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
  16 DEFINE_PER_CPU_SHARED_ALIGNED(struct tlb_state, cpu_tlbstate)
  17                         = { &init_mm, 0, };
  18
  19 /*
  20  *      Smarter SMP flushing macros.
  21  *              c/o Linus Torvalds.
  22  *
  23  *      These mean you can really definitely utterly forget about
  24  *      writing to user space from interrupts. (Its not allowed anyway).
  25  *
  26  *      Optimizations Manfred Spraul <manfred@colorfullife.com>
  27  *
  28  *      More scalable flush, from Andi Kleen
  29  *
  30  *      To avoid global state use 8 different call vectors.
  31  *      Each CPU uses a specific vector to trigger flushes on other
  32  *      CPUs. Depending on the received vector the target CPUs look into
  33  *      the right array slot for the flush data.
  34  *
  35  *      With more than 8 CPUs they are hashed to the 8 available
  36  *      vectors. The limited global vector space forces us to this right now.
  37  *      In future when interrupts are split into per CPU domains this could be
  38  *      fixed, at the cost of triggering multiple IPIs in some cases.
  39  */
  40
  41 union smp_flush_state {
  42         struct {
  43                 struct mm_struct *flush_mm;
  44                 unsigned long flush_va;
  45                 raw_spinlock_t tlbstate_lock;
  46                 DECLARE_BITMAP(flush_cpumask, NR_CPUS);
  47         };
  48         char pad[INTERNODE_CACHE_BYTES];
  49 } ____cacheline_internodealigned_in_smp;
  50
  51 /* State is put into the per CPU data section, but padded
  52    to a full cache line because other CPUs can access it and we don't
  53    want false sharing in the per cpu data segment. */
  54 static union smp_flush_state flush_state[NUM_INVALIDATE_TLB_VECTORS];
  55
  56 static DEFINE_PER_CPU_READ_MOSTLY(int, tlb_vector_offset);
  57
  58 /*
  59  * We cannot call mmdrop() because we are in interrupt context,
  60  * instead update mm->cpu_vm_mask.
  61  */
  62 void leave_mm(int cpu)
  63 {
  64         if (percpu_read(cpu_tlbstate.state) == TLBSTATE_OK)
  65                 BUG();
  66         cpumask_clear_cpu(cpu,
  67                           mm_cpumask(percpu_read(cpu_tlbstate.active_mm)));
  68         load_cr3(swapper_pg_dir);
  69 }
  70 EXPORT_SYMBOL_GPL(leave_mm);
  71
  72 /*
  73  *
  74  * The flush IPI assumes that a thread switch happens in this order:
  75  * [cpu0: the cpu that switches]
  76  * 1) switch_mm() either 1a) or 1b)
  77  * 1a) thread switch to a different mm
  78  * 1a1) cpu_clear(cpu, old_mm->cpu_vm_mask);
  79  *      Stop ipi delivery for the old mm. This is not synchronized with
  80  *      the other cpus, but smp_invalidate_interrupt ignore flush ipis
  81  *      for the wrong mm, and in the worst case we perform a superfluous
  82  *      tlb flush.
  83  * 1a2) set cpu mmu_state to TLBSTATE_OK
  84  *      Now the smp_invalidate_interrupt won't call leave_mm if cpu0
  85  *      was in lazy tlb mode.
  86  * 1a3) update cpu active_mm
  87  *      Now cpu0 accepts tlb flushes for the new mm.
  88  * 1a4) cpu_set(cpu, new_mm->cpu_vm_mask);
  89  *      Now the other cpus will send tlb flush ipis.
  90  * 1a4) change cr3.
  91  * 1b) thread switch without mm change
  92  *      cpu active_mm is correct, cpu0 already handles
  93  *      flush ipis.
  94  * 1b1) set cpu mmu_state to TLBSTATE_OK
  95  * 1b2) test_and_set the cpu bit in cpu_vm_mask.
  96  *      Atomically set the bit [other cpus will start sending flush ipis],
  97  *      and test the bit.
  98  * 1b3) if the bit was 0: leave_mm was called, flush the tlb.
  99  * 2) switch %%esp, ie current
 100  *
 101  * The interrupt must handle 2 special cases:
 102  * - cr3 is changed before %%esp, ie. it cannot use current->{active_,}mm.
 103  * - the cpu performs speculative tlb reads, i.e. even if the cpu only
 104  *   runs in kernel space, the cpu could load tlb entries for user space
 105  *   pages.
