2 * Copyright 2005, Paul Mackerras, IBM Corporation.
3 * Copyright 2009, Benjamin Herrenschmidt, IBM Corporation.
4 * Copyright 2015-2016, Aneesh Kumar K.V, IBM Corporation.
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.
12 #include <linux/sched.h>
13 #include <linux/mm_types.h>
16 #include <asm/pgalloc.h>
17 #include <asm/pgtable.h>
18 #include <asm/sections.h>
24 #define CREATE_TRACE_POINTS
25 #include <trace/events/thp.h>
27 #if H_PGTABLE_RANGE > (USER_VSID_RANGE * (TASK_SIZE_USER64 / TASK_CONTEXT_SIZE))
28 #warning Limited user VSID range means pagetable space is wasted
31 #ifdef CONFIG_SPARSEMEM_VMEMMAP
33 * vmemmap is the starting address of the virtual address space where
34 * struct pages are allocated for all possible PFNs present on the system
35 * including holes and bad memory (hence sparse). These virtual struct
36 * pages are stored in sequence in this virtual address space irrespective
37 * of the fact whether the corresponding PFN is valid or not. This achieves
38 * constant relationship between address of struct page and its PFN.
40 * During boot or memory hotplug operation when a new memory section is
41 * added, physical memory allocation (including hash table bolting) will
42 * be performed for the set of struct pages which are part of the memory
43 * section. This saves memory by not allocating struct pages for PFNs
44 * which are not valid.
46 * ----------------------------------------------
47 * | PHYSICAL ALLOCATION OF VIRTUAL STRUCT PAGES|
48 * ----------------------------------------------
50 * f000000000000000 c000000000000000
51 * vmemmap +--------------+ +--------------+
52 * + | page struct | +--------------> | page struct |
53 * | +--------------+ +--------------+
54 * | | page struct | +--------------> | page struct |
55 * | +--------------+ | +--------------+
56 * | | page struct | + +------> | page struct |
57 * | +--------------+ | +--------------+
58 * | | page struct | | +--> | page struct |
59 * | +--------------+ | | +--------------+
60 * | | page struct | | |
61 * | +--------------+ | |
62 * | | page struct | | |
63 * | +--------------+ | |
64 * | | page struct | | |
65 * | +--------------+ | |
66 * | | page struct | | |
67 * | +--------------+ | |
68 * | | page struct | +-------+ |
69 * | +--------------+ |
70 * | | page struct | +-----------+
72 * | | page struct | No mapping
74 * | | page struct | No mapping
77 * -----------------------------------------
78 * | RELATION BETWEEN STRUCT PAGES AND PFNS|
79 * -----------------------------------------
81 * vmemmap +--------------+ +---------------+
82 * + | page struct | +-------------> | PFN |
83 * | +--------------+ +---------------+
84 * | | page struct | +-------------> | PFN |
85 * | +--------------+ +---------------+
86 * | | page struct | +-------------> | PFN |
87 * | +--------------+ +---------------+
88 * | | page struct | +-------------> | PFN |
89 * | +--------------+ +---------------+
95 * | +--------------+ +---------------+
96 * | | page struct | +-------------> | PFN |
97 * | +--------------+ +---------------+
101 * | +--------------+ +---------------+
102 * | | page struct | +-------------> | PFN |
103 * | +--------------+ +---------------+
104 * | | page struct | +-------------> | PFN |
105 * v +--------------+ +---------------+
108 * On hash-based CPUs, the vmemmap is bolted in the hash table.
