1 // SPDX-License-Identifier: GPL-2.0-or-later
3 * Copyright 2015-2016, Aneesh Kumar K.V, IBM Corporation.
6 #include <linux/sched.h>
7 #include <linux/mm_types.h>
8 #include <linux/memblock.h>
9 #include <misc/cxl-base.h>
11 #include <asm/pgalloc.h>
13 #include <asm/trace.h>
14 #include <asm/powernv.h>
16 #include <mm/mmu_decl.h>
17 #include <trace/events/thp.h>
19 unsigned long __pmd_frag_nr;
20 EXPORT_SYMBOL(__pmd_frag_nr);
21 unsigned long __pmd_frag_size_shift;
22 EXPORT_SYMBOL(__pmd_frag_size_shift);
24 int (*register_process_table)(unsigned long base, unsigned long page_size,
25 unsigned long tbl_size);
27 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
29 * This is called when relaxing access to a hugepage. It's also called in the page
30 * fault path when we don't hit any of the major fault cases, ie, a minor
31 * update of _PAGE_ACCESSED, _PAGE_DIRTY, etc... The generic code will have
32 * handled those two for us, we additionally deal with missing execute
33 * permission here on some processors
35 int pmdp_set_access_flags(struct vm_area_struct *vma, unsigned long address,
36 pmd_t *pmdp, pmd_t entry, int dirty)
39 #ifdef CONFIG_DEBUG_VM
40 WARN_ON(!pmd_trans_huge(*pmdp) && !pmd_devmap(*pmdp));
41 assert_spin_locked(pmd_lockptr(vma->vm_mm, pmdp));
43 changed = !pmd_same(*(pmdp), entry);
46 * We can use MMU_PAGE_2M here, because only radix
47 * path look at the psize.
49 __ptep_set_access_flags(vma, pmdp_ptep(pmdp),
50 pmd_pte(entry), address, MMU_PAGE_2M);
55 int pmdp_test_and_clear_young(struct vm_area_struct *vma,
56 unsigned long address, pmd_t *pmdp)
58 return __pmdp_test_and_clear_young(vma->vm_mm, address, pmdp);
61 * set a new huge pmd. We should not be called for updating
62 * an existing pmd entry. That should go via pmd_hugepage_update.
64 void set_pmd_at(struct mm_struct *mm, unsigned long addr,
65 pmd_t *pmdp, pmd_t pmd)
67 #ifdef CONFIG_DEBUG_VM
69 * Make sure hardware valid bit is not set. We don't do
70 * tlb flush for this update.
73 WARN_ON(pte_hw_valid(pmd_pte(*pmdp)) && !pte_protnone(pmd_pte(*pmdp)));
74 assert_spin_locked(pmd_lockptr(mm, pmdp));
75 WARN_ON(!(pmd_large(pmd) || pmd_devmap(pmd)));
77 trace_hugepage_set_pmd(addr, pmd_val(pmd));
78 return set_pte_at(mm, addr, pmdp_ptep(pmdp), pmd_pte(pmd));
81 static void do_nothing(void *unused)
86 * Serialize against find_current_mm_pte which does lock-less
87 * lookup in page tables with local interrupts disabled. For huge pages
88 * it casts pmd_t to pte_t. Since format of pte_t is different from
89 * pmd_t we want to prevent transit from pmd pointing to page table
90 * to pmd pointing to huge page (and back) while interrupts are disabled.
91 * We clear pmd to possibly replace it with page table pointer in
92 * different code paths. So make sure we wait for the parallel
93 * find_current_mm_pte to finish.
95 void serialize_against_pte_lookup(struct mm_struct *mm)
98 smp_call_function_many(mm_cpumask(mm), do_nothing, NULL, 1);
102 * We use this to invalidate a pmdp entry before switching from a
103 * hugepte to regular pmd entry.
105 pmd_t pmdp_invalidate(struct vm_area_struct *vma, unsigned long address,
108 unsigned long old_pmd;
110 old_pmd = pmd_hugepage_update(vma->vm_mm, address, pmdp, _PAGE_PRESENT, _PAGE_INVALID);
111 flush_pmd_tlb_range(vma, address, address + HPAGE_PMD_SIZE);
113 * This ensures that generic code that rely on IRQ disabling
114 * to prevent a parallel THP split work as expected.
116 * Marking the entry with _PAGE_INVALID && ~_PAGE_PRESENT requires
117 * a special case check in pmd_access_permitted.
119 serialize_against_pte_lookup(vma->vm_mm);
120 return __pmd(old_pmd);
123 static pmd_t pmd_set_protbits(pmd_t pmd, pgprot_t pgprot)
125 return __pmd(pmd_val(pmd) | pgprot_val(pgprot));
128 pmd_t pfn_pmd(unsigned long pfn, pgprot_t pgprot)
132 pmdv = (pfn << PAGE_SHIFT) & PTE_RPN_MASK;
133 return pmd_set_protbits(__pmd(pmdv), pgprot);
136 pmd_t mk_pmd(struct page *page, pgprot_t pgprot)
138 return pfn_pmd(page_to_pfn(page), pgprot);
141 pmd_t pmd_modify(pmd_t pmd, pgprot_t newprot)
146 pmdv &= _HPAGE_CHG_MASK;
147 return pmd_set_protbits(__pmd(pmdv), newprot);
151 * This is called at the end of handling a user page fault, when the
152 * fault has been handled by updating a HUGE PMD entry in the linux page tables.
