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 <linux/memremap.h>
10 #include <linux/pkeys.h>
11 #include <linux/debugfs.h>
12 #include <misc/cxl-base.h>
14 #include <asm/pgalloc.h>
16 #include <asm/trace.h>
17 #include <asm/powernv.h>
18 #include <asm/firmware.h>
19 #include <asm/ultravisor.h>
20 #include <asm/kexec.h>
22 #include <mm/mmu_decl.h>
23 #include <trace/events/thp.h>
27 struct mmu_psize_def mmu_psize_defs[MMU_PAGE_COUNT];
28 EXPORT_SYMBOL_GPL(mmu_psize_defs);
30 #ifdef CONFIG_SPARSEMEM_VMEMMAP
31 int mmu_vmemmap_psize = MMU_PAGE_4K;
34 unsigned long __pmd_frag_nr;
35 EXPORT_SYMBOL(__pmd_frag_nr);
36 unsigned long __pmd_frag_size_shift;
37 EXPORT_SYMBOL(__pmd_frag_size_shift);
39 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
41 * This is called when relaxing access to a hugepage. It's also called in the page
42 * fault path when we don't hit any of the major fault cases, ie, a minor
43 * update of _PAGE_ACCESSED, _PAGE_DIRTY, etc... The generic code will have
44 * handled those two for us, we additionally deal with missing execute
45 * permission here on some processors
47 int pmdp_set_access_flags(struct vm_area_struct *vma, unsigned long address,
48 pmd_t *pmdp, pmd_t entry, int dirty)
51 #ifdef CONFIG_DEBUG_VM
52 WARN_ON(!pmd_trans_huge(*pmdp) && !pmd_devmap(*pmdp));
53 assert_spin_locked(pmd_lockptr(vma->vm_mm, pmdp));
55 changed = !pmd_same(*(pmdp), entry);
58 * We can use MMU_PAGE_2M here, because only radix
59 * path look at the psize.
61 __ptep_set_access_flags(vma, pmdp_ptep(pmdp),
62 pmd_pte(entry), address, MMU_PAGE_2M);
67 int pmdp_test_and_clear_young(struct vm_area_struct *vma,
68 unsigned long address, pmd_t *pmdp)
70 return __pmdp_test_and_clear_young(vma->vm_mm, address, pmdp);
73 * set a new huge pmd. We should not be called for updating
74 * an existing pmd entry. That should go via pmd_hugepage_update.
76 void set_pmd_at(struct mm_struct *mm, unsigned long addr,
77 pmd_t *pmdp, pmd_t pmd)
79 #ifdef CONFIG_DEBUG_VM
81 * Make sure hardware valid bit is not set. We don't do
82 * tlb flush for this update.
85 WARN_ON(pte_hw_valid(pmd_pte(*pmdp)) && !pte_protnone(pmd_pte(*pmdp)));
86 assert_spin_locked(pmd_lockptr(mm, pmdp));
87 WARN_ON(!(pmd_large(pmd)));
89 trace_hugepage_set_pmd(addr, pmd_val(pmd));
90 return set_pte_at(mm, addr, pmdp_ptep(pmdp), pmd_pte(pmd));
93 static void do_serialize(void *arg)
95 /* We've taken the IPI, so try to trim the mask while here */
96 if (radix_enabled()) {
97 struct mm_struct *mm = arg;
98 exit_lazy_flush_tlb(mm, false);
103 * Serialize against __find_linux_pte() which does lock-less
104 * lookup in page tables with local interrupts disabled. For huge pages
105 * it casts pmd_t to pte_t. Since format of pte_t is different from
106 * pmd_t we want to prevent transit from pmd pointing to page table
107 * to pmd pointing to huge page (and back) while interrupts are disabled.
108 * We clear pmd to possibly replace it with page table pointer in
109 * different code paths. So make sure we wait for the parallel
110 * __find_linux_pte() to finish.
112 void serialize_against_pte_lookup(struct mm_struct *mm)
115 smp_call_function_many(mm_cpumask(mm), do_serialize, mm, 1);
119 * We use this to invalidate a pmdp entry before switching from a
120 * hugepte to regular pmd entry.
