arch/powerpc/mm/pgtable-hash64.c

   1 /*
   2  * Copyright 2005, Paul Mackerras, IBM Corporation.
   3  * Copyright 2009, Benjamin Herrenschmidt, IBM Corporation.
   4  * Copyright 2015-2016, Aneesh Kumar K.V, IBM Corporation.
   5  *
   6  * This program is free software; you can redistribute it and/or
   7  * modify it under the terms of the GNU General Public License
   8  * as published by the Free Software Foundation; either version
   9  * 2 of the License, or (at your option) any later version.
  10  */
  11
  12 #include <linux/sched.h>
  13 #include <linux/mm_types.h>
  14 #include <linux/mm.h>
  15
  16 #include <asm/pgalloc.h>
  17 #include <asm/pgtable.h>
  18 #include <asm/sections.h>
  19 #include <asm/mmu.h>
  20 #include <asm/tlb.h>
  21
  22 #include "mmu_decl.h"
  23
  24 #define CREATE_TRACE_POINTS
  25 #include <trace/events/thp.h>
  26
  27 #ifdef CONFIG_SPARSEMEM_VMEMMAP
  28 /*
  29  * vmemmap is the starting address of the virtual address space where
  30  * struct pages are allocated for all possible PFNs present on the system
  31  * including holes and bad memory (hence sparse). These virtual struct
  32  * pages are stored in sequence in this virtual address space irrespective
  33  * of the fact whether the corresponding PFN is valid or not. This achieves
  34  * constant relationship between address of struct page and its PFN.
  35  *
  36  * During boot or memory hotplug operation when a new memory section is
  37  * added, physical memory allocation (including hash table bolting) will
  38  * be performed for the set of struct pages which are part of the memory
  39  * section. This saves memory by not allocating struct pages for PFNs
  40  * which are not valid.
  41  *
  42  *              ----------------------------------------------
  43  *              | PHYSICAL ALLOCATION OF VIRTUAL STRUCT PAGES|
  44  *              ----------------------------------------------
  45  *
  46  *         f000000000000000                  c000000000000000
  47  * vmemmap +--------------+                  +--------------+
  48  *  +      |  page struct | +--------------> |  page struct |
  49  *  |      +--------------+                  +--------------+
  50  *  |      |  page struct | +--------------> |  page struct |
  51  *  |      +--------------+ |                +--------------+
  52  *  |      |  page struct | +       +------> |  page struct |
  53  *  |      +--------------+         |        +--------------+
  54  *  |      |  page struct |         |   +--> |  page struct |
  55  *  |      +--------------+         |   |    +--------------+
  56  *  |      |  page struct |         |   |
  57  *  |      +--------------+         |   |
  58  *  |      |  page struct |         |   |
  59  *  |      +--------------+         |   |
  60  *  |      |  page struct |         |   |
  61  *  |      +--------------+         |   |
  62  *  |      |  page struct |         |   |
  63  *  |      +--------------+         |   |
  64  *  |      |  page struct | +-------+   |
  65  *  |      +--------------+             |
  66  *  |      |  page struct | +-----------+
  67  *  |      +--------------+
  68  *  |      |  page struct | No mapping
  69  *  |      +--------------+
  70  *  |      |  page struct | No mapping
  71  *  v      +--------------+
  72  *
  73  *              -----------------------------------------
  74  *              | RELATION BETWEEN STRUCT PAGES AND PFNS|
  75  *              -----------------------------------------
  76  *
  77  * vmemmap +--------------+                 +---------------+
  78  *  +      |  page struct | +-------------> |      PFN      |
  79  *  |      +--------------+                 +---------------+
  80  *  |      |  page struct | +-------------> |      PFN      |
  81  *  |      +--------------+                 +---------------+
  82  *  |      |  page struct | +-------------> |      PFN      |
  83  *  |      +--------------+                 +---------------+
  84  *  |      |  page struct | +-------------> |      PFN      |
  85  *  |      +--------------+                 +---------------+
  86  *  |      |              |
  87  *  |      +--------------+
  88  *  |      |              |
  89  *  |      +--------------+
  90  *  |      |              |
  91  *  |      +--------------+                 +---------------+
  92  *  |      |  page struct | +-------------> |      PFN      |
  93  *  |      +--------------+                 +---------------+
  94  *  |      |              |
  95  *  |      +--------------+
  96  *  |      |              |
  97  *  |      +--------------+                 +---------------+
  98  *  |      |  page struct | +-------------> |      PFN      |
  99  *  |      +--------------+                 +---------------+
 100  *  |      |  page struct | +-------------> |      PFN      |
 101  *  v      +--------------+                 +---------------+
 102  */
 103 /*
 104  * On hash-based CPUs, the vmemmap is bolted in the hash table.
