arch/tile/mm/homecache.c

   1 /*
   2  * Copyright 2010 Tilera Corporation. All Rights Reserved.
   3  *
   4  *   This program is free software; you can redistribute it and/or
   5  *   modify it under the terms of the GNU General Public License
   6  *   as published by the Free Software Foundation, version 2.
   7  *
   8  *   This program is distributed in the hope that it will be useful, but
   9  *   WITHOUT ANY WARRANTY; without even the implied warranty of
  10  *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
  11  *   NON INFRINGEMENT.  See the GNU General Public License for
  12  *   more details.
  13  *
  14  * This code maintains the "home" for each page in the system.
  15  */
  16
  17 #include <linux/kernel.h>
  18 #include <linux/mm.h>
  19 #include <linux/spinlock.h>
  20 #include <linux/list.h>
  21 #include <linux/bootmem.h>
  22 #include <linux/rmap.h>
  23 #include <linux/pagemap.h>
  24 #include <linux/mutex.h>
  25 #include <linux/interrupt.h>
  26 #include <linux/sysctl.h>
  27 #include <linux/pagevec.h>
  28 #include <linux/ptrace.h>
  29 #include <linux/timex.h>
  30 #include <linux/cache.h>
  31 #include <linux/smp.h>
  32 #include <linux/module.h>
  33 #include <linux/hugetlb.h>
  34
  35 #include <asm/page.h>
  36 #include <asm/sections.h>
  37 #include <asm/tlbflush.h>
  38 #include <asm/pgalloc.h>
  39 #include <asm/homecache.h>
  40
  41 #include <arch/sim.h>
  42
  43 #include "migrate.h"
  44
  45
  46 /*
  47  * The noallocl2 option suppresses all use of the L2 cache to cache
  48  * locally from a remote home.
  49  */
  50 static int __write_once noallocl2;
  51 static int __init set_noallocl2(char *str)
  52 {
  53         noallocl2 = 1;
  54         return 0;
  55 }
  56 early_param("noallocl2", set_noallocl2);
  57
  58
  59 /*
  60  * Update the irq_stat for cpus that we are going to interrupt
  61  * with TLB or cache flushes.  Also handle removing dataplane cpus
  62  * from the TLB flush set, and setting dataplane_tlb_state instead.
  63  */
  64 static void hv_flush_update(const struct cpumask *cache_cpumask,
  65                             struct cpumask *tlb_cpumask,
  66                             unsigned long tlb_va, unsigned long tlb_length,
  67                             HV_Remote_ASID *asids, int asidcount)
  68 {
  69         struct cpumask mask;
  70         int i, cpu;
  71
  72         cpumask_clear(&mask);
  73         if (cache_cpumask)
  74                 cpumask_or(&mask, &mask, cache_cpumask);
  75         if (tlb_cpumask && tlb_length) {
  76                 cpumask_or(&mask, &mask, tlb_cpumask);
  77         }
  78
  79         for (i = 0; i < asidcount; ++i)
  80                 cpumask_set_cpu(asids[i].y * smp_width + asids[i].x, &mask);
  81
  82         /*
  83          * Don't bother to update atomically; losing a count
  84          * here is not that critical.
  85          */
  86         for_each_cpu(cpu, &mask)
  87                 ++per_cpu(irq_stat, cpu).irq_hv_flush_count;
  88 }
  89
  90 /*
  91  * This wrapper function around hv_flush_remote() does several things:
  92  *
  93  *  - Provides a return value error-checking panic path, since
  94  *    there's never any good reason for hv_flush_remote() to fail.
  95  *  - Accepts a 32-bit PFN rather than a 64-bit PA, which generally
  96  *    is the type that Linux wants to pass around anyway.
  97  *  - Canonicalizes that lengths of zero make cpumasks NULL.
  98  *  - Handles deferring TLB flushes for dataplane tiles.
  99  *  - Tracks remote interrupts in the per-cpu irq_cpustat_t.
 100  *
 101  * Note that we have to wait until the cache flush completes before
 102  * updating the per-cpu last_cache_flush word, since otherwise another
 103  * concurrent flush can race, conclude the flush has already
 104  * completed, and start to use the page while it's still dirty
 105  * remotely (running concurrently with the actual evict, presumably).
 106  */
 107 void flush_remote(unsigned long cache_pfn, unsigned long cache_control,
 108                   const struct cpumask *cache_cpumask_orig,
 109                   HV_VirtAddr tlb_va, unsigned long tlb_length,
 110                   unsigned long tlb_pgsize,
 111                   const struct cpumask *tlb_cpumask_orig,
 112                   HV_Remote_ASID *asids, int asidcount)
 113 {
 114         int rc;
 115         struct cpumask cache_cpumask_copy, tlb_cpumask_copy;
 116         struct cpumask *cache_cpumask, *tlb_cpumask;
 117         HV_PhysAddr cache_pa;
 118         char cache_buf[NR_CPUS*5], tlb_buf[NR_CPUS*5];
 119
 120         mb();   /* provided just to simplify "magic hypervisor" mode */
 121
 122         /*
 123          * Canonicalize and copy the cpumasks.
