arch/sparc/mm/tsb.c

   1 /* arch/sparc64/mm/tsb.c
   2  *
   3  * Copyright (C) 2006, 2008 David S. Miller <davem@davemloft.net>
   4  */
   5
   6 #include <linux/kernel.h>
   7 #include <linux/preempt.h>
   8 #include <linux/slab.h>
   9 #include <asm/page.h>
  10 #include <asm/pgtable.h>
  11 #include <asm/mmu_context.h>
  12 #include <asm/setup.h>
  13 #include <asm/tsb.h>
  14 #include <asm/tlb.h>
  15 #include <asm/oplib.h>
  16
  17 extern struct tsb swapper_tsb[KERNEL_TSB_NENTRIES];
  18
  19 static inline unsigned long tsb_hash(unsigned long vaddr, unsigned long hash_shift, unsigned long nentries)
  20 {
  21         vaddr >>= hash_shift;
  22         return vaddr & (nentries - 1);
  23 }
  24
  25 static inline int tag_compare(unsigned long tag, unsigned long vaddr)
  26 {
  27         return (tag == (vaddr >> 22));
  28 }
  29
  30 static void flush_tsb_kernel_range_scan(unsigned long start, unsigned long end)
  31 {
  32         unsigned long idx;
  33
  34         for (idx = 0; idx < KERNEL_TSB_NENTRIES; idx++) {
  35                 struct tsb *ent = &swapper_tsb[idx];
  36                 unsigned long match = idx << 13;
  37
  38                 match |= (ent->tag << 22);
  39                 if (match >= start && match < end)
  40                         ent->tag = (1UL << TSB_TAG_INVALID_BIT);
  41         }
  42 }
  43
  44 /* TSB flushes need only occur on the processor initiating the address
  45  * space modification, not on each cpu the address space has run on.
  46  * Only the TLB flush needs that treatment.
  47  */
  48
  49 void flush_tsb_kernel_range(unsigned long start, unsigned long end)
  50 {
  51         unsigned long v;
  52
  53         if ((end - start) >> PAGE_SHIFT >= 2 * KERNEL_TSB_NENTRIES)
  54                 return flush_tsb_kernel_range_scan(start, end);
  55
  56         for (v = start; v < end; v += PAGE_SIZE) {
  57                 unsigned long hash = tsb_hash(v, PAGE_SHIFT,
  58                                               KERNEL_TSB_NENTRIES);
  59                 struct tsb *ent = &swapper_tsb[hash];
  60
  61                 if (tag_compare(ent->tag, v))
  62                         ent->tag = (1UL << TSB_TAG_INVALID_BIT);
  63         }
  64 }
  65
  66 static void __flush_tsb_one_entry(unsigned long tsb, unsigned long v,
  67                                   unsigned long hash_shift,
  68                                   unsigned long nentries)
  69 {
  70         unsigned long tag, ent, hash;
  71
  72         v &= ~0x1UL;
  73         hash = tsb_hash(v, hash_shift, nentries);
  74         ent = tsb + (hash * sizeof(struct tsb));
  75         tag = (v >> 22UL);
  76
  77         tsb_flush(ent, tag);
  78 }
  79
  80 static void __flush_tsb_one(struct tlb_batch *tb, unsigned long hash_shift,
  81                             unsigned long tsb, unsigned long nentries)
  82 {
  83         unsigned long i;
  84
  85         for (i = 0; i < tb->tlb_nr; i++)
  86                 __flush_tsb_one_entry(tsb, tb->vaddrs[i], hash_shift, nentries);
  87 }
  88
  89 void flush_tsb_user(struct tlb_batch *tb)
  90 {
  91         struct mm_struct *mm = tb->mm;
  92         unsigned long nentries, base, flags;
  93
  94         spin_lock_irqsave(&mm->context.lock, flags);
  95
  96         if (!tb->huge) {
  97                 base = (unsigned long) mm->context.tsb_block[MM_TSB_BASE].tsb;