 105  *
 106  */
 107 int __meminit hash__vmemmap_create_mapping(unsigned long start,
 108                                        unsigned long page_size,
 109                                        unsigned long phys)
 110 {
 111         int rc = htab_bolt_mapping(start, start + page_size, phys,
 112                                    pgprot_val(PAGE_KERNEL),
 113                                    mmu_vmemmap_psize, mmu_kernel_ssize);
 114         if (rc < 0) {
 115                 int rc2 = htab_remove_mapping(start, start + page_size,
 116                                               mmu_vmemmap_psize,
 117                                               mmu_kernel_ssize);
 118                 BUG_ON(rc2 && (rc2 != -ENOENT));
 119         }
 120         return rc;
 121 }
 122
 123 #ifdef CONFIG_MEMORY_HOTPLUG
 124 void hash__vmemmap_remove_mapping(unsigned long start,
 125                               unsigned long page_size)
 126 {
 127         int rc = htab_remove_mapping(start, start + page_size,
 128                                      mmu_vmemmap_psize,
 129                                      mmu_kernel_ssize);
 130         BUG_ON((rc < 0) && (rc != -ENOENT));
 131         WARN_ON(rc == -ENOENT);
 132 }
 133 #endif
 134 #endif /* CONFIG_SPARSEMEM_VMEMMAP */
 135
 136 /*
 137  * map_kernel_page currently only called by __ioremap
 138  * map_kernel_page adds an entry to the ioremap page table
 139  * and adds an entry to the HPT, possibly bolting it
 140  */
 141 int hash__map_kernel_page(unsigned long ea, unsigned long pa, unsigned long flags)
 142 {
 143         pgd_t *pgdp;
 144         pud_t *pudp;
 145         pmd_t *pmdp;
 146         pte_t *ptep;
 147
 148         BUILD_BUG_ON(TASK_SIZE_USER64 > H_PGTABLE_RANGE);
 149         if (slab_is_available()) {
 150                 pgdp = pgd_offset_k(ea);
 151                 pudp = pud_alloc(&init_mm, pgdp, ea);
 152                 if (!pudp)
 153                         return -ENOMEM;
 154                 pmdp = pmd_alloc(&init_mm, pudp, ea);
 155                 if (!pmdp)
 156                         return -ENOMEM;
 157                 ptep = pte_alloc_kernel(pmdp, ea);
 158                 if (!ptep)
 159                         return -ENOMEM;
 160                 set_pte_at(&init_mm, ea, ptep, pfn_pte(pa >> PAGE_SHIFT,
 161                                                           __pgprot(flags)));
 162         } else {
 163                 /*
 164                  * If the mm subsystem is not fully up, we cannot create a
 165                  * linux page table entry for this mapping.  Simply bolt an
 166                  * entry in the hardware page table.
