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 atomic64_t last_mm_ctx_id = ATOMIC64_INIT(1);
  32
  33 void leave_mm(int cpu)
  34 {
  35         struct mm_struct *loaded_mm = this_cpu_read(cpu_tlbstate.loaded_mm);
  36
  37         /*
  38          * It's plausible that we're in lazy TLB mode while our mm is init_mm.
  39          * If so, our callers still expect us to flush the TLB, but there
  40          * aren't any user TLB entries in init_mm to worry about.
  41          *
  42          * This needs to happen before any other sanity checks due to
  43          * intel_idle's shenanigans.
  44          */
  45         if (loaded_mm == &init_mm)
  46                 return;
  47
  48         /* Warn if we're not lazy. */
  49         WARN_ON(cpumask_test_cpu(smp_processor_id(), mm_cpumask(loaded_mm)));
  50
  51         switch_mm(NULL, &init_mm, NULL);
  52 }
  53
  54 void switch_mm(struct mm_struct *prev, struct mm_struct *next,
  55                struct task_struct *tsk)
  56 {
  57         unsigned long flags;
  58
  59         local_irq_save(flags);
  60         switch_mm_irqs_off(prev, next, tsk);
  61         local_irq_restore(flags);
  62 }
  63
  64 void switch_mm_irqs_off(struct mm_struct *prev, struct mm_struct *next,
  65                         struct task_struct *tsk)
  66 {
  67         struct mm_struct *real_prev = this_cpu_read(cpu_tlbstate.loaded_mm);
  68         unsigned cpu = smp_processor_id();
  69         u64 next_tlb_gen;
  70
  71         /*
  72          * NB: The scheduler will call us with prev == next when switching
  73          * from lazy TLB mode to normal mode if active_mm isn't changing.
  74          * When this happens, we don't assume that CR3 (and hence
  75          * cpu_tlbstate.loaded_mm) matches next.
  76          *
  77          * NB: leave_mm() calls us with prev == NULL and tsk == NULL.
  78          */
  79
  80         /* We don't want flush_tlb_func_* to run concurrently with us. */
  81         if (IS_ENABLED(CONFIG_PROVE_LOCKING))
  82                 WARN_ON_ONCE(!irqs_disabled());
  83
  84         /*
  85          * Verify that CR3 is what we think it is.  This will catch
  86          * hypothetical buggy code that directly switches to swapper_pg_dir
  87          * without going through leave_mm() / switch_mm_irqs_off().
  88          */
  89         VM_BUG_ON(read_cr3_pa() != __pa(real_prev->pgd));
  90
  91         if (real_prev == next) {
  92                 VM_BUG_ON(this_cpu_read(cpu_tlbstate.ctxs[0].ctx_id) !=
  93                           next->context.ctx_id);
  94
  95                 if (cpumask_test_cpu(cpu, mm_cpumask(next))) {
  96                         /*
  97                          * There's nothing to do: we weren't lazy, and we
  98                          * aren't changing our mm.  We don't need to flush
  99                          * anything, nor do we need to update CR3, CR4, or
 100                          * LDTR.
 101                          */
 102                         return;
 103                 }
 104
 105                 /* Resume remote flushes and then read tlb_gen. */
 106                 cpumask_set_cpu(cpu, mm_cpumask(next));
 107                 next_tlb_gen = atomic64_read(&next->context.tlb_gen);
 108
 109                 if (this_cpu_read(cpu_tlbstate.ctxs[0].tlb_gen) < next_tlb_gen) {
 110                         /*
 111                          * Ideally, we'd have a flush_tlb() variant that
 112                          * takes the known CR3 value as input.  This would
 113                          * be faster on Xen PV and on hypothetical CPUs
 114                          * on which INVPCID is fast.
 115                          */
 116                         this_cpu_write(cpu_tlbstate.ctxs[0].tlb_gen,
 117                                        next_tlb_gen);
 118                         write_cr3(__pa(next->pgd));
 119                         trace_tlb_flush(TLB_FLUSH_ON_TASK_SWITCH,
 120                                         TLB_FLUSH_ALL);
 121                 }
 122
 123                 /*
 124                  * We just exited lazy mode, which means that CR4 and/or LDTR
 125                  * may be stale.  (Changes to the required CR4 and LDTR states
 126                  * are not reflected in tlb_gen.)