  98                 nentries = mm->context.tsb_block[MM_TSB_BASE].tsb_nentries;
  99                 if (tlb_type == cheetah_plus || tlb_type == hypervisor)
 100                         base = __pa(base);
 101                 __flush_tsb_one(tb, PAGE_SHIFT, base, nentries);
 102         }
 103 #if defined(CONFIG_HUGETLB_PAGE) || defined(CONFIG_TRANSPARENT_HUGEPAGE)
 104         if (tb->huge && mm->context.tsb_block[MM_TSB_HUGE].tsb) {
 105                 base = (unsigned long) mm->context.tsb_block[MM_TSB_HUGE].tsb;
 106                 nentries = mm->context.tsb_block[MM_TSB_HUGE].tsb_nentries;
 107                 if (tlb_type == cheetah_plus || tlb_type == hypervisor)
 108                         base = __pa(base);
 109                 __flush_tsb_one(tb, REAL_HPAGE_SHIFT, base, nentries);
 110         }
 111 #endif
 112         spin_unlock_irqrestore(&mm->context.lock, flags);
 113 }
 114
 115 void flush_tsb_user_page(struct mm_struct *mm, unsigned long vaddr, bool huge)
 116 {
 117         unsigned long nentries, base, flags;
 118
 119         spin_lock_irqsave(&mm->context.lock, flags);
 120
 121         if (!huge) {
 122                 base = (unsigned long) mm->context.tsb_block[MM_TSB_BASE].tsb;
 123                 nentries = mm->context.tsb_block[MM_TSB_BASE].tsb_nentries;
 124                 if (tlb_type == cheetah_plus || tlb_type == hypervisor)
 125                         base = __pa(base);
 126                 __flush_tsb_one_entry(base, vaddr, PAGE_SHIFT, nentries);
 127         }
 128 #if defined(CONFIG_HUGETLB_PAGE) || defined(CONFIG_TRANSPARENT_HUGEPAGE)
 129         if (huge && mm->context.tsb_block[MM_TSB_HUGE].tsb) {
 130                 base = (unsigned long) mm->context.tsb_block[MM_TSB_HUGE].tsb;
 131                 nentries = mm->context.tsb_block[MM_TSB_HUGE].tsb_nentries;
 132                 if (tlb_type == cheetah_plus || tlb_type == hypervisor)
 133                         base = __pa(base);
 134                 __flush_tsb_one_entry(base, vaddr, REAL_HPAGE_SHIFT, nentries);
 135         }
 136 #endif
 137         spin_unlock_irqrestore(&mm->context.lock, flags);
 138 }
 139
 140 #define HV_PGSZ_IDX_BASE        HV_PGSZ_IDX_8K
 141 #define HV_PGSZ_MASK_BASE       HV_PGSZ_MASK_8K
 142
 143 #if defined(CONFIG_HUGETLB_PAGE) || defined(CONFIG_TRANSPARENT_HUGEPAGE)
 144 #define HV_PGSZ_IDX_HUGE        HV_PGSZ_IDX_4MB
 145 #define HV_PGSZ_MASK_HUGE       HV_PGSZ_MASK_4MB
 146 #endif
 147
 148 static void setup_tsb_params(struct mm_struct *mm, unsigned long tsb_idx, unsigned long tsb_bytes)
 149 {
 150         unsigned long tsb_reg, base, tsb_paddr;
 151         unsigned long page_sz, tte;
 152
 153         mm->context.tsb_block[tsb_idx].tsb_nentries =
 154                 tsb_bytes / sizeof(struct tsb);
 155
 156         switch (tsb_idx) {
 157         case MM_TSB_BASE:
 158                 base = TSBMAP_8K_BASE;
 159                 break;
 160 #if defined(CONFIG_HUGETLB_PAGE) || defined(CONFIG_TRANSPARENT_HUGEPAGE)
 161         case MM_TSB_HUGE:
 162                 base = TSBMAP_4M_BASE;
 163                 break;
 164 #endif
 165         default:
 166                 BUG();
 167         }
 168
 169         tte = pgprot_val(PAGE_KERNEL_LOCKED);
 170         tsb_paddr = __pa(mm->context.tsb_block[tsb_idx].tsb);
 171         BUG_ON(tsb_paddr & (tsb_bytes - 1UL));
 172
 173         /* Use the smallest page size that can map the whole TSB
 174          * in one TLB entry.
