kernel/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 struct flush_tlb_info {
  32         struct mm_struct *flush_mm;
  33         unsigned long flush_start;
  34         unsigned long flush_end;
  35 };
  36
  37 /*
  38  * We cannot call mmdrop() because we are in interrupt context,
  39  * instead update mm->cpu_vm_mask.
  40  */
  41 void leave_mm(int cpu)
  42 {
  43         struct mm_struct *active_mm = this_cpu_read(cpu_tlbstate.active_mm);
  44         if (this_cpu_read(cpu_tlbstate.state) == TLBSTATE_OK)
  45                 BUG();
  46         if (cpumask_test_cpu(cpu, mm_cpumask(active_mm))) {
  47                 cpumask_clear_cpu(cpu, mm_cpumask(active_mm));
  48                 load_cr3(swapper_pg_dir);
  49                 /*
  50                  * This gets called in the idle path where RCU
  51                  * functions differently.  Tracing normally
  52                  * uses RCU, so we have to call the tracepoint
  53                  * specially here.
  54                  */
  55                 trace_tlb_flush_rcuidle(TLB_FLUSH_ON_TASK_SWITCH, TLB_FLUSH_ALL);
  56         }
  57 }
  58 EXPORT_SYMBOL_GPL(leave_mm);
  59
  60 /*
  61  * The flush IPI assumes that a thread switch happens in this order:
  62  * [cpu0: the cpu that switches]
  63  * 1) switch_mm() either 1a) or 1b)
  64  * 1a) thread switch to a different mm
  65  * 1a1) set cpu_tlbstate to TLBSTATE_OK
  66  *      Now the tlb flush NMI handler flush_tlb_func won't call leave_mm
  67  *      if cpu0 was in lazy tlb mode.
  68  * 1a2) update cpu active_mm
  69  *      Now cpu0 accepts tlb flushes for the new mm.
  70  * 1a3) cpu_set(cpu, new_mm->cpu_vm_mask);
  71  *      Now the other cpus will send tlb flush ipis.
  72  * 1a4) change cr3.
  73  * 1a5) cpu_clear(cpu, old_mm->cpu_vm_mask);
  74  *      Stop ipi delivery for the old mm. This is not synchronized with
  75  *      the other cpus, but flush_tlb_func ignore flush ipis for the wrong
  76  *      mm, and in the worst case we perform a superfluous tlb flush.
  77  * 1b) thread switch without mm change
  78  *      cpu active_mm is correct, cpu0 already handles flush ipis.
  79  * 1b1) set cpu_tlbstate to TLBSTATE_OK
  80  * 1b2) test_and_set the cpu bit in cpu_vm_mask.
  81  *      Atomically set the bit [other cpus will start sending flush ipis],
  82  *      and test the bit.
  83  * 1b3) if the bit was 0: leave_mm was called, flush the tlb.
  84  * 2) switch %%esp, ie current
  85  *
  86  * The interrupt must handle 2 special cases:
  87  * - cr3 is changed before %%esp, ie. it cannot use current->{active_,}mm.
  88  * - the cpu performs speculative tlb reads, i.e. even if the cpu only
  89  *   runs in kernel space, the cpu could load tlb entries for user space
  90  *   pages.
  91  *
  92  * The good news is that cpu_tlbstate is local to each cpu, no
  93  * write/read ordering problems.
  94  */
  95
  96 /*
  97  * TLB flush funcation:
  98  * 1) Flush the tlb entries if the cpu uses the mm that's being flushed.
  99  * 2) Leave the mm if we are in the lazy tlb mode.
