| 1 | /* |
| 2 | * Copyright 2015-2016, Aneesh Kumar K.V, IBM Corporation. |
| 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; either version |
| 7 | * 2 of the License, or (at your option) any later version. |
| 8 | */ |
| 9 | |
| 10 | #include <linux/sched.h> |
| 11 | #include <linux/mm_types.h> |
| 12 | #include <linux/memblock.h> |
| 13 | #include <misc/cxl-base.h> |
| 14 | |
| 15 | #include <asm/pgalloc.h> |
| 16 | #include <asm/tlb.h> |
| 17 | #include <asm/trace.h> |
| 18 | #include <asm/powernv.h> |
| 19 | |
| 20 | #include "mmu_decl.h" |
| 21 | #include <trace/events/thp.h> |
| 22 | |
| 23 | unsigned long __pmd_frag_nr; |
| 24 | EXPORT_SYMBOL(__pmd_frag_nr); |
| 25 | unsigned long __pmd_frag_size_shift; |
| 26 | EXPORT_SYMBOL(__pmd_frag_size_shift); |
| 27 | |
| 28 | int (*register_process_table)(unsigned long base, unsigned long page_size, |
| 29 | unsigned long tbl_size); |
| 30 | |
| 31 | #ifdef CONFIG_TRANSPARENT_HUGEPAGE |
| 32 | /* |
| 33 | * This is called when relaxing access to a hugepage. It's also called in the page |
| 34 | * fault path when we don't hit any of the major fault cases, ie, a minor |
| 35 | * update of _PAGE_ACCESSED, _PAGE_DIRTY, etc... The generic code will have |
| 36 | * handled those two for us, we additionally deal with missing execute |
| 37 | * permission here on some processors |
| 38 | */ |
| 39 | int pmdp_set_access_flags(struct vm_area_struct *vma, unsigned long address, |
| 40 | pmd_t *pmdp, pmd_t entry, int dirty) |
| 41 | { |
| 42 | int changed; |
| 43 | #ifdef CONFIG_DEBUG_VM |
| 44 | WARN_ON(!pmd_trans_huge(*pmdp) && !pmd_devmap(*pmdp)); |
| 45 | assert_spin_locked(pmd_lockptr(vma->vm_mm, pmdp)); |
| 46 | #endif |
| 47 | changed = !pmd_same(*(pmdp), entry); |
| 48 | if (changed) { |
| 49 | /* |
| 50 | * We can use MMU_PAGE_2M here, because only radix |
| 51 | * path look at the psize. |
| 52 | */ |
| 53 | __ptep_set_access_flags(vma, pmdp_ptep(pmdp), |
| 54 | pmd_pte(entry), address, MMU_PAGE_2M); |
| 55 | } |
| 56 | return changed; |
| 57 | } |
| 58 | |
| 59 | int pmdp_test_and_clear_young(struct vm_area_struct *vma, |
| 60 | unsigned long address, pmd_t *pmdp) |
| 61 | { |
| 62 | return __pmdp_test_and_clear_young(vma->vm_mm, address, pmdp); |
| 63 | } |
| 64 | /* |
| 65 | * set a new huge pmd. We should not be called for updating |
| 66 | * an existing pmd entry. That should go via pmd_hugepage_update. |
| 67 | */ |
| 68 | void set_pmd_at(struct mm_struct *mm, unsigned long addr, |
| 69 | pmd_t *pmdp, pmd_t pmd) |
| 70 | { |
| 71 | #ifdef CONFIG_DEBUG_VM |
| 72 | /* |
| 73 | * Make sure hardware valid bit is not set. We don't do |
| 74 | * tlb flush for this update. |
| 75 | */ |
| 76 | |
| 77 | WARN_ON(pte_hw_valid(pmd_pte(*pmdp)) && !pte_protnone(pmd_pte(*pmdp))); |
| 78 | assert_spin_locked(pmd_lockptr(mm, pmdp)); |
| 79 | WARN_ON(!(pmd_large(pmd) || pmd_devmap(pmd))); |
| 80 | #endif |
| 81 | trace_hugepage_set_pmd(addr, pmd_val(pmd)); |
| 82 | return set_pte_at(mm, addr, pmdp_ptep(pmdp), pmd_pte(pmd)); |
| 83 | } |