122 pmd_t pmdp_invalidate(struct vm_area_struct *vma, unsigned long address,
125 unsigned long old_pmd;
127 old_pmd = pmd_hugepage_update(vma->vm_mm, address, pmdp, _PAGE_PRESENT, _PAGE_INVALID);
128 flush_pmd_tlb_range(vma, address, address + HPAGE_PMD_SIZE);
129 return __pmd(old_pmd);
132 pmd_t pmdp_huge_get_and_clear_full(struct vm_area_struct *vma,
133 unsigned long addr, pmd_t *pmdp, int full)
136 VM_BUG_ON(addr & ~HPAGE_PMD_MASK);
137 VM_BUG_ON((pmd_present(*pmdp) && !pmd_trans_huge(*pmdp) &&
138 !pmd_devmap(*pmdp)) || !pmd_present(*pmdp));
139 pmd = pmdp_huge_get_and_clear(vma->vm_mm, addr, pmdp);
141 * if it not a fullmm flush, then we can possibly end up converting
142 * this PMD pte entry to a regular level 0 PTE by a parallel page fault.
143 * Make sure we flush the tlb in this case.
146 flush_pmd_tlb_range(vma, addr, addr + HPAGE_PMD_SIZE);
150 static pmd_t pmd_set_protbits(pmd_t pmd, pgprot_t pgprot)
152 return __pmd(pmd_val(pmd) | pgprot_val(pgprot));
156 * At some point we should be able to get rid of
157 * pmd_mkhuge() and mk_huge_pmd() when we update all the
158 * other archs to mark the pmd huge in pfn_pmd()
160 pmd_t pfn_pmd(unsigned long pfn, pgprot_t pgprot)
164 pmdv = (pfn << PAGE_SHIFT) & PTE_RPN_MASK;
166 return __pmd_mkhuge(pmd_set_protbits(__pmd(pmdv), pgprot));
169 pmd_t mk_pmd(struct page *page, pgprot_t pgprot)
171 return pfn_pmd(page_to_pfn(page), pgprot);
174 pmd_t pmd_modify(pmd_t pmd, pgprot_t newprot)
179 pmdv &= _HPAGE_CHG_MASK;
180 return pmd_set_protbits(__pmd(pmdv), newprot);
182 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
184 /* For use by kexec, called with MMU off */
185 notrace void mmu_cleanup_all(void)
188 radix__mmu_cleanup_all();
189 else if (mmu_hash_ops.hpte_clear_all)
190 mmu_hash_ops.hpte_clear_all();
195 #ifdef CONFIG_MEMORY_HOTPLUG
196 int __meminit create_section_mapping(unsigned long start, unsigned long end,
197 int nid, pgprot_t prot)
200 return radix__create_section_mapping(start, end, nid, prot);
202 return hash__create_section_mapping(start, end, nid, prot);
205 int __meminit remove_section_mapping(unsigned long start, unsigned long end)
208 return radix__remove_section_mapping(start, end);
210 return hash__remove_section_mapping(start, end);
212 #endif /* CONFIG_MEMORY_HOTPLUG */
214 void __init mmu_partition_table_init(void)
216 unsigned long patb_size = 1UL << PATB_SIZE_SHIFT;
219 /* Initialize the Partition Table with no entries */
220 partition_tb = memblock_alloc(patb_size, patb_size);
222 panic("%s: Failed to allocate %lu bytes align=0x%lx\n",
223 __func__, patb_size, patb_size);
225 ptcr = __pa(partition_tb) | (PATB_SIZE_SHIFT - 12);
226 set_ptcr_when_no_uv(ptcr);
227 powernv_set_nmmu_ptcr(ptcr);
230 static void flush_partition(unsigned int lpid, bool radix)
233 radix__flush_all_lpid(lpid);
234 radix__flush_all_lpid_guest(lpid);
236 asm volatile("ptesync" : : : "memory");
237 asm volatile(PPC_TLBIE_5(%0,%1,2,0,0) : :
238 "r" (TLBIEL_INVAL_SET_LPID), "r" (lpid));
239 /* do we need fixup here ?*/
240 asm volatile("eieio; tlbsync; ptesync" : : : "memory");
241 trace_tlbie(lpid, 0, TLBIEL_INVAL_SET_LPID, lpid, 2, 0, 0);
245 void mmu_partition_table_set_entry(unsigned int lpid, unsigned long dw0,
246 unsigned long dw1, bool flush)
248 unsigned long old = be64_to_cpu(partition_tb[lpid].patb0);
251 * When ultravisor is enabled, the partition table is stored in secure
252 * memory and can only be accessed doing an ultravisor call. However, we
253 * maintain a copy of the partition table in normal memory to allow Nest
254 * MMU translations to occur (for normal VMs).
256 * Therefore, here we always update partition_tb, regardless of whether
257 * we are running under an ultravisor or not.
259 partition_tb[lpid].patb0 = cpu_to_be64(dw0);
260 partition_tb[lpid].patb1 = cpu_to_be64(dw1);
263 * If ultravisor is enabled, we do an ultravisor call to register the
264 * partition table entry (PATE), which also do a global flush of TLBs
265 * and partition table caches for the lpid. Otherwise, just do the
266 * flush. The type of flush (hash or radix) depends on what the previous
267 * use of the partition ID was, not the new use.