 100  */
 101 static void flush_tlb_func(void *info)
 102 {
 103         struct flush_tlb_info *f = info;
 104
 105         inc_irq_stat(irq_tlb_count);
 106
 107         if (f->flush_mm != this_cpu_read(cpu_tlbstate.active_mm))
 108                 return;
 109         if (!f->flush_end)
 110                 f->flush_end = f->flush_start + PAGE_SIZE;
 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         info.flush_mm = mm;
 139         info.flush_start = start;
 140         info.flush_end = end;
 141
 142         count_vm_tlb_event(NR_TLB_REMOTE_FLUSH);
 143         trace_tlb_flush(TLB_REMOTE_SEND_IPI, end - start);
 144         if (is_uv_system()) {
 145                 unsigned int cpu;
 146
 147                 cpu = smp_processor_id();
 148                 cpumask = uv_flush_tlb_others(cpumask, mm, start, end, cpu);
 149                 if (cpumask)
 150                         smp_call_function_many(cpumask, flush_tlb_func,
 151                                                                 &info, 1);
 152                 return;
 153         }
 154         smp_call_function_many(cpumask, flush_tlb_func, &info, 1);
 155 }
 156
 157 void flush_tlb_current_task(void)
 158 {
 159         struct mm_struct *mm = current->mm;
 160
 161         preempt_disable();
 162
 163         count_vm_tlb_event(NR_TLB_LOCAL_FLUSH_ALL);
 164
 165         /* This is an implicit full barrier that synchronizes with switch_mm. */
 166         local_flush_tlb();
 167
 168         trace_tlb_flush(TLB_LOCAL_SHOOTDOWN, TLB_FLUSH_ALL);
 169         if (cpumask_any_but(mm_cpumask(mm), smp_processor_id()) < nr_cpu_ids)
 170                 flush_tlb_others(mm_cpumask(mm), mm, 0UL, TLB_FLUSH_ALL);
 171         preempt_enable();
 172 }
 173
 174 /*
 175  * See Documentation/x86/tlb.txt for details.  We choose 33
 176  * because it is large enough to cover the vast majority (at
 177  * least 95%) of allocations, and is small enough that we are
 178  * confident it will not cause too much overhead.  Each single
 179  * flush is about 100 ns, so this caps the maximum overhead at
 180  * _about_ 3,000 ns.
 181  *
 182  * This is in units of pages.
 183  */
 184 static unsigned long tlb_single_page_flush_ceiling __read_mostly = 33;
 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         /* do a global flush by default */
 191         unsigned long base_pages_to_flush = TLB_FLUSH_ALL;
 192
 193         preempt_disable();
 194         if (current->active_mm != mm) {
 195                 /* Synchronize with switch_mm. */
 196                 smp_mb();
 197
 198                 goto out;
 199         }
 200
 201         if (!current->mm) {
 202                 leave_mm(smp_processor_id());
 203
 204                 /* Synchronize with switch_mm. */
 205                 smp_mb();
 206
 207                 goto out;
 208         }
 209
 210         if ((end != TLB_FLUSH_ALL) && !(vmflag & VM_HUGETLB))
 211                 base_pages_to_flush = (end - start) >> PAGE_SHIFT;
 212
 213         /*
 214          * Both branches below are implicit full barriers (MOV to CR or
 215          * INVLPG) that synchronize with switch_mm.
 216          */
 217         if (base_pages_to_flush > tlb_single_page_flush_ceiling) {
 218                 base_pages_to_flush = TLB_FLUSH_ALL;
 219                 count_vm_tlb_event(NR_TLB_LOCAL_FLUSH_ALL);
 220                 local_flush_tlb();
 221         } else {
 222                 /* flush range by one by one 'invlpg' */
 223                 for (addr = start; addr < end;  addr += PAGE_SIZE) {
 224                         count_vm_tlb_event(NR_TLB_LOCAL_FLUSH_ONE);
 225                         __flush_tlb_single(addr);
 226                 }
 227         }
 228         trace_tlb_flush(TLB_LOCAL_MM_SHOOTDOWN, base_pages_to_flush);
 229 out:
 230         if (base_pages_to_flush == TLB_FLUSH_ALL) {
 231                 start = 0UL;
 232                 end = TLB_FLUSH_ALL;
 233         }
 234         if (cpumask_any_but(mm_cpumask(mm), smp_processor_id()) < nr_cpu_ids)
 235                 flush_tlb_others(mm_cpumask(mm), mm, start, end);
 236         preempt_enable();
 237 }
 238
 239 void flush_tlb_page(struct vm_area_struct *vma, unsigned long start)
 240 {
 241         struct mm_struct *mm = vma->vm_mm;
 242
 243         preempt_disable();
 244
 245         if (current->active_mm == mm) {
 246                 if (current->mm) {
 247                         /*
 248                          * Implicit full barrier (INVLPG) that synchronizes
 249                          * with switch_mm.