 106  *
 107  * The good news is that cpu mmu_state is local to each cpu, no
 108  * write/read ordering problems.
 109  */
 110
 111 /*
 112  * TLB flush IPI:
 113  *
 114  * 1) Flush the tlb entries if the cpu uses the mm that's being flushed.
 115  * 2) Leave the mm if we are in the lazy tlb mode.
 116  *
 117  * Interrupts are disabled.
 118  */
 119
 120 /*
 121  * FIXME: use of asmlinkage is not consistent.  On x86_64 it's noop
 122  * but still used for documentation purpose but the usage is slightly
 123  * inconsistent.  On x86_32, asmlinkage is regparm(0) but interrupt
 124  * entry calls in with the first parameter in %eax.  Maybe define
 125  * intrlinkage?
 126  */
 127 #ifdef CONFIG_X86_64
 128 asmlinkage
 129 #endif
 130 void smp_invalidate_interrupt(struct pt_regs *regs)
 131 {
 132         unsigned int cpu;
 133         unsigned int sender;
 134         union smp_flush_state *f;
 135
 136         cpu = smp_processor_id();
 137         /*
 138          * orig_rax contains the negated interrupt vector.
 139          * Use that to determine where the sender put the data.
 140          */
 141         sender = ~regs->orig_ax - INVALIDATE_TLB_VECTOR_START;
 142         f = &flush_state[sender];
 143
 144         if (!cpumask_test_cpu(cpu, to_cpumask(f->flush_cpumask)))
 145                 goto out;
 146                 /*
 147                  * This was a BUG() but until someone can quote me the
 148                  * line from the intel manual that guarantees an IPI to
 149                  * multiple CPUs is retried _only_ on the erroring CPUs
 150                  * its staying as a return
 151                  *
 152                  * BUG();
 153                  */
 154
 155         if (f->flush_mm == percpu_read(cpu_tlbstate.active_mm)) {
 156                 if (percpu_read(cpu_tlbstate.state) == TLBSTATE_OK) {
 157                         if (f->flush_va == TLB_FLUSH_ALL)
 158                                 local_flush_tlb();
 159                         else
 160                                 __flush_tlb_one(f->flush_va);
 161                 } else
 162                         leave_mm(cpu);
 163         }
 164 out:
 165         ack_APIC_irq();
 166         smp_mb__before_clear_bit();
 167         cpumask_clear_cpu(cpu, to_cpumask(f->flush_cpumask));
 168         smp_mb__after_clear_bit();
 169         inc_irq_stat(irq_tlb_count);
 170 }
 171
 172 static void flush_tlb_others_ipi(const struct cpumask *cpumask,
 173                                  struct mm_struct *mm, unsigned long va)
 174 {
 175         unsigned int sender;
 176         union smp_flush_state *f;
 177
 178         /* Caller has disabled preemption */
 179         sender = this_cpu_read(tlb_vector_offset);
 180         f = &flush_state[sender];
 181
 182         /*
 183          * Could avoid this lock when
 184          * num_online_cpus() <= NUM_INVALIDATE_TLB_VECTORS, but it is
 185          * probably not worth checking this for a cache-hot lock.
 186          */
 187         raw_spin_lock(&f->tlbstate_lock);
 188
 189         f->flush_mm = mm;
 190         f->flush_va = va;
 191         if (cpumask_andnot(to_cpumask(f->flush_cpumask), cpumask, cpumask_of(smp_processor_id()))) {
 192                 /*
 193                  * We have to send the IPI only to
 194                  * CPUs affected.
 195                  */
 196                 apic->send_IPI_mask(to_cpumask(f->flush_cpumask),
 197                               INVALIDATE_TLB_VECTOR_START + sender);
 198
 199                 while (!cpumask_empty(to_cpumask(f->flush_cpumask)))
 200                         cpu_relax();
 201         }
 202
 203         f->flush_mm = NULL;
 204         f->flush_va = 0;
 205         raw_spin_unlock(&f->tlbstate_lock);
 206 }
 207
 208 void native_flush_tlb_others(const struct cpumask *cpumask,
 209                              struct mm_struct *mm, unsigned long va)
 210 {
 211         if (is_uv_system()) {
 212                 unsigned int cpu;
 213
 214                 cpu = get_cpu();
 215                 cpumask = uv_flush_tlb_others(cpumask, mm, va, cpu);
 216                 if (cpumask)
 217                         flush_tlb_others_ipi(cpumask, mm, va);
 218                 put_cpu();
 219                 return;
 220         }
 221         flush_tlb_others_ipi(cpumask, mm, va);
 222 }
 223
 224 static void __cpuinit calculate_tlb_offset(void)
 225 {
 226         int cpu, node, nr_node_vecs, idx = 0;
 227         /*
 228          * we are changing tlb_vector_offset for each CPU in runtime, but this
 229          * will not cause inconsistency, as the write is atomic under X86. we
 230          * might see more lock contentions in a short time, but after all CPU's
 231          * tlb_vector_offset are changed, everything should go normal
 232          *
 233          * Note: if NUM_INVALIDATE_TLB_VECTORS % nr_online_nodes !=0, we might
 234          * waste some vectors.