| 84 | |
| 85 | static void do_nothing(void *unused) |
| 86 | { |
| 87 | |
| 88 | } |
| 89 | /* |
| 90 | * Serialize against find_current_mm_pte which does lock-less |
| 91 | * lookup in page tables with local interrupts disabled. For huge pages |
| 92 | * it casts pmd_t to pte_t. Since format of pte_t is different from |
| 93 | * pmd_t we want to prevent transit from pmd pointing to page table |
| 94 | * to pmd pointing to huge page (and back) while interrupts are disabled. |
| 95 | * We clear pmd to possibly replace it with page table pointer in |
| 96 | * different code paths. So make sure we wait for the parallel |
| 97 | * find_current_mm_pte to finish. |
| 98 | */ |
| 99 | void serialize_against_pte_lookup(struct mm_struct *mm) |
| 100 | { |
| 101 | smp_mb(); |
| 102 | smp_call_function_many(mm_cpumask(mm), do_nothing, NULL, 1); |
| 103 | } |
| 104 | |
| 105 | /* |
| 106 | * We use this to invalidate a pmdp entry before switching from a |
| 107 | * hugepte to regular pmd entry. |
| 108 | */ |
| 109 | pmd_t pmdp_invalidate(struct vm_area_struct *vma, unsigned long address, |
| 110 | pmd_t *pmdp) |
| 111 | { |
| 112 | unsigned long old_pmd; |
| 113 | |
| 114 | old_pmd = pmd_hugepage_update(vma->vm_mm, address, pmdp, _PAGE_PRESENT, _PAGE_INVALID); |
| 115 | flush_pmd_tlb_range(vma, address, address + HPAGE_PMD_SIZE); |
| 116 | /* |
| 117 | * This ensures that generic code that rely on IRQ disabling |
| 118 | * to prevent a parallel THP split work as expected. |
| 119 | */ |
| 120 | serialize_against_pte_lookup(vma->vm_mm); |
| 121 | return __pmd(old_pmd); |
| 122 | } |
| 123 | |
| 124 | static pmd_t pmd_set_protbits(pmd_t pmd, pgprot_t pgprot) |
| 125 | { |
| 126 | return __pmd(pmd_val(pmd) | pgprot_val(pgprot)); |
| 127 | } |
| 128 | |
| 129 | pmd_t pfn_pmd(unsigned long pfn, pgprot_t pgprot) |
| 130 | { |
| 131 | unsigned long pmdv; |
| 132 | |
| 133 | pmdv = (pfn << PAGE_SHIFT) & PTE_RPN_MASK; |
| 134 | return pmd_set_protbits(__pmd(pmdv), pgprot); |
| 135 | } |
| 136 | |
| 137 | pmd_t mk_pmd(struct page *page, pgprot_t pgprot) |
| 138 | { |
| 139 | return pfn_pmd(page_to_pfn(page), pgprot); |
| 140 | } |
| 141 | |
| 142 | pmd_t pmd_modify(pmd_t pmd, pgprot_t newprot) |
| 143 | { |
| 144 | unsigned long pmdv; |
| 145 | |
| 146 | pmdv = pmd_val(pmd); |
| 147 | pmdv &= _HPAGE_CHG_MASK; |
| 148 | return pmd_set_protbits(__pmd(pmdv), newprot); |
| 149 | } |
| 150 | |
| 151 | /* |
| 152 | * This is called at the end of handling a user page fault, when the |
| 153 | * fault has been handled by updating a HUGE PMD entry in the linux page tables. |
| 154 | * We use it to preload an HPTE into the hash table corresponding to |
| 155 | * the updated linux HUGE PMD entry. |
| 156 | */ |
| 157 | void update_mmu_cache_pmd(struct vm_area_struct *vma, unsigned long addr, |
| 158 | pmd_t *pmd) |
| 159 | { |
| 160 | if (radix_enabled()) |
| 161 | prefetch((void *)addr); |
| 162 | } |
| 163 | #endif /* CONFIG_TRANSPARENT_HUGEPAGE */ |
| 164 | |
| 165 | /* For use by kexec */ |
