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
   9 #include <asm/tlbflush.h>
  10 #include <asm/mmu_context.h>
  11 #include <asm/cache.h>
  12 #include <asm/apic.h>
  13 #include <asm/uv/uv.h>
  14
  15 DEFINE_PER_CPU_SHARED_ALIGNED(struct tlb_state, cpu_tlbstate)
  16                         = { &init_mm, 0, };
  17
  18 /*
  19  *      Smarter SMP flushing macros.
  20  *              c/o Linus Torvalds.
  21  *
  22  *      These mean you can really definitely utterly forget about
  23  *      writing to user space from interrupts. (Its not allowed anyway).
  24  *
  25  *      Optimizations Manfred Spraul <manfred@colorfullife.com>
  26  *
  27  *      More scalable flush, from Andi Kleen
  28  *
  29  *      To avoid global state use 8 different call vectors.
  30  *      Each CPU uses a specific vector to trigger flushes on other
  31  *      CPUs. Depending on the received vector the target CPUs look into
  32  *      the right array slot for the flush data.
  33  *
  34  *      With more than 8 CPUs they are hashed to the 8 available
  35  *      vectors. The limited global vector space forces us to this right now.
  36  *      In future when interrupts are split into per CPU domains this could be
  37  *      fixed, at the cost of triggering multiple IPIs in some cases.
  38  */
  39
  40 union smp_flush_state {
  41         struct {
  42                 struct mm_struct *flush_mm;
  43                 unsigned long flush_va;
  44                 spinlock_t tlbstate_lock;
  45                 DECLARE_BITMAP(flush_cpumask, NR_CPUS);
  46         };
  47         char pad[INTERNODE_CACHE_BYTES];
  48 } ____cacheline_internodealigned_in_smp;
  49
  50 /* State is put into the per CPU data section, but padded
  51    to a full cache line because other CPUs can access it and we don't
  52    want false sharing in the per cpu data segment. */
  53 static union smp_flush_state flush_state[NUM_INVALIDATE_TLB_VECTORS];
  54
  55 /*
  56  * We cannot call mmdrop() because we are in interrupt context,
  57  * instead update mm->cpu_vm_mask.
  58  */
  59 void leave_mm(int cpu)
  60 {
  61         if (percpu_read(cpu_tlbstate.state) == TLBSTATE_OK)
  62                 BUG();
  63         cpumask_clear_cpu(cpu,
  64                           mm_cpumask(percpu_read(cpu_tlbstate.active_mm)));
  65         load_cr3(swapper_pg_dir);
  66 }
  67 EXPORT_SYMBOL_GPL(leave_mm);
  68
  69 /*
  70  *
  71  * The flush IPI assumes that a thread switch happens in this order:
  72  * [cpu0: the cpu that switches]
  73  * 1) switch_mm() either 1a) or 1b)
  74  * 1a) thread switch to a different mm
  75  * 1a1) cpu_clear(cpu, old_mm->cpu_vm_mask);
  76  *      Stop ipi delivery for the old mm. This is not synchronized with
  77  *      the other cpus, but smp_invalidate_interrupt ignore flush ipis
  78  *      for the wrong mm, and in the worst case we perform a superfluous
  79  *      tlb flush.
  80  * 1a2) set cpu mmu_state to TLBSTATE_OK
  81  *      Now the smp_invalidate_interrupt won't call leave_mm if cpu0
  82  *      was in lazy tlb mode.
  83  * 1a3) update cpu active_mm
  84  *      Now cpu0 accepts tlb flushes for the new mm.
  85  * 1a4) cpu_set(cpu, new_mm->cpu_vm_mask);
  86  *      Now the other cpus will send tlb flush ipis.
  87  * 1a4) change cr3.
  88  * 1b) thread switch without mm change
  89  *      cpu active_mm is correct, cpu0 already handles
  90  *      flush ipis.