 235          **/
 236         if (nr_online_nodes > NUM_INVALIDATE_TLB_VECTORS)
 237                 nr_node_vecs = 1;
 238         else
 239                 nr_node_vecs = NUM_INVALIDATE_TLB_VECTORS/nr_online_nodes;
 240
 241         for_each_online_node(node) {
 242                 int node_offset = (idx % NUM_INVALIDATE_TLB_VECTORS) *
 243                         nr_node_vecs;
 244                 int cpu_offset = 0;
 245                 for_each_cpu(cpu, cpumask_of_node(node)) {
 246                         per_cpu(tlb_vector_offset, cpu) = node_offset +
 247                                 cpu_offset;
 248                         cpu_offset++;
 249                         cpu_offset = cpu_offset % nr_node_vecs;
 250                 }
 251                 idx++;
 252         }
 253 }
 254
 255 static int __cpuinit tlb_cpuhp_notify(struct notifier_block *n,
 256                 unsigned long action, void *hcpu)
 257 {
 258         switch (action & 0xf) {
 259         case CPU_ONLINE:
 260         case CPU_DEAD:
 261                 calculate_tlb_offset();
 262         }
 263         return NOTIFY_OK;
 264 }
 265
 266 static int __cpuinit init_smp_flush(void)
 267 {
 268         int i;
 269
 270         for (i = 0; i < ARRAY_SIZE(flush_state); i++)
 271                 raw_spin_lock_init(&flush_state[i].tlbstate_lock);
 272
 273         calculate_tlb_offset();
 274         hotcpu_notifier(tlb_cpuhp_notify, 0);
 275         return 0;
 276 }
 277 core_initcall(init_smp_flush);
 278
 279 void flush_tlb_current_task(void)
 280 {
 281         struct mm_struct *mm = current->mm;
 282
 283         preempt_disable();
 284
 285         local_flush_tlb();
 286         if (cpumask_any_but(mm_cpumask(mm), smp_processor_id()) < nr_cpu_ids)
 287                 flush_tlb_others(mm_cpumask(mm), mm, TLB_FLUSH_ALL);
 288         preempt_enable();
 289 }
 290
 291 void flush_tlb_mm(struct mm_struct *mm)
 292 {
 293         preempt_disable();
 294
 295         if (current->active_mm == mm) {
 296                 if (current->mm)
 297                         local_flush_tlb();
 298                 else
 299                         leave_mm(smp_processor_id());
 300         }
 301         if (cpumask_any_but(mm_cpumask(mm), smp_processor_id()) < nr_cpu_ids)
 302                 flush_tlb_others(mm_cpumask(mm), mm, TLB_FLUSH_ALL);
 303
 304         preempt_enable();
 305 }
 306
 307 void flush_tlb_page(struct vm_area_struct *vma, unsigned long va)
 308 {
 309         struct mm_struct *mm = vma->vm_mm;
 310
 311         preempt_disable();
 312
 313         if (current->active_mm == mm) {
 314                 if (current->mm)
 315                         __flush_tlb_one(va);
 316                 else
 317                         leave_mm(smp_processor_id());
 318         }
 319
 320         if (cpumask_any_but(mm_cpumask(mm), smp_processor_id()) < nr_cpu_ids)
 321                 flush_tlb_others(mm_cpumask(mm), mm, va);
 322
 323         preempt_enable();
 324 }
 325
 326 static void do_flush_tlb_all(void *info)
 327 {
 328         __flush_tlb_all();
 329         if (percpu_read(cpu_tlbstate.state) == TLBSTATE_LAZY)
 330                 leave_mm(smp_processor_id());
 331 }
 332
 333 void flush_tlb_all(void)
 334 {
 335         on_each_cpu(do_flush_tlb_all, NULL, 1);
 336 }