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         if (is_uv_system()) {
 144                 unsigned int cpu;
 145
 146                 cpu = smp_processor_id();
 147                 cpumask = uv_flush_tlb_others(cpumask, mm, start, end, cpu);
 148                 if (cpumask)
 149                         smp_call_function_many(cpumask, flush_tlb_func,
 150                                                                 &info, 1);
 151                 return;
 152         }
 153         smp_call_function_many(cpumask, flush_tlb_func, &info, 1);
 154 }
 155
 156 void flush_tlb_current_task(void)
 157 {
 158         struct mm_struct *mm = current->mm;
 159
 160         preempt_disable();
 161
 162         count_vm_tlb_event(NR_TLB_LOCAL_FLUSH_ALL);
 163         local_flush_tlb();
 164         trace_tlb_flush(TLB_LOCAL_SHOOTDOWN, TLB_FLUSH_ALL);
 165         if (cpumask_any_but(mm_cpumask(mm), smp_processor_id()) < nr_cpu_ids)
 166                 flush_tlb_others(mm_cpumask(mm), mm, 0UL, TLB_FLUSH_ALL);
 167         preempt_enable();
 168 }
 169
 170 /*
 171  * See Documentation/x86/tlb.txt for details.  We choose 33
 172  * because it is large enough to cover the vast majority (at
 173  * least 95%) of allocations, and is small enough that we are
 174  * confident it will not cause too much overhead.  Each single
 175  * flush is about 100 ns, so this caps the maximum overhead at
 176  * _about_ 3,000 ns.
 177  *
 178  * This is in units of pages.
 179  */
 180 static unsigned long tlb_single_page_flush_ceiling __read_mostly = 33;
 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         /* do a global flush by default */
 187         unsigned long base_pages_to_flush = TLB_FLUSH_ALL;
 188
 189         preempt_disable();
 190         if (current->active_mm != mm)
 191                 goto out;
 192
 193         if (!current->mm) {
 194                 leave_mm(smp_processor_id());
 195                 goto out;
 196         }
 197
 198         if ((end != TLB_FLUSH_ALL) && !(vmflag & VM_HUGETLB))
 199                 base_pages_to_flush = (end - start) >> PAGE_SHIFT;
 200
 201         if (base_pages_to_flush > tlb_single_page_flush_ceiling) {
 202                 base_pages_to_flush = TLB_FLUSH_ALL;
 203                 count_vm_tlb_event(NR_TLB_LOCAL_FLUSH_ALL);
 204                 local_flush_tlb();
 205         } else {
 206                 /* flush range by one by one 'invlpg' */
 207                 for (addr = start; addr < end;  addr += PAGE_SIZE) {
 208                         count_vm_tlb_event(NR_TLB_LOCAL_FLUSH_ONE);
 209                         __flush_tlb_single(addr);
 210                 }
 211         }
 212         trace_tlb_flush(TLB_LOCAL_MM_SHOOTDOWN, base_pages_to_flush);
 213 out:
 214         if (base_pages_to_flush == TLB_FLUSH_ALL) {
 215                 start = 0UL;
 216                 end = TLB_FLUSH_ALL;
 217         }
 218         if (cpumask_any_but(mm_cpumask(mm), smp_processor_id()) < nr_cpu_ids)
 219                 flush_tlb_others(mm_cpumask(mm), mm, start, end);
 220         preempt_enable();
 221 }
 222
 223 void flush_tlb_page(struct vm_area_struct *vma, unsigned long start)
 224 {
 225         struct mm_struct *mm = vma->vm_mm;
 226
 227         preempt_disable();
 228
 229         if (current->active_mm == mm) {
 230                 if (current->mm)
 231                         __flush_tlb_one(start);
 232                 else
 233                         leave_mm(smp_processor_id());
 234         }
 235
 236         if (cpumask_any_but(mm_cpumask(mm), smp_processor_id()) < nr_cpu_ids)
 237                 flush_tlb_others(mm_cpumask(mm), mm, start, 0UL);
 238
 239         preempt_enable();
 240 }
 241
 242 static void do_flush_tlb_all(void *info)
 243 {
 244         count_vm_tlb_event(NR_TLB_REMOTE_FLUSH_RECEIVED);
 245         __flush_tlb_all();
 246         if (this_cpu_read(cpu_tlbstate.state) == TLBSTATE_LAZY)
 247                 leave_mm(smp_processor_id());
 248 }
 249
 250 void flush_tlb_all(void)
 251 {
 252         count_vm_tlb_event(NR_TLB_REMOTE_FLUSH);
 253         on_each_cpu(do_flush_tlb_all, NULL, 1);
 254 }
 255
 256 static void do_kernel_range_flush(void *info)
 257 {
 258         struct flush_tlb_info *f = info;
 259         unsigned long addr;
 260
 261         /* flush range by one by one 'invlpg' */
 262         for (addr = f->flush_start; addr < f->flush_end; addr += PAGE_SIZE)
 263                 __flush_tlb_single(addr);
 264 }
 265
 266 void flush_tlb_kernel_range(unsigned long start, unsigned long end)
 267 {
 268
 269         /* Balance as user space task's flush, a bit conservative */
 270         if (end == TLB_FLUSH_ALL ||
 271             (end - start) > tlb_single_page_flush_ceiling * PAGE_SIZE) {
 272                 on_each_cpu(do_flush_tlb_all, NULL, 1);
 273         } else {
 274                 struct flush_tlb_info info;
 275                 info.flush_start = start;
 276                 info.flush_end = end;
 277                 on_each_cpu(do_kernel_range_flush, &info, 1);
 278         }
 279 }
 280
 281 static ssize_t tlbflush_read_file(struct file *file, char __user *user_buf,
 282                              size_t count, loff_t *ppos)
 283 {
 284         char buf[32];
 285         unsigned int len;
 286
 287         len = sprintf(buf, "%ld\n", tlb_single_page_flush_ceiling);
 288         return simple_read_from_buffer(user_buf, count, ppos, buf, len);
 289 }
 290
 291 static ssize_t tlbflush_write_file(struct file *file,
 292                  const char __user *user_buf, size_t count, loff_t *ppos)
 293 {
 294         char buf[32];
 295         ssize_t len;
 296         int ceiling;
 297
 298         len = min(count, sizeof(buf) - 1);
 299         if (copy_from_user(buf, user_buf, len))
 300                 return -EFAULT;
 301
 302         buf[len] = '\0';
 303         if (kstrtoint(buf, 0, &ceiling))
 304                 return -EINVAL;
 305
 306         if (ceiling < 0)
 307                 return -EINVAL;
 308
 309         tlb_single_page_flush_ceiling = ceiling;
 310         return count;
 311 }
 312
 313 static const struct file_operations fops_tlbflush = {
 314         .read = tlbflush_read_file,
 315         .write = tlbflush_write_file,
 316         .llseek = default_llseek,
 317 };
 318
 319 static int __init create_tlb_single_page_flush_ceiling(void)
 320 {
 321         debugfs_create_file("tlb_single_page_flush_ceiling", S_IRUSR | S_IWUSR,
 322                             arch_debugfs_dir, NULL, &fops_tlbflush);
 323         return 0;
 324 }
 325 late_initcall(create_tlb_single_page_flush_ceiling);