2 * Copyright 2015-2016, Aneesh Kumar K.V, IBM Corporation.
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.
10 #include <linux/sched.h>
11 #include <linux/mm_types.h>
12 #include <linux/memblock.h>
13 #include <misc/cxl-base.h>
15 #include <asm/pgalloc.h>
17 #include <asm/trace.h>
18 #include <asm/powernv.h>
21 #include <trace/events/thp.h>
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);
28 int (*register_process_table)(unsigned long base, unsigned long page_size,
29 unsigned long tbl_size);
31 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
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
39 int pmdp_set_access_flags(struct vm_area_struct *vma, unsigned long address,
40 pmd_t *pmdp, pmd_t entry, int dirty)
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));
47 changed = !pmd_same(*(pmdp), entry);
49 __ptep_set_access_flags(vma->vm_mm, pmdp_ptep(pmdp),
50 pmd_pte(entry), address);
51 flush_pmd_tlb_range(vma, address, address + HPAGE_PMD_SIZE);
56 int pmdp_test_and_clear_young(struct vm_area_struct *vma,
57 unsigned long address, pmd_t *pmdp)
59 return __pmdp_test_and_clear_young(vma->vm_mm, address, pmdp);
62 * set a new huge pmd. We should not be called for updating
63 * an existing pmd entry. That should go via pmd_hugepage_update.
65 void set_pmd_at(struct mm_struct *mm, unsigned long addr,
66 pmd_t *pmdp, pmd_t pmd)
68 #ifdef CONFIG_DEBUG_VM
69 WARN_ON(pte_present(pmd_pte(*pmdp)) && !pte_protnone(pmd_pte(*pmdp)));
70 assert_spin_locked(pmd_lockptr(mm, pmdp));
71 WARN_ON(!(pmd_trans_huge(pmd) || pmd_devmap(pmd)));
73 trace_hugepage_set_pmd(addr, pmd_val(pmd));
74 return set_pte_at(mm, addr, pmdp_ptep(pmdp), pmd_pte(pmd));
77 static void do_nothing(void *unused)
82 * Serialize against find_current_mm_pte which does lock-less
83 * lookup in page tables with local interrupts disabled. For huge pages
84 * it casts pmd_t to pte_t. Since format of pte_t is different from
85 * pmd_t we want to prevent transit from pmd pointing to page table
86 * to pmd pointing to huge page (and back) while interrupts are disabled.
87 * We clear pmd to possibly replace it with page table pointer in
88 * different code paths. So make sure we wait for the parallel
89 * find_current_mm_pte to finish.
91 void serialize_against_pte_lookup(struct mm_struct *mm)
94 smp_call_function_many(mm_cpumask(mm), do_nothing, NULL, 1);
98 * We use this to invalidate a pmdp entry before switching from a
99 * hugepte to regular pmd entry.
101 pmd_t pmdp_invalidate(struct vm_area_struct *vma, unsigned long address,
104 unsigned long old_pmd;
106 old_pmd = pmd_hugepage_update(vma->vm_mm, address, pmdp, _PAGE_PRESENT, 0);
107 flush_pmd_tlb_range(vma, address, address + HPAGE_PMD_SIZE);
109 * This ensures that generic code that rely on IRQ disabling
110 * to prevent a parallel THP split work as expected.
112 serialize_against_pte_lookup(vma->vm_mm);
113 return __pmd(old_pmd);
116 static pmd_t pmd_set_protbits(pmd_t pmd, pgprot_t pgprot)
118 return __pmd(pmd_val(pmd) | pgprot_val(pgprot));
121 pmd_t pfn_pmd(unsigned long pfn, pgprot_t pgprot)
125 pmdv = (pfn << PAGE_SHIFT) & PTE_RPN_MASK;
126 return pmd_set_protbits(__pmd(pmdv), pgprot);
129 pmd_t mk_pmd(struct page *page, pgprot_t pgprot)
131 return pfn_pmd(page_to_pfn(page), pgprot);
134 pmd_t pmd_modify(pmd_t pmd, pgprot_t newprot)
139 pmdv &= _HPAGE_CHG_MASK;
140 return pmd_set_protbits(__pmd(pmdv), newprot);
144 * This is called at the end of handling a user page fault, when the
145 * fault has been handled by updating a HUGE PMD entry in the linux page tables.