153 * We use it to preload an HPTE into the hash table corresponding to
154 * the updated linux HUGE PMD entry.
156 void update_mmu_cache_pmd(struct vm_area_struct *vma, unsigned long addr,
160 prefetch((void *)addr);
162 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
164 /* For use by kexec */
165 void mmu_cleanup_all(void)
168 radix__mmu_cleanup_all();
169 else if (mmu_hash_ops.hpte_clear_all)
170 mmu_hash_ops.hpte_clear_all();
173 #ifdef CONFIG_MEMORY_HOTPLUG
174 int __meminit create_section_mapping(unsigned long start, unsigned long end, int nid)
177 return radix__create_section_mapping(start, end, nid);
179 return hash__create_section_mapping(start, end, nid);
182 int __meminit remove_section_mapping(unsigned long start, unsigned long end)
185 return radix__remove_section_mapping(start, end);
187 return hash__remove_section_mapping(start, end);
189 #endif /* CONFIG_MEMORY_HOTPLUG */
191 void __init mmu_partition_table_init(void)
193 unsigned long patb_size = 1UL << PATB_SIZE_SHIFT;
196 BUILD_BUG_ON_MSG((PATB_SIZE_SHIFT > 36), "Partition table size too large.");
197 /* Initialize the Partition Table with no entries */
198 partition_tb = memblock_alloc(patb_size, patb_size);
200 panic("%s: Failed to allocate %lu bytes align=0x%lx\n",
201 __func__, patb_size, patb_size);
204 * update partition table control register,
207 ptcr = __pa(partition_tb) | (PATB_SIZE_SHIFT - 12);
208 mtspr(SPRN_PTCR, ptcr);
209 powernv_set_nmmu_ptcr(ptcr);
212 void mmu_partition_table_set_entry(unsigned int lpid, unsigned long dw0,
215 unsigned long old = be64_to_cpu(partition_tb[lpid].patb0);
217 partition_tb[lpid].patb0 = cpu_to_be64(dw0);
218 partition_tb[lpid].patb1 = cpu_to_be64(dw1);
221 * Global flush of TLBs and partition table caches for this lpid.
222 * The type of flush (hash or radix) depends on what the previous
223 * use of this partition ID was, not the new use.
225 asm volatile("ptesync" : : : "memory");
227 asm volatile(PPC_TLBIE_5(%0,%1,2,0,1) : :
228 "r" (TLBIEL_INVAL_SET_LPID), "r" (lpid));
229 asm volatile(PPC_TLBIE_5(%0,%1,2,1,1) : :
230 "r" (TLBIEL_INVAL_SET_LPID), "r" (lpid));
231 trace_tlbie(lpid, 0, TLBIEL_INVAL_SET_LPID, lpid, 2, 0, 1);
233 asm volatile(PPC_TLBIE_5(%0,%1,2,0,0) : :
234 "r" (TLBIEL_INVAL_SET_LPID), "r" (lpid));
235 trace_tlbie(lpid, 0, TLBIEL_INVAL_SET_LPID, lpid, 2, 0, 0);
237 /* do we need fixup here ?*/
238 asm volatile("eieio; tlbsync; ptesync" : : : "memory");
240 EXPORT_SYMBOL_GPL(mmu_partition_table_set_entry);
242 static pmd_t *get_pmd_from_cache(struct mm_struct *mm)
244 void *pmd_frag, *ret;
246 if (PMD_FRAG_NR == 1)
249 spin_lock(&mm->page_table_lock);
250 ret = mm->context.pmd_frag;
252 pmd_frag = ret + PMD_FRAG_SIZE;
254 * If we have taken up all the fragments mark PTE page NULL
256 if (((unsigned long)pmd_frag & ~PAGE_MASK) == 0)
258 mm->context.pmd_frag = pmd_frag;
260 spin_unlock(&mm->page_table_lock);
264 static pmd_t *__alloc_for_pmdcache(struct mm_struct *mm)
268 gfp_t gfp = GFP_KERNEL_ACCOUNT | __GFP_ZERO;
271 gfp &= ~__GFP_ACCOUNT;
272 page = alloc_page(gfp);
275 if (!pgtable_pmd_page_ctor(page)) {
276 __free_pages(page, 0);
280 atomic_set(&page->pt_frag_refcount, 1);
282 ret = page_address(page);
284 * if we support only one fragment just return the
287 if (PMD_FRAG_NR == 1)
290 spin_lock(&mm->page_table_lock);
292 * If we find pgtable_page set, we return
293 * the allocated page with single fragement
296 if (likely(!mm->context.pmd_frag)) {
297 atomic_set(&page->pt_frag_refcount, PMD_FRAG_NR);
298 mm->context.pmd_frag = ret + PMD_FRAG_SIZE;
