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