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                 trace_tlb_flush(TLB_FLUSH_ON_TASK_SWITCH, TLB_FLUSH_ALL);
  53         }
  54 }
  55 EXPORT_SYMBOL_GPL(leave_mm);
  56
  57 /*
  58  * The flush IPI assumes that a thread switch happens in this order:
  59  * [cpu0: the cpu that switches]
  60  * 1) switch_mm() either 1a) or 1b)
  61  * 1a) thread switch to a different mm
  62  * 1a1) set cpu_tlbstate to TLBSTATE_OK
  63  *      Now the tlb flush NMI handler flush_tlb_func won't call leave_mm
  64  *      if cpu0 was in lazy tlb mode.
  65  * 1a2) update cpu active_mm
  66  *      Now cpu0 accepts tlb flushes for the new mm.
  67  * 1a3) cpu_set(cpu, new_mm->cpu_vm_mask);
  68  *      Now the other cpus will send tlb flush ipis.
  69  * 1a4) change cr3.
  70  * 1a5) cpu_clear(cpu, old_mm->cpu_vm_mask);
  71  *      Stop ipi delivery for the old mm. This is not synchronized with
  72  *      the other cpus, but flush_tlb_func ignore flush ipis for the wrong
  73  *      mm, and in the worst case we perform a superfluous tlb flush.
  74  * 1b) thread switch without mm change
  75  *      cpu active_mm is correct, cpu0 already handles flush ipis.
  76  * 1b1) set cpu_tlbstate to TLBSTATE_OK
  77  * 1b2) test_and_set the cpu bit in cpu_vm_mask.
  78  *      Atomically set the bit [other cpus will start sending flush ipis],
  79  *      and test the bit.
  80  * 1b3) if the bit was 0: leave_mm was called, flush the tlb.
  81  * 2) switch %%esp, ie current
  82  *
  83  * The interrupt must handle 2 special cases:
  84  * - cr3 is changed before %%esp, ie. it cannot use current->{active_,}mm.
  85  * - the cpu performs speculative tlb reads, i.e. even if the cpu only
  86  *   runs in kernel space, the cpu could load tlb entries for user space
  87  *   pages.
  88  *
  89  * The good news is that cpu_tlbstate is local to each cpu, no
  90  * write/read ordering problems.
  91  */
  92
  93 /*
  94  * TLB flush funcation:
  95  * 1) Flush the tlb entries if the cpu uses the mm that's being flushed.
  96  * 2) Leave the mm if we are in the lazy tlb mode.
  97  */
  98 static void flush_tlb_func(void *info)
  99 {
 100         struct flush_tlb_info *f = info;
 101
 102         inc_irq_stat(irq_tlb_count);
 103
 104         if (f->flush_mm != this_cpu_read(cpu_tlbstate.active_mm))
 105                 return;
 106         if (!f->flush_end)
 107                 f->flush_end = f->flush_start + PAGE_SIZE;
 108
 109         count_vm_tlb_event(NR_TLB_REMOTE_FLUSH_RECEIVED);
 110         if (this_cpu_read(cpu_tlbstate.state) == TLBSTATE_OK) {
 111                 if (f->flush_end == TLB_FLUSH_ALL) {
 112                         local_flush_tlb();
 113                         trace_tlb_flush(TLB_REMOTE_SHOOTDOWN, TLB_FLUSH_ALL);
 114                 } else {
 115                         unsigned long addr;
 116                         unsigned long nr_pages =
 117                                 f->flush_end - f->flush_start / PAGE_SIZE;
 118                         addr = f->flush_start;
 119                         while (addr < f->flush_end) {
 120                                 __flush_tlb_single(addr);
 121                                 addr += PAGE_SIZE;
 122                         }
 123                         trace_tlb_flush(TLB_REMOTE_SHOOTDOWN, nr_pages);
 124                 }
 125         } else
 126                 leave_mm(smp_processor_id());
 127
 128 }
 129
 130 void native_flush_tlb_others(const struct cpumask *cpumask,