146 * We use it to preload an HPTE into the hash table corresponding to
147 * the updated linux HUGE PMD entry.
149 void update_mmu_cache_pmd(struct vm_area_struct *vma, unsigned long addr,
154 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
156 /* For use by kexec */
157 void mmu_cleanup_all(void)
160 radix__mmu_cleanup_all();
161 else if (mmu_hash_ops.hpte_clear_all)
162 mmu_hash_ops.hpte_clear_all();
165 #ifdef CONFIG_MEMORY_HOTPLUG
166 int __meminit create_section_mapping(unsigned long start, unsigned long end, int nid)
169 return radix__create_section_mapping(start, end, nid);
171 return hash__create_section_mapping(start, end, nid);
174 int __meminit remove_section_mapping(unsigned long start, unsigned long end)
177 return radix__remove_section_mapping(start, end);
179 return hash__remove_section_mapping(start, end);
181 #endif /* CONFIG_MEMORY_HOTPLUG */
183 void __init mmu_partition_table_init(void)
185 unsigned long patb_size = 1UL << PATB_SIZE_SHIFT;
188 BUILD_BUG_ON_MSG((PATB_SIZE_SHIFT > 36), "Partition table size too large.");
189 partition_tb = __va(memblock_alloc_base(patb_size, patb_size,
190 MEMBLOCK_ALLOC_ANYWHERE));
192 /* Initialize the Partition Table with no entries */
193 memset((void *)partition_tb, 0, patb_size);
196 * update partition table control register,
199 ptcr = __pa(partition_tb) | (PATB_SIZE_SHIFT - 12);
200 mtspr(SPRN_PTCR, ptcr);
201 powernv_set_nmmu_ptcr(ptcr);
204 void mmu_partition_table_set_entry(unsigned int lpid, unsigned long dw0,
207 unsigned long old = be64_to_cpu(partition_tb[lpid].patb0);
209 partition_tb[lpid].patb0 = cpu_to_be64(dw0);
210 partition_tb[lpid].patb1 = cpu_to_be64(dw1);
213 * Global flush of TLBs and partition table caches for this lpid.
214 * The type of flush (hash or radix) depends on what the previous
215 * use of this partition ID was, not the new use.
217 asm volatile("ptesync" : : : "memory");
219 asm volatile(PPC_TLBIE_5(%0,%1,2,0,1) : :
220 "r" (TLBIEL_INVAL_SET_LPID), "r" (lpid));
221 asm volatile(PPC_TLBIE_5(%0,%1,2,1,1) : :
222 "r" (TLBIEL_INVAL_SET_LPID), "r" (lpid));
223 trace_tlbie(lpid, 0, TLBIEL_INVAL_SET_LPID, lpid, 2, 0, 1);
225 asm volatile(PPC_TLBIE_5(%0,%1,2,0,0) : :
226 "r" (TLBIEL_INVAL_SET_LPID), "r" (lpid));
227 trace_tlbie(lpid, 0, TLBIEL_INVAL_SET_LPID, lpid, 2, 0, 0);
229 /* do we need fixup here ?*/
230 asm volatile("eieio; tlbsync; ptesync" : : : "memory");
232 EXPORT_SYMBOL_GPL(mmu_partition_table_set_entry);
234 static pmd_t *get_pmd_from_cache(struct mm_struct *mm)
236 void *pmd_frag, *ret;
238 spin_lock(&mm->page_table_lock);
239 ret = mm->context.pmd_frag;
241 pmd_frag = ret + PMD_FRAG_SIZE;
243 * If we have taken up all the fragments mark PTE page NULL
245 if (((unsigned long)pmd_frag & ~PAGE_MASK) == 0)
247 mm->context.pmd_frag = pmd_frag;
249 spin_unlock(&mm->page_table_lock);
253 static pmd_t *__alloc_for_pmdcache(struct mm_struct *mm)
257 gfp_t gfp = GFP_KERNEL_ACCOUNT | __GFP_ZERO;
260 gfp &= ~__GFP_ACCOUNT;
