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/system.h>
  10 #include <asm/page.h>
  11 #include <asm/pgtable.h>
  12 #include <asm/mmu_context.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         base = (unsigned long) mm->context.tsb_block[MM_TSB_BASE].tsb;
  80         nentries = mm->context.tsb_block[MM_TSB_BASE].tsb_nentries;
  81         if (tlb_type == cheetah_plus || tlb_type == hypervisor)
  82                 base = __pa(base);
  83         __flush_tsb_one(tb, PAGE_SHIFT, base, nentries);
  84
  85 #ifdef CONFIG_HUGETLB_PAGE
  86         if (mm->context.tsb_block[MM_TSB_HUGE].tsb) {
  87                 base = (unsigned long) mm->context.tsb_block[MM_TSB_HUGE].tsb;
  88                 nentries = mm->context.tsb_block[MM_TSB_HUGE].tsb_nentries;
  89                 if (tlb_type == cheetah_plus || tlb_type == hypervisor)
  90                         base = __pa(base);
  91                 __flush_tsb_one(tb, HPAGE_SHIFT, base, nentries);
  92         }
  93 #endif
  94         spin_unlock_irqrestore(&mm->context.lock, flags);
  95 }
  96
  97 void flush_tsb_user_page(struct mm_struct *mm, unsigned long vaddr)
  98 {
  99         unsigned long nentries, base, flags;
 100
 101         spin_lock_irqsave(&mm->context.lock, flags);
 102
 103         base = (unsigned long) mm->context.tsb_block[MM_TSB_BASE].tsb;
 104         nentries = mm->context.tsb_block[MM_TSB_BASE].tsb_nentries;
 105         if (tlb_type == cheetah_plus || tlb_type == hypervisor)
 106                 base = __pa(base);
 107         __flush_tsb_one_entry(base, vaddr, PAGE_SHIFT, nentries);
 108
 109 #if defined(CONFIG_HUGETLB_PAGE) || defined(CONFIG_TRANSPARENT_HUGEPAGE)
 110         if (mm->context.tsb_block[MM_TSB_HUGE].tsb) {
 111                 base = (unsigned long) mm->context.tsb_block[MM_TSB_HUGE].tsb;
 112                 nentries = mm->context.tsb_block[MM_TSB_HUGE].tsb_nentries;
 113                 if (tlb_type == cheetah_plus || tlb_type == hypervisor)
 114                         base = __pa(base);
 115                 __flush_tsb_one_entry(base, vaddr, HPAGE_SHIFT, nentries);
 116         }
 117 #endif
 118         spin_unlock_irqrestore(&mm->context.lock, flags);
 119 }
 120
 121 #if defined(CONFIG_SPARC64_PAGE_SIZE_8KB)
 122 #define HV_PGSZ_IDX_BASE        HV_PGSZ_IDX_8K
 123 #define HV_PGSZ_MASK_BASE       HV_PGSZ_MASK_8K
 124 #elif defined(CONFIG_SPARC64_PAGE_SIZE_64KB)
 125 #define HV_PGSZ_IDX_BASE        HV_PGSZ_IDX_64K
 126 #define HV_PGSZ_MASK_BASE       HV_PGSZ_MASK_64K
 127 #else
 128 #error Broken base page size setting...
 129 #endif
 130
 131 #ifdef CONFIG_HUGETLB_PAGE
 132 #if defined(CONFIG_HUGETLB_PAGE_SIZE_64K)
 133 #define HV_PGSZ_IDX_HUGE        HV_PGSZ_IDX_64K
 134 #define HV_PGSZ_MASK_HUGE       HV_PGSZ_MASK_64K
 135 #elif defined(CONFIG_HUGETLB_PAGE_SIZE_512K)
 136 #define HV_PGSZ_IDX_HUGE        HV_PGSZ_IDX_512K
 137 #define HV_PGSZ_MASK_HUGE       HV_PGSZ_MASK_512K
 138 #elif defined(CONFIG_HUGETLB_PAGE_SIZE_4MB)
 139 #define HV_PGSZ_IDX_HUGE        HV_PGSZ_IDX_4MB
 140 #define HV_PGSZ_MASK_HUGE       HV_PGSZ_MASK_4MB
 141 #else
 142 #error Broken huge page size setting...