| 166 | void mmu_cleanup_all(void) |
| 167 | { |
| 168 | if (radix_enabled()) |
| 169 | radix__mmu_cleanup_all(); |
| 170 | else if (mmu_hash_ops.hpte_clear_all) |
| 171 | mmu_hash_ops.hpte_clear_all(); |
| 172 | } |
| 173 | |
| 174 | #ifdef CONFIG_MEMORY_HOTPLUG |
| 175 | int __meminit create_section_mapping(unsigned long start, unsigned long end, int nid) |
| 176 | { |
| 177 | if (radix_enabled()) |
| 178 | return radix__create_section_mapping(start, end, nid); |
| 179 | |
| 180 | return hash__create_section_mapping(start, end, nid); |
| 181 | } |
| 182 | |
| 183 | int __meminit remove_section_mapping(unsigned long start, unsigned long end) |
| 184 | { |
| 185 | if (radix_enabled()) |
| 186 | return radix__remove_section_mapping(start, end); |
| 187 | |
| 188 | return hash__remove_section_mapping(start, end); |
| 189 | } |
| 190 | #endif /* CONFIG_MEMORY_HOTPLUG */ |
| 191 | |
| 192 | void __init mmu_partition_table_init(void) |
| 193 | { |
| 194 | unsigned long patb_size = 1UL << PATB_SIZE_SHIFT; |
| 195 | unsigned long ptcr; |
| 196 | |
| 197 | BUILD_BUG_ON_MSG((PATB_SIZE_SHIFT > 36), "Partition table size too large."); |
| 198 | partition_tb = __va(memblock_alloc_base(patb_size, patb_size, |
| 199 | MEMBLOCK_ALLOC_ANYWHERE)); |
| 200 | |
| 201 | /* Initialize the Partition Table with no entries */ |
| 202 | memset((void *)partition_tb, 0, patb_size); |
| 203 | |
| 204 | /* |
| 205 | * update partition table control register, |
| 206 | * 64 K size. |
| 207 | */ |
| 208 | ptcr = __pa(partition_tb) | (PATB_SIZE_SHIFT - 12); |
| 209 | mtspr(SPRN_PTCR, ptcr); |
| 210 | powernv_set_nmmu_ptcr(ptcr); |
| 211 | } |
| 212 | |
| 213 | void mmu_partition_table_set_entry(unsigned int lpid, unsigned long dw0, |
| 214 | unsigned long dw1) |
| 215 | { |
| 216 | unsigned long old = be64_to_cpu(partition_tb[lpid].patb0); |
| 217 | |
| 218 | partition_tb[lpid].patb0 = cpu_to_be64(dw0); |
| 219 | partition_tb[lpid].patb1 = cpu_to_be64(dw1); |
| 220 | |
| 221 | /* |
| 222 | * Global flush of TLBs and partition table caches for this lpid. |
| 223 | * The type of flush (hash or radix) depends on what the previous |
| 224 | * use of this partition ID was, not the new use. |
| 225 | */ |
| 226 | asm volatile("ptesync" : : : "memory"); |
| 227 | if (old & PATB_HR) { |
| 228 | asm volatile(PPC_TLBIE_5(%0,%1,2,0,1) : : |
| 229 | "r" (TLBIEL_INVAL_SET_LPID), "r" (lpid)); |
| 230 | asm volatile(PPC_TLBIE_5(%0,%1,2,1,1) : : |
| 231 | "r" (TLBIEL_INVAL_SET_LPID), "r" (lpid)); |
| 232 | trace_tlbie(lpid, 0, TLBIEL_INVAL_SET_LPID, lpid, 2, 0, 1); |
| 233 | } else { |
| 234 | asm volatile(PPC_TLBIE_5(%0,%1,2,0,0) : : |
| 235 | "r" (TLBIEL_INVAL_SET_LPID), "r" (lpid)); |
| 236 | trace_tlbie(lpid, 0, TLBIEL_INVAL_SET_LPID, lpid, 2, 0, 0); |
| 237 | } |
| 238 | /* do we need fixup here ?*/ |
| 239 | asm volatile("eieio; tlbsync; ptesync" : : : "memory"); |
| 240 | } |
| 241 | EXPORT_SYMBOL_GPL(mmu_partition_table_set_entry); |
| 242 | |
| 243 | static pmd_t *get_pmd_from_cache(struct mm_struct *mm) |
| 244 | { |
| 245 | void *pmd_frag, *ret; |