269 if (firmware_has_feature(FW_FEATURE_ULTRAVISOR)) {
270 uv_register_pate(lpid, dw0, dw1);
271 pr_info("PATE registered by ultravisor: dw0 = 0x%lx, dw1 = 0x%lx\n",
275 * Boot does not need to flush, because MMU is off and each
276 * CPU does a tlbiel_all() before switching them on, which
277 * flushes everything.
279 flush_partition(lpid, (old & PATB_HR));
282 EXPORT_SYMBOL_GPL(mmu_partition_table_set_entry);
284 static pmd_t *get_pmd_from_cache(struct mm_struct *mm)
286 void *pmd_frag, *ret;
288 if (PMD_FRAG_NR == 1)
291 spin_lock(&mm->page_table_lock);
292 ret = mm->context.pmd_frag;
294 pmd_frag = ret + PMD_FRAG_SIZE;
296 * If we have taken up all the fragments mark PTE page NULL
298 if (((unsigned long)pmd_frag & ~PAGE_MASK) == 0)
300 mm->context.pmd_frag = pmd_frag;
302 spin_unlock(&mm->page_table_lock);
306 static pmd_t *__alloc_for_pmdcache(struct mm_struct *mm)
310 gfp_t gfp = GFP_KERNEL_ACCOUNT | __GFP_ZERO;
313 gfp &= ~__GFP_ACCOUNT;
314 page = alloc_page(gfp);
317 if (!pgtable_pmd_page_ctor(page)) {
318 __free_pages(page, 0);
322 atomic_set(&page->pt_frag_refcount, 1);
324 ret = page_address(page);
326 * if we support only one fragment just return the
329 if (PMD_FRAG_NR == 1)
332 spin_lock(&mm->page_table_lock);
334 * If we find pgtable_page set, we return
335 * the allocated page with single fragment
338 if (likely(!mm->context.pmd_frag)) {
339 atomic_set(&page->pt_frag_refcount, PMD_FRAG_NR);
340 mm->context.pmd_frag = ret + PMD_FRAG_SIZE;
342 spin_unlock(&mm->page_table_lock);
347 pmd_t *pmd_fragment_alloc(struct mm_struct *mm, unsigned long vmaddr)
351 pmd = get_pmd_from_cache(mm);
355 return __alloc_for_pmdcache(mm);
358 void pmd_fragment_free(unsigned long *pmd)
360 struct page *page = virt_to_page(pmd);
362 if (PageReserved(page))
363 return free_reserved_page(page);
365 BUG_ON(atomic_read(&page->pt_frag_refcount) <= 0);
366 if (atomic_dec_and_test(&page->pt_frag_refcount)) {
367 pgtable_pmd_page_dtor(page);
372 static inline void pgtable_free(void *table, int index)
376 pte_fragment_free(table, 0);
379 pmd_fragment_free(table);
384 #if defined(CONFIG_PPC_4K_PAGES) && defined(CONFIG_HUGETLB_PAGE)
385 /* 16M hugepd directory at pud level */
387 BUILD_BUG_ON(H_16M_CACHE_INDEX <= 0);
388 kmem_cache_free(PGT_CACHE(H_16M_CACHE_INDEX), table);
390 /* 16G hugepd directory at the pgd level */
392 BUILD_BUG_ON(H_16G_CACHE_INDEX <= 0);
393 kmem_cache_free(PGT_CACHE(H_16G_CACHE_INDEX), table);
396 /* We don't free pgd table via RCU callback */
402 void pgtable_free_tlb(struct mmu_gather *tlb, void *table, int index)
404 unsigned long pgf = (unsigned long)table;
406 BUG_ON(index > MAX_PGTABLE_INDEX_SIZE);
408 tlb_remove_table(tlb, (void *)pgf);
411 void __tlb_remove_table(void *_table)
413 void *table = (void *)((unsigned long)_table & ~MAX_PGTABLE_INDEX_SIZE);
414 unsigned int index = (unsigned long)_table & MAX_PGTABLE_INDEX_SIZE;
416 return pgtable_free(table, index);
419 #ifdef CONFIG_PROC_FS
420 atomic_long_t direct_pages_count[MMU_PAGE_COUNT];
422 void arch_report_meminfo(struct seq_file *m)
425 * Hash maps the memory with one size mmu_linear_psize.