 143 #endif
 144 #endif
 145
 146 static void setup_tsb_params(struct mm_struct *mm, unsigned long tsb_idx, unsigned long tsb_bytes)
 147 {
 148         unsigned long tsb_reg, base, tsb_paddr;
 149         unsigned long page_sz, tte;
 150
 151         mm->context.tsb_block[tsb_idx].tsb_nentries =
 152                 tsb_bytes / sizeof(struct tsb);
 153
 154         switch (tsb_idx) {
 155         case MM_TSB_BASE:
 156                 base = TSBMAP_8K_BASE;
 157                 break;
 158 #if defined(CONFIG_HUGETLB_PAGE) || defined(CONFIG_TRANSPARENT_HUGEPAGE)
 159         case MM_TSB_HUGE:
 160                 base = TSBMAP_4M_BASE;
 161                 break;
 162 #endif
 163         default:
 164                 BUG();
 165         }
 166
 167         tte = pgprot_val(PAGE_KERNEL_LOCKED);
 168         tsb_paddr = __pa(mm->context.tsb_block[tsb_idx].tsb);
 169         BUG_ON(tsb_paddr & (tsb_bytes - 1UL));
 170
 171         /* Use the smallest page size that can map the whole TSB
 172          * in one TLB entry.
 173          */
 174         switch (tsb_bytes) {
 175         case 8192 << 0:
 176                 tsb_reg = 0x0UL;
 177 #ifdef DCACHE_ALIASING_POSSIBLE
 178                 base += (tsb_paddr & 8192);
 179 #endif
 180                 page_sz = 8192;
 181                 break;
 182
 183         case 8192 << 1:
 184                 tsb_reg = 0x1UL;
 185                 page_sz = 64 * 1024;
 186                 break;
 187
 188         case 8192 << 2:
 189                 tsb_reg = 0x2UL;
 190                 page_sz = 64 * 1024;
 191                 break;
 192
 193         case 8192 << 3:
 194                 tsb_reg = 0x3UL;
 195                 page_sz = 64 * 1024;
 196                 break;
 197
 198         case 8192 << 4:
 199                 tsb_reg = 0x4UL;
 200                 page_sz = 512 * 1024;
 201                 break;
 202
 203         case 8192 << 5:
 204                 tsb_reg = 0x5UL;
 205                 page_sz = 512 * 1024;
 206                 break;
 207
 208         case 8192 << 6:
 209                 tsb_reg = 0x6UL;
 210                 page_sz = 512 * 1024;
 211                 break;
 212
 213         case 8192 << 7:
 214                 tsb_reg = 0x7UL;
 215                 page_sz = 4 * 1024 * 1024;
 216                 break;
 217
 218         default:
 219                 printk(KERN_ERR "TSB[%s:%d]: Impossible TSB size %lu, killing process.\n",
 220                        current->comm, current->pid, tsb_bytes);
 221                 do_exit(SIGSEGV);
 222         }
 223         tte |= pte_sz_bits(page_sz);
 224
 225         if (tlb_type == cheetah_plus || tlb_type == hypervisor) {
 226                 /* Physical mapping, no locked TLB entry for TSB.  */
 227                 tsb_reg |= tsb_paddr;
 228
 229                 mm->context.tsb_block[tsb_idx].tsb_reg_val = tsb_reg;
 230                 mm->context.tsb_block[tsb_idx].tsb_map_vaddr = 0;
 231                 mm->context.tsb_block[tsb_idx].tsb_map_pte = 0;
 232         } else {
 233                 tsb_reg |= base;
 234                 tsb_reg |= (tsb_paddr & (page_sz - 1UL));
 235                 tte |= (tsb_paddr & ~(page_sz - 1UL));
 236
 237                 mm->context.tsb_block[tsb_idx].tsb_reg_val = tsb_reg;
 238                 mm->context.tsb_block[tsb_idx].tsb_map_vaddr = base;
 239                 mm->context.tsb_block[tsb_idx].tsb_map_pte = tte;
 240         }
 241
 242         /* Setup the Hypervisor TSB descriptor.  */
 243         if (tlb_type == hypervisor) {
 244                 struct hv_tsb_descr *hp = &mm->context.tsb_descr[tsb_idx];
 245
 246                 switch (tsb_idx) {
 247                 case MM_TSB_BASE:
 248                         hp->pgsz_idx = HV_PGSZ_IDX_BASE;
 249                         break;
 250 #ifdef CONFIG_HUGETLB_PAGE
 251                 case MM_TSB_HUGE:
 252                         hp->pgsz_idx = HV_PGSZ_IDX_HUGE;
 253                         break;
 254 #endif
 255                 default:
 256                         BUG();
 257                 }
 258                 hp->assoc = 1;
 259                 hp->num_ttes = tsb_bytes / 16;
 260                 hp->ctx_idx = 0;
 261                 switch (tsb_idx) {
 262                 case MM_TSB_BASE:
 263                         hp->pgsz_mask = HV_PGSZ_MASK_BASE;
 264                         break;
 265 #ifdef CONFIG_HUGETLB_PAGE
 266                 case MM_TSB_HUGE:
 267                         hp->pgsz_mask = HV_PGSZ_MASK_HUGE;
 268                         break;
 269 #endif
 270                 default:
 271                         BUG();
 272                 }
 273                 hp->tsb_base = tsb_paddr;
 274                 hp->resv = 0;
 275         }
 276 }
 277
 278 struct kmem_cache *pgtable_cache __read_mostly;
 279
 280 static struct kmem_cache *tsb_caches[8] __read_mostly;
 281
 282 static const char *tsb_cache_names[8] = {
 283         "tsb_8KB",
 284         "tsb_16KB",
 285         "tsb_32KB",
 286         "tsb_64KB",
 287         "tsb_128KB",
 288         "tsb_256KB",
 289         "tsb_512KB",
 290         "tsb_1MB",
 291 };
 292
 293 void __init pgtable_cache_init(void)
 294 {
 295         unsigned long i;
 296
 297         pgtable_cache = kmem_cache_create("pgtable_cache",
 298                                           PAGE_SIZE, PAGE_SIZE,
 299                                           0,
 300                                           _clear_page);
 301         if (!pgtable_cache) {
 302                 prom_printf("pgtable_cache_init(): Could not create!\n");
 303                 prom_halt();
 304         }
 305
 306         for (i = 0; i < 8; i++) {
 307                 unsigned long size = 8192 << i;
 308                 const char *name = tsb_cache_names[i];
 309
 310                 tsb_caches[i] = kmem_cache_create(name,
 311                                                   size, size,
 312                                                   0, NULL);
 313                 if (!tsb_caches[i]) {
 314                         prom_printf("Could not create %s cache\n", name);
 315                         prom_halt();
 316                 }
 317         }
 318 }
 319
 320 int sysctl_tsb_ratio = -2;
 321
 322 static unsigned long tsb_size_to_rss_limit(unsigned long new_size)
 323 {
 324         unsigned long num_ents = (new_size / sizeof(struct tsb));
 325
 326         if (sysctl_tsb_ratio < 0)
 327                 return num_ents - (num_ents >> -sysctl_tsb_ratio);
 328         else
 329                 return num_ents + (num_ents >> sysctl_tsb_ratio);
 330 }
 331
 332 /* When the RSS of an address space exceeds tsb_rss_limit for a TSB,
 333  * do_sparc64_fault() invokes this routine to try and grow it.
 334  *
 335  * When we reach the maximum TSB size supported, we stick ~0UL into
 336  * tsb_rss_limit for that TSB so the grow checks in do_sparc64_fault()
 337  * will not trigger any longer.
 338  *
 339  * The TSB can be anywhere from 8K to 1MB in size, in increasing powers
 340  * of two.  The TSB must be aligned to it's size, so f.e. a 512K TSB
 341  * must be 512K aligned.  It also must be physically contiguous, so we
 342  * cannot use vmalloc().
 343  *
 344  * The idea here is to grow the TSB when the RSS of the process approaches
 345  * the number of entries that the current TSB can hold at once.  Currently,
 346  * we trigger when the RSS hits 3/4 of the TSB capacity.