 175          */
 176         switch (tsb_bytes) {
 177         case 8192 << 0:
 178                 tsb_reg = 0x0UL;
 179 #ifdef DCACHE_ALIASING_POSSIBLE
 180                 base += (tsb_paddr & 8192);
 181 #endif
 182                 page_sz = 8192;
 183                 break;
 184
 185         case 8192 << 1:
 186                 tsb_reg = 0x1UL;
 187                 page_sz = 64 * 1024;
 188                 break;
 189
 190         case 8192 << 2:
 191                 tsb_reg = 0x2UL;
 192                 page_sz = 64 * 1024;
 193                 break;
 194
 195         case 8192 << 3:
 196                 tsb_reg = 0x3UL;
 197                 page_sz = 64 * 1024;
 198                 break;
 199
 200         case 8192 << 4:
 201                 tsb_reg = 0x4UL;
 202                 page_sz = 512 * 1024;
 203                 break;
 204
 205         case 8192 << 5:
 206                 tsb_reg = 0x5UL;
 207                 page_sz = 512 * 1024;
 208                 break;
 209
 210         case 8192 << 6:
 211                 tsb_reg = 0x6UL;
 212                 page_sz = 512 * 1024;
 213                 break;
 214
 215         case 8192 << 7:
 216                 tsb_reg = 0x7UL;
 217                 page_sz = 4 * 1024 * 1024;
 218                 break;
 219
 220         default:
 221                 printk(KERN_ERR "TSB[%s:%d]: Impossible TSB size %lu, killing process.\n",
 222                        current->comm, current->pid, tsb_bytes);
 223                 do_exit(SIGSEGV);
 224         }
 225         tte |= pte_sz_bits(page_sz);
 226
 227         if (tlb_type == cheetah_plus || tlb_type == hypervisor) {
 228                 /* Physical mapping, no locked TLB entry for TSB.  */
 229                 tsb_reg |= tsb_paddr;
 230
 231                 mm->context.tsb_block[tsb_idx].tsb_reg_val = tsb_reg;
 232                 mm->context.tsb_block[tsb_idx].tsb_map_vaddr = 0;
 233                 mm->context.tsb_block[tsb_idx].tsb_map_pte = 0;
 234         } else {
 235                 tsb_reg |= base;
 236                 tsb_reg |= (tsb_paddr & (page_sz - 1UL));
 237                 tte |= (tsb_paddr & ~(page_sz - 1UL));
 238
 239                 mm->context.tsb_block[tsb_idx].tsb_reg_val = tsb_reg;
 240                 mm->context.tsb_block[tsb_idx].tsb_map_vaddr = base;
 241                 mm->context.tsb_block[tsb_idx].tsb_map_pte = tte;
 242         }
 243
 244         /* Setup the Hypervisor TSB descriptor.  */
 245         if (tlb_type == hypervisor) {
 246                 struct hv_tsb_descr *hp = &mm->context.tsb_descr[tsb_idx];
 247
 248                 switch (tsb_idx) {
 249                 case MM_TSB_BASE:
 250                         hp->pgsz_idx = HV_PGSZ_IDX_BASE;
 251                         break;
 252 #if defined(CONFIG_HUGETLB_PAGE) || defined(CONFIG_TRANSPARENT_HUGEPAGE)
 253                 case MM_TSB_HUGE:
 254                         hp->pgsz_idx = HV_PGSZ_IDX_HUGE;
 255                         break;
 256 #endif
 257                 default:
 258                         BUG();
 259                 }
 260                 hp->assoc = 1;
 261                 hp->num_ttes = tsb_bytes / 16;
 262                 hp->ctx_idx = 0;
 263                 switch (tsb_idx) {
 264                 case MM_TSB_BASE:
 265                         hp->pgsz_mask = HV_PGSZ_MASK_BASE;
 266                         break;
 267 #if defined(CONFIG_HUGETLB_PAGE) || defined(CONFIG_TRANSPARENT_HUGEPAGE)
 268                 case MM_TSB_HUGE:
 269                         hp->pgsz_mask = HV_PGSZ_MASK_HUGE;
 270                         break;
 271 #endif
 272                 default:
 273                         BUG();
 274                 }
 275                 hp->tsb_base = tsb_paddr;
 276                 hp->resv = 0;
 277         }
 278 }
 279
 280 struct kmem_cache *pgtable_cache __read_mostly;
 281
 282 static struct kmem_cache *tsb_caches[8] __read_mostly;
 283
 284 static const char *tsb_cache_names[8] = {
 285         "tsb_8KB",
 286         "tsb_16KB",
