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 #ifdef CONFIG_SMP
  32
  33 struct flush_tlb_info {
  34         struct mm_struct *flush_mm;
  35         unsigned long flush_start;
  36         unsigned long flush_end;
  37 };
  38
  39 /*
  40  * We cannot call mmdrop() because we are in interrupt context,
  41  * instead update mm->cpu_vm_mask.
  42  */
  43 void leave_mm(int cpu)
  44 {
  45         struct mm_struct *active_mm = this_cpu_read(cpu_tlbstate.active_mm);
  46         if (this_cpu_read(cpu_tlbstate.state) == TLBSTATE_OK)
  47                 BUG();
  48         if (cpumask_test_cpu(cpu, mm_cpumask(active_mm))) {
  49                 cpumask_clear_cpu(cpu, mm_cpumask(active_mm));
  50                 load_cr3(swapper_pg_dir);
  51                 /*
  52                  * This gets called in the idle path where RCU
  53                  * functions differently.  Tracing normally
  54                  * uses RCU, so we have to call the tracepoint
  55                  * specially here.
  56                  */
  57                 trace_tlb_flush_rcuidle(TLB_FLUSH_ON_TASK_SWITCH, TLB_FLUSH_ALL);
  58         }
  59 }
  60 EXPORT_SYMBOL_GPL(leave_mm);
  61
  62 /*
  63  * The flush IPI assumes that a thread switch happens in this order:
  64  * [cpu0: the cpu that switches]
  65  * 1) switch_mm() either 1a) or 1b)
  66  * 1a) thread switch to a different mm
  67  * 1a1) set cpu_tlbstate to TLBSTATE_OK
  68  *      Now the tlb flush NMI handler flush_tlb_func won't call leave_mm
  69  *      if cpu0 was in lazy tlb mode.
  70  * 1a2) update cpu active_mm
  71  *      Now cpu0 accepts tlb flushes for the new mm.
  72  * 1a3) cpu_set(cpu, new_mm->cpu_vm_mask);
  73  *      Now the other cpus will send tlb flush ipis.
  74  * 1a4) change cr3.
  75  * 1a5) cpu_clear(cpu, old_mm->cpu_vm_mask);
  76  *      Stop ipi delivery for the old mm. This is not synchronized with
  77  *      the other cpus, but flush_tlb_func ignore flush ipis for the wrong
  78  *      mm, and in the worst case we perform a superfluous tlb flush.
  79  * 1b) thread switch without mm change
  80  *      cpu active_mm is correct, cpu0 already handles flush ipis.
  81  * 1b1) set cpu_tlbstate to TLBSTATE_OK
  82  * 1b2) test_and_set the cpu bit in cpu_vm_mask.
  83  *      Atomically set the bit [other cpus will start sending flush ipis],
  84  *      and test the bit.
  85  * 1b3) if the bit was 0: leave_mm was called, flush the tlb.
  86  * 2) switch %%esp, ie current
  87  *
  88  * The interrupt must handle 2 special cases:
  89  * - cr3 is changed before %%esp, ie. it cannot use current->{active_,}mm.
  90  * - the cpu performs speculative tlb reads, i.e. even if the cpu only
  91  *   runs in kernel space, the cpu could load tlb entries for user space
  92  *   pages.
  93  *
  94  * The good news is that cpu_tlbstate is local to each cpu, no
  95  * write/read ordering problems.
  96  */
  97
  98 /*
  99  * TLB flush funcation:
 100  * 1) Flush the tlb entries if the cpu uses the mm that's being flushed.
 101  * 2) Leave the mm if we are in the lazy tlb mode.