  91  * 1b1) set cpu mmu_state to TLBSTATE_OK
  92  * 1b2) test_and_set the cpu bit in cpu_vm_mask.
  93  *      Atomically set the bit [other cpus will start sending flush ipis],
  94  *      and test the bit.
  95  * 1b3) if the bit was 0: leave_mm was called, flush the tlb.
  96  * 2) switch %%esp, ie current
  97  *
  98  * The interrupt must handle 2 special cases:
  99  * - cr3 is changed before %%esp, ie. it cannot use current->{active_,}mm.
 100  * - the cpu performs speculative tlb reads, i.e. even if the cpu only
 101  *   runs in kernel space, the cpu could load tlb entries for user space
 102  *   pages.
 103  *
 104  * The good news is that cpu mmu_state is local to each cpu, no
 105  * write/read ordering problems.
 106  */
 107
 108 /*
 109  * TLB flush IPI:
 110  *
 111  * 1) Flush the tlb entries if the cpu uses the mm that's being flushed.
 112  * 2) Leave the mm if we are in the lazy tlb mode.
 113  *
 114  * Interrupts are disabled.
 115  */
 116
 117 /*
 118  * FIXME: use of asmlinkage is not consistent.  On x86_64 it's noop
 119  * but still used for documentation purpose but the usage is slightly
 120  * inconsistent.  On x86_32, asmlinkage is regparm(0) but interrupt
 121  * entry calls in with the first parameter in %eax.  Maybe define
 122  * intrlinkage?
 123  */
 124 #ifdef CONFIG_X86_64
 125 asmlinkage
 126 #endif
 127 void smp_invalidate_interrupt(struct pt_regs *regs)
 128 {
 129         unsigned int cpu;
 130         unsigned int sender;
 131         union smp_flush_state *f;
 132
 133         cpu = smp_processor_id();
 134         /*
 135          * orig_rax contains the negated interrupt vector.
 136          * Use that to determine where the sender put the data.
 137          */
 138         sender = ~regs->orig_ax - INVALIDATE_TLB_VECTOR_START;
 139         f = &flush_state[sender];
 140
 141         if (!cpumask_test_cpu(cpu, to_cpumask(f->flush_cpumask)))
 142                 goto out;
 143                 /*
 144                  * This was a BUG() but until someone can quote me the
 145                  * line from the intel manual that guarantees an IPI to
 146                  * multiple CPUs is retried _only_ on the erroring CPUs
 147                  * its staying as a return
 148                  *
 149                  * BUG();
 150                  */
 151
 152         if (f->flush_mm == percpu_read(cpu_tlbstate.active_mm)) {
 153                 if (percpu_read(cpu_tlbstate.state) == TLBSTATE_OK) {
 154                         if (f->flush_va == TLB_FLUSH_ALL)
 155                                 local_flush_tlb();
 156                         else
 157                                 __flush_tlb_one(f->flush_va);
 158                 } else
 159                         leave_mm(cpu);
 160         }
 161 out:
 162         ack_APIC_irq();
 163         smp_mb__before_clear_bit();
 164         cpumask_clear_cpu(cpu, to_cpumask(f->flush_cpumask));
 165         smp_mb__after_clear_bit();
 166         inc_irq_stat(irq_tlb_count);
 167 }
 168
 169 static void flush_tlb_others_ipi(const struct cpumask *cpumask,
 170                                  struct mm_struct *mm, unsigned long va)
 171 {
 172         unsigned int sender;
 173         union smp_flush_state *f;
 174
 175         /* Caller has disabled preemption */
 176         sender = smp_processor_id() % NUM_INVALIDATE_TLB_VECTORS;
 177         f = &flush_state[sender];
 178
 179         /*
 180          * Could avoid this lock when
 181          * num_online_cpus() <= NUM_INVALIDATE_TLB_VECTORS, but it is
 182          * probably not worth checking this for a cache-hot lock.
