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_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_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_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 /* 162 * It can find out the THP large page, or 163 * HUGETLB page in tlb_flush when THP disabled 164 */ 165 static inline unsigned long has_large_page(struct mm_struct *mm, 166 unsigned long start, unsigned long end) 167 { 168 pgd_t *pgd; 169 pud_t *pud; 170 pmd_t *pmd; 171 unsigned long addr = ALIGN(start, HPAGE_SIZE); 172 for (; addr < end; addr += HPAGE_SIZE) { 173 pgd = pgd_offset(mm, addr); 174 if (likely(!pgd_none(*pgd))) { 175 pud = pud_offset(pgd, addr); 176 if (likely(!pud_none(*pud))) { 177 pmd = pmd_offset(pud, addr); 178 if (likely(!pmd_none(*pmd))) 179 if (pmd_large(*pmd)) 180 return addr; 181 } 182 } 183 } 184 return 0; 185 } 186 187 void flush_tlb_mm_range(struct mm_struct *mm, unsigned long start, 188 unsigned long end, unsigned long vmflag) 189 { 190 unsigned long addr; 191 unsigned act_entries, tlb_entries = 0; 192 193 preempt_disable(); 194 if (current->active_mm != mm) 195 goto flush_all; 196 197 if (!current->mm) { 198 leave_mm(smp_processor_id()); 199 goto flush_all; 200 } 201 202 if (end == TLB_FLUSH_ALL || tlb_flushall_shift == -1 203 || vmflag & VM_HUGETLB) { 204 local_flush_tlb(); 205 goto flush_all; 206 } 207 208 /* In modern CPU, last level tlb used for both data/ins */ 209 if (vmflag & VM_EXEC) 210 tlb_entries = tlb_lli_4k[ENTRIES]; 211 else 212 tlb_entries = tlb_lld_4k[ENTRIES]; 213 /* Assume all of TLB entries was occupied by this task */ 214 act_entries = mm->total_vm > tlb_entries ? tlb_entries : mm->total_vm; 215 216 /* tlb_flushall_shift is on balance point, details in commit log */ 217 if ((end - start) >> PAGE_SHIFT > act_entries >> tlb_flushall_shift) { 218 count_vm_event(NR_TLB_LOCAL_FLUSH_ALL); 219 local_flush_tlb(); 220 } else { 221 if (has_large_page(mm, start, end)) { 222 local_flush_tlb(); 223 goto flush_all; 224 } 225 /* flush range by one by one 'invlpg' */ 226 for (addr = start; addr < end; addr += PAGE_SIZE) { 227 count_vm_event(NR_TLB_LOCAL_FLUSH_ONE); 228 __flush_tlb_single(addr); 229 } 230 231 if (cpumask_any_but(mm_cpumask(mm), 232 smp_processor_id()) < nr_cpu_ids) 233 flush_tlb_others(mm_cpumask(mm), mm, start, end); 234 preempt_enable(); 235 return; 236 } 237 238 flush_all: 239 if (cpumask_any_but(mm_cpumask(mm), smp_processor_id()) < nr_cpu_ids) 240 flush_tlb_others(mm_cpumask(mm), mm, 0UL, TLB_FLUSH_ALL); 241 preempt_enable(); 242 } 243 244 void flush_tlb_page(struct vm_area_struct *vma, unsigned long start) 245 { 246 struct mm_struct *mm = vma->vm_mm; 247 248 preempt_disable(); 249 250 if (current->active_mm == mm) { 251 if (current->mm) 252 __flush_tlb_one(start); 253 else 254 leave_mm(smp_processor_id()); 255 } 256 257 if (cpumask_any_but(mm_cpumask(mm), smp_processor_id()) < nr_cpu_ids) 258 flush_tlb_others(mm_cpumask(mm), mm, start, 0UL); 259 260 preempt_enable(); 261 } 262 263 static void do_flush_tlb_all(void *info) 264 { 265 count_vm_event(NR_TLB_REMOTE_FLUSH_RECEIVED); 266 __flush_tlb_all(); 267 if (this_cpu_read(cpu_tlbstate.state) == TLBSTATE_LAZY) 268 leave_mm(smp_processor_id()); 269 } 270 271 void flush_tlb_all(void) 272 { 273 count_vm_event(NR_TLB_REMOTE_FLUSH); 274 on_each_cpu(do_flush_tlb_all, NULL, 1); 275 } 276 277 static void do_kernel_range_flush(void *info) 278 { 279 struct flush_tlb_info *f = info; 280 unsigned long addr; 281 282 /* flush range by one by one 'invlpg' */ 283 for (addr = f->flush_start; addr < f->flush_end; addr += PAGE_SIZE) 284 __flush_tlb_single(addr); 285 } 286 287 void flush_tlb_kernel_range(unsigned long start, unsigned long end) 288 { 289 unsigned act_entries; 290 struct flush_tlb_info info; 291 292 /* In modern CPU, last level tlb used for both data/ins */ 293 act_entries = tlb_lld_4k[ENTRIES]; 294 295 /* Balance as user space task's flush, a bit conservative */ 296 if (end == TLB_FLUSH_ALL || tlb_flushall_shift == -1 || 297 (end - start) >> PAGE_SHIFT > act_entries >> tlb_flushall_shift) 298 299 on_each_cpu(do_flush_tlb_all, NULL, 1); 300 else { 301 info.flush_start = start; 302 info.flush_end = end; 303 on_each_cpu(do_kernel_range_flush, &info, 1); 304 } 305 } 306 307 #ifdef CONFIG_DEBUG_TLBFLUSH 308 static ssize_t tlbflush_read_file(struct file *file, char __user *user_buf, 309 size_t count, loff_t *ppos) 310 { 311 char buf[32]; 312 unsigned int len; 313 314 len = sprintf(buf, "%hd\n", tlb_flushall_shift); 315 return simple_read_from_buffer(user_buf, count, ppos, buf, len); 316 } 317 318 static ssize_t tlbflush_write_file(struct file *file, 319 const char __user *user_buf, size_t count, loff_t *ppos) 320 { 321 char buf[32]; 322 ssize_t len; 323 s8 shift; 324 325 len = min(count, sizeof(buf) - 1); 326 if (copy_from_user(buf, user_buf, len)) 327 return -EFAULT; 328 329 buf[len] = '\0'; 330 if (kstrtos8(buf, 0, &shift)) 331 return -EINVAL; 332 333 if (shift < -1 || shift >= BITS_PER_LONG) 334 return -EINVAL; 335 336 tlb_flushall_shift = shift; 337 return count; 338 } 339 340 static const struct file_operations fops_tlbflush = { 341 .read = tlbflush_read_file, 342 .write = tlbflush_write_file, 343 .llseek = default_llseek, 344 }; 345 346 static int __init create_tlb_flushall_shift(void) 347 { 348 debugfs_create_file("tlb_flushall_shift", S_IRUSR | S_IWUSR, 349 arch_debugfs_dir, NULL, &fops_tlbflush); 350 return 0; 351 } 352 late_initcall(create_tlb_flushall_shift); 353 #endif 354