111 int __meminit hash__vmemmap_create_mapping(unsigned long start,
112 unsigned long page_size,
115 int rc = htab_bolt_mapping(start, start + page_size, phys,
116 pgprot_val(PAGE_KERNEL),
117 mmu_vmemmap_psize, mmu_kernel_ssize);
119 int rc2 = htab_remove_mapping(start, start + page_size,
122 BUG_ON(rc2 && (rc2 != -ENOENT));
127 #ifdef CONFIG_MEMORY_HOTPLUG
128 void hash__vmemmap_remove_mapping(unsigned long start,
129 unsigned long page_size)
131 int rc = htab_remove_mapping(start, start + page_size,
134 BUG_ON((rc < 0) && (rc != -ENOENT));
135 WARN_ON(rc == -ENOENT);
138 #endif /* CONFIG_SPARSEMEM_VMEMMAP */
141 * map_kernel_page currently only called by __ioremap
142 * map_kernel_page adds an entry to the ioremap page table
143 * and adds an entry to the HPT, possibly bolting it
145 int hash__map_kernel_page(unsigned long ea, unsigned long pa, pgprot_t prot)
152 BUILD_BUG_ON(TASK_SIZE_USER64 > H_PGTABLE_RANGE);
153 if (slab_is_available()) {
154 pgdp = pgd_offset_k(ea);
155 pudp = pud_alloc(&init_mm, pgdp, ea);
158 pmdp = pmd_alloc(&init_mm, pudp, ea);
161 ptep = pte_alloc_kernel(pmdp, ea);
164 set_pte_at(&init_mm, ea, ptep, pfn_pte(pa >> PAGE_SHIFT, prot));
167 * If the mm subsystem is not fully up, we cannot create a
168 * linux page table entry for this mapping. Simply bolt an
169 * entry in the hardware page table.
172 if (htab_bolt_mapping(ea, ea + PAGE_SIZE, pa, pgprot_val(prot),
173 mmu_io_psize, mmu_kernel_ssize)) {
174 printk(KERN_ERR "Failed to do bolted mapping IO "
175 "memory at %016lx !\n", pa);
184 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
186 unsigned long hash__pmd_hugepage_update(struct mm_struct *mm, unsigned long addr,
187 pmd_t *pmdp, unsigned long clr,
193 #ifdef CONFIG_DEBUG_VM
194 WARN_ON(!hash__pmd_trans_huge(*pmdp) && !pmd_devmap(*pmdp));
195 assert_spin_locked(pmd_lockptr(mm, pmdp));
198 __asm__ __volatile__(
206 : "=&r" (old_be), "=&r" (tmp), "=m" (*pmdp)
207 : "r" (pmdp), "r" (cpu_to_be64(clr)), "m" (*pmdp),
208 "r" (cpu_to_be64(H_PAGE_BUSY)), "r" (cpu_to_be64(set))
211 old = be64_to_cpu(old_be);
213 trace_hugepage_update(addr, old, clr, set);
214 if (old & H_PAGE_HASHPTE)
215 hpte_do_hugepage_flush(mm, addr, pmdp, old);
219 pmd_t hash__pmdp_collapse_flush(struct vm_area_struct *vma, unsigned long address,
224 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
225 VM_BUG_ON(pmd_trans_huge(*pmdp));
226 VM_BUG_ON(pmd_devmap(*pmdp));
231 * Wait for all pending hash_page to finish. This is needed
232 * in case of subpage collapse. When we collapse normal pages
233 * to hugepage, we first clear the pmd, then invalidate all
234 * the PTE entries. The assumption here is that any low level
235 * page fault will see a none pmd and take the slow path that
236 * will wait on mmap_sem. But we could very well be in a
237 * hash_page with local ptep pointer value. Such a hash page
238 * can result in adding new HPTE entries for normal subpages.
239 * That means we could be modifying the page content as we
240 * copy them to a huge page. So wait for parallel hash_page
241 * to finish before invalidating HPTE entries. We can do this
242 * by sending an IPI to all the cpus and executing a dummy
245 serialize_against_pte_lookup(vma->vm_mm);
247 * Now invalidate the hpte entries in the range
248 * covered by pmd. This make sure we take a
249 * fault and will find the pmd as none, which will
250 * result in a major fault which takes mmap_sem and
251 * hence wait for collapse to complete. Without this
252 * the __collapse_huge_page_copy can result in copying
255 flush_tlb_pmd_range(vma->vm_mm, &pmd, address);
260 * We want to put the pgtable in pmd and use pgtable for tracking
261 * the base page size hptes
263 void hash__pgtable_trans_huge_deposit(struct mm_struct *mm, pmd_t *pmdp,
266 pgtable_t *pgtable_slot;
268 assert_spin_locked(pmd_lockptr(mm, pmdp));
270 * we store the pgtable in the second half of PMD
272 pgtable_slot = (pgtable_t *)pmdp + PTRS_PER_PMD;
273 *pgtable_slot = pgtable;
275 * expose the deposited pgtable to other cpus.
276 * before we set the hugepage PTE at pmd level
277 * hash fault code looks at the deposted pgtable
278 * to store hash index values.