 287         "tsb_32KB",
 288         "tsb_64KB",
 289         "tsb_128KB",
 290         "tsb_256KB",
 291         "tsb_512KB",
 292         "tsb_1MB",
 293 };
 294
 295 void __init pgtable_cache_init(void)
 296 {
 297         unsigned long i;
 298
 299         pgtable_cache = kmem_cache_create("pgtable_cache",
 300                                           PAGE_SIZE, PAGE_SIZE,
 301                                           0,
 302                                           _clear_page);
 303         if (!pgtable_cache) {
 304                 prom_printf("pgtable_cache_init(): Could not create!\n");
 305                 prom_halt();
 306         }
 307
 308         for (i = 0; i < ARRAY_SIZE(tsb_cache_names); i++) {
 309                 unsigned long size = 8192 << i;
 310                 const char *name = tsb_cache_names[i];
 311
 312                 tsb_caches[i] = kmem_cache_create(name,
 313                                                   size, size,
 314                                                   0, NULL);
 315                 if (!tsb_caches[i]) {
 316                         prom_printf("Could not create %s cache\n", name);
 317                         prom_halt();
 318                 }
 319         }
 320 }
 321
 322 int sysctl_tsb_ratio = -2;
 323
 324 static unsigned long tsb_size_to_rss_limit(unsigned long new_size)
 325 {
 326         unsigned long num_ents = (new_size / sizeof(struct tsb));
 327
 328         if (sysctl_tsb_ratio < 0)
 329                 return num_ents - (num_ents >> -sysctl_tsb_ratio);
 330         else
 331                 return num_ents + (num_ents >> sysctl_tsb_ratio);
 332 }
 333
 334 /* When the RSS of an address space exceeds tsb_rss_limit for a TSB,
 335  * do_sparc64_fault() invokes this routine to try and grow it.
 336  *
 337  * When we reach the maximum TSB size supported, we stick ~0UL into
 338  * tsb_rss_limit for that TSB so the grow checks in do_sparc64_fault()
 339  * will not trigger any longer.
 340  *
 341  * The TSB can be anywhere from 8K to 1MB in size, in increasing powers
 342  * of two.  The TSB must be aligned to it's size, so f.e. a 512K TSB
 343  * must be 512K aligned.  It also must be physically contiguous, so we
 344  * cannot use vmalloc().
 345  *
 346  * The idea here is to grow the TSB when the RSS of the process approaches
 347  * the number of entries that the current TSB can hold at once.  Currently,
 348  * we trigger when the RSS hits 3/4 of the TSB capacity.
 349  */
 350 void tsb_grow(struct mm_struct *mm, unsigned long tsb_index, unsigned long rss)
 351 {
 352         unsigned long max_tsb_size = 1 * 1024 * 1024;
 353         unsigned long new_size, old_size, flags;
 354         struct tsb *old_tsb, *new_tsb;
 355         unsigned long new_cache_index, old_cache_index;
 356         unsigned long new_rss_limit;
 357         gfp_t gfp_flags;
 358
 359         if (max_tsb_size > (PAGE_SIZE << MAX_ORDER))
 360                 max_tsb_size = (PAGE_SIZE << MAX_ORDER);
 361
 362         new_cache_index = 0;
 363         for (new_size = 8192; new_size < max_tsb_size; new_size <<= 1UL) {
 364                 new_rss_limit = tsb_size_to_rss_limit(new_size);
 365                 if (new_rss_limit > rss)
 366                         break;
 367                 new_cache_index++;
 368         }
 369
 370         if (new_size == max_tsb_size)
 371                 new_rss_limit = ~0UL;
 372
 373 retry_tsb_alloc:
 374         gfp_flags = GFP_KERNEL;
 375         if (new_size > (PAGE_SIZE * 2))
 376                 gfp_flags |= __GFP_NOWARN | __GFP_NORETRY;
 377
 378         new_tsb = kmem_cache_alloc_node(tsb_caches[new_cache_index],
 379                                         gfp_flags, numa_node_id());
 380         if (unlikely(!new_tsb)) {
 381                 /* Not being able to fork due to a high-order TSB
 382                  * allocation failure is very bad behavior.  Just back
 383                  * down to a 0-order allocation and force no TSB
 384                  * growing for this address space.
