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 DEFINE_PER_CPU_SHARED_ALIGNED(struct tlb_state, cpu_tlbstate)
  18                         = { &init_mm, 0, };
  19
  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  *
  31  *      Implement flush IPI by CALL_FUNCTION_VECTOR, Alex Shi
  32  */
  33
  34 struct flush_tlb_info {
  35         struct mm_struct *flush_mm;
  36         unsigned long flush_start;
  37         unsigned long flush_end;
  38 };
  39
  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 {
  46         struct mm_struct *active_mm = this_cpu_read(cpu_tlbstate.active_mm);
  47         if (this_cpu_read(cpu_tlbstate.state) == TLBSTATE_OK)
  48                 BUG();
  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);
  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);
  59         }
  60 }
  61 EXPORT_SYMBOL_GPL(leave_mm);
  62
  63 /*
  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
  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
  72  *      Now cpu0 accepts tlb flushes for the new mm.
  73  * 1a3) cpu_set(cpu, new_mm->cpu_vm_mask);
  74  *      Now the other cpus will send tlb flush ipis.
  75  * 1a4) change cr3.
  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.
  80  * 1b) thread switch without mm change
  81  *      cpu active_mm is correct, cpu0 already handles flush ipis.
  82  * 1b1) set cpu_tlbstate to TLBSTATE_OK
  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  *
  95  * The good news is that cpu_tlbstate is local to each cpu, no
  96  * write/read ordering problems.
  97  */
  98
  99 /*
 100  * TLB flush funcation:
 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.
 103  */
 104 static void flush_tlb_func(void *info)
 105 {
 106         struct flush_tlb_info *f = info;
 107
 108         inc_irq_stat(irq_tlb_count);
 109
 110         if (f->flush_mm != this_cpu_read(cpu_tlbstate.active_mm))
 111                 return;
 112         if (!f->flush_end)
 113                 f->flush_end = f->flush_start + PAGE_SIZE;
 114
 115         count_vm_tlb_event(NR_TLB_REMOTE_FLUSH_RECEIVED);
 116         if (this_cpu_read(cpu_tlbstate.state) == TLBSTATE_OK) {
 117                 if (f->flush_end == TLB_FLUSH_ALL) {
 118                         local_flush_tlb();
 119                         trace_tlb_flush(TLB_REMOTE_SHOOTDOWN, TLB_FLUSH_ALL);
 120                 } else {
 121                         unsigned long addr;
 122                         unsigned long nr_pages =
 123                                 f->flush_end - f->flush_start / PAGE_SIZE;
 124                         addr = f->flush_start;
 125                         while (addr < f->flush_end) {
 126                                 __flush_tlb_single(addr);
 127                                 addr += PAGE_SIZE;
 128                         }
 129                         trace_tlb_flush(TLB_REMOTE_SHOOTDOWN, nr_pages);
 130                 }
 131         } else
 132                 leave_mm(smp_processor_id());
 133
 134 }
 135
 136 void native_flush_tlb_others(const struct cpumask *cpumask,
 137                                  struct mm_struct *mm, unsigned long start,
 138                                  unsigned long end)
 139 {
 140         struct flush_tlb_info info;
 141         info.flush_mm = mm;
 142         info.flush_start = start;
 143         info.flush_end = end;
 144
 145         count_vm_tlb_event(NR_TLB_REMOTE_FLUSH);
 146         if (is_uv_system()) {
 147                 unsigned int cpu;
 148
 149                 cpu = smp_processor_id();
 150                 cpumask = uv_flush_tlb_others(cpumask, mm, start, end, cpu);
 151                 if (cpumask)
 152                         smp_call_function_many(cpumask, flush_tlb_func,
 153                                                                 &info, 1);
 154                 return;
 155         }
 156         smp_call_function_many(cpumask, flush_tlb_func, &info, 1);
 157 }
 158
 159 void flush_tlb_current_task(void)
 160 {
