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/export.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  *      TLB flushing, formerly SMP-only
  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 void leave_mm(int cpu)
  32 {
  33         struct mm_struct *loaded_mm = this_cpu_read(cpu_tlbstate.loaded_mm);
  34
  35         /*
  36          * It's plausible that we're in lazy TLB mode while our mm is init_mm.
  37          * If so, our callers still expect us to flush the TLB, but there
  38          * aren't any user TLB entries in init_mm to worry about.
  39          *
  40          * This needs to happen before any other sanity checks due to
  41          * intel_idle's shenanigans.
  42          */
  43         if (loaded_mm == &init_mm)
  44                 return;
  45
  46         if (this_cpu_read(cpu_tlbstate.state) == TLBSTATE_OK)
  47                 BUG();
  48
  49         switch_mm(NULL, &init_mm, NULL);
  50 }
  51 EXPORT_SYMBOL_GPL(leave_mm);
  52
  53 void switch_mm(struct mm_struct *prev, struct mm_struct *next,
  54                struct task_struct *tsk)
  55 {
  56         unsigned long flags;
  57
  58         local_irq_save(flags);
  59         switch_mm_irqs_off(prev, next, tsk);
  60         local_irq_restore(flags);
  61 }
  62
  63 void switch_mm_irqs_off(struct mm_struct *prev, struct mm_struct *next,
  64                         struct task_struct *tsk)
  65 {
  66         unsigned cpu = smp_processor_id();
  67         struct mm_struct *real_prev = this_cpu_read(cpu_tlbstate.loaded_mm);
  68
  69         /*
  70          * NB: The scheduler will call us with prev == next when
  71          * switching from lazy TLB mode to normal mode if active_mm
  72          * isn't changing.  When this happens, there is no guarantee
  73          * that CR3 (and hence cpu_tlbstate.loaded_mm) matches next.
  74          *
  75          * NB: leave_mm() calls us with prev == NULL and tsk == NULL.
  76          */
  77
  78         this_cpu_write(cpu_tlbstate.state, TLBSTATE_OK);
  79
  80         if (real_prev == next) {
  81                 /*
  82                  * There's nothing to do: we always keep the per-mm control
  83                  * regs in sync with cpu_tlbstate.loaded_mm.  Just
  84                  * sanity-check mm_cpumask.
  85                  */
  86                 if (WARN_ON_ONCE(!cpumask_test_cpu(cpu, mm_cpumask(next))))
  87                         cpumask_set_cpu(cpu, mm_cpumask(next));
  88                 return;
  89         }
  90
  91         if (IS_ENABLED(CONFIG_VMAP_STACK)) {
  92                 /*
  93                  * If our current stack is in vmalloc space and isn't
  94                  * mapped in the new pgd, we'll double-fault.  Forcibly
  95                  * map it.
  96                  */
  97                 unsigned int stack_pgd_index = pgd_index(current_stack_pointer());
  98
  99                 pgd_t *pgd = next->pgd + stack_pgd_index;
 100
 101                 if (unlikely(pgd_none(*pgd)))
 102                         set_pgd(pgd, init_mm.pgd[stack_pgd_index]);
 103         }
 104
 105         this_cpu_write(cpu_tlbstate.loaded_mm, next);
 106
 107         WARN_ON_ONCE(cpumask_test_cpu(cpu, mm_cpumask(next)));
 108         cpumask_set_cpu(cpu, mm_cpumask(next));
 109
 110         /*
 111          * Re-load page tables.
 112          *
 113          * This logic has an ordering constraint:
 114          *
 115          *  CPU 0: Write to a PTE for 'next'
 116          *  CPU 0: load bit 1 in mm_cpumask.  if nonzero, send IPI.
 117          *  CPU 1: set bit 1 in next's mm_cpumask
 118          *  CPU 1: load from the PTE that CPU 0 writes (implicit)
 119          *
 120          * We need to prevent an outcome in which CPU 1 observes
 121          * the new PTE value and CPU 0 observes bit 1 clear in
 122          * mm_cpumask.  (If that occurs, then the IPI will never
 123          * be sent, and CPU 0's TLB will contain a stale entry.)
 124          *
 125          * The bad outcome can occur if either CPU's load is
 126          * reordered before that CPU's store, so both CPUs must
 127          * execute full barriers to prevent this from happening.
 128          *
 129          * Thus, switch_mm needs a full barrier between the
 130          * store to mm_cpumask and any operation that could load
 131          * from next->pgd.  TLB fills are special and can happen
 132          * due to instruction fetches or for no reason at all,
 133          * and neither LOCK nor MFENCE orders them.
 134          * Fortunately, load_cr3() is serializing and gives the
 135          * ordering guarantee we need.
 136          */
 137         load_cr3(next->pgd);
 138
 139         /*
 140          * This gets called via leave_mm() in the idle path where RCU
 141          * functions differently.  Tracing normally uses RCU, so we have to
 142          * call the tracepoint specially here.