 131                                  struct mm_struct *mm, unsigned long start,
 132                                  unsigned long end)
 133 {
 134         struct flush_tlb_info info;
 135         info.flush_mm = mm;
 136         info.flush_start = start;
 137         info.flush_end = end;
 138
 139         count_vm_tlb_event(NR_TLB_REMOTE_FLUSH);
 140         if (is_uv_system()) {
 141                 unsigned int cpu;
 142
 143                 cpu = smp_processor_id();
 144                 cpumask = uv_flush_tlb_others(cpumask, mm, start, end, cpu);
 145                 if (cpumask)
 146                         smp_call_function_many(cpumask, flush_tlb_func,
 147                                                                 &info, 1);
 148                 return;
 149         }
 150         smp_call_function_many(cpumask, flush_tlb_func, &info, 1);
 151 }
 152
 153 void flush_tlb_current_task(void)
 154 {
 155         struct mm_struct *mm = current->mm;
 156
 157         preempt_disable();
 158
 159         count_vm_tlb_event(NR_TLB_LOCAL_FLUSH_ALL);
 160         local_flush_tlb();
 161         trace_tlb_flush(TLB_LOCAL_SHOOTDOWN, TLB_FLUSH_ALL);
 162         if (cpumask_any_but(mm_cpumask(mm), smp_processor_id()) < nr_cpu_ids)
 163                 flush_tlb_others(mm_cpumask(mm), mm, 0UL, TLB_FLUSH_ALL);
 164         preempt_enable();
 165 }
 166
 167 /*
 168  * See Documentation/x86/tlb.txt for details.  We choose 33
 169  * because it is large enough to cover the vast majority (at
 170  * least 95%) of allocations, and is small enough that we are
 171  * confident it will not cause too much overhead.  Each single
 172  * flush is about 100 ns, so this caps the maximum overhead at
 173  * _about_ 3,000 ns.
 174  *
 175  * This is in units of pages.
 176  */
 177 unsigned long tlb_single_page_flush_ceiling = 33;
 178
 179 void flush_tlb_mm_range(struct mm_struct *mm, unsigned long start,
 180                                 unsigned long end, unsigned long vmflag)
 181 {
 182         unsigned long addr;
 183         /* do a global flush by default */
 184         unsigned long base_pages_to_flush = TLB_FLUSH_ALL;
 185
 186         preempt_disable();
 187         if (current->active_mm != mm)
 188                 goto out;
 189
 190         if (!current->mm) {
 191                 leave_mm(smp_processor_id());
 192                 goto out;
 193         }
 194
 195         if ((end != TLB_FLUSH_ALL) && !(vmflag & VM_HUGETLB))
 196                 base_pages_to_flush = (end - start) >> PAGE_SHIFT;
 197
 198         if (base_pages_to_flush > tlb_single_page_flush_ceiling) {
 199                 base_pages_to_flush = TLB_FLUSH_ALL;
 200                 count_vm_tlb_event(NR_TLB_LOCAL_FLUSH_ALL);
 201                 local_flush_tlb();
 202         } else {
 203                 /* flush range by one by one 'invlpg' */
 204                 for (addr = start; addr < end;  addr += PAGE_SIZE) {
 205                         count_vm_tlb_event(NR_TLB_LOCAL_FLUSH_ONE);
 206                         __flush_tlb_single(addr);
 207                 }
 208         }
 209         trace_tlb_flush(TLB_LOCAL_MM_SHOOTDOWN, base_pages_to_flush);
 210 out:
 211         if (base_pages_to_flush == TLB_FLUSH_ALL) {
 212                 start = 0UL;
 213                 end = TLB_FLUSH_ALL;
 214         }
 215         if (cpumask_any_but(mm_cpumask(mm), smp_processor_id()) < nr_cpu_ids)
 216                 flush_tlb_others(mm_cpumask(mm), mm, start, end);
 217         preempt_enable();
 218 }
 219
