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 }
  54 EXPORT_SYMBOL_GPL(leave_mm);
  55
  56 /*
  57  * The flush IPI assumes that a thread switch happens in this order:
  58  * [cpu0: the cpu that switches]
  59  * 1) switch_mm() either 1a) or 1b)
  60  * 1a) thread switch to a different mm
  61  * 1a1) set cpu_tlbstate to TLBSTATE_OK
  62  *      Now the tlb flush NMI handler flush_tlb_func won't call leave_mm
  63  *      if cpu0 was in lazy tlb mode.
  64  * 1a2) update cpu active_mm
  65  *      Now cpu0 accepts tlb flushes for the new mm.
  66  * 1a3) cpu_set(cpu, new_mm->cpu_vm_mask);
  67  *      Now the other cpus will send tlb flush ipis.
  68  * 1a4) change cr3.
  69  * 1a5) cpu_clear(cpu, old_mm->cpu_vm_mask);
  70  *      Stop ipi delivery for the old mm. This is not synchronized with
  71  *      the other cpus, but flush_tlb_func ignore flush ipis for the wrong
  72  *      mm, and in the worst case we perform a superfluous tlb flush.
  73  * 1b) thread switch without mm change
  74  *      cpu active_mm is correct, cpu0 already handles flush ipis.
  75  * 1b1) set cpu_tlbstate to TLBSTATE_OK
  76  * 1b2) test_and_set the cpu bit in cpu_vm_mask.
  77  *      Atomically set the bit [other cpus will start sending flush ipis],
  78  *      and test the bit.
  79  * 1b3) if the bit was 0: leave_mm was called, flush the tlb.
  80  * 2) switch %%esp, ie current
  81  *
  82  * The interrupt must handle 2 special cases:
  83  * - cr3 is changed before %%esp, ie. it cannot use current->{active_,}mm.
  84  * - the cpu performs speculative tlb reads, i.e. even if the cpu only
  85  *   runs in kernel space, the cpu could load tlb entries for user space
  86  *   pages.
  87  *
  88  * The good news is that cpu_tlbstate is local to each cpu, no
  89  * write/read ordering problems.
  90  */
  91
  92 /*
  93  * TLB flush funcation:
  94  * 1) Flush the tlb entries if the cpu uses the mm that's being flushed.
  95  * 2) Leave the mm if we are in the lazy tlb mode.
  96  */
  97 static void flush_tlb_func(void *info)
  98 {
  99         struct flush_tlb_info *f = info;
 100
 101         inc_irq_stat(irq_tlb_count);
 102
 103         if (f->flush_mm != this_cpu_read(cpu_tlbstate.active_mm))
 104                 return;
 105
 106         if (this_cpu_read(cpu_tlbstate.state) == TLBSTATE_OK) {
 107                 if (f->flush_end == TLB_FLUSH_ALL)
 108                         local_flush_tlb();
 109                 else if (!f->flush_end)
 110                         __flush_tlb_single(f->flush_start);
 111                 else {
 112                         unsigned long addr;
 113                         addr = f->flush_start;
 114                         while (addr < f->flush_end) {
 115                                 __flush_tlb_single(addr);
 116                                 addr += PAGE_SIZE;
 117                         }
 118                 }
 119         } else
 120                 leave_mm(smp_processor_id());
 121
 122 }
 123
 124 void native_flush_tlb_others(const struct cpumask *cpumask,
 125                                  struct mm_struct *mm, unsigned long start,
 126                                  unsigned long end)
 127 {
 128         struct flush_tlb_info info;
 129         info.flush_mm = mm;
 130         info.flush_start = start;
 131         info.flush_end = end;
 132
 133         if (is_uv_system()) {
 134                 unsigned int cpu;
 135
 136                 cpu = smp_processor_id();
 137                 cpumask = uv_flush_tlb_others(cpumask, mm, start, end, cpu);
 138                 if (cpumask)
 139                         smp_call_function_many(cpumask, flush_tlb_func,
 140                                                                 &info, 1);
 141                 return;
 142         }
 143         smp_call_function_many(cpumask, flush_tlb_func, &info, 1);
 144 }
 145
 146 void flush_tlb_current_task(void)
 147 {
 148         struct mm_struct *mm = current->mm;
 149
 150         preempt_disable();
 151
 152         /* This is an implicit full barrier that synchronizes with switch_mm. */
 153         local_flush_tlb();
