stop ipi delivery for the old mm. This is not synchronized with
75  *      the other cpus, but flush_tlb_func ignore flush ipis for the wrong
76  *      mm, and in the worst case we perform a superfluous tlb flush.
77  * 1b) thread switch without mm change
78  *      cpu active_mm is correct, cpu0 already handles flush ipis.
79  * 1b1) set cpu_tlbstate to TLBSTATE_OK
80  * 1b2) test_and_set the cpu bit in cpu_vm_mask.
81  *      Atomically set the bit [other cpus will start sending flush ipis],
82  *      and test the bit.
83  * 1b3) if the bit was 0: leave_mm was called, flush the tlb.
84  * 2) switch %%esp, ie current
85  *
86  * The interrupt must handle 2 special cases:
87  * - cr3 is changed before %%esp, ie. it cannot use current->{active_,}mm.
88  * - the cpu performs speculative tlb reads, i.e. even if the cpu only
89  *   runs in kernel space, the cpu could load tlb entries for user space
90  *   pages.
91  *
92  * The good news is that cpu_tlbstate is local to each cpu, no
93  * write/read ordering problems.
94  */
95
96 /*
97  * TLB flush funcation:
98  * 1) Flush the tlb entries if the cpu uses the mm that's being flushed.
99  * 2) Leave the mm if we are in the lazy tlb mode.
100  */
101 static void flush_tlb_func(void *info)
102 {
103         struct flush_tlb_info *f = info;
104
105         inc_irq_stat(irq_tlb_count);
106
107         if (f->flush_mm != this_cpu_read(cpu_tlbstate.active_mm))
108                 return;
109
110         count_vm_tlb_event(NR_TLB_REMOTE_FLUSH_RECEIVED);
111         if (this_cpu_read(cpu_tlbstate.state) == TLBSTATE_OK) {
112                 if (f->flush_end == TLB_FLUSH_ALL) {
113                         local_flush_tlb();
114                         trace_tlb_flush(TLB_REMOTE_SHOOTDOWN, TLB_FLUSH_ALL);
115                 } else {
116                         unsigned long addr;
117                         unsigned long nr_pages =
118                                 (f->flush_end - f->flush_start) / PAGE_SIZE;
119                         addr = f->flush_start;
120                         while (addr < f->flush_end) {
121                                 __flush_tlb_single(addr);
122                                 addr += PAGE_SIZE;
123                         }
124                         trace_tlb_flush(TLB_REMOTE_SHOOTDOWN, nr_pages);
125                 }
126         } else
127                 leave_mm(smp_processor_id());
128
129 }
130
131 void native_flush_tlb_others(const struct cpumask *cpumask,
132                                  struct mm_struct *mm, unsigned long start,
133                                  unsigned long end)
134 {
135         struct flush_tlb_info info;
136
137         if (end == 0)
138                 end = start + PAGE_SIZE;
139         info.flush_mm = mm;
140         info.flush_start = start;
141         info.flush_end = end;
142
143         count_vm_tlb_event(NR_TLB_REMOTE_FLUSH);
144         if (end == TLB_FLUSH_ALL)
145                 trace_tlb_flush(TLB_REMOTE_SEND_IPI, TLB_FLUSH_ALL);
146         else
147                 trace_tlb_flush(TLB_REMOTE_SEND_IPI,
148                                 (end - start) >> PAGE_SHIFT);
149
150         if (is_uv_system()) {
151                 unsigned int cpu;
152
153                 cpu = smp_processor_id();
154                 cpumask = uv_flush_tlb_others(cpumask, mm, start, end, cpu);
155                 if (cpumask)
156                         smp_call_function_many(cpumask, flush_tlb_func,
157                                                                 &info, 1);
158                 return;
159         }
160         smp_call_function_many(cpumask, flush_tlb_func, &info, 1);
161 }
162
163 void flush_tlb_current_task(void)
164 {
165         struct mm_struct *mm = current->mm;
166
167         preempt_disable();
168
169         count_vm_tlb_event(NR_TLB_LOCAL_FLUSH_ALL);
170
171         /* This is an implicit full barrier that synchronizes with switch_mm. */
172         local_flush_tlb();
173
174         trace_tlb_flush(TLB_LOCAL_SHOOTDOWN, TLB_FLUSH_ALL);
175         if (cpumask_any_but(mm_cpumask(mm), smp_processor_id()) < nr_cpu_ids)
176                 flush_tlb_others(mm_cpumask(mm), mm, 0UL, TLB_FLUSH_ALL);
177         preempt_enable();
178 }
179
180 /*
181  * See Documentation/x86/tlb.txt for details.  We choose 33
182  * because it is large enough to cover the vast majority (at
183  * least 95%) of allocations, and is small enough that we are
184  * confident it will not cause too much overhead.  Each single
185  * flush is about 100 ns, so this caps the maximum overhead at
186  * _about_ 3,000 ns.
187  *
188  * This is in units of pages.
189  */
190 static unsigned long tlb_single_page_flush_ceiling __read_mostly = 33;
191
192 void flush_tlb_mm_range(struct mm_struct *mm, unsigned long start,
193                                 unsigned long end, unsigned long vmflag)
194 {
195         unsigned long addr;
196         /* do a global flush by default */
197         unsigned long base_pages_to_flush = TLB_FLUSH_ALL;
198
199         preempt_disable();
200         if (current->active_mm != mm) {
201                 /* Synchronize with switch_mm. */
202                 smp_mb();
203
204                 goto out;
205         }
206
207         if (!current->mm) {
208                 leave_mm(smp_processor_id());
209
210                 /* Synchronize with switch_mm. */
211                 smp_mb();
212
213                 goto out;
214         }
215
216         if ((end != TLB_FLUSH_ALL) && !(vmflag & VM_HUGETLB))
217                 base_pages_to_flush = (end - start) >> PAGE_SHIFT;
218
219         /*
220          * Both branches below are implicit full barriers (MOV to CR or
221          * INVLPG) that synchronize with switch_mm.