 143          */
 144         trace_tlb_flush_rcuidle(TLB_FLUSH_ON_TASK_SWITCH, TLB_FLUSH_ALL);
 145
 146         /* Stop flush ipis for the previous mm */
 147         WARN_ON_ONCE(!cpumask_test_cpu(cpu, mm_cpumask(real_prev)) &&
 148                      real_prev != &init_mm);
 149         cpumask_clear_cpu(cpu, mm_cpumask(real_prev));
 150
 151         /* Load per-mm CR4 and LDTR state */
 152         load_mm_cr4(next);
 153         switch_ldt(real_prev, next);
 154 }
 155
 156 static void flush_tlb_func_common(const struct flush_tlb_info *f,
 157                                   bool local, enum tlb_flush_reason reason)
 158 {
 159         /* This code cannot presently handle being reentered. */
 160         VM_WARN_ON(!irqs_disabled());
 161
 162         if (this_cpu_read(cpu_tlbstate.state) != TLBSTATE_OK) {
 163                 leave_mm(smp_processor_id());
 164                 return;
 165         }
 166
 167         if (f->end == TLB_FLUSH_ALL) {
 168                 local_flush_tlb();
 169                 if (local)
 170                         count_vm_tlb_event(NR_TLB_LOCAL_FLUSH_ALL);
 171                 trace_tlb_flush(reason, TLB_FLUSH_ALL);
 172         } else {
 173                 unsigned long addr;
 174                 unsigned long nr_pages = (f->end - f->start) >> PAGE_SHIFT;
 175                 addr = f->start;
 176                 while (addr < f->end) {
 177                         __flush_tlb_single(addr);
 178                         addr += PAGE_SIZE;
 179                 }
 180                 if (local)
 181                         count_vm_tlb_events(NR_TLB_LOCAL_FLUSH_ONE, nr_pages);
 182                 trace_tlb_flush(reason, nr_pages);
 183         }
 184 }
 185
 186 static void flush_tlb_func_local(void *info, enum tlb_flush_reason reason)
 187 {
 188         const struct flush_tlb_info *f = info;
 189
 190         flush_tlb_func_common(f, true, reason);
 191 }
 192
 193 static void flush_tlb_func_remote(void *info)
 194 {
 195         const struct flush_tlb_info *f = info;
 196
 197         inc_irq_stat(irq_tlb_count);
 198
 199         if (f->mm && f->mm != this_cpu_read(cpu_tlbstate.loaded_mm))
 200                 return;
 201
 202         count_vm_tlb_event(NR_TLB_REMOTE_FLUSH_RECEIVED);
 203         flush_tlb_func_common(f, false, TLB_REMOTE_SHOOTDOWN);
 204 }
 205
 206 void native_flush_tlb_others(const struct cpumask *cpumask,
 207                              const struct flush_tlb_info *info)
 208 {
 209         count_vm_tlb_event(NR_TLB_REMOTE_FLUSH);
 210         if (info->end == TLB_FLUSH_ALL)
 211                 trace_tlb_flush(TLB_REMOTE_SEND_IPI, TLB_FLUSH_ALL);
 212         else
 213                 trace_tlb_flush(TLB_REMOTE_SEND_IPI,
 214                                 (info->end - info->start) >> PAGE_SHIFT);
 215
 216         if (is_uv_system()) {
 217                 unsigned int cpu;
 218
 219                 cpu = smp_processor_id();
 220                 cpumask = uv_flush_tlb_others(cpumask, info);
 221                 if (cpumask)
 222                         smp_call_function_many(cpumask, flush_tlb_func_remote,
 223                                                (void *)info, 1);
 224                 return;
 225         }
 226         smp_call_function_many(cpumask, flush_tlb_func_remote,
 227                                (void *)info, 1);
 228 }
 229
 230 /*
 231  * See Documentation/x86/tlb.txt for details.  We choose 33
 232  * because it is large enough to cover the vast majority (at
 233  * least 95%) of allocations, and is small enough that we are
 234  * confident it will not cause too much overhead.  Each single
 235  * flush is about 100 ns, so this caps the maximum overhead at
 236  * _about_ 3,000 ns.
 237  *
 238  * This is in units of pages.
