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