 124          */
 125         if (cache_cpumask_orig && cache_control) {
 126                 cpumask_copy(&cache_cpumask_copy, cache_cpumask_orig);
 127                 cache_cpumask = &cache_cpumask_copy;
 128         } else {
 129                 cpumask_clear(&cache_cpumask_copy);
 130                 cache_cpumask = NULL;
 131         }
 132         if (cache_cpumask == NULL)
 133                 cache_control = 0;
 134         if (tlb_cpumask_orig && tlb_length) {
 135                 cpumask_copy(&tlb_cpumask_copy, tlb_cpumask_orig);
 136                 tlb_cpumask = &tlb_cpumask_copy;
 137         } else {
 138                 cpumask_clear(&tlb_cpumask_copy);
 139                 tlb_cpumask = NULL;
 140         }
 141
 142         hv_flush_update(cache_cpumask, tlb_cpumask, tlb_va, tlb_length,
 143                         asids, asidcount);
 144         cache_pa = (HV_PhysAddr)cache_pfn << PAGE_SHIFT;
 145         rc = hv_flush_remote(cache_pa, cache_control,
 146                              cpumask_bits(cache_cpumask),
 147                              tlb_va, tlb_length, tlb_pgsize,
 148                              cpumask_bits(tlb_cpumask),
 149                              asids, asidcount);
 150         if (rc == 0)
 151                 return;
 152         cpumask_scnprintf(cache_buf, sizeof(cache_buf), &cache_cpumask_copy);
 153         cpumask_scnprintf(tlb_buf, sizeof(tlb_buf), &tlb_cpumask_copy);
 154
 155         pr_err("hv_flush_remote(%#llx, %#lx, %p [%s],"
 156                " %#lx, %#lx, %#lx, %p [%s], %p, %d) = %d\n",
 157                cache_pa, cache_control, cache_cpumask, cache_buf,
 158                (unsigned long)tlb_va, tlb_length, tlb_pgsize,
 159                tlb_cpumask, tlb_buf,
 160                asids, asidcount, rc);
 161         panic("Unsafe to continue.");
 162 }
 163
 164 static void homecache_finv_page_va(void* va, int home)
 165 {
 166         int cpu = get_cpu();
 167         if (home == cpu) {
 168                 finv_buffer_local(va, PAGE_SIZE);
 169         } else if (home == PAGE_HOME_HASH) {
 170                 finv_buffer_remote(va, PAGE_SIZE, 1);
 171         } else {
 172                 BUG_ON(home < 0 || home >= NR_CPUS);
 173                 finv_buffer_remote(va, PAGE_SIZE, 0);
 174         }
 175         put_cpu();
 176 }
 177
 178 void homecache_finv_map_page(struct page *page, int home)
 179 {
 180         unsigned long flags;
 181         unsigned long va;
 182         pte_t *ptep;
 183         pte_t pte;
 184
 185         if (home == PAGE_HOME_UNCACHED)
 186                 return;
 187         local_irq_save(flags);
 188 #ifdef CONFIG_HIGHMEM
 189         va = __fix_to_virt(FIX_KMAP_BEGIN + kmap_atomic_idx_push() +
 190                            (KM_TYPE_NR * smp_processor_id()));
 191 #else
 192         va = __fix_to_virt(FIX_HOMECACHE_BEGIN + smp_processor_id());
 193 #endif
 194         ptep = virt_to_kpte(va);
 195         pte = pfn_pte(page_to_pfn(page), PAGE_KERNEL);
 196         __set_pte(ptep, pte_set_home(pte, home));
 197         homecache_finv_page_va((void *)va, home);
 198         __pte_clear(ptep);
 199         hv_flush_page(va, PAGE_SIZE);
 200 #ifdef CONFIG_HIGHMEM
 201         kmap_atomic_idx_pop();
 202 #endif
 203         local_irq_restore(flags);
 204 }
 205
 206 static void homecache_finv_page_home(struct page *page, int home)
 207 {
 208         if (!PageHighMem(page) && home == page_home(page))
 209                 homecache_finv_page_va(page_address(page), home);
 210         else
 211                 homecache_finv_map_page(page, home);