 239  */
 240 static unsigned long tlb_single_page_flush_ceiling __read_mostly = 33;
 241
 242 void flush_tlb_mm_range(struct mm_struct *mm, unsigned long start,
 243                                 unsigned long end, unsigned long vmflag)
 244 {
 245         int cpu;
 246
 247         struct flush_tlb_info info = {
 248                 .mm = mm,
 249         };
 250
 251         cpu = get_cpu();
 252
 253         /* Synchronize with switch_mm. */
 254         smp_mb();
 255
 256         /* Should we flush just the requested range? */
 257         if ((end != TLB_FLUSH_ALL) &&
 258             !(vmflag & VM_HUGETLB) &&
 259             ((end - start) >> PAGE_SHIFT) <= tlb_single_page_flush_ceiling) {
 260                 info.start = start;
 261                 info.end = end;
 262         } else {
 263                 info.start = 0UL;
 264                 info.end = TLB_FLUSH_ALL;
 265         }
 266
 267         if (mm == this_cpu_read(cpu_tlbstate.loaded_mm)) {
 268                 VM_WARN_ON(irqs_disabled());
 269                 local_irq_disable();
 270                 flush_tlb_func_local(&info, TLB_LOCAL_MM_SHOOTDOWN);
 271                 local_irq_enable();
 272         }
 273
 274         if (cpumask_any_but(mm_cpumask(mm), cpu) < nr_cpu_ids)
 275                 flush_tlb_others(mm_cpumask(mm), &info);
 276         put_cpu();
 277 }
 278
 279
 280 static void do_flush_tlb_all(void *info)
 281 {
 282         count_vm_tlb_event(NR_TLB_REMOTE_FLUSH_RECEIVED);
 283         __flush_tlb_all();
 284         if (this_cpu_read(cpu_tlbstate.state) == TLBSTATE_LAZY)
 285                 leave_mm(smp_processor_id());
 286 }
 287
 288 void flush_tlb_all(void)
 289 {
 290         count_vm_tlb_event(NR_TLB_REMOTE_FLUSH);
 291         on_each_cpu(do_flush_tlb_all, NULL, 1);
 292 }
 293
 294 static void do_kernel_range_flush(void *info)
 295 {
 296         struct flush_tlb_info *f = info;
 297         unsigned long addr;
 298
 299         /* flush range by one by one 'invlpg' */
 300         for (addr = f->start; addr < f->end; addr += PAGE_SIZE)
 301                 __flush_tlb_single(addr);
 302 }
 303
 304 void flush_tlb_kernel_range(unsigned long start, unsigned long end)
 305 {
 306
 307         /* Balance as user space task's flush, a bit conservative */
 308         if (end == TLB_FLUSH_ALL ||
 309             (end - start) > tlb_single_page_flush_ceiling << PAGE_SHIFT) {
 310                 on_each_cpu(do_flush_tlb_all, NULL, 1);
 311         } else {
 312                 struct flush_tlb_info info;
 313                 info.start = start;
 314                 info.end = end;
 315                 on_each_cpu(do_kernel_range_flush, &info, 1);
 316         }
 317 }
 318
 319 void arch_tlbbatch_flush(struct arch_tlbflush_unmap_batch *batch)
 320 {
 321         struct flush_tlb_info info = {
 322                 .mm = NULL,
 323                 .start = 0UL,
 324                 .end = TLB_FLUSH_ALL,
 325         };
 326
 327         int cpu = get_cpu();
 328
 329         if (cpumask_test_cpu(cpu, &batch->cpumask)) {
 330                 VM_WARN_ON(irqs_disabled());
 331                 local_irq_disable();
 332                 flush_tlb_func_local(&info, TLB_LOCAL_SHOOTDOWN);
 333                 local_irq_enable();
 334         }
 335
 336         if (cpumask_any_but(&batch->cpumask, cpu) < nr_cpu_ids)
 337                 flush_tlb_others(&batch->cpumask, &info);
 338         cpumask_clear(&batch->cpumask);
 339
 340         put_cpu();
 341 }
 342
 343 static ssize_t tlbflush_read_file(struct file *file, char __user *user_buf,
 344                              size_t count, loff_t *ppos)
 345 {
 346         char buf[32];
 347         unsigned int len;
 348
 349         len = sprintf(buf, "%ld\n", tlb_single_page_flush_ceiling);
 350         return simple_read_from_buffer(user_buf, count, ppos, buf, len);
 351 }
 352
 353 static ssize_t tlbflush_write_file(struct file *file,
 354                  const char __user *user_buf, size_t count, loff_t *ppos)
 355 {
 356         char buf[32];
 357         ssize_t len;
 358         int ceiling;
 359
 360         len = min(count, sizeof(buf) - 1);
 361         if (copy_from_user(buf, user_buf, len))
 362                 return -EFAULT;
 363
 364         buf[len] = '\0';
 365         if (kstrtoint(buf, 0, &ceiling))
 366                 return -EINVAL;
 367
 368         if (ceiling < 0)
 369                 return -EINVAL;
 370
 371         tlb_single_page_flush_ceiling = ceiling;
 372         return count;
 373 }
 374
 375 static const struct file_operations fops_tlbflush = {
 376         .read = tlbflush_read_file,
 377         .write = tlbflush_write_file,
 378         .llseek = default_llseek,
 379 };
 380
 381 static int __init create_tlb_single_page_flush_ceiling(void)
 382 {
 383         debugfs_create_file("tlb_single_page_flush_ceiling", S_IRUSR | S_IWUSR,
 384                             arch_debugfs_dir, NULL, &fops_tlbflush);
 385         return 0;
 386 }
 387 late_initcall(create_tlb_single_page_flush_ceiling);