 161         struct mm_struct *mm = current->mm;
 162
 163         preempt_disable();
 164
 165         count_vm_tlb_event(NR_TLB_LOCAL_FLUSH_ALL);
 166         local_flush_tlb();
 167         trace_tlb_flush(TLB_LOCAL_SHOOTDOWN, TLB_FLUSH_ALL);
 168         if (cpumask_any_but(mm_cpumask(mm), smp_processor_id()) < nr_cpu_ids)
 169                 flush_tlb_others(mm_cpumask(mm), mm, 0UL, TLB_FLUSH_ALL);
 170         preempt_enable();
 171 }
 172
 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  */
 183 static unsigned long tlb_single_page_flush_ceiling __read_mostly = 33;
 184
 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;
 189         /* do a global flush by default */
 190         unsigned long base_pages_to_flush = TLB_FLUSH_ALL;
 191
 192         preempt_disable();
 193         if (current->active_mm != mm)
 194                 goto out;
 195
 196         if (!current->mm) {
 197                 leave_mm(smp_processor_id());
 198                 goto out;
 199         }
 200
 201         if ((end != TLB_FLUSH_ALL) && !(vmflag & VM_HUGETLB))
 202                 base_pages_to_flush = (end - start) >> PAGE_SHIFT;
 203
 204         if (base_pages_to_flush > tlb_single_page_flush_ceiling) {
 205                 base_pages_to_flush = TLB_FLUSH_ALL;
 206                 count_vm_tlb_event(NR_TLB_LOCAL_FLUSH_ALL);
 207                 local_flush_tlb();
 208         } else {
 209                 /* flush range by one by one 'invlpg' */
 210                 for (addr = start; addr < end;  addr += PAGE_SIZE) {
 211                         count_vm_tlb_event(NR_TLB_LOCAL_FLUSH_ONE);
 212                         __flush_tlb_single(addr);
 213                 }
 214         }
 215         trace_tlb_flush(TLB_LOCAL_MM_SHOOTDOWN, base_pages_to_flush);
 216 out:
 217         if (base_pages_to_flush == TLB_FLUSH_ALL) {
 218                 start = 0UL;
 219                 end = TLB_FLUSH_ALL;
 220         }
 221         if (cpumask_any_but(mm_cpumask(mm), smp_processor_id()) < nr_cpu_ids)
 222                 flush_tlb_others(mm_cpumask(mm), mm, start, end);
 223         preempt_enable();
 224 }
 225
 226 void flush_tlb_page(struct vm_area_struct *vma, unsigned long start)
 227 {
 228         struct mm_struct *mm = vma->vm_mm;
 229
 230         preempt_disable();
 231
 232         if (current->active_mm == mm) {
 233                 if (current->mm)
 234                         __flush_tlb_one(start);
 235                 else
 236                         leave_mm(smp_processor_id());
 237         }
 238
 239         if (cpumask_any_but(mm_cpumask(mm), smp_processor_id()) < nr_cpu_ids)
 240                 flush_tlb_others(mm_cpumask(mm), mm, start, 0UL);
 241
 242         preempt_enable();
 243 }
 244
 245 static void do_flush_tlb_all(void *info)
 246 {
 247         count_vm_tlb_event(NR_TLB_REMOTE_FLUSH_RECEIVED);
 248         __flush_tlb_all();
 249         if (this_cpu_read(cpu_tlbstate.state) == TLBSTATE_LAZY)
 250                 leave_mm(smp_processor_id());
 251 }
 252
 253 void flush_tlb_all(void)
 254 {
 255         count_vm_tlb_event(NR_TLB_REMOTE_FLUSH);
 256         on_each_cpu(do_flush_tlb_all, NULL, 1);
 257 }
 258
 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' */
 265         for (addr = f->flush_start; addr < f->flush_end; addr += PAGE_SIZE)
 266                 __flush_tlb_single(addr);
 267 }
 268
 269 void flush_tlb_kernel_range(unsigned long start, unsigned long end)
 270 {
 271
 272         /* Balance as user space task's flush, a bit conservative */
 273         if (end == TLB_FLUSH_ALL ||
 274             (end - start) > tlb_single_page_flush_ceiling * PAGE_SIZE) {
 275                 on_each_cpu(do_flush_tlb_all, NULL, 1);
 276         } else {
 277                 struct flush_tlb_info info;
 278                 info.flush_start = start;
 279                 info.flush_end = end;
 280                 on_each_cpu(do_kernel_range_flush, &info, 1);
 281         }
 282 }
 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);