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 } 53 } 54 EXPORT_SYMBOL_GPL(leave_mm); 55 56 /* 57 * The flush IPI assumes that a thread switch happens in this order: 58 * [cpu0: the cpu that switches] 59 * 1) switch_mm() either 1a) or 1b) 60 * 1a) thread switch to a different mm 61 * 1a1) set cpu_tlbstate to TLBSTATE_OK 62 * Now the tlb flush NMI handler flush_tlb_func won't call leave_mm 63 * if cpu0 was in lazy tlb mode. 64 * 1a2) update cpu active_mm 65 * Now cpu0 accepts tlb flushes for the new mm. 66 * 1a3) cpu_set(cpu, new_mm->cpu_vm_mask); 67 * Now the other cpus will send tlb flush ipis. 68 * 1a4) change cr3. 69 * 1a5) cpu_clear(cpu, old_mm->cpu_vm_mask); 70 * Stop ipi delivery for the old mm. This is not synchronized with 71 * the other cpus, but flush_tlb_func ignore flush ipis for the wrong 72 * mm, and in the worst case we perform a superfluous tlb flush. 73 * 1b) thread switch without mm change 74 * cpu active_mm is correct, cpu0 already handles flush ipis. 75 * 1b1) set cpu_tlbstate to TLBSTATE_OK 76 * 1b2) test_and_set the cpu bit in cpu_vm_mask. 77 * Atomically set the bit [other cpus will start sending flush ipis], 78 * and test the bit. 79 * 1b3) if the bit was 0: leave_mm was called, flush the tlb. 80 * 2) switch %%esp, ie current 81 * 82 * The interrupt must handle 2 special cases: 83 * - cr3 is changed before %%esp, ie. it cannot use current->{active_,}mm. 84 * - the cpu performs speculative tlb reads, i.e. even if the cpu only 85 * runs in kernel space, the cpu could load tlb entries for user space 86 * pages. 87 * 88 * The good news is that cpu_tlbstate is local to each cpu, no 89 * write/read ordering problems. 90 */ 91 92 /* 93 * TLB flush funcation: 94 * 1) Flush the tlb entries if the cpu uses the mm that's being flushed. 95 * 2) Leave the mm if we are in the lazy tlb mode. 96 */ 97 static void flush_tlb_func(void *info) 98 { 99 struct flush_tlb_info *f = info; 100 101 inc_irq_stat(irq_tlb_count); 102 103 if (f->flush_mm != this_cpu_read(cpu_tlbstate.active_mm)) 104 return; 105 106 count_vm_tlb_event(NR_TLB_REMOTE_FLUSH_RECEIVED); 107 if (this_cpu_read(cpu_tlbstate.state) == TLBSTATE_OK) { 108 if (f->flush_end == TLB_FLUSH_ALL) 109 local_flush_tlb(); 110 else if (!f->flush_end) 111 __flush_tlb_single(f->flush_start); 112 else { 113 unsigned long addr; 114 addr = f->flush_start; 115 while (addr < f->flush_end) { 116 __flush_tlb_single(addr); 117 addr += PAGE_SIZE; 118 } 119 } 120 } else 121 leave_mm(smp_processor_id()); 122 123 } 124 125 void native_flush_tlb_others(const struct cpumask *cpumask, 126 struct mm_struct *mm, unsigned long start, 127 unsigned long end) 128 { 129 struct flush_tlb_info info; 130 info.flush_mm = mm; 131 info.flush_start = start; 132 info.flush_end = end; 133 134 count_vm_tlb_event(NR_TLB_REMOTE_FLUSH); 135 if (is_uv_system()) { 136 unsigned int cpu; 137 138 cpu = smp_processor_id(); 139 cpumask = uv_flush_tlb_others(cpumask, mm, start, end, cpu); 140 if (cpumask) 141 smp_call_function_many(cpumask, flush_tlb_func, 142 &info, 1); 143 return; 144 } 145 smp_call_function_many(cpumask, flush_tlb_func, &info, 1); 146 } 147 148 void flush_tlb_current_task(void) 149 { 150 struct mm_struct *mm = current->mm; 151 152 preempt_disable(); 153 154 count_vm_tlb_event(NR_TLB_LOCAL_FLUSH_ALL); 155 local_flush_tlb(); 156 if (cpumask_any_but(mm_cpumask(mm), smp_processor_id()) < nr_cpu_ids) 157 flush_tlb_others(mm_cpumask(mm), mm, 0UL, TLB_FLUSH_ALL); 158 preempt_enable(); 159 } 160 161 void flush_tlb_mm_range(struct mm_struct *mm, unsigned long start, 162 unsigned long end, unsigned long vmflag) 163 { 164 unsigned long addr; 165 unsigned act_entries, tlb_entries = 0; 166 unsigned long nr_base_pages; 167 168 preempt_disable(); 169 if (current->active_mm != mm) 170 goto flush_all; 171 172 if (!current->mm) { 173 leave_mm(smp_processor_id()); 174 goto flush_all; 175 } 176 177 if (end == TLB_FLUSH_ALL || tlb_flushall_shift == -1 178 || vmflag & VM_HUGETLB) { 179 local_flush_tlb(); 180 goto flush_all; 181 } 182 183 /* In modern CPU, last level tlb used for