 347  */
 348 void tsb_grow(struct mm_struct *mm, unsigned long tsb_index, unsigned long rss)
 349 {
 350         unsigned long max_tsb_size = 1 * 1024 * 1024;
 351         unsigned long new_size, old_size, flags;
 352         struct tsb *old_tsb, *new_tsb;
 353         unsigned long new_cache_index, old_cache_index;
 354         unsigned long new_rss_limit;
 355         gfp_t gfp_flags;
 356
 357         if (max_tsb_size > (PAGE_SIZE << MAX_ORDER))
 358                 max_tsb_size = (PAGE_SIZE << MAX_ORDER);
 359
 360         new_cache_index = 0;
 361         for (new_size = 8192; new_size < max_tsb_size; new_size <<= 1UL) {
 362                 new_rss_limit = tsb_size_to_rss_limit(new_size);
 363                 if (new_rss_limit > rss)
 364                         break;
 365                 new_cache_index++;
 366         }
 367
 368         if (new_size == max_tsb_size)
 369                 new_rss_limit = ~0UL;
 370
 371 retry_tsb_alloc:
 372         gfp_flags = GFP_KERNEL;
 373         if (new_size > (PAGE_SIZE * 2))
 374                 gfp_flags = __GFP_NOWARN | __GFP_NORETRY;
 375
 376         new_tsb = kmem_cache_alloc_node(tsb_caches[new_cache_index],
 377                                         gfp_flags, numa_node_id());
 378         if (unlikely(!new_tsb)) {
 379                 /* Not being able to fork due to a high-order TSB
 380                  * allocation failure is very bad behavior.  Just back
 381                  * down to a 0-order allocation and force no TSB
 382                  * growing for this address space.
 383                  */
 384                 if (mm->context.tsb_block[tsb_index].tsb == NULL &&
 385                     new_cache_index > 0) {
 386                         new_cache_index = 0;
 387                         new_size = 8192;
 388                         new_rss_limit = ~0UL;
 389                         goto retry_tsb_alloc;
 390                 }
 391
 392                 /* If we failed on a TSB grow, we are under serious
 393                  * memory pressure so don't try to grow any more.
 394                  */
 395                 if (mm->context.tsb_block[tsb_index].tsb != NULL)
 396                         mm->context.tsb_block[tsb_index].tsb_rss_limit = ~0UL;
 397                 return;
 398         }
 399
 400         /* Mark all tags as invalid.  */
 401         tsb_init(new_tsb, new_size);
 402
 403         /* Ok, we are about to commit the changes.  If we are
 404          * growing an existing TSB the locking is very tricky,
 405          * so WATCH OUT!
 406          *
 407          * We have to hold mm->context.lock while committing to the
 408          * new TSB, this synchronizes us with processors in
 409          * flush_tsb_user() and switch_mm() for this address space.
 410          *
 411          * But even with that lock held, processors run asynchronously
 412          * accessing the old TSB via TLB miss handling.  This is OK
 413          * because those actions are just propagating state from the
 414          * Linux page tables into the TSB, page table mappings are not
 415          * being changed.  If a real fault occurs, the processor will
 416          * synchronize with us when it hits flush_tsb_user(), this is
 417          * also true for the case where vmscan is modifying the page
 418          * tables.  The only thing we need to be careful with is to
 419          * skip any locked TSB entries during copy_tsb().
 420          *
 421          * When we finish committing to the new TSB, we have to drop
 422          * the lock and ask all other cpus running this address space
 423          * to run tsb_context_switch() to see the new TSB table.
 424          */
 425         spin_lock_irqsave(&mm->context.lock, flags);
 426
 427         old_tsb = mm->context.tsb_block[tsb_index].tsb;
 428         old_cache_index =
 429                 (mm->context.tsb_block[tsb_index].tsb_reg_val & 0x7UL);
 430         old_size = (mm->context.tsb_block[tsb_index].tsb_nentries *
 431                     sizeof(struct tsb));
 432
 433
 434         /* Handle multiple threads trying to grow the TSB at the same time.
 435          * One will get in here first, and bump the size and the RSS limit.
 436          * The others will get in here next and hit this check.