 250                          */
 251                         __flush_tlb_one(start);
 252                 } else {
 253                         leave_mm(smp_processor_id());
 254
 255                         /* Synchronize with switch_mm. */
 256                         smp_mb();
 257                 }
 258         }
 259
 260         if (cpumask_any_but(mm_cpumask(mm), smp_processor_id()) < nr_cpu_ids)
 261                 flush_tlb_others(mm_cpumask(mm), mm, start, 0UL);
 262
 263         preempt_enable();
 264 }
 265
 266 static void do_flush_tlb_all(void *info)
 267 {
 268         count_vm_tlb_event(NR_TLB_REMOTE_FLUSH_RECEIVED);
 269         __flush_tlb_all();
 270         if (this_cpu_read(cpu_tlbstate.state) == TLBSTATE_LAZY)
 271                 leave_mm(smp_processor_id());
 272 }
 273
 274 void flush_tlb_all(void)
 275 {
 276         count_vm_tlb_event(NR_TLB_REMOTE_FLUSH);
 277         on_each_cpu(do_flush_tlb_all, NULL, 1);
 278 }
 279
 280 static void do_kernel_range_flush(void *info)
 281 {
 282         struct flush_tlb_info *f = info;
 283         unsigned long addr;
 284
 285         /* flush range by one by one 'invlpg' */
 286         for (addr = f->flush_start; addr < f->flush_end; addr += PAGE_SIZE)
 287                 __flush_tlb_single(addr);
 288 }
 289
 290 void flush_tlb_kernel_range(unsigned long start, unsigned long end)
 291 {
 292
 293         /* Balance as user space task's flush, a bit conservative */
 294         if (end == TLB_FLUSH_ALL ||
 295             (end - start) > tlb_single_page_flush_ceiling * PAGE_SIZE) {
 296                 on_each_cpu(do_flush_tlb_all, NULL, 1);
 297         } else {
 298                 struct flush_tlb_info info;
 299                 info.flush_start = start;
 300                 info.flush_end = end;
 301                 on_each_cpu(do_kernel_range_flush, &info, 1);
 302         }
 303 }
 304
 305 static ssize_t tlbflush_read_file(struct file *file, char __user *user_buf,
 306                              size_t count, loff_t *ppos)
 307 {
 308         char buf[32];
 309         unsigned int len;
 310
 311         len = sprintf(buf, "%ld\n", tlb_single_page_flush_ceiling);
 312         return simple_read_from_buffer(user_buf, count, ppos, buf, len);
 313 }
 314
 315 static ssize_t tlbflush_write_file(struct file *file,
 316                  const char __user *user_buf, size_t count, loff_t *ppos)
 317 {
 318         char buf[32];
 319         ssize_t len;
 320         int ceiling;
 321
 322         len = min(count, sizeof(buf) - 1);
 323         if (copy_from_user(buf, user_buf, len))
 324                 return -EFAULT;
 325
 326         buf[len] = '\0';
 327         if (kstrtoint(buf, 0, &ceiling))
 328                 return -EINVAL;
 329
 330         if (ceiling < 0)
 331                 return -EINVAL;
 332
 333         tlb_single_page_flush_ceiling = ceiling;
 334         return count;
 335 }
 336
 337 static const struct file_operations fops_tlbflush = {
 338         .read = tlbflush_read_file,
 339         .write = tlbflush_write_file,
 340         .llseek = default_llseek,
 341 };
 342
 343 static int __init create_tlb_single_page_flush_ceiling(void)
 344 {
 345         debugfs_create_file("tlb_single_page_flush_ceiling", S_IRUSR | S_IWUSR,
 346                             arch_debugfs_dir, NULL, &fops_tlbflush);
 347         return 0;
 348 }
 349 late_initcall(create_tlb_single_page_flush_ceiling);