300 spin_unlock(&mm->page_table_lock);
305 pmd_t *pmd_fragment_alloc(struct mm_struct *mm, unsigned long vmaddr)
309 pmd = get_pmd_from_cache(mm);
313 return __alloc_for_pmdcache(mm);
316 void pmd_fragment_free(unsigned long *pmd)
318 struct page *page = virt_to_page(pmd);
320 BUG_ON(atomic_read(&page->pt_frag_refcount) <= 0);
321 if (atomic_dec_and_test(&page->pt_frag_refcount)) {
322 pgtable_pmd_page_dtor(page);
327 static inline void pgtable_free(void *table, int index)
331 pte_fragment_free(table, 0);
334 pmd_fragment_free(table);
337 kmem_cache_free(PGT_CACHE(PUD_CACHE_INDEX), table);
339 #if defined(CONFIG_PPC_4K_PAGES) && defined(CONFIG_HUGETLB_PAGE)
340 /* 16M hugepd directory at pud level */
342 BUILD_BUG_ON(H_16M_CACHE_INDEX <= 0);
343 kmem_cache_free(PGT_CACHE(H_16M_CACHE_INDEX), table);
345 /* 16G hugepd directory at the pgd level */
347 BUILD_BUG_ON(H_16G_CACHE_INDEX <= 0);
348 kmem_cache_free(PGT_CACHE(H_16G_CACHE_INDEX), table);
351 /* We don't free pgd table via RCU callback */
358 void pgtable_free_tlb(struct mmu_gather *tlb, void *table, int index)
360 unsigned long pgf = (unsigned long)table;
362 BUG_ON(index > MAX_PGTABLE_INDEX_SIZE);
364 tlb_remove_table(tlb, (void *)pgf);
367 void __tlb_remove_table(void *_table)
369 void *table = (void *)((unsigned long)_table & ~MAX_PGTABLE_INDEX_SIZE);
370 unsigned int index = (unsigned long)_table & MAX_PGTABLE_INDEX_SIZE;
372 return pgtable_free(table, index);
375 void pgtable_free_tlb(struct mmu_gather *tlb, void *table, int index)
377 return pgtable_free(table, index);
381 #ifdef CONFIG_PROC_FS
382 atomic_long_t direct_pages_count[MMU_PAGE_COUNT];
384 void arch_report_meminfo(struct seq_file *m)
387 * Hash maps the memory with one size mmu_linear_psize.
388 * So don't bother to print these on hash
390 if (!radix_enabled())
392 seq_printf(m, "DirectMap4k: %8lu kB\n",
393 atomic_long_read(&direct_pages_count[MMU_PAGE_4K]) << 2);
394 seq_printf(m, "DirectMap64k: %8lu kB\n",
395 atomic_long_read(&direct_pages_count[MMU_PAGE_64K]) << 6);
396 seq_printf(m, "DirectMap2M: %8lu kB\n",
397 atomic_long_read(&direct_pages_count[MMU_PAGE_2M]) << 11);
398 seq_printf(m, "DirectMap1G: %8lu kB\n",
399 atomic_long_read(&direct_pages_count[MMU_PAGE_1G]) << 20);
401 #endif /* CONFIG_PROC_FS */
403 pte_t ptep_modify_prot_start(struct vm_area_struct *vma, unsigned long addr,
406 unsigned long pte_val;
409 * Clear the _PAGE_PRESENT so that no hardware parallel update is
410 * possible. Also keep the pte_present true so that we don't take
413 pte_val = pte_update(vma->vm_mm, addr, ptep, _PAGE_PRESENT, _PAGE_INVALID, 0);
415 return __pte(pte_val);
419 void ptep_modify_prot_commit(struct vm_area_struct *vma, unsigned long addr,
420 pte_t *ptep, pte_t old_pte, pte_t pte)
423 return radix__ptep_modify_prot_commit(vma, addr,
425 set_pte_at(vma->vm_mm, addr, ptep, pte);
429 * For hash translation mode, we use the deposited table to store hash slot
430 * information and they are stored at PTRS_PER_PMD offset from related pmd
431 * location. Hence a pmd move requires deposit and withdraw.
433 * For radix translation with split pmd ptl, we store the deposited table in the
434 * pmd page. Hence if we have different pmd page we need to withdraw during pmd
437 * With hash we use deposited table always irrespective of anon or not.
438 * With radix we use deposited table only for anonymous mapping.
440 int pmd_move_must_withdraw(struct spinlock *new_pmd_ptl,
441 struct spinlock *old_pmd_ptl,
442 struct vm_area_struct *vma)
445 return (new_pmd_ptl != old_pmd_ptl) && vma_is_anonymous(vma);