both data/ins */ 184 if (vmflag & VM_EXEC) 185 tlb_entries = tlb_lli_4k[ENTRIES]; 186 else 187 tlb_entries = tlb_lld_4k[ENTRIES]; 188 189 /* Assume all of TLB entries was occupied by this task */ 190 act_entries = tlb_entries >> tlb_flushall_shift; 191 act_entries = mm->total_vm > act_entries ? act_entries : mm->total_vm; 192 nr_base_pages = (end - start) >> PAGE_SHIFT; 193 194 /* tlb_flushall_shift is on balance point, details in commit log */ 195 if (nr_base_pages > act_entries) { 196 count_vm_tlb_event(NR_TLB_LOCAL_FLUSH_ALL); 197 local_flush_tlb(); 198 } else { 199 /* flush range by one by one 'invlpg' */ 200 for (addr = start; addr < end; addr += PAGE_SIZE) { 201 count_vm_tlb_event(NR_TLB_LOCAL_FLUSH_ONE); 202 __flush_tlb_single(addr); 203 } 204 205 if (cpumask_any_but(mm_cpumask(mm), 206 smp_processor_id()) < nr_cpu_ids) 207 flush_tlb_others(mm_cpumask(mm), mm, start, end); 208 preempt_enable(); 209 return; 210 } 211 212 flush_all: 213 if (cpumask_any_but(mm_cpumask(mm), smp_processor_id()) < nr_cpu_ids) 214 flush_tlb_others(mm_cpumask(mm), mm, 0UL, TLB_FLUSH_ALL); 215 preempt_enable(); 216 } 217 218 void flush_tlb_page(struct vm_area_struct *vma, unsigned long start) 219 { 220 struct mm_struct *mm = vma->vm_mm; 221 222 preempt_disable(); 223 224 if (current->active_mm == mm) { 225 if (current->mm) 226 __flush_tlb_one(start); 227 else 228 leave_mm(smp_processor_id()); 229 } 230 231 if (cpumask_any_but(mm_cpumask(mm), smp_processor_id()) < nr_cpu_ids) 232 flush_tlb_others(mm_cpumask(mm), mm, start, 0UL); 233 234 preempt_enable(); 235 } 236 237 static void do_flush_tlb_all(void *info) 238 { 239 count_vm_tlb_event(NR_TLB_REMOTE_FLUSH_RECEIVED); 240 __flush_tlb_all(); 241 if (this_cpu_read(cpu_tlbstate.state) == TLBSTATE_LAZY) 242 leave_mm(smp_processor_id()); 243 } 244 245 void flush_tlb_all(void) 246 { 247 count_vm_tlb_event(NR_TLB_REMOTE_FLUSH); 248 on_each_cpu(do_flush_tlb_all, NULL, 1); 249 } 250 251 static void do_kernel_range_flush(void *info) 252 { 253 struct flush_tlb_info *f = info; 254 unsigned long addr; 255 256 /* flush range by one by one 'invlpg' */ 257 for (addr = f->flush_start; addr < f->flush_end; addr += PAGE_SIZE) 258 __flush_tlb_single(addr); 259 } 260 261 void flush_tlb_kernel_range(unsigned long start, unsigned long end) 262 { 263 unsigned act_entries; 264 struct flush_tlb_info info; 265 266 /* In modern CPU, last level tlb used for both data/ins */ 267 act_entries = tlb_lld_4k[ENTRIES]; 268 269 /* Balance as user space task's flush, a bit conservative */ 270 if (end == TLB_FLUSH_ALL || tlb_flushall_shift == -1 || 271 (end - start) >> PAGE_SHIFT > act_entries >> tlb_flushall_shift) 272 273 on_each_cpu(do_flush_tlb_all, NULL, 1); 274 else { 275 info.flush_start = start; 276 info.flush_end = end; 277 on_each_cpu(do_kernel_range_flush, &info, 1); 278 } 279 } 280 281 #ifdef CONFIG_DEBUG_TLBFLUSH 282 static ssize_t tlbflush_read_file(struct file *file, char __user *user_buf, 283 size_t count, loff_t *ppos) 284 { 285 char buf[32]; 286 unsigned int len; 287 288 len = sprintf(buf, "%hd\n", tlb_flushall_shift); 289 return simple_read_from_buffer(user_buf, count, ppos, buf, len); 290 } 291 292 static ssize_t tlbflush_write_file(struct file *file, 293 const char __user *user_buf, size_t count, loff_t *ppos) 294 { 295 char buf[32]; 296 ssize_t len; 297 s8 shift; 298 299 len = min(count, sizeof(buf) - 1); 300 if (copy_from_user(buf, user_buf, len)) 301 return -EFAULT; 302 303 buf[len] = '\0'; 304 if (kstrtos8(buf, 0, &shift)) 305 return -EINVAL; 306 307 if (shift < -1 || shift >= BITS_PER_LONG) 308 return -EINVAL; 309 310 tlb_flushall_shift = shift; 311 return count; 312 } 313 314 static const struct file_operations fops_tlbflush = { 315 .read = tlbflush_read_file, 316 .write = tlbflush_write_file, 317 .llseek = default_llseek, 318 }; 319 320 static int __init create_tlb_flushall_shift(void) 321 { 322 debugfs_create_file("tlb_flushall_shift", S_IRUSR | S_IWUSR, 323 arch_debugfs_dir, NULL, &fops_tlbflush); 324 return 0; 325 } 326 late_initcall(create_tlb_flushall_shift); 327 #endif 328