 183          */
 184         spin_lock(&f->tlbstate_lock);
 185
 186         f->flush_mm = mm;
 187         f->flush_va = va;
 188         if (cpumask_andnot(to_cpumask(f->flush_cpumask), cpumask, cpumask_of(smp_processor_id()))) {
 189                 /*
 190                  * We have to send the IPI only to
 191                  * CPUs affected.
 192                  */
 193                 apic->send_IPI_mask(to_cpumask(f->flush_cpumask),
 194                               INVALIDATE_TLB_VECTOR_START + sender);
 195
 196                 while (!cpumask_empty(to_cpumask(f->flush_cpumask)))
 197                         cpu_relax();
 198         }
 199
 200         f->flush_mm = NULL;
 201         f->flush_va = 0;
 202         spin_unlock(&f->tlbstate_lock);
 203 }
 204
 205 void native_flush_tlb_others(const struct cpumask *cpumask,
 206                              struct mm_struct *mm, unsigned long va)
 207 {
 208         if (is_uv_system()) {
 209                 unsigned int cpu;
 210
 211                 cpu = get_cpu();
 212                 cpumask = uv_flush_tlb_others(cpumask, mm, va, cpu);
 213                 if (cpumask)
 214                         flush_tlb_others_ipi(cpumask, mm, va);
 215                 put_cpu();
 216                 return;
 217         }
 218         flush_tlb_others_ipi(cpumask, mm, va);
 219 }
 220
 221 static int __cpuinit init_smp_flush(void)
 222 {
 223         int i;
 224
 225         for (i = 0; i < ARRAY_SIZE(flush_state); i++)
 226                 spin_lock_init(&flush_state[i].tlbstate_lock);
 227
 228         return 0;
 229 }
 230 core_initcall(init_smp_flush);
 231
 232 void flush_tlb_current_task(void)
 233 {
 234         struct mm_struct *mm = current->mm;
 235
 236         preempt_disable();
 237
 238         local_flush_tlb();
 239         if (cpumask_any_but(mm_cpumask(mm), smp_processor_id()) < nr_cpu_ids)
 240                 flush_tlb_others(mm_cpumask(mm), mm, TLB_FLUSH_ALL);
 241         preempt_enable();
 242 }
 243
 244 void flush_tlb_mm(struct mm_struct *mm)
 245 {
 246         preempt_disable();
 247
 248         if (current->active_mm == mm) {
 249                 if (current->mm)
 250                         local_flush_tlb();
 251                 else
 252                         leave_mm(smp_processor_id());
 253         }
 254         if (cpumask_any_but(mm_cpumask(mm), smp_processor_id()) < nr_cpu_ids)
 255                 flush_tlb_others(mm_cpumask(mm), mm, TLB_FLUSH_ALL);
 256
 257         preempt_enable();
 258 }
 259
 260 void flush_tlb_page(struct vm_area_struct *vma, unsigned long va)
 261 {
 262         struct mm_struct *mm = vma->vm_mm;
 263
 264         preempt_disable();
 265
 266         if (current->active_mm == mm) {
 267                 if (current->mm)
 268                         __flush_tlb_one(va);
 269                 else
 270                         leave_mm(smp_processor_id());
 271         }
 272
 273         if (cpumask_any_but(mm_cpumask(mm), smp_processor_id()) < nr_cpu_ids)
 274                 flush_tlb_others(mm_cpumask(mm), mm, va);
 275
 276         preempt_enable();
 277 }
 278
 279 static void do_flush_tlb_all(void *info)
 280 {
 281         unsigned long cpu = smp_processor_id();
 282
 283         __flush_tlb_all();
 284         if (percpu_read(cpu_tlbstate.state) == TLBSTATE_LAZY)
 285                 leave_mm(cpu);
 286 }
 287
 288 void flush_tlb_all(void)
 289 {
 290         on_each_cpu(do_flush_tlb_all, NULL, 1);
 291 }