261 page = alloc_page(gfp);
264 if (!pgtable_pmd_page_ctor(page)) {
265 __free_pages(page, 0);
269 ret = page_address(page);
271 * if we support only one fragment just return the
274 if (PMD_FRAG_NR == 1)
277 spin_lock(&mm->page_table_lock);
279 * If we find pgtable_page set, we return
280 * the allocated page with single fragement
283 if (likely(!mm->context.pmd_frag)) {
284 set_page_count(page, PMD_FRAG_NR);
285 mm->context.pmd_frag = ret + PMD_FRAG_SIZE;
287 spin_unlock(&mm->page_table_lock);
292 pmd_t *pmd_fragment_alloc(struct mm_struct *mm, unsigned long vmaddr)
296 pmd = get_pmd_from_cache(mm);
300 return __alloc_for_pmdcache(mm);
303 void pmd_fragment_free(unsigned long *pmd)
305 struct page *page = virt_to_page(pmd);
307 if (put_page_testzero(page)) {
308 pgtable_pmd_page_dtor(page);
309 free_unref_page(page);
313 static pte_t *get_pte_from_cache(struct mm_struct *mm)
315 void *pte_frag, *ret;
317 spin_lock(&mm->page_table_lock);
318 ret = mm->context.pte_frag;
320 pte_frag = ret + PTE_FRAG_SIZE;
322 * If we have taken up all the fragments mark PTE page NULL
324 if (((unsigned long)pte_frag & ~PAGE_MASK) == 0)
326 mm->context.pte_frag = pte_frag;
328 spin_unlock(&mm->page_table_lock);
332 static pte_t *__alloc_for_ptecache(struct mm_struct *mm, int kernel)
338 page = alloc_page(PGALLOC_GFP | __GFP_ACCOUNT);
341 if (!pgtable_page_ctor(page)) {
346 page = alloc_page(PGALLOC_GFP);
352 ret = page_address(page);
354 * if we support only one fragment just return the
357 if (PTE_FRAG_NR == 1)
359 spin_lock(&mm->page_table_lock);
361 * If we find pgtable_page set, we return
362 * the allocated page with single fragement
365 if (likely(!mm->context.pte_frag)) {
366 set_page_count(page, PTE_FRAG_NR);
367 mm->context.pte_frag = ret + PTE_FRAG_SIZE;
369 spin_unlock(&mm->page_table_lock);
374 pte_t *pte_fragment_alloc(struct mm_struct *mm, unsigned long vmaddr, int kernel)
378 pte = get_pte_from_cache(mm);
382 return __alloc_for_ptecache(mm, kernel);
385 void pte_fragment_free(unsigned long *table, int kernel)
387 struct page *page = virt_to_page(table);
389 if (put_page_testzero(page)) {
391 pgtable_page_dtor(page);
392 free_unref_page(page);
396 static inline void pgtable_free(void *table, int index)
400 pte_fragment_free(table, 0);
403 pmd_fragment_free(table);
406 kmem_cache_free(PGT_CACHE(PUD_CACHE_INDEX), table);
408 /* We don't free pgd table via RCU callback */
415 void pgtable_free_tlb(struct mmu_gather *tlb, void *table, int index)
417 unsigned long pgf = (unsigned long)table;
419 BUG_ON(index > MAX_PGTABLE_INDEX_SIZE);
421 tlb_remove_table(tlb, (void *)pgf);
424 void __tlb_remove_table(void *_table)
426 void *table = (void *)((unsigned long)_table & ~MAX_PGTABLE_INDEX_SIZE);
427 unsigned int index = (unsigned long)_table & MAX_PGTABLE_INDEX_SIZE;
429 return pgtable_free(table, index);
432 void pgtable_free_tlb(struct mmu_gather *tlb, void *table, int index)
434 return pgtable_free(table, index);