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         if (f->flush_mm != this_cpu_read(cpu_tlbstate.active_mm))
 102                 return;
 103
 104         if (this_cpu_read(cpu_tlbstate.state) == TLBSTATE_OK) {
 105                 if (f->flush_end == TLB_FLUSH_ALL || !cpu_has_invlpg)
 106                         local_flush_tlb();
 107                 else if (!f->flush_end)
 108                         __flush_tlb_single(f->flush_start);
 109                 else {
 110                         unsigned long addr;
 111                         addr = f->flush_start;
 112                         while (addr < f->flush_end) {
 113                                 __flush_tlb_single(addr);
 114                                 addr += PAGE_SIZE;
 115                         }
 116                 }
 117         } else
 118                 leave_mm(smp_processor_id());
 119
 120 }
 121
 122 void native_flush_tlb_others(const struct cpumask *cpumask,
 123                                  struct mm_struct *mm, unsigned long start,
 124                                  unsigned long end)
 125 {
 126         struct flush_tlb_info info;
 127         info.flush_mm = mm;
 128         info.flush_start = start;
 129         info.flush_end = end;
 130
 131         if (is_uv_system()) {
 132                 unsigned int cpu;
 133
 134                 cpu = smp_processor_id();
 135                 cpumask = uv_flush_tlb_others(cpumask, mm, start, end, cpu);
 136                 if (cpumask)
 137                         smp_call_function_many(cpumask, flush_tlb_func,
 138                                                                 &info, 1);
 139                 return;
 140         }
 141         smp_call_function_many(cpumask, flush_tlb_func, &info, 1);
 142 }
 143
 144 void flush_tlb_current_task(void)
 145 {
 146         struct mm_struct *mm = current->mm;
 147
 148         preempt_disable();
 149
 150         local_flush_tlb();
 151         if (cpumask_any_but(mm_cpumask(mm), smp_processor_id()) < nr_cpu_ids)
 152                 flush_tlb_others(mm_cpumask(mm), mm, 0UL, TLB_FLUSH_ALL);
 153         preempt_enable();
 154 }
 155
 156 /*
 157  * It can find out the THP large page, or
 158  * HUGETLB page in tlb_flush when THP disabled
 159  */
 160 static inline unsigned long has_large_page(struct mm_struct *mm,
 161                                  unsigned long start, unsigned long end)
 162 {
 163         pgd_t *pgd;
 164         pud_t *pud;
 165         pmd_t *pmd;
 166         unsigned long addr = ALIGN(start, HPAGE_SIZE);
 167         for (; addr < end; addr += HPAGE_SIZE) {
 168                 pgd = pgd_offset(mm, addr);
 169                 if (likely(!pgd_none(*pgd))) {
 170                         pud = pud_offset(pgd, addr);
 171                         if (likely(!pud_none(*pud))) {
 172                                 pmd = pmd_offset(pud, addr);
 173                                 if (likely(!pmd_none(*pmd)))
 174                                         if (pmd_large(*pmd))
 175                                                 return addr;
 176                         }
 177                 }
 178         }
 179         return 0;
 180 }
 181
 182 void flush_tlb_mm_range(struct mm_struct *mm, unsigned long start,
 183                                 unsigned long end, unsigned long vmflag)
 184 {
 185         unsigned long addr;
 186         unsigned act_entries, tlb_entries = 0;
 187
 188         preempt_disable();
 189         if (current->active_mm != mm)
 190                 goto flush_all;
 191
 192         if (!current->mm) {
 193                 leave_mm(smp_processor_id());
 194                 goto flush_all;
 195         }
 196
 197         if (end == TLB_FLUSH_ALL || tlb_flushall_shift == -1
 198                                         || vmflag == VM_HUGETLB) {
 199                 local_flush_tlb();
 200                 goto flush_all;
 201         }
 202
