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