 154
 155         if (cpumask_any_but(mm_cpumask(mm), smp_processor_id()) < nr_cpu_ids)
 156                 flush_tlb_others(mm_cpumask(mm), mm, 0UL, TLB_FLUSH_ALL);
 157         preempt_enable();
 158 }
 159
 160 /*
 161  * It can find out the THP large page, or
 162  * HUGETLB page in tlb_flush when THP disabled
 163  */
 164 static inline unsigned long has_large_page(struct mm_struct *mm,
 165                                  unsigned long start, unsigned long end)
 166 {
 167         pgd_t *pgd;
 168         pud_t *pud;
 169         pmd_t *pmd;
 170         unsigned long addr = ALIGN(start, HPAGE_SIZE);
 171         for (; addr < end; addr += HPAGE_SIZE) {
 172                 pgd = pgd_offset(mm, addr);
 173                 if (likely(!pgd_none(*pgd))) {
 174                         pud = pud_offset(pgd, addr);
 175                         if (likely(!pud_none(*pud))) {
 176                                 pmd = pmd_offset(pud, addr);
 177                                 if (likely(!pmd_none(*pmd)))
 178                                         if (pmd_large(*pmd))
 179                                                 return addr;
 180                         }
 181                 }
 182         }
 183         return 0;
 184 }
 185
 186 void flush_tlb_mm_range(struct mm_struct *mm, unsigned long start,
 187                                 unsigned long end, unsigned long vmflag)
 188 {
 189         unsigned long addr;
 190         unsigned act_entries, tlb_entries = 0;
 191
 192         preempt_disable();
 193         if (current->active_mm != mm) {
 194                 /* Synchronize with switch_mm. */
 195                 smp_mb();
 196
 197                 goto flush_all;
 198         }
 199
 200         if (!current->mm) {
 201                 leave_mm(smp_processor_id());
 202
 203                 /* Synchronize with switch_mm. */
 204                 smp_mb();
 205
 206                 goto flush_all;
 207         }
 208
 209         if (end == TLB_FLUSH_ALL || tlb_flushall_shift == -1
 210                                         || vmflag & VM_HUGETLB) {
 211                 local_flush_tlb();
 212                 goto flush_all;
 213         }
 214
 215         /* In modern CPU, last level tlb used for both data/ins */
 216         if (vmflag & VM_EXEC)
 217                 tlb_entries = tlb_lli_4k[ENTRIES];
 218         else
 219                 tlb_entries = tlb_lld_4k[ENTRIES];
 220         /* Assume all of TLB entries was occupied by this task */
 221         act_entries = mm->total_vm > tlb_entries ? tlb_entries : mm->total_vm;
 222
 223         /* tlb_flushall_shift is on balance point, details in commit log */
 224         if ((end - start) >> PAGE_SHIFT > act_entries >> tlb_flushall_shift)
 225                 local_flush_tlb();
 226         else {
 227                 if (has_large_page(mm, start, end)) {
 228                         local_flush_tlb();
 229                         goto flush_all;
 230                 }
 231                 /* flush range by one by one 'invlpg' */
 232                 for (addr = start; addr < end;  addr += PAGE_SIZE)
 233                         __flush_tlb_single(addr);
 234
 235                 if (cpumask_any_but(mm_cpumask(mm),
 236                                 smp_processor_id()) < nr_cpu_ids)
 237                         flush_tlb_others(mm_cpumask(mm), mm, start, end);
 238                 preempt_enable();
 239                 return;
 240         }
 241
 242 flush_all:
 243         if (cpumask_any_but(mm_cpumask(mm), smp_processor_id()) < nr_cpu_ids)
 244                 flush_tlb_others(mm_cpumask(mm), mm, 0UL, TLB_FLUSH_ALL);
 245         preempt_enable();
 246 }
 247
 248 void flush_tlb_page(struct vm_area_struct *vma, unsigned long start)
 249 {
 250         struct mm_struct *mm = vma->vm_mm;
 251
 252         preempt_disable();
 253
 254         if (current->active_mm == mm) {
 255                 if (current->mm) {
 256                         /*
 257                          * Implicit full barrier (INVLPG) that synchronizes
 258                          * with switch_mm.