222          */
223         if (base_pages_to_flush > tlb_single_page_flush_ceiling) {
224                 base_pages_to_flush = TLB_FLUSH_ALL;
225                 count_vm_tlb_event(NR_TLB_LOCAL_FLUSH_ALL);
226                 local_flush_tlb();
227         } else {
228                 /* flush range by one by one 'invlpg' */
229                 for (addr = start; addr < end;  addr += PAGE_SIZE) {
230                         count_vm_tlb_event(NR_TLB_LOCAL_FLUSH_ONE);
231                         __flush_tlb_single(addr);
232                 }
233         }
234         trace_tlb_flush(TLB_LOCAL_MM_SHOOTDOWN, base_pages_to_flush);
235 out:
236         if (base_pages_to_flush == TLB_FLUSH_ALL) {
237                 start = 0UL;
238                 end = TLB_FLUSH_ALL;
239         }
240         if (cpumask_any_but(mm_cpumask(mm), smp_processor_id()) < nr_cpu_ids)
241                 flush_tlb_others(mm_cpumask(mm), mm, start, end);
242         preempt_enable();
243 }
244
245 void flush_tlb_page(struct vm_area_struct *vma, unsigned long start)
246 {
247         struct mm_struct *mm = vma->vm_mm;
248
249         preempt_disable();
250
251         if (current->active_mm == mm) {
252                 if (current->mm) {
253                         /*
254                          * Implicit full barrier (INVLPG) that synchronizes
255                          * with switch_mm.
256                          */
257                         __flush_tlb_one(start);
258                 } else {
259                         leave_mm(smp_processor_id());
260
261                         /* Synchronize with switch_mm. */
262                         smp_mb();
263                 }
264         }
265
266         if (cpumask_any_but(mm_cpumask(mm), smp_processor_id()) < nr_cpu_ids)
267                 flush_tlb_others(mm_cpumask(mm), mm, start, 0UL);
268
269         preempt_enable();
270 }
271
272 static void do_flush_tlb_all(void *info)
273 {
274         count_vm_tlb_event(NR_TLB_REMOTE_FLUSH_RECEIVED);
275         __flush_tlb_all();
276         if (this_cpu_read(cpu_tlbstate.state) == TLBSTATE_LAZY)
277                 leave_mm(smp_processor_id());
278 }
279
280 void flush_tlb_all(void)
281 {
282         count_vm_tlb_event(NR_TLB_REMOTE_FLUSH);
283         on_each_cpu(do_flush_tlb_all, NULL, 1);
284 }
285
286 static void do_kernel_range_flush(void *info)
287 {
288         struct flush_tlb_info *f = info;
289         unsigned long addr;
290
291         /* flush range by one by one 'invlpg' */
292         for (addr = f->flush_start; addr < f->flush_end; addr += PAGE_SIZE)
293                 __flush_tlb_single(addr);
294 }
295
296 void flush_tlb_kernel_range(unsigned long start, unsigned long end)
297 {
298
299         /* Balance as user space task's flush, a bit conservative */
300         if (end == TLB_FLUSH_ALL ||
301             (end - start) > tlb_single_page_flush_ceiling * PAGE_SIZE) {
302                 on_each_cpu(do_flush_tlb_all, NULL, 1);
303         } else {
304                 struct flush_tlb_info info;
305                 info.flush_start = start;
306                 info.flush_end = end;
307                 on_each_cpu(do_kernel_range_flush, &info, 1);
308         }
309 }
310
311 static ssize_t tlbflush_read_file(struct file *file, char __user *user_buf,
312                              size_t count, loff_t *ppos)
313 {
314         char buf[32];
315         unsigned int len;
316
317         len = sprintf(buf, "%ld\n", tlb_single_page_flush_ceiling);
318         return simple_read_from_buffer(user_buf, count, ppos, buf, len);
319 }
320
321 static ssize_t tlbflush_write_file(struct file *file,
322                  const char __user *user_buf, size_t count, loff_t *ppos)
323 {
324         char buf[32];
325         ssize_t len;
326         int ceiling;
327
328         len = min(count, sizeof(buf) - 1);
329         if (copy_from_user(buf, user_buf, len))
330                 return -EFAULT;
331
332         buf[len] = '\0';
333         if (kstrtoint(buf, 0, &ceiling))
334                 return -EINVAL;
335
336         if (ceiling < 0)
337                 return -EINVAL;
338
339         tlb_single_page_flush_ceiling = ceiling;
340         return count;
341 }
342
343 static const struct file_operations fops_tlbflush = {
344         .read = tlbflush_read_file,
345         .write = tlbflush_write_file,
346         .llseek = default_llseek,
347 };
348
349 static int __init create_tlb_single_page_flush_ceiling(void)
350 {
351         debugfs_create_file("tlb_single_page_flush_ceiling", S_IRUSR | S_IWUSR,
352                             arch_debugfs_dir, NULL, &fops_tlbflush);
353         return 0;
354 }
355 late_initcall(create_tlb_single_page_flush_ceiling);