| 246 | |
| 247 | if (PMD_FRAG_NR == 1) |
| 248 | return NULL; |
| 249 | |
| 250 | spin_lock(&mm->page_table_lock); |
| 251 | ret = mm->context.pmd_frag; |
| 252 | if (ret) { |
| 253 | pmd_frag = ret + PMD_FRAG_SIZE; |
| 254 | /* |
| 255 | * If we have taken up all the fragments mark PTE page NULL |
| 256 | */ |
| 257 | if (((unsigned long)pmd_frag & ~PAGE_MASK) == 0) |
| 258 | pmd_frag = NULL; |
| 259 | mm->context.pmd_frag = pmd_frag; |
| 260 | } |
| 261 | spin_unlock(&mm->page_table_lock); |
| 262 | return (pmd_t *)ret; |
| 263 | } |
| 264 | |
| 265 | static pmd_t *__alloc_for_pmdcache(struct mm_struct *mm) |
| 266 | { |
| 267 | void *ret = NULL; |
| 268 | struct page *page; |
| 269 | gfp_t gfp = GFP_KERNEL_ACCOUNT | __GFP_ZERO; |
| 270 | |
| 271 | if (mm == &init_mm) |
| 272 | gfp &= ~__GFP_ACCOUNT; |
| 273 | page = alloc_page(gfp); |
| 274 | if (!page) |
| 275 | return NULL; |
| 276 | if (!pgtable_pmd_page_ctor(page)) { |
| 277 | __free_pages(page, 0); |
| 278 | return NULL; |
| 279 | } |
| 280 | |
| 281 | atomic_set(&page->pt_frag_refcount, 1); |
| 282 | |
| 283 | ret = page_address(page); |
| 284 | /* |
| 285 | * if we support only one fragment just return the |
| 286 | * allocated page. |
| 287 | */ |
| 288 | if (PMD_FRAG_NR == 1) |
| 289 | return ret; |
| 290 | |
| 291 | spin_lock(&mm->page_table_lock); |
| 292 | /* |
| 293 | * If we find pgtable_page set, we return |
| 294 | * the allocated page with single fragement |
| 295 | * count. |
| 296 | */ |
| 297 | if (likely(!mm->context.pmd_frag)) { |
| 298 | atomic_set(&page->pt_frag_refcount, PMD_FRAG_NR); |
| 299 | mm->context.pmd_frag = ret + PMD_FRAG_SIZE; |
| 300 | } |
| 301 | spin_unlock(&mm->page_table_lock); |
| 302 | |
| 303 | return (pmd_t *)ret; |
| 304 | } |
| 305 | |
| 306 | pmd_t *pmd_fragment_alloc(struct mm_struct *mm, unsigned long vmaddr) |
| 307 | { |
| 308 | pmd_t *pmd; |
| 309 | |
| 310 | pmd = get_pmd_from_cache(mm); |
| 311 | if (pmd) |
| 312 | return pmd; |
| 313 | |
| 314 | return __alloc_for_pmdcache(mm); |
| 315 | } |
| 316 | |
| 317 | void pmd_fragment_free(unsigned long *pmd) |
| 318 | { |
| 319 | struct page *page = virt_to_page(pmd); |
| 320 | |
| 321 | BUG_ON(atomic_read(&page->pt_frag_refcount) <= 0); |
| 322 | if (atomic_dec_and_test(&page->pt_frag_refcount)) { |
| 323 | pgtable_pmd_page_dtor(page); |
| 324 | __free_page(page); |
| 325 | } |
| 326 | } |
| 327 | |
| 328 | static inline void pgtable_free(void *table, int index) |
| 329 | { |
| 330 | switch (index) { |
| 331 | case PTE_INDEX: |
| 332 | pte_fragment_free(table, 0); |
| 333 | break; |
| 334 | case PMD_INDEX: |
| 335 | pmd_fragment_free(table); |
| 336 | break; |
| 337 | case PUD_INDEX: |
| 338 | kmem_cache_free(PGT_CACHE(PUD_CACHE_INDEX), table); |
| 339 | break; |
| 340 | #if defined(CONFIG_PPC_4K_PAGES) && defined(CONFIG_HUGETLB_PAGE) |
| 341 | /* 16M hugepd directory at pud level */ |
| 342 | case HTLB_16M_INDEX: |
| 343 | BUILD_BUG_ON(H_16M_CACHE_INDEX <= 0); |
| 344 | kmem_cache_free(PGT_CACHE(H_16M_CACHE_INDEX), table); |