 212 }
 213
 214 static inline bool incoherent_home(int home)
 215 {
 216         return home == PAGE_HOME_IMMUTABLE || home == PAGE_HOME_INCOHERENT;
 217 }
 218
 219 static void homecache_finv_page_internal(struct page *page, int force_map)
 220 {
 221         int home = page_home(page);
 222         if (home == PAGE_HOME_UNCACHED)
 223                 return;
 224         if (incoherent_home(home)) {
 225                 int cpu;
 226                 for_each_cpu(cpu, &cpu_cacheable_map)
 227                         homecache_finv_map_page(page, cpu);
 228         } else if (force_map) {
 229                 /* Force if, e.g., the normal mapping is migrating. */
 230                 homecache_finv_map_page(page, home);
 231         } else {
 232                 homecache_finv_page_home(page, home);
 233         }
 234         sim_validate_lines_evicted(PFN_PHYS(page_to_pfn(page)), PAGE_SIZE);
 235 }
 236
 237 void homecache_finv_page(struct page *page)
 238 {
 239         homecache_finv_page_internal(page, 0);
 240 }
 241
 242 void homecache_evict(const struct cpumask *mask)
 243 {
 244         flush_remote(0, HV_FLUSH_EVICT_L2, mask, 0, 0, 0, NULL, NULL, 0);
 245 }
 246
 247 /* Report the home corresponding to a given PTE. */
 248 static int pte_to_home(pte_t pte)
 249 {
 250         if (hv_pte_get_nc(pte))
 251                 return PAGE_HOME_IMMUTABLE;
 252         switch (hv_pte_get_mode(pte)) {
 253         case HV_PTE_MODE_CACHE_TILE_L3:
 254                 return get_remote_cache_cpu(pte);
 255         case HV_PTE_MODE_CACHE_NO_L3:
 256                 return PAGE_HOME_INCOHERENT;
 257         case HV_PTE_MODE_UNCACHED:
 258                 return PAGE_HOME_UNCACHED;
 259         case HV_PTE_MODE_CACHE_HASH_L3:
 260                 return PAGE_HOME_HASH;
 261         }
 262         panic("Bad PTE %#llx\n", pte.val);
 263 }
 264
 265 /* Update the home of a PTE if necessary (can also be used for a pgprot_t). */
 266 pte_t pte_set_home(pte_t pte, int home)
 267 {
 268         /* Check for non-linear file mapping "PTEs" and pass them through. */
 269         if (pte_file(pte))
 270                 return pte;
 271
 272 #if CHIP_HAS_MMIO()
 273         /* Check for MMIO mappings and pass them through. */
 274         if (hv_pte_get_mode(pte) == HV_PTE_MODE_MMIO)
 275                 return pte;
 276 #endif
 277
 278
 279         /*
 280          * Only immutable pages get NC mappings.  If we have a
 281          * non-coherent PTE, but the underlying page is not
 282          * immutable, it's likely the result of a forced
 283          * caching setting running up against ptrace setting
 284          * the page to be writable underneath.  In this case,
 285          * just keep the PTE coherent.
 286          */
 287         if (hv_pte_get_nc(pte) && home != PAGE_HOME_IMMUTABLE) {
 288                 pte = hv_pte_clear_nc(pte);
 289                 pr_err("non-immutable page incoherently referenced: %#llx\n",
 290                        pte.val);
 291         }
 292
 293         switch (home) {
 294
 295         case PAGE_HOME_UNCACHED:
 296                 pte = hv_pte_set_mode(pte, HV_PTE_MODE_UNCACHED);
 297                 break;
 298
 299         case PAGE_HOME_INCOHERENT:
 300                 pte = hv_pte_set_mode(pte, HV_PTE_MODE_CACHE_NO_L3);
 301                 break;
 302
 303         case PAGE_HOME_IMMUTABLE:
 304                 /*
 305                  * We could home this page anywhere, since it's immutable,
 306                  * but by default just home it to follow "hash_default".