 167                  *
 168                  */
 169                 if (htab_bolt_mapping(ea, ea + PAGE_SIZE, pa, flags,
 170                                       mmu_io_psize, mmu_kernel_ssize)) {
 171                         printk(KERN_ERR "Failed to do bolted mapping IO "
 172                                "memory at %016lx !\n", pa);
 173                         return -ENOMEM;
 174                 }
 175         }
 176
 177         smp_wmb();
 178         return 0;
 179 }
 180
 181 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 182
 183 unsigned long hash__pmd_hugepage_update(struct mm_struct *mm, unsigned long addr,
 184                                     pmd_t *pmdp, unsigned long clr,
 185                                     unsigned long set)
 186 {
 187         __be64 old_be, tmp;
 188         unsigned long old;
 189
 190 #ifdef CONFIG_DEBUG_VM
 191         WARN_ON(!hash__pmd_trans_huge(*pmdp) && !pmd_devmap(*pmdp));
 192         assert_spin_locked(&mm->page_table_lock);
 193 #endif
 194
 195         __asm__ __volatile__(
 196         "1:     ldarx   %0,0,%3\n\
 197                 and.    %1,%0,%6\n\
 198                 bne-    1b \n\
 199                 andc    %1,%0,%4 \n\
 200                 or      %1,%1,%7\n\
 201                 stdcx.  %1,0,%3 \n\
 202                 bne-    1b"
 203         : "=&r" (old_be), "=&r" (tmp), "=m" (*pmdp)
 204         : "r" (pmdp), "r" (cpu_to_be64(clr)), "m" (*pmdp),
 205           "r" (cpu_to_be64(H_PAGE_BUSY)), "r" (cpu_to_be64(set))
 206         : "cc" );
 207
 208         old = be64_to_cpu(old_be);
 209
 210         trace_hugepage_update(addr, old, clr, set);
 211         if (old & H_PAGE_HASHPTE)
 212                 hpte_do_hugepage_flush(mm, addr, pmdp, old);
 213         return old;
 214 }
 215
 216 pmd_t hash__pmdp_collapse_flush(struct vm_area_struct *vma, unsigned long address,
 217                             pmd_t *pmdp)
 218 {
 219         pmd_t pmd;
 220
 221         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
 222         VM_BUG_ON(pmd_trans_huge(*pmdp));
 223         VM_BUG_ON(pmd_devmap(*pmdp));
 224
 225         pmd = *pmdp;
 226         pmd_clear(pmdp);
 227         /*
 228          * Wait for all pending hash_page to finish. This is needed
 229          * in case of subpage collapse. When we collapse normal pages
 230          * to hugepage, we first clear the pmd, then invalidate all
 231          * the PTE entries. The assumption here is that any low level
 232          * page fault will see a none pmd and take the slow path that
 233          * will wait on mmap_sem. But we could very well be in a
 234          * hash_page with local ptep pointer value. Such a hash page
 235          * can result in adding new HPTE entries for normal subpages.
 236          * That means we could be modifying the page content as we
 237          * copy them to a huge page. So wait for parallel hash_page
 238          * to finish before invalidating HPTE entries. We can do this
 239          * by sending an IPI to all the cpus and executing a dummy
 240          * function there.
 241          */
 242         serialize_against_pte_lookup(vma->vm_mm);
 243         /*
 244          * Now invalidate the hpte entries in the range
 245          * covered by pmd. This make sure we take a
 246          * fault and will find the pmd as none, which will
 247          * result in a major fault which takes mmap_sem and
 248          * hence wait for collapse to complete. Without this
 249          * the __collapse_huge_page_copy can result in copying
 250          * the old content.
 251          */
 252         flush_tlb_pmd_range(vma->vm_mm, &pmd, address);
 253         return pmd;
 254 }
 255
 256 /*
 257  * We want to put the pgtable in pmd and use pgtable for tracking
 258  * the base page size hptes
 259  */
 260 void hash__pgtable_trans_huge_deposit(struct mm_struct *mm, pmd_t *pmdp,
 261                                   pgtable_t pgtable)
 262 {
 263         pgtable_t *pgtable_slot;
 264         assert_spin_locked(&mm->page_table_lock);
 265         /*
 266          * we store the pgtable in the second half of PMD
 267          */
 268         pgtable_slot = (pgtable_t *)pmdp + PTRS_PER_PMD;
 269         *pgtable_slot = pgtable;
 270         /*
 271          * expose the deposited pgtable to other cpus.
 272          * before we set the hugepage PTE at pmd level
 273          * hash fault code looks at the deposted pgtable
 274          * to store hash index values.