| 345 | break; |
| 346 | /* 16G hugepd directory at the pgd level */ |
| 347 | case HTLB_16G_INDEX: |
| 348 | BUILD_BUG_ON(H_16G_CACHE_INDEX <= 0); |
| 349 | kmem_cache_free(PGT_CACHE(H_16G_CACHE_INDEX), table); |
| 350 | break; |
| 351 | #endif |
| 352 | /* We don't free pgd table via RCU callback */ |
| 353 | default: |
| 354 | BUG(); |
| 355 | } |
| 356 | } |
| 357 | |
| 358 | #ifdef CONFIG_SMP |
| 359 | void pgtable_free_tlb(struct mmu_gather *tlb, void *table, int index) |
| 360 | { |
| 361 | unsigned long pgf = (unsigned long)table; |
| 362 | |
| 363 | BUG_ON(index > MAX_PGTABLE_INDEX_SIZE); |
| 364 | pgf |= index; |
| 365 | tlb_remove_table(tlb, (void *)pgf); |
| 366 | } |
| 367 | |
| 368 | void __tlb_remove_table(void *_table) |
| 369 | { |
| 370 | void *table = (void *)((unsigned long)_table & ~MAX_PGTABLE_INDEX_SIZE); |
| 371 | unsigned int index = (unsigned long)_table & MAX_PGTABLE_INDEX_SIZE; |
| 372 | |
| 373 | return pgtable_free(table, index); |
| 374 | } |
| 375 | #else |
| 376 | void pgtable_free_tlb(struct mmu_gather *tlb, void *table, int index) |
| 377 | { |
| 378 | return pgtable_free(table, index); |
| 379 | } |
| 380 | #endif |
| 381 | |
| 382 | #ifdef CONFIG_PROC_FS |
| 383 | atomic_long_t direct_pages_count[MMU_PAGE_COUNT]; |
| 384 | |
| 385 | void arch_report_meminfo(struct seq_file *m) |
| 386 | { |
| 387 | /* |
| 388 | * Hash maps the memory with one size mmu_linear_psize. |
| 389 | * So don't bother to print these on hash |
| 390 | */ |
| 391 | if (!radix_enabled()) |
| 392 | return; |
| 393 | seq_printf(m, "DirectMap4k: %8lu kB\n", |
| 394 | atomic_long_read(&direct_pages_count[MMU_PAGE_4K]) << 2); |
| 395 | seq_printf(m, "DirectMap64k: %8lu kB\n", |
| 396 | atomic_long_read(&direct_pages_count[MMU_PAGE_64K]) << 6); |
| 397 | seq_printf(m, "DirectMap2M: %8lu kB\n", |
| 398 | atomic_long_read(&direct_pages_count[MMU_PAGE_2M]) << 11); |
| 399 | seq_printf(m, "DirectMap1G: %8lu kB\n", |
| 400 | atomic_long_read(&direct_pages_count[MMU_PAGE_1G]) << 20); |
| 401 | } |
| 402 | #endif /* CONFIG_PROC_FS */ |
| 403 | |
| 404 | /* |
| 405 | * For hash translation mode, we use the deposited table to store hash slot |
| 406 | * information and they are stored at PTRS_PER_PMD offset from related pmd |
| 407 | * location. Hence a pmd move requires deposit and withdraw. |
| 408 | * |
| 409 | * For radix translation with split pmd ptl, we store the deposited table in the |
| 410 | * pmd page. Hence if we have different pmd page we need to withdraw during pmd |
| 411 | * move. |
| 412 | * |
| 413 | * With hash we use deposited table always irrespective of anon or not. |
| 414 | * With radix we use deposited table only for anonymous mapping. |
| 415 | */ |
| 416 | int pmd_move_must_withdraw(struct spinlock *new_pmd_ptl, |
| 417 | struct spinlock *old_pmd_ptl, |
| 418 | struct vm_area_struct *vma) |
| 419 | { |
| 420 | if (radix_enabled()) |
| 421 | return (new_pmd_ptl != old_pmd_ptl) && vma_is_anonymous(vma); |
| 422 | |
| 423 | return true; |
| 424 | } |