 259                          */
 260                         __flush_tlb_one(start);
 261                 } else {
 262                         leave_mm(smp_processor_id());
 263
 264                         /* Synchronize with switch_mm. */
 265                         smp_mb();
 266                 }
 267         }
 268
 269         if (cpumask_any_but(mm_cpumask(mm), smp_processor_id()) < nr_cpu_ids)
 270                 flush_tlb_others(mm_cpumask(mm), mm, start, 0UL);
 271
 272         preempt_enable();
 273 }
 274
 275 static void do_flush_tlb_all(void *info)
 276 {
 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         on_each_cpu(do_flush_tlb_all, NULL, 1);
 285 }
 286
 287 static void do_kernel_range_flush(void *info)
 288 {
 289         struct flush_tlb_info *f = info;
 290         unsigned long addr;
 291
 292         /* flush range by one by one 'invlpg' */
 293         for (addr = f->flush_start; addr < f->flush_end; addr += PAGE_SIZE)
 294                 __flush_tlb_single(addr);
 295 }
 296
 297 void flush_tlb_kernel_range(unsigned long start, unsigned long end)
 298 {
 299         unsigned act_entries;
 300         struct flush_tlb_info info;
 301
 302         /* In modern CPU, last level tlb used for both data/ins */
 303         act_entries = tlb_lld_4k[ENTRIES];
 304
 305         /* Balance as user space task's flush, a bit conservative */
 306         if (end == TLB_FLUSH_ALL || tlb_flushall_shift == -1 ||
 307                 (end - start) >> PAGE_SHIFT > act_entries >> tlb_flushall_shift)
 308
 309                 on_each_cpu(do_flush_tlb_all, NULL, 1);
 310         else {
 311                 info.flush_start = start;
 312                 info.flush_end = end;
 313                 on_each_cpu(do_kernel_range_flush, &info, 1);
 314         }
 315 }
 316
 317 #ifdef CONFIG_DEBUG_TLBFLUSH
 318 static ssize_t tlbflush_read_file(struct file *file, char __user *user_buf,
 319                              size_t count, loff_t *ppos)
 320 {
 321         char buf[32];
 322         unsigned int len;
 323
 324         len = sprintf(buf, "%hd\n", tlb_flushall_shift);
 325         return simple_read_from_buffer(user_buf, count, ppos, buf, len);
 326 }
 327
 328 static ssize_t tlbflush_write_file(struct file *file,
 329                  const char __user *user_buf, size_t count, loff_t *ppos)
 330 {
 331         char buf[32];
 332         ssize_t len;
 333         s8 shift;
 334
 335         len = min(count, sizeof(buf) - 1);
 336         if (copy_from_user(buf, user_buf, len))
 337                 return -EFAULT;
 338
 339         buf[len] = '\0';
 340         if (kstrtos8(buf, 0, &shift))
 341                 return -EINVAL;
 342
 343         if (shift < -1 || shift >= BITS_PER_LONG)
 344                 return -EINVAL;
 345
 346         tlb_flushall_shift = shift;
 347         return count;
 348 }
 349
 350 static const struct file_operations fops_tlbflush = {
 351         .read = tlbflush_read_file,
 352         .write = tlbflush_write_file,
 353         .llseek = default_llseek,
 354 };
 355
 356 static int __init create_tlb_flushall_shift(void)
 357 {
 358         debugfs_create_file("tlb_flushall_shift", S_IRUSR | S_IWUSR,
 359                             arch_debugfs_dir, NULL, &fops_tlbflush);
 360         return 0;
 361 }
 362 late_initcall(create_tlb_flushall_shift);
 363 #endif