 220 void flush_tlb_page(struct vm_area_struct *vma, unsigned long start)
 221 {
 222         struct mm_struct *mm = vma->vm_mm;
 223
 224         preempt_disable();
 225
 226         if (current->active_mm == mm) {
 227                 if (current->mm)
 228                         __flush_tlb_one(start);
 229                 else
 230                         leave_mm(smp_processor_id());
 231         }
 232
 233         if (cpumask_any_but(mm_cpumask(mm), smp_processor_id()) < nr_cpu_ids)
 234                 flush_tlb_others(mm_cpumask(mm), mm, start, 0UL);
 235
 236         preempt_enable();
 237 }
 238
 239 static void do_flush_tlb_all(void *info)
 240 {
 241         count_vm_tlb_event(NR_TLB_REMOTE_FLUSH_RECEIVED);
 242         __flush_tlb_all();
 243         if (this_cpu_read(cpu_tlbstate.state) == TLBSTATE_LAZY)
 244                 leave_mm(smp_processor_id());
 245 }
 246
 247 void flush_tlb_all(void)
 248 {
 249         count_vm_tlb_event(NR_TLB_REMOTE_FLUSH);
 250         on_each_cpu(do_flush_tlb_all, NULL, 1);
 251 }
 252
 253 static void do_kernel_range_flush(void *info)
 254 {
 255         struct flush_tlb_info *f = info;
 256         unsigned long addr;
 257
 258         /* flush range by one by one 'invlpg' */
 259         for (addr = f->flush_start; addr < f->flush_end; addr += PAGE_SIZE)
 260                 __flush_tlb_single(addr);
 261 }
 262
 263 void flush_tlb_kernel_range(unsigned long start, unsigned long end)
 264 {
 265
 266         /* Balance as user space task's flush, a bit conservative */
 267         if (end == TLB_FLUSH_ALL ||
 268             (end - start) > tlb_single_page_flush_ceiling * PAGE_SIZE) {
 269                 on_each_cpu(do_flush_tlb_all, NULL, 1);
 270         } else {
 271                 struct flush_tlb_info info;
 272                 info.flush_start = start;
 273                 info.flush_end = end;
 274                 on_each_cpu(do_kernel_range_flush, &info, 1);
 275         }
 276 }
 277
 278 static ssize_t tlbflush_read_file(struct file *file, char __user *user_buf,
 279                              size_t count, loff_t *ppos)
 280 {
 281         char buf[32];
 282         unsigned int len;
 283
 284         len = sprintf(buf, "%ld\n", tlb_single_page_flush_ceiling);
 285         return simple_read_from_buffer(user_buf, count, ppos, buf, len);
 286 }
 287
 288 static ssize_t tlbflush_write_file(struct file *file,
 289                  const char __user *user_buf, size_t count, loff_t *ppos)
 290 {
 291         char buf[32];
 292         ssize_t len;
 293         int ceiling;
 294
 295         len = min(count, sizeof(buf) - 1);
 296         if (copy_from_user(buf, user_buf, len))
 297                 return -EFAULT;
 298
 299         buf[len] = '\0';
 300         if (kstrtoint(buf, 0, &ceiling))
 301                 return -EINVAL;
 302
 303         if (ceiling < 0)
 304                 return -EINVAL;
 305
 306         tlb_single_page_flush_ceiling = ceiling;
 307         return count;
 308 }
 309
 310 static const struct file_operations fops_tlbflush = {
 311         .read = tlbflush_read_file,
 312         .write = tlbflush_write_file,
 313         .llseek = default_llseek,
 314 };
 315
 316 static int __init create_tlb_single_page_flush_ceiling(void)
 317 {
 318         debugfs_create_file("tlb_single_page_flush_ceiling", S_IRUSR | S_IWUSR,
 319                             arch_debugfs_dir, NULL, &fops_tlbflush);
 320         return 0;
 321 }
 322 late_initcall(create_tlb_single_page_flush_ceiling);