 127                  */
 128         } else {
 129                 VM_BUG_ON(this_cpu_read(cpu_tlbstate.ctxs[0].ctx_id) ==
 130                           next->context.ctx_id);
 131
 132                 if (IS_ENABLED(CONFIG_VMAP_STACK)) {
 133                         /*
 134                          * If our current stack is in vmalloc space and isn't
 135                          * mapped in the new pgd, we'll double-fault.  Forcibly
 136                          * map it.
 137                          */
 138                         unsigned int index = pgd_index(current_stack_pointer());
 139                         pgd_t *pgd = next->pgd + index;
 140
 141                         if (unlikely(pgd_none(*pgd)))
 142                                 set_pgd(pgd, init_mm.pgd[index]);
 143                 }
 144
 145                 /* Stop remote flushes for the previous mm */
 146                 if (cpumask_test_cpu(cpu, mm_cpumask(real_prev)))
 147                         cpumask_clear_cpu(cpu, mm_cpumask(real_prev));
 148
 149                 VM_WARN_ON_ONCE(cpumask_test_cpu(cpu, mm_cpumask(next)));
 150
 151                 /*
 152                  * Start remote flushes and then read tlb_gen.
 153                  */
 154                 cpumask_set_cpu(cpu, mm_cpumask(next));
 155                 next_tlb_gen = atomic64_read(&next->context.tlb_gen);
 156
 157                 this_cpu_write(cpu_tlbstate.ctxs[0].ctx_id, next->context.ctx_id);
 158                 this_cpu_write(cpu_tlbstate.ctxs[0].tlb_gen, next_tlb_gen);
 159                 this_cpu_write(cpu_tlbstate.loaded_mm, next);
 160                 write_cr3(__pa(next->pgd));
 161
 162                 trace_tlb_flush(TLB_FLUSH_ON_TASK_SWITCH, TLB_FLUSH_ALL);
 163         }
 164
 165         load_mm_cr4(next);
 166         switch_ldt(real_prev, next);
 167 }
 168
 169 /*
 170  * flush_tlb_func_common()'s memory ordering requirement is that any
 171  * TLB fills that happen after we flush the TLB are ordered after we
 172  * read active_mm's tlb_gen.  We don't need any explicit barriers
 173  * because all x86 flush operations are serializing and the
 174  * atomic64_read operation won't be reordered by the compiler.
 175  */
 176 static void flush_tlb_func_common(const struct flush_tlb_info *f,
 177                                   bool local, enum tlb_flush_reason reason)
 178 {
 179         /*
 180          * We have three different tlb_gen values in here.  They are:
 181          *
 182          * - mm_tlb_gen:     the latest generation.
 183          * - local_tlb_gen:  the generation that this CPU has already caught
 184          *                   up to.
 185          * - f->new_tlb_gen: the generation that the requester of the flush
 186          *                   wants us to catch up to.
 187          */
 188         struct mm_struct *loaded_mm = this_cpu_read(cpu_tlbstate.loaded_mm);
 189         u64 mm_tlb_gen = atomic64_read(&loaded_mm->context.tlb_gen);
 190         u64 local_tlb_gen = this_cpu_read(cpu_tlbstate.ctxs[0].tlb_gen);
 191
 192         /* This code cannot presently handle being reentered. */
 193         VM_WARN_ON(!irqs_disabled());
 194
 195         VM_WARN_ON(this_cpu_read(cpu_tlbstate.ctxs[0].ctx_id) !=
 196                    loaded_mm->context.ctx_id);
 197
 198         if (!cpumask_test_cpu(smp_processor_id(), mm_cpumask(loaded_mm))) {
 199                 /*
 200                  * We're in lazy mode -- don't flush.  We can get here on
 201                  * remote flushes due to races and on local flushes if a
 202                  * kernel thread coincidentally flushes the mm it's lazily
 203                  * still using.
 204                  */
 205                 return;
 206         }
 207
 208         if (unlikely(local_tlb_gen == mm_tlb_gen)) {
 209                 /*
 210                  * There's nothing to do: we're already up to date.  This can
 211                  * happen if two concurrent flushes happen -- the first flush to
 212                  * be handled can catch us all the way up, leaving no work for
 213                  * the second flush.