 275          */
 276         smp_wmb();
 277 }
 278
 279 pgtable_t hash__pgtable_trans_huge_withdraw(struct mm_struct *mm, pmd_t *pmdp)
 280 {
 281         pgtable_t pgtable;
 282         pgtable_t *pgtable_slot;
 283
 284         assert_spin_locked(&mm->page_table_lock);
 285         pgtable_slot = (pgtable_t *)pmdp + PTRS_PER_PMD;
 286         pgtable = *pgtable_slot;
 287         /*
 288          * Once we withdraw, mark the entry NULL.
 289          */
 290         *pgtable_slot = NULL;
 291         /*
 292          * We store HPTE information in the deposited PTE fragment.
 293          * zero out the content on withdraw.
 294          */
 295         memset(pgtable, 0, PTE_FRAG_SIZE);
 296         return pgtable;
 297 }
 298
 299 void hash__pmdp_huge_split_prepare(struct vm_area_struct *vma,
 300                                unsigned long address, pmd_t *pmdp)
 301 {
 302         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
 303         VM_BUG_ON(REGION_ID(address) != USER_REGION_ID);
 304         VM_BUG_ON(pmd_devmap(*pmdp));
 305
 306         /*
 307          * We can't mark the pmd none here, because that will cause a race
 308          * against exit_mmap. We need to continue mark pmd TRANS HUGE, while
 309          * we spilt, but at the same time we wan't rest of the ppc64 code
 310          * not to insert hash pte on this, because we will be modifying
 311          * the deposited pgtable in the caller of this function. Hence
 312          * clear the _PAGE_USER so that we move the fault handling to
 313          * higher level function and that will serialize against ptl.
 314          * We need to flush existing hash pte entries here even though,
 315          * the translation is still valid, because we will withdraw
 316          * pgtable_t after this.
 317          */
 318         pmd_hugepage_update(vma->vm_mm, address, pmdp, 0, _PAGE_PRIVILEGED);
 319 }
 320
 321 /*
 322  * A linux hugepage PMD was changed and the corresponding hash table entries
 323  * neesd to be flushed.
 324  */
 325 void hpte_do_hugepage_flush(struct mm_struct *mm, unsigned long addr,
 326                             pmd_t *pmdp, unsigned long old_pmd)
 327 {
 328         int ssize;
 329         unsigned int psize;
 330         unsigned long vsid;
 331         unsigned long flags = 0;
 332
 333         /* get the base page size,vsid and segment size */
 334 #ifdef CONFIG_DEBUG_VM
 335         psize = get_slice_psize(mm, addr);
 336         BUG_ON(psize == MMU_PAGE_16M);
 337 #endif
 338         if (old_pmd & H_PAGE_COMBO)
 339                 psize = MMU_PAGE_4K;
 340         else
 341                 psize = MMU_PAGE_64K;
 342
 343         if (!is_kernel_addr(addr)) {
 344                 ssize = user_segment_size(addr);
 345                 vsid = get_vsid(mm->context.id, addr, ssize);
 346                 WARN_ON(vsid == 0);
 347         } else {
 348                 vsid = get_kernel_vsid(addr, mmu_kernel_ssize);
 349                 ssize = mmu_kernel_ssize;
 350         }
 351
 352         if (mm_is_thread_local(mm))
 353                 flags |= HPTE_LOCAL_UPDATE;
 354
 355         return flush_hash_hugepage(vsid, addr, pmdp, psize, ssize, flags);
 356 }
 357
 358 pmd_t hash__pmdp_huge_get_and_clear(struct mm_struct *mm,
 359                                 unsigned long addr, pmd_t *pmdp)
 360 {
 361         pmd_t old_pmd;
 362         pgtable_t pgtable;
 363         unsigned long old;
 364         pgtable_t *pgtable_slot;
 365
 366         old = pmd_hugepage_update(mm, addr, pmdp, ~0UL, 0);
 367         old_pmd = __pmd(old);
 368         /*
 369          * We have pmd == none and we are holding page_table_lock.