 203         /* In modern CPU, last level tlb used for both data/ins */
 204         if (vmflag & VM_EXEC)
 205                 tlb_entries = tlb_lli_4k[ENTRIES];
 206         else
 207                 tlb_entries = tlb_lld_4k[ENTRIES];
 208         /* Assume all of TLB entries was occupied by this task */
 209         act_entries = mm->total_vm > tlb_entries ? tlb_entries : mm->total_vm;
 210
 211         /* tlb_flushall_shift is on balance point, details in commit log */
 212         if ((end - start) >> PAGE_SHIFT > act_entries >> tlb_flushall_shift)
 213                 local_flush_tlb();
 214         else {
 215                 if (has_large_page(mm, start, end)) {
 216                         local_flush_tlb();
 217                         goto flush_all;
 218                 }
 219                 /* flush range by one by one 'invlpg' */
 220                 for (addr = start; addr < end;  addr += PAGE_SIZE)
 221                         __flush_tlb_single(addr);
 222
 223                 if (cpumask_any_but(mm_cpumask(mm),
 224                                 smp_processor_id()) < nr_cpu_ids)
 225                         flush_tlb_others(mm_cpumask(mm), mm, start, end);
 226                 preempt_enable();
 227                 return;
 228         }
 229
 230 flush_all:
 231         if (cpumask_any_but(mm_cpumask(mm), smp_processor_id()) < nr_cpu_ids)
 232                 flush_tlb_others(mm_cpumask(mm), mm, 0UL, TLB_FLUSH_ALL);
 233         preempt_enable();
 234 }
 235
 236 void flush_tlb_page(struct vm_area_struct *vma, unsigned long start)
 237 {
 238         struct mm_struct *mm = vma->vm_mm;
 239
 240         preempt_disable();
 241
 242         if (current->active_mm == mm) {
 243                 if (current->mm)
 244                         __flush_tlb_one(start);
 245                 else
 246                         leave_mm(smp_processor_id());
 247         }
 248
 249         if (cpumask_any_but(mm_cpumask(mm), smp_processor_id()) < nr_cpu_ids)
 250                 flush_tlb_others(mm_cpumask(mm), mm, start, 0UL);
 251
 252         preempt_enable();
 253 }
 254
 255 static void do_flush_tlb_all(void *info)
 256 {
 257         __flush_tlb_all();
 258         if (this_cpu_read(cpu_tlbstate.state) == TLBSTATE_LAZY)
 259                 leave_mm(smp_processor_id());
 260 }
 261
 262 void flush_tlb_all(void)
 263 {
 264         on_each_cpu(do_flush_tlb_all, NULL, 1);
 265 }
 266
 267 #ifdef CONFIG_DEBUG_TLBFLUSH
 268 static ssize_t tlbflush_read_file(struct file *file, char __user *user_buf,
 269                              size_t count, loff_t *ppos)
 270 {
 271         char buf[32];
 272         unsigned int len;
 273
 274         len = sprintf(buf, "%hd\n", tlb_flushall_shift);
 275         return simple_read_from_buffer(user_buf, count, ppos, buf, len);
 276 }
 277
 278 static ssize_t tlbflush_write_file(struct file *file,
 279                  const char __user *user_buf, size_t count, loff_t *ppos)
 280 {
 281         char buf[32];
 282         ssize_t len;
 283         s8 shift;
 284
 285         len = min(count, sizeof(buf) - 1);
 286         if (copy_from_user(buf, user_buf, len))
 287                 return -EFAULT;
 288
 289         buf[len] = '\0';
 290         if (kstrtos8(buf, 0, &shift))
 291                 return -EINVAL;
 292
 293         if (shift > 64)
 294                 return -EINVAL;
 295
 296         tlb_flushall_shift = shift;
 297         return count;
 298 }
 299
 300 static const struct file_operations fops_tlbflush = {
 301         .read = tlbflush_read_file,
 302         .write = tlbflush_write_file,
 303         .llseek = default_llseek,
 304 };
 305
 306 static int __cpuinit create_tlb_flushall_shift(void)
 307 {
 308         if (cpu_has_invlpg) {
 309                 debugfs_create_file("tlb_flushall_shift", S_IRUSR | S_IWUSR,
 310                         arch_debugfs_dir, NULL, &fops_tlbflush);
 311         }
 312         return 0;
 313 }
 314 late_initcall(create_tlb_flushall_shift);
 315 #endif