 214                  */
 215                 trace_tlb_flush(reason, 0);
 216                 return;
 217         }
 218
 219         WARN_ON_ONCE(local_tlb_gen > mm_tlb_gen);
 220         WARN_ON_ONCE(f->new_tlb_gen > mm_tlb_gen);
 221
 222         /*
 223          * If we get to this point, we know that our TLB is out of date.
 224          * This does not strictly imply that we need to flush (it's
 225          * possible that f->new_tlb_gen <= local_tlb_gen), but we're
 226          * going to need to flush in the very near future, so we might
 227          * as well get it over with.
 228          *
 229          * The only question is whether to do a full or partial flush.
 230          *
 231          * We do a partial flush if requested and two extra conditions
 232          * are met:
 233          *
 234          * 1. f->new_tlb_gen == local_tlb_gen + 1.  We have an invariant that
 235          *    we've always done all needed flushes to catch up to
 236          *    local_tlb_gen.  If, for example, local_tlb_gen == 2 and
 237          *    f->new_tlb_gen == 3, then we know that the flush needed to bring
 238          *    us up to date for tlb_gen 3 is the partial flush we're
 239          *    processing.
 240          *
 241          *    As an example of why this check is needed, suppose that there
 242          *    are two concurrent flushes.  The first is a full flush that
 243          *    changes context.tlb_gen from 1 to 2.  The second is a partial
 244          *    flush that changes context.tlb_gen from 2 to 3.  If they get
 245          *    processed on this CPU in reverse order, we'll see
 246          *     local_tlb_gen == 1, mm_tlb_gen == 3, and end != TLB_FLUSH_ALL.
 247          *    If we were to use __flush_tlb_single() and set local_tlb_gen to
 248          *    3, we'd be break the invariant: we'd update local_tlb_gen above
 249          *    1 without the full flush that's needed for tlb_gen 2.
 250          *
 251          * 2. f->new_tlb_gen == mm_tlb_gen.  This is purely an optimiation.
 252          *    Partial TLB flushes are not all that much cheaper than full TLB
 253          *    flushes, so it seems unlikely that it would be a performance win
 254          *    to do a partial flush if that won't bring our TLB fully up to
 255          *    date.  By doing a full flush instead, we can increase
 256          *    local_tlb_gen all the way to mm_tlb_gen and we can probably
 257          *    avoid another flush in the very near future.
 258          */
 259         if (f->end != TLB_FLUSH_ALL &&
 260             f->new_tlb_gen == local_tlb_gen + 1 &&
 261             f->new_tlb_gen == mm_tlb_gen) {
 262                 /* Partial flush */
 263                 unsigned long addr;
 264                 unsigned long nr_pages = (f->end - f->start) >> PAGE_SHIFT;
 265
 266                 addr = f->start;
 267                 while (addr < f->end) {
 268                         __flush_tlb_single(addr);
 269                         addr += PAGE_SIZE;
 270                 }
 271                 if (local)
 272                         count_vm_tlb_events(NR_TLB_LOCAL_FLUSH_ONE, nr_pages);
 273                 trace_tlb_flush(reason, nr_pages);
 274         } else {
 275                 /* Full flush. */
 276                 local_flush_tlb();
 277                 if (local)
 278                         count_vm_tlb_event(NR_TLB_LOCAL_FLUSH_ALL);
 279                 trace_tlb_flush(reason, TLB_FLUSH_ALL);
 280         }
 281
 282         /* Both paths above update our state to mm_tlb_gen. */
 283         this_cpu_write(cpu_tlbstate.ctxs[0].tlb_gen, mm_tlb_gen);
 284 }
 285
 286 static void flush_tlb_func_local(void *info, enum tlb_flush_reason reason)
 287 {
 288         const struct flush_tlb_info *f = info;
 289
 290         flush_tlb_func_common(f, true, reason);
 291 }
 292
 293 static void flush_tlb_func_remote(void *info)
 294 {
 295         const struct flush_tlb_info *f = info;
 296
 297         inc_irq_stat(irq_tlb_count);
 298
 299         if (f->mm && f->mm != this_cpu_read(cpu_tlbstate.loaded_mm))
 300                 return;
 301
 302         count_vm_tlb_event(NR_TLB_REMOTE_FLUSH_RECEIVED);
 303         flush_tlb_func_common(f, false, TLB_REMOTE_SHOOTDOWN);
 304 }
 305
 306 void native_flush_tlb_others(const struct cpumask *cpumask,
 307                              const struct flush_tlb_info *info)
 308 {
 309         count_vm_tlb_event(NR_TLB_REMOTE_FLUSH);
 310         if (info->end == TLB_FLUSH_ALL)
 311                 trace_tlb_flush(TLB_REMOTE_SEND_IPI, TLB_FLUSH_ALL);
 312         else
 313                 trace_tlb_flush(TLB_REMOTE_SEND_IPI,
 314                                 (info->end - info->start) >> PAGE_SHIFT);
 315
 316         if (is_uv_system()) {
 317                 /*
 318                  * This whole special case is confused.  UV has a "Broadcast
 319                  * Assist Unit", which seems to be a fancy way to send IPIs.