 370          * So we can safely go and clear the pgtable hash
 371          * index info.
 372          */
 373         pgtable_slot = (pgtable_t *)pmdp + PTRS_PER_PMD;
 374         pgtable = *pgtable_slot;
 375         /*
 376          * Let's zero out old valid and hash index details
 377          * hash fault look at them.
 378          */
 379         memset(pgtable, 0, PTE_FRAG_SIZE);
 380         /*
 381          * Serialize against find_current_mm_pte variants which does lock-less
 382          * lookup in page tables with local interrupts disabled. For huge pages
 383          * it casts pmd_t to pte_t. Since format of pte_t is different from
 384          * pmd_t we want to prevent transit from pmd pointing to page table
 385          * to pmd pointing to huge page (and back) while interrupts are disabled.
 386          * We clear pmd to possibly replace it with page table pointer in
 387          * different code paths. So make sure we wait for the parallel
 388          * find_curren_mm_pte to finish.
 389          */
 390         serialize_against_pte_lookup(mm);
 391         return old_pmd;
 392 }
 393
 394 int hash__has_transparent_hugepage(void)
 395 {
 396
 397         if (!mmu_has_feature(MMU_FTR_16M_PAGE))
 398                 return 0;
 399         /*
 400          * We support THP only if PMD_SIZE is 16MB.
 401          */
 402         if (mmu_psize_defs[MMU_PAGE_16M].shift != PMD_SHIFT)
 403                 return 0;
 404         /*
 405          * We need to make sure that we support 16MB hugepage in a segement
 406          * with base page size 64K or 4K. We only enable THP with a PAGE_SIZE
 407          * of 64K.
 408          */
 409         /*
 410          * If we have 64K HPTE, we will be using that by default
 411          */
 412         if (mmu_psize_defs[MMU_PAGE_64K].shift &&
 413             (mmu_psize_defs[MMU_PAGE_64K].penc[MMU_PAGE_16M] == -1))
 414                 return 0;
 415         /*
 416          * Ok we only have 4K HPTE
 417          */
 418         if (mmu_psize_defs[MMU_PAGE_4K].penc[MMU_PAGE_16M] == -1)
 419                 return 0;
 420
 421         return 1;
 422 }
 423 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
 424
 425 #ifdef CONFIG_STRICT_KERNEL_RWX
 426 static bool hash__change_memory_range(unsigned long start, unsigned long end,
 427                                       unsigned long newpp)
 428 {
 429         unsigned long idx;
 430         unsigned int step, shift;
 431
 432         shift = mmu_psize_defs[mmu_linear_psize].shift;
 433         step = 1 << shift;
 434
 435         start = ALIGN_DOWN(start, step);
 436         end = ALIGN(end, step); // aligns up
 437
 438         if (start >= end)
 439                 return false;
 440
 441         pr_debug("Changing page protection on range 0x%lx-0x%lx, to 0x%lx, step 0x%x\n",
 442                  start, end, newpp, step);
 443
 444         for (idx = start; idx < end; idx += step)
 445                 /* Not sure if we can do much with the return value */
 446                 mmu_hash_ops.hpte_updateboltedpp(newpp, idx, mmu_linear_psize,
 447                                                         mmu_kernel_ssize);
 448
 449         return true;
 450 }
 451
 452 void hash__mark_rodata_ro(void)
 453 {
 454         unsigned long start, end;
 455
 456         start = (unsigned long)_stext;
 457         end = (unsigned long)__init_begin;
 458
 459         WARN_ON(!hash__change_memory_range(start, end, PP_RXXX));
 460 }
 461
 462 void hash__mark_initmem_nx(void)
 463 {
 464         unsigned long start, end, pp;
 465
 466         start = (unsigned long)__init_begin;
 467         end = (unsigned long)__init_end;
 468
 469         pp = htab_convert_pte_flags(pgprot_val(PAGE_KERNEL));
 470
 471         WARN_ON(!hash__change_memory_range(start, end, pp));
 472 }
 473 #endif