 102  */
 103 static void flush_tlb_func(void *info)
 104 {
 105         struct flush_tlb_info *f = info;
 106
 107         inc_irq_stat(irq_tlb_count);
 108
 109         if (f->flush_mm && f->flush_mm != this_cpu_read(cpu_tlbstate.active_mm))
 110                 return;
 111
 112         count_vm_tlb_event(NR_TLB_REMOTE_FLUSH_RECEIVED);
 113         if (this_cpu_read(cpu_tlbstate.state) == TLBSTATE_OK) {
 114                 if (f->flush_end == TLB_FLUSH_ALL) {
 115                         local_flush_tlb();
 116                         trace_tlb_flush(TLB_REMOTE_SHOOTDOWN, TLB_FLUSH_ALL);
 117                 } else {
 118                         unsigned long addr;
 119                         unsigned long nr_pages =
 120                                 (f->flush_end - f->flush_start) / PAGE_SIZE;
 121                         addr = f->flush_start;
 122                         while (addr < f->flush_end) {
 123                                 __flush_tlb_single(addr);
 124                                 addr += PAGE_SIZE;
 125                         }
 126                         trace_tlb_flush(TLB_REMOTE_SHOOTDOWN, nr_pages);
 127                 }
 128         } else
 129                 leave_mm(smp_processor_id());
 130
 131 }
 132
 133 void native_flush_tlb_others(const struct cpumask *cpumask,
 134                                  struct mm_struct *mm, unsigned long start,
 135                                  unsigned long end)
 136 {
 137         struct flush_tlb_info info;
 138
 139         if (end == 0)
 140                 end = start + PAGE_SIZE;
 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 (end == TLB_FLUSH_ALL)
 147                 trace_tlb_flush(TLB_REMOTE_SEND_IPI, TLB_FLUSH_ALL);
 148         else
 149                 trace_tlb_flush(TLB_REMOTE_SEND_IPI,
 150                                 (end - start) >> PAGE_SHIFT);
 151
 152         if (is_uv_system()) {
 153                 unsigned int cpu;
 154
 155                 cpu = smp_processor_id();
 156                 cpumask = uv_flush_tlb_others(cpumask, mm, start, end, cpu);
 157                 if (cpumask)
 158                         smp_call_function_many(cpumask, flush_tlb_func,
 159                                                                 &info, 1);
 160                 return;
 161         }
 162         smp_call_function_many(cpumask, flush_tlb_func, &info, 1);
 163 }
 164
 165 void flush_tlb_current_task(void)
 166 {
 167         struct mm_struct *mm = current->mm;
 168
 169         preempt_disable();
 170
 171         count_vm_tlb_event(NR_TLB_LOCAL_FLUSH_ALL);
 172
 173         /* This is an implicit full barrier that synchronizes with switch_mm. */
 174         local_flush_tlb();
 175
 176         trace_tlb_flush(TLB_LOCAL_SHOOTDOWN, TLB_FLUSH_ALL);
 177         if (cpumask_any_but(mm_cpumask(mm), smp_processor_id()) < nr_cpu_ids)
 178                 flush_tlb_others(mm_cpumask(mm), mm, 0UL, TLB_FLUSH_ALL);
 179         preempt_enable();
 180 }
 181
 182 /*
 183  * See Documentation/x86/tlb.txt for details.  We choose 33
 184  * because it is large enough to cover the vast majority (at
 185  * least 95%) of allocations, and is small enough that we are
 186  * confident it will not cause too much overhead.  Each single
 187  * flush is about 100 ns, so this caps the maximum overhead at
 188  * _about_ 3,000 ns.
 189  *
 190  * This is in units of pages.
 191  */
 192 static unsigned long tlb_single_page_flush_ceiling __read_mostly = 33;
 193
 194 void flush_tlb_mm_range(struct mm_struct *mm, unsigned long start,
 195                                 unsigned long end, unsigned long vmflag)
 196 {
 197         unsigned long addr;
 198         /* do a global flush by default */
 199         unsigned long base_pages_to_flush = TLB_FLUSH_ALL;
 200
 201         preempt_disable();
 202         if (current->active_mm != mm) {
 203                 /* Synchronize with switch_mm. */
 204                 smp_mb();
 205
 206                 goto out;
 207         }
 208
 209         if (!current->mm) {
 210                 leave_mm(smp_processor_id());
 211
 212                 /* Synchronize with switch_mm. */
 213                 smp_mb();
 214
 215                 goto out;
 216         }
 217
 218         if ((end != TLB_FLUSH_ALL) && !(vmflag & VM_HUGETLB))
 219                 base_pages_to_flush = (end - start) >> PAGE_SHIFT;
 220
 221         /*
 222          * Both branches below are implicit full barriers (MOV to CR or
 223          * INVLPG) that synchronize with switch_mm.
 224          */
 225         if (base_pages_to_flush > tlb_single_page_flush_ceiling) {
 226                 base_pages_to_flush = TLB_FLUSH_ALL;
 227                 count_vm_tlb_event(NR_TLB_LOCAL_FLUSH_ALL);
 228                 local_flush_tlb();
 229         } else {
 230                 /* flush range by one by one 'invlpg' */
 231                 for (addr = start; addr < end;  addr += PAGE_SIZE) {
 232                         count_vm_tlb_event(NR_TLB_LOCAL_FLUSH_ONE);
 233                         __flush_tlb_single(addr);
 234                 }
 235         }
 236         trace_tlb_flush(TLB_LOCAL_MM_SHOOTDOWN, base_pages_to_flush);
 237 out:
 238         if (base_pages_to_flush == TLB_FLUSH_ALL) {
 239                 start = 0UL;
 240                 end = TLB_FLUSH_ALL;
 241         }
 242         if (cpumask_any_but(mm_cpumask(mm), smp_processor_id()) < nr_cpu_ids)
 243                 flush_tlb_others(mm_cpumask(mm), mm, start, end);
 244         preempt_enable();
 245 }
 246
 247 void flush_tlb_page(struct vm_area_struct *vma, unsigned long start)
 248 {
 249         struct mm_struct *mm = vma->vm_mm;
 250
 251         preempt_disable();
 252
 253         if (current->active_mm == mm) {
 254                 if (current->mm) {
 255                         /*
 256                          * Implicit full barrier (INVLPG) that synchronizes
 257                          * with switch_mm.