 307                  */
 308                 BUG_ON(hv_pte_get_writable(pte));
 309                 if (pte_get_forcecache(pte)) {
 310                         /* Upgrade "force any cpu" to "No L3" for immutable. */
 311                         if (hv_pte_get_mode(pte) == HV_PTE_MODE_CACHE_TILE_L3
 312                             && pte_get_anyhome(pte)) {
 313                                 pte = hv_pte_set_mode(pte,
 314                                                       HV_PTE_MODE_CACHE_NO_L3);
 315                         }
 316                 } else
 317                 if (hash_default)
 318                         pte = hv_pte_set_mode(pte, HV_PTE_MODE_CACHE_HASH_L3);
 319                 else
 320                         pte = hv_pte_set_mode(pte, HV_PTE_MODE_CACHE_NO_L3);
 321                 pte = hv_pte_set_nc(pte);
 322                 break;
 323
 324         case PAGE_HOME_HASH:
 325                 pte = hv_pte_set_mode(pte, HV_PTE_MODE_CACHE_HASH_L3);
 326                 break;
 327
 328         default:
 329                 BUG_ON(home < 0 || home >= NR_CPUS ||
 330                        !cpu_is_valid_lotar(home));
 331                 pte = hv_pte_set_mode(pte, HV_PTE_MODE_CACHE_TILE_L3);
 332                 pte = set_remote_cache_cpu(pte, home);
 333                 break;
 334         }
 335
 336         if (noallocl2)
 337                 pte = hv_pte_set_no_alloc_l2(pte);
 338
 339         /* Simplify "no local and no l3" to "uncached" */
 340         if (hv_pte_get_no_alloc_l2(pte) && hv_pte_get_no_alloc_l1(pte) &&
 341             hv_pte_get_mode(pte) == HV_PTE_MODE_CACHE_NO_L3) {
 342                 pte = hv_pte_set_mode(pte, HV_PTE_MODE_UNCACHED);
 343         }
 344
 345         /* Checking this case here gives a better panic than from the hv. */
 346         BUG_ON(hv_pte_get_mode(pte) == 0);
 347
 348         return pte;
 349 }
 350 EXPORT_SYMBOL(pte_set_home);
 351
 352 /*
 353  * The routines in this section are the "static" versions of the normal
 354  * dynamic homecaching routines; they just set the home cache
 355  * of a kernel page once, and require a full-chip cache/TLB flush,
 356  * so they're not suitable for anything but infrequent use.
 357  */
 358
 359 int page_home(struct page *page)
 360 {
 361         if (PageHighMem(page)) {
 362                 return PAGE_HOME_HASH;
 363         } else {
 364                 unsigned long kva = (unsigned long)page_address(page);
 365                 return pte_to_home(*virt_to_kpte(kva));
 366         }
 367 }
 368 EXPORT_SYMBOL(page_home);
 369
 370 void homecache_change_page_home(struct page *page, int order, int home)
 371 {
 372         int i, pages = (1 << order);
 373         unsigned long kva;
 374
 375         BUG_ON(PageHighMem(page));
 376         BUG_ON(page_count(page) > 1);
 377         BUG_ON(page_mapcount(page) != 0);
 378         kva = (unsigned long) page_address(page);
 379         flush_remote(0, HV_FLUSH_EVICT_L2, &cpu_cacheable_map,
 380                      kva, pages * PAGE_SIZE, PAGE_SIZE, cpu_online_mask,
 381                      NULL, 0);
 382
 383         for (i = 0; i < pages; ++i, kva += PAGE_SIZE) {
 384                 pte_t *ptep = virt_to_kpte(kva);
 385                 pte_t pteval = *ptep;
 386                 BUG_ON(!pte_present(pteval) || pte_huge(pteval));
 387                 __set_pte(ptep, pte_set_home(pteval, home));
 388         }
 389 }
 390 EXPORT_SYMBOL(homecache_change_page_home);
 391
 392 struct page *homecache_alloc_pages(gfp_t gfp_mask,
 393                                    unsigned int order, int home)
 394 {
 395         struct page *page;
 396         BUG_ON(gfp_mask & __GFP_HIGHMEM);   /* must be lowmem */
 397         page = alloc_pages(gfp_mask, order);
 398         if (page)
 399                 homecache_change_page_home(page, order, home);
 400         return page;
 401 }
 402 EXPORT_SYMBOL(homecache_alloc_pages);
 403
 404 struct page *homecache_alloc_pages_node(int nid, gfp_t gfp_mask,
 405                                         unsigned int order, int home)
 406 {
 407         struct page *page;
 408         BUG_ON(gfp_mask & __GFP_HIGHMEM);   /* must be lowmem */
 409         page = alloc_pages_node(nid, gfp_mask, order);
 410         if (page)
 411                 homecache_change_page_home(page, order, home);
 412         return page;
 413 }
 414
 415 void __homecache_free_pages(struct page *page, unsigned int order)
 416 {
 417         if (put_page_testzero(page)) {
 418                 homecache_change_page_home(page, order, PAGE_HOME_HASH);
 419                 if (order == 0) {
 420                         free_hot_cold_page(page, 0);
 421                 } else {
 422                         init_page_count(page);
 423                         __free_pages(page, order);
 424                 }
 425         }
 426 }
 427 EXPORT_SYMBOL(__homecache_free_pages);
 428
 429 void homecache_free_pages(unsigned long addr, unsigned int order)
 430 {
 431         if (addr != 0) {
 432                 VM_BUG_ON(!virt_addr_valid((void *)addr));
 433                 __homecache_free_pages(virt_to_page((void *)addr), order);
 434         }
 435 }
 436 EXPORT_SYMBOL(homecache_free_pages);