 385                  */
 386                 if (mm->context.tsb_block[tsb_index].tsb == NULL &&
 387                     new_cache_index > 0) {
 388                         new_cache_index = 0;
 389                         new_size = 8192;
 390                         new_rss_limit = ~0UL;
 391                         goto retry_tsb_alloc;
 392                 }
 393
 394                 /* If we failed on a TSB grow, we are under serious
 395                  * memory pressure so don't try to grow any more.
 396                  */
 397                 if (mm->context.tsb_block[tsb_index].tsb != NULL)
 398                         mm->context.tsb_block[tsb_index].tsb_rss_limit = ~0UL;
 399                 return;
 400         }
 401
 402         /* Mark all tags as invalid.  */
 403         tsb_init(new_tsb, new_size);
 404
 405         /* Ok, we are about to commit the changes.  If we are
 406          * growing an existing TSB the locking is very tricky,
 407          * so WATCH OUT!
 408          *
 409          * We have to hold mm->context.lock while committing to the
 410          * new TSB, this synchronizes us with processors in
 411          * flush_tsb_user() and switch_mm() for this address space.
 412          *
 413          * But even with that lock held, processors run asynchronously
 414          * accessing the old TSB via TLB miss handling.  This is OK
 415          * because those actions are just propagating state from the
 416          * Linux page tables into the TSB, page table mappings are not
 417          * being changed.  If a real fault occurs, the processor will
 418          * synchronize with us when it hits flush_tsb_user(), this is
 419          * also true for the case where vmscan is modifying the page
 420          * tables.  The only thing we need to be careful with is to
 421          * skip any locked TSB entries during copy_tsb().
 422          *
 423          * When we finish committing to the new TSB, we have to drop
 424          * the lock and ask all other cpus running this address space
 425          * to run tsb_context_switch() to see the new TSB table.
 426          */
 427         spin_lock_irqsave(&mm->context.lock, flags);
 428
 429         old_tsb = mm->context.tsb_block[tsb_index].tsb;
 430         old_cache_index =
 431                 (mm->context.tsb_block[tsb_index].tsb_reg_val & 0x7UL);
 432         old_size = (mm->context.tsb_block[tsb_index].tsb_nentries *
 433                     sizeof(struct tsb));
 434
 435
 436         /* Handle multiple threads trying to grow the TSB at the same time.
 437          * One will get in here first, and bump the size and the RSS limit.
 438          * The others will get in here next and hit this check.
 439          */
 440         if (unlikely(old_tsb &&
 441                      (rss < mm->context.tsb_block[tsb_index].tsb_rss_limit))) {
 442                 spin_unlock_irqrestore(&mm->context.lock, flags);
 443
 444                 kmem_cache_free(tsb_caches[new_cache_index], new_tsb);
 445                 return;
 446         }
 447
 448         mm->context.tsb_block[tsb_index].tsb_rss_limit = new_rss_limit;
 449
 450         if (old_tsb) {
 451                 extern void copy_tsb(unsigned long old_tsb_base,
 452                                      unsigned long old_tsb_size,
 453                                      unsigned long new_tsb_base,
 454                                      unsigned long new_tsb_size);
 455                 unsigned long old_tsb_base = (unsigned long) old_tsb;
 456                 unsigned long new_tsb_base = (unsigned long) new_tsb;
 457
 458                 if (tlb_type == cheetah_plus || tlb_type == hypervisor) {
 459                         old_tsb_base = __pa(old_tsb_base);
 460                         new_tsb_base = __pa(new_tsb_base);
 461                 }
 462                 copy_tsb(old_tsb_base, old_size, new_tsb_base, new_size);
 463         }
 464
 465         mm->context.tsb_block[tsb_index].tsb = new_tsb;
 466         setup_tsb_params(mm, tsb_index, new_size);
 467
 468         spin_unlock_irqrestore(&mm->context.lock, flags);
 469
 470         /* If old_tsb is NULL, we're being invoked for the first time
 471          * from init_new_context().