 437          */
 438         if (unlikely(old_tsb &&
 439                      (rss < mm->context.tsb_block[tsb_index].tsb_rss_limit))) {
 440                 spin_unlock_irqrestore(&mm->context.lock, flags);
 441
 442                 kmem_cache_free(tsb_caches[new_cache_index], new_tsb);
 443                 return;
 444         }
 445
 446         mm->context.tsb_block[tsb_index].tsb_rss_limit = new_rss_limit;
 447
 448         if (old_tsb) {
 449                 extern void copy_tsb(unsigned long old_tsb_base,
 450                                      unsigned long old_tsb_size,
 451                                      unsigned long new_tsb_base,
 452                                      unsigned long new_tsb_size);
 453                 unsigned long old_tsb_base = (unsigned long) old_tsb;
 454                 unsigned long new_tsb_base = (unsigned long) new_tsb;
 455
 456                 if (tlb_type == cheetah_plus || tlb_type == hypervisor) {
 457                         old_tsb_base = __pa(old_tsb_base);
 458                         new_tsb_base = __pa(new_tsb_base);
 459                 }
 460                 copy_tsb(old_tsb_base, old_size, new_tsb_base, new_size);
 461         }
 462
 463         mm->context.tsb_block[tsb_index].tsb = new_tsb;
 464         setup_tsb_params(mm, tsb_index, new_size);
 465
 466         spin_unlock_irqrestore(&mm->context.lock, flags);
 467
 468         /* If old_tsb is NULL, we're being invoked for the first time
 469          * from init_new_context().
 470          */
 471         if (old_tsb) {
 472                 /* Reload it on the local cpu.  */
 473                 tsb_context_switch(mm);
 474
 475                 /* Now force other processors to do the same.  */
 476                 preempt_disable();
 477                 smp_tsb_sync(mm);
 478                 preempt_enable();
 479
 480                 /* Now it is safe to free the old tsb.  */
 481                 kmem_cache_free(tsb_caches[old_cache_index], old_tsb);
 482         }
 483 }
 484
 485 int init_new_context(struct task_struct *tsk, struct mm_struct *mm)
 486 {
 487 #ifdef CONFIG_HUGETLB_PAGE
 488         unsigned long huge_pte_count;
 489 #endif
 490         unsigned int i;
 491
 492         spin_lock_init(&mm->context.lock);
 493
 494         mm->context.sparc64_ctx_val = 0UL;
 495
 496 #ifdef CONFIG_HUGETLB_PAGE
 497         /* We reset it to zero because the fork() page copying
 498          * will re-increment the counters as the parent PTEs are
 499          * copied into the child address space.
 500          */
 501         huge_pte_count = mm->context.huge_pte_count;
 502         mm->context.huge_pte_count = 0;
 503 #endif
 504
 505         /* copy_mm() copies over the parent's mm_struct before calling
 506          * us, so we need to zero out the TSB pointer or else tsb_grow()
 507          * will be confused and think there is an older TSB to free up.
 508          */
 509         for (i = 0; i < MM_NUM_TSBS; i++)
 510                 mm->context.tsb_block[i].tsb = NULL;
 511
 512         /* If this is fork, inherit the parent's TSB size.  We would
 513          * grow it to that size on the first page fault anyways.
 514          */
 515         tsb_grow(mm, MM_TSB_BASE, get_mm_rss(mm));
 516
 517 #ifdef CONFIG_HUGETLB_PAGE
 518         if (unlikely(huge_pte_count))
 519                 tsb_grow(mm, MM_TSB_HUGE, huge_pte_count);
 520 #endif
 521
 522         if (unlikely(!mm->context.tsb_block[MM_TSB_BASE].tsb))
 523                 return -ENOMEM;
 524
 525         return 0;
 526 }
 527
 528 static void tsb_destroy_one(struct tsb_config *tp)
 529 {
 530         unsigned long cache_index;
 531
 532         if (!tp->tsb)
 533                 return;
 534         cache_index = tp->tsb_reg_val & 0x7UL;
 535         kmem_cache_free(tsb_caches[cache_index], tp->tsb);
 536         tp->tsb = NULL;
 537         tp->tsb_reg_val = 0UL;
 538 }
 539
 540 void destroy_context(struct mm_struct *mm)
 541 {
 542         unsigned long flags, i;
 543
 544         for (i = 0; i < MM_NUM_TSBS; i++)
 545                 tsb_destroy_one(&mm->context.tsb_block[i]);
 546
 547         spin_lock_irqsave(&ctx_alloc_lock, flags);
 548
 549         if (CTX_VALID(mm->context)) {
 550                 unsigned long nr = CTX_NRBITS(mm->context);
 551                 mmu_context_bmap[nr>>6] &= ~(1UL << (nr & 63));
 552         }
 553
 554         spin_unlock_irqrestore(&ctx_alloc_lock, flags);
 555 }