426 * So don't bother to print these on hash
428 if (!radix_enabled())
430 seq_printf(m, "DirectMap4k: %8lu kB\n",
431 atomic_long_read(&direct_pages_count[MMU_PAGE_4K]) << 2);
432 seq_printf(m, "DirectMap64k: %8lu kB\n",
433 atomic_long_read(&direct_pages_count[MMU_PAGE_64K]) << 6);
434 seq_printf(m, "DirectMap2M: %8lu kB\n",
435 atomic_long_read(&direct_pages_count[MMU_PAGE_2M]) << 11);
436 seq_printf(m, "DirectMap1G: %8lu kB\n",
437 atomic_long_read(&direct_pages_count[MMU_PAGE_1G]) << 20);
439 #endif /* CONFIG_PROC_FS */
441 pte_t ptep_modify_prot_start(struct vm_area_struct *vma, unsigned long addr,
444 unsigned long pte_val;
447 * Clear the _PAGE_PRESENT so that no hardware parallel update is
448 * possible. Also keep the pte_present true so that we don't take
451 pte_val = pte_update(vma->vm_mm, addr, ptep, _PAGE_PRESENT, _PAGE_INVALID, 0);
453 return __pte(pte_val);
457 void ptep_modify_prot_commit(struct vm_area_struct *vma, unsigned long addr,
458 pte_t *ptep, pte_t old_pte, pte_t pte)
461 return radix__ptep_modify_prot_commit(vma, addr,
463 set_pte_at(vma->vm_mm, addr, ptep, pte);
467 * For hash translation mode, we use the deposited table to store hash slot
468 * information and they are stored at PTRS_PER_PMD offset from related pmd
469 * location. Hence a pmd move requires deposit and withdraw.
471 * For radix translation with split pmd ptl, we store the deposited table in the
472 * pmd page. Hence if we have different pmd page we need to withdraw during pmd
475 * With hash we use deposited table always irrespective of anon or not.
476 * With radix we use deposited table only for anonymous mapping.
478 int pmd_move_must_withdraw(struct spinlock *new_pmd_ptl,
479 struct spinlock *old_pmd_ptl,
480 struct vm_area_struct *vma)
483 return (new_pmd_ptl != old_pmd_ptl) && vma_is_anonymous(vma);
489 * Does the CPU support tlbie?
491 bool tlbie_capable __read_mostly = true;
492 EXPORT_SYMBOL(tlbie_capable);
495 * Should tlbie be used for management of CPU TLBs, for kernel and process
496 * address spaces? tlbie may still be used for nMMU accelerators, and for KVM
497 * guest address spaces.
499 bool tlbie_enabled __read_mostly = true;
501 static int __init setup_disable_tlbie(char *str)
503 if (!radix_enabled()) {
504 pr_err("disable_tlbie: Unable to disable TLBIE with Hash MMU.\n");
508 tlbie_capable = false;
509 tlbie_enabled = false;
513 __setup("disable_tlbie", setup_disable_tlbie);
515 static int __init pgtable_debugfs_setup(void)
521 * There is no locking vs tlb flushing when changing this value.
522 * The tlb flushers will see one value or another, and use either
523 * tlbie or tlbiel with IPIs. In both cases the TLBs will be
524 * invalidated as expected.
526 debugfs_create_bool("tlbie_enabled", 0600,
532 arch_initcall(pgtable_debugfs_setup);
534 #if defined(CONFIG_ZONE_DEVICE) && defined(CONFIG_ARCH_HAS_MEMREMAP_COMPAT_ALIGN)
536 * Override the generic version in mm/memremap.c.
538 * With hash translation, the direct-map range is mapped with just one
539 * page size selected by htab_init_page_sizes(). Consult
540 * mmu_psize_defs[] to determine the minimum page size alignment.
542 unsigned long memremap_compat_align(void)
544 if (!radix_enabled()) {
545 unsigned int shift = mmu_psize_defs[mmu_linear_psize].shift;
546 return max(SUBSECTION_SIZE, 1UL << shift);
549 return SUBSECTION_SIZE;
551 EXPORT_SYMBOL_GPL(memremap_compat_align);
554 pgprot_t vm_get_page_prot(unsigned long vm_flags)
558 /* Radix supports execute-only, but protection_map maps X -> RX */
559 if (radix_enabled() && ((vm_flags & VM_ACCESS_FLAGS) == VM_EXEC)) {
560 prot = pgprot_val(PAGE_EXECONLY);
562 prot = pgprot_val(protection_map[vm_flags &
563 (VM_ACCESS_FLAGS | VM_SHARED)]);
566 if (vm_flags & VM_SAO)
569 #ifdef CONFIG_PPC_MEM_KEYS
570 prot |= vmflag_to_pte_pkey_bits(vm_flags);
573 return __pgprot(prot);
575 EXPORT_SYMBOL(vm_get_page_prot);