 472          */
 473         if (old_tsb) {
 474                 /* Reload it on the local cpu.  */
 475                 tsb_context_switch(mm);
 476
 477                 /* Now force other processors to do the same.  */
 478                 preempt_disable();
 479                 smp_tsb_sync(mm);
 480                 preempt_enable();
 481
 482                 /* Now it is safe to free the old tsb.  */
 483                 kmem_cache_free(tsb_caches[old_cache_index], old_tsb);
 484         }
 485 }
 486
 487 int init_new_context(struct task_struct *tsk, struct mm_struct *mm)
 488 {
 489         unsigned long mm_rss = get_mm_rss(mm);
 490 #if defined(CONFIG_HUGETLB_PAGE) || defined(CONFIG_TRANSPARENT_HUGEPAGE)
 491         unsigned long saved_hugetlb_pte_count;
 492         unsigned long saved_thp_pte_count;
 493 #endif
 494         unsigned int i;
 495
 496         spin_lock_init(&mm->context.lock);
 497
 498         mm->context.sparc64_ctx_val = 0UL;
 499
 500 #if defined(CONFIG_HUGETLB_PAGE) || defined(CONFIG_TRANSPARENT_HUGEPAGE)
 501         /* We reset them to zero because the fork() page copying
 502          * will re-increment the counters as the parent PTEs are
 503          * copied into the child address space.
 504          */
 505         saved_hugetlb_pte_count = mm->context.hugetlb_pte_count;
 506         saved_thp_pte_count = mm->context.thp_pte_count;
 507         mm->context.hugetlb_pte_count = 0;
 508         mm->context.thp_pte_count = 0;
 509
 510         mm_rss -= saved_thp_pte_count * (HPAGE_SIZE / PAGE_SIZE);
 511 #endif
 512
 513         /* copy_mm() copies over the parent's mm_struct before calling
 514          * us, so we need to zero out the TSB pointer or else tsb_grow()
 515          * will be confused and think there is an older TSB to free up.
 516          */
 517         for (i = 0; i < MM_NUM_TSBS; i++)
 518                 mm->context.tsb_block[i].tsb = NULL;
 519
 520         /* If this is fork, inherit the parent's TSB size.  We would
 521          * grow it to that size on the first page fault anyways.
 522          */
 523         tsb_grow(mm, MM_TSB_BASE, mm_rss);
 524
 525 #if defined(CONFIG_HUGETLB_PAGE) || defined(CONFIG_TRANSPARENT_HUGEPAGE)
 526         if (unlikely(saved_hugetlb_pte_count + saved_thp_pte_count))
 527                 tsb_grow(mm, MM_TSB_HUGE,
 528                          (saved_hugetlb_pte_count + saved_thp_pte_count) *
 529                          REAL_HPAGE_PER_HPAGE);
 530 #endif
 531
 532         if (unlikely(!mm->context.tsb_block[MM_TSB_BASE].tsb))
 533                 return -ENOMEM;
 534
 535         return 0;
 536 }
 537
 538 static void tsb_destroy_one(struct tsb_config *tp)
 539 {
 540         unsigned long cache_index;
 541
 542         if (!tp->tsb)
 543                 return;
 544         cache_index = tp->tsb_reg_val & 0x7UL;
 545         kmem_cache_free(tsb_caches[cache_index], tp->tsb);
 546         tp->tsb = NULL;
 547         tp->tsb_reg_val = 0UL;
 548 }
 549
 550 void destroy_context(struct mm_struct *mm)
 551 {
 552         unsigned long flags, i;
 553
 554         for (i = 0; i < MM_NUM_TSBS; i++)
 555                 tsb_destroy_one(&mm->context.tsb_block[i]);
 556
 557         spin_lock_irqsave(&ctx_alloc_lock, flags);
 558
 559         if (CTX_VALID(mm->context)) {
 560                 unsigned long nr = CTX_NRBITS(mm->context);
 561                 mmu_context_bmap[nr>>6] &= ~(1UL << (nr & 63));
 562         }
 563
 564         spin_unlock_irqrestore(&ctx_alloc_lock, flags);
 565 }