 320                  * Back when x86 used an explicit TLB flush IPI, UV was
 321                  * optimized to use its own mechanism.  These days, x86 uses
 322                  * smp_call_function_many(), but UV still uses a manual IPI,
 323                  * and that IPI's action is out of date -- it does a manual
 324                  * flush instead of calling flush_tlb_func_remote().  This
 325                  * means that the percpu tlb_gen variables won't be updated
 326                  * and we'll do pointless flushes on future context switches.
 327                  *
 328                  * Rather than hooking native_flush_tlb_others() here, I think
 329                  * that UV should be updated so that smp_call_function_many(),
 330                  * etc, are optimal on UV.
 331                  */
 332                 unsigned int cpu;
 333
 334                 cpu = smp_processor_id();
 335                 cpumask = uv_flush_tlb_others(cpumask, info);
 336                 if (cpumask)
 337                         smp_call_function_many(cpumask, flush_tlb_func_remote,
 338                                                (void *)info, 1);
 339                 return;
 340         }
 341         smp_call_function_many(cpumask, flush_tlb_func_remote,
 342                                (void *)info, 1);
 343 }
 344
 345 /*
 346  * See Documentation/x86/tlb.txt for details.  We choose 33
 347  * because it is large enough to cover the vast majority (at
 348  * least 95%) of allocations, and is small enough that we are
 349  * confident it will not cause too much overhead.  Each single
 350  * flush is about 100 ns, so this caps the maximum overhead at
 351  * _about_ 3,000 ns.
 352  *
 353  * This is in units of pages.
 354  */
 355 static unsigned long tlb_single_page_flush_ceiling __read_mostly = 33;
 356
 357 void flush_tlb_mm_range(struct mm_struct *mm, unsigned long start,
 358                                 unsigned long end, unsigned long vmflag)
 359 {
 360         int cpu;
 361
 362         struct flush_tlb_info info = {
 363                 .mm = mm,
 364         };
 365
 366         cpu = get_cpu();
 367
 368         /* This is also a barrier that synchronizes with switch_mm(). */
 369         info.new_tlb_gen = inc_mm_tlb_gen(mm);
 370
 371         /* Should we flush just the requested range? */
 372         if ((end != TLB_FLUSH_ALL) &&
 373             !(vmflag & VM_HUGETLB) &&
 374             ((end - start) >> PAGE_SHIFT) <= tlb_single_page_flush_ceiling) {
 375                 info.start = start;
 376                 info.end = end;
 377         } else {
 378                 info.start = 0UL;
 379                 info.end = TLB_FLUSH_ALL;
 380         }
 381
 382         if (mm == this_cpu_read(cpu_tlbstate.loaded_mm)) {
 383                 VM_WARN_ON(irqs_disabled());
 384                 local_irq_disable();
 385                 flush_tlb_func_local(&info, TLB_LOCAL_MM_SHOOTDOWN);
 386                 local_irq_enable();
 387         }
 388
 389         if (cpumask_any_but(mm_cpumask(mm), cpu) < nr_cpu_ids)
 390                 flush_tlb_others(mm_cpumask(mm), &info);
 391
 392         put_cpu();
 393 }
 394
 395
 396 static void do_flush_tlb_all(void *info)
 397 {
 398         count_vm_tlb_event(NR_TLB_REMOTE_FLUSH_RECEIVED);
 399         __flush_tlb_all();
 400 }
 401
 402 void flush_tlb_all(void)