 258                          */
 259                         __flush_tlb_one(start);
 260                 } else {
 261                         leave_mm(smp_processor_id());
 262
 263                         /* Synchronize with switch_mm. */
 264                         smp_mb();
 265                 }
 266         }
 267
 268         if (cpumask_any_but(mm_cpumask(mm), smp_processor_id()) < nr_cpu_ids)
 269                 flush_tlb_others(mm_cpumask(mm), mm, start, 0UL);
 270
 271         preempt_enable();
 272 }
 273
 274 static void do_flush_tlb_all(void *info)
 275 {
 276         count_vm_tlb_event(NR_TLB_REMOTE_FLUSH_RECEIVED);
 277         __flush_tlb_all();
 278         if (this_cpu_read(cpu_tlbstate.state) == TLBSTATE_LAZY)
 279                 leave_mm(smp_processor_id());
 280 }
 281
 282 void flush_tlb_all(void)
 283 {
 284         count_vm_tlb_event(NR_TLB_REMOTE_FLUSH);
 285         on_each_cpu(do_flush_tlb_all, NULL, 1);
 286 }
 287
 288 static void do_kernel_range_flush(void *info)
 289 {
 290         struct flush_tlb_info *f = info;
 291         unsigned long addr;
 292
 293         /* flush range by one by one 'invlpg' */
 294         for (addr = f->flush_start; addr < f->flush_end; addr += PAGE_SIZE)
 295                 __flush_tlb_single(addr);
 296 }
 297
 298 void flush_tlb_kernel_range(unsigned long start, unsigned long end)
 299 {
 300
 301         /* Balance as user space task's flush, a bit conservative */
 302         if (end == TLB_FLUSH_ALL ||
 303             (end - start) > tlb_single_page_flush_ceiling * PAGE_SIZE) {
 304                 on_each_cpu(do_flush_tlb_all, NULL, 1);
 305         } else {
 306                 struct flush_tlb_info info;
 307                 info.flush_start = start;
 308                 info.flush_end = end;
 309                 on_each_cpu(do_kernel_range_flush, &info, 1);
 310         }
 311 }
 312
 313 static ssize_t tlbflush_read_file(struct file *file, char __user *user_buf,
 314                              size_t count, loff_t *ppos)
 315 {
 316         char buf[32];
 317         unsigned int len;
 318
 319         len = sprintf(buf, "%ld\n", tlb_single_page_flush_ceiling);
 320         return simple_read_from_buffer(user_buf, count, ppos, buf, len);
 321 }
 322
 323 static ssize_t tlbflush_write_file(struct file *file,
 324                  const char __user *user_buf, size_t count, loff_t *ppos)
 325 {
 326         char buf[32];
 327         ssize_t len;
 328         int ceiling;
 329
 330         len = min(count, sizeof(buf) - 1);
 331         if (copy_from_user(buf, user_buf, len))
 332                 return -EFAULT;
 333
 334         buf[len] = '\0';
 335         if (kstrtoint(buf, 0, &ceiling))
 336                 return -EINVAL;
 337
 338         if (ceiling < 0)
 339                 return -EINVAL;
 340
 341         tlb_single_page_flush_ceiling = ceiling;
 342         return count;
 343 }
 344
 345 static const struct file_operations fops_tlbflush = {
 346         .read = tlbflush_read_file,
 347         .write = tlbflush_write_file,
 348         .llseek = default_llseek,
 349 };
 350
 351 static int __init create_tlb_single_page_flush_ceiling(void)
 352 {
 353         debugfs_create_file("tlb_single_page_flush_ceiling", S_IRUSR | S_IWUSR,
 354                             arch_debugfs_dir, NULL, &fops_tlbflush);
 355         return 0;
 356 }
 357 late_initcall(create_tlb_single_page_flush_ceiling);
 358
 359 #endif /* CONFIG_SMP */