 403 {
 404         count_vm_tlb_event(NR_TLB_REMOTE_FLUSH);
 405         on_each_cpu(do_flush_tlb_all, NULL, 1);
 406 }
 407
 408 static void do_kernel_range_flush(void *info)
 409 {
 410         struct flush_tlb_info *f = info;
 411         unsigned long addr;
 412
 413         /* flush range by one by one 'invlpg' */
 414         for (addr = f->start; addr < f->end; addr += PAGE_SIZE)
 415                 __flush_tlb_single(addr);
 416 }
 417
 418 void flush_tlb_kernel_range(unsigned long start, unsigned long end)
 419 {
 420
 421         /* Balance as user space task's flush, a bit conservative */
 422         if (end == TLB_FLUSH_ALL ||
 423             (end - start) > tlb_single_page_flush_ceiling << PAGE_SHIFT) {
 424                 on_each_cpu(do_flush_tlb_all, NULL, 1);
 425         } else {
 426                 struct flush_tlb_info info;
 427                 info.start = start;
 428                 info.end = end;
 429                 on_each_cpu(do_kernel_range_flush, &info, 1);
 430         }
 431 }
 432
 433 void arch_tlbbatch_flush(struct arch_tlbflush_unmap_batch *batch)
 434 {
 435         struct flush_tlb_info info = {
 436                 .mm = NULL,
 437                 .start = 0UL,
 438                 .end = TLB_FLUSH_ALL,
 439         };
 440
 441         int cpu = get_cpu();
 442
 443         if (cpumask_test_cpu(cpu, &batch->cpumask)) {
 444                 VM_WARN_ON(irqs_disabled());
 445                 local_irq_disable();
 446                 flush_tlb_func_local(&info, TLB_LOCAL_SHOOTDOWN);
 447                 local_irq_enable();
 448         }
 449
 450         if (cpumask_any_but(&batch->cpumask, cpu) < nr_cpu_ids)
 451                 flush_tlb_others(&batch->cpumask, &info);
 452
 453         cpumask_clear(&batch->cpumask);
 454
 455         put_cpu();
 456 }
 457
 458 static ssize_t tlbflush_read_file(struct file *file, char __user *user_buf,
 459                              size_t count, loff_t *ppos)
 460 {
 461         char buf[32];
 462         unsigned int len;
 463
 464         len = sprintf(buf, "%ld\n", tlb_single_page_flush_ceiling);
 465         return simple_read_from_buffer(user_buf, count, ppos, buf, len);
 466 }
 467
 468 static ssize_t tlbflush_write_file(struct file *file,
 469                  const char __user *user_buf, size_t count, loff_t *ppos)
 470 {
 471         char buf[32];
 472         ssize_t len;
 473         int ceiling;
 474
 475         len = min(count, sizeof(buf) - 1);
 476         if (copy_from_user(buf, user_buf, len))
 477                 return -EFAULT;
 478
 479         buf[len] = '\0';
 480         if (kstrtoint(buf, 0, &ceiling))
 481                 return -EINVAL;
 482
 483         if (ceiling < 0)
 484                 return -EINVAL;
 485
 486         tlb_single_page_flush_ceiling = ceiling;
 487         return count;
 488 }
 489
 490 static const struct file_operations fops_tlbflush = {
 491         .read = tlbflush_read_file,
 492         .write = tlbflush_write_file,
 493         .llseek = default_llseek,
 494 };
 495
 496 static int __init create_tlb_single_page_flush_ceiling(void)
 497 {
 498         debugfs_create_file("tlb_single_page_flush_ceiling", S_IRUSR | S_IWUSR,
 499                             arch_debugfs_dir, NULL, &fops_tlbflush);
 500         return 0;
 501 }
 502 late_initcall(create_tlb_single_page_flush_ceiling);