283 pgtable_t hash__pgtable_trans_huge_withdraw(struct mm_struct *mm, pmd_t *pmdp)
286 pgtable_t *pgtable_slot;
288 assert_spin_locked(pmd_lockptr(mm, pmdp));
290 pgtable_slot = (pgtable_t *)pmdp + PTRS_PER_PMD;
291 pgtable = *pgtable_slot;
293 * Once we withdraw, mark the entry NULL.
295 *pgtable_slot = NULL;
297 * We store HPTE information in the deposited PTE fragment.
298 * zero out the content on withdraw.
300 memset(pgtable, 0, PTE_FRAG_SIZE);
305 * A linux hugepage PMD was changed and the corresponding hash table entries
306 * neesd to be flushed.
308 void hpte_do_hugepage_flush(struct mm_struct *mm, unsigned long addr,
309 pmd_t *pmdp, unsigned long old_pmd)
314 unsigned long flags = 0;
316 /* get the base page size,vsid and segment size */
317 #ifdef CONFIG_DEBUG_VM
318 psize = get_slice_psize(mm, addr);
319 BUG_ON(psize == MMU_PAGE_16M);
321 if (old_pmd & H_PAGE_COMBO)
324 psize = MMU_PAGE_64K;
326 if (!is_kernel_addr(addr)) {
327 ssize = user_segment_size(addr);
328 vsid = get_user_vsid(&mm->context, addr, ssize);
331 vsid = get_kernel_vsid(addr, mmu_kernel_ssize);
332 ssize = mmu_kernel_ssize;
335 if (mm_is_thread_local(mm))
336 flags |= HPTE_LOCAL_UPDATE;
338 return flush_hash_hugepage(vsid, addr, pmdp, psize, ssize, flags);
341 pmd_t hash__pmdp_huge_get_and_clear(struct mm_struct *mm,
342 unsigned long addr, pmd_t *pmdp)
347 pgtable_t *pgtable_slot;
349 old = pmd_hugepage_update(mm, addr, pmdp, ~0UL, 0);
350 old_pmd = __pmd(old);
352 * We have pmd == none and we are holding page_table_lock.
353 * So we can safely go and clear the pgtable hash
356 pgtable_slot = (pgtable_t *)pmdp + PTRS_PER_PMD;
357 pgtable = *pgtable_slot;
359 * Let's zero out old valid and hash index details
360 * hash fault look at them.
362 memset(pgtable, 0, PTE_FRAG_SIZE);
364 * Serialize against find_current_mm_pte variants which does lock-less
365 * lookup in page tables with local interrupts disabled. For huge pages
366 * it casts pmd_t to pte_t. Since format of pte_t is different from
367 * pmd_t we want to prevent transit from pmd pointing to page table
368 * to pmd pointing to huge page (and back) while interrupts are disabled.
369 * We clear pmd to possibly replace it with page table pointer in
370 * different code paths. So make sure we wait for the parallel
371 * find_curren_mm_pte to finish.
373 serialize_against_pte_lookup(mm);
377 int hash__has_transparent_hugepage(void)
380 if (!mmu_has_feature(MMU_FTR_16M_PAGE))
383 * We support THP only if PMD_SIZE is 16MB.
385 if (mmu_psize_defs[MMU_PAGE_16M].shift != PMD_SHIFT)
388 * We need to make sure that we support 16MB hugepage in a segement
389 * with base page size 64K or 4K. We only enable THP with a PAGE_SIZE
393 * If we have 64K HPTE, we will be using that by default
395 if (mmu_psize_defs[MMU_PAGE_64K].shift &&
396 (mmu_psize_defs[MMU_PAGE_64K].penc[MMU_PAGE_16M] == -1))
399 * Ok we only have 4K HPTE
401 if (mmu_psize_defs[MMU_PAGE_4K].penc[MMU_PAGE_16M] == -1)
406 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
408 #ifdef CONFIG_STRICT_KERNEL_RWX
409 static bool hash__change_memory_range(unsigned long start, unsigned long end,
413 unsigned int step, shift;
415 shift = mmu_psize_defs[mmu_linear_psize].shift;
418 start = ALIGN_DOWN(start, step);
419 end = ALIGN(end, step); // aligns up
424 pr_debug("Changing page protection on range 0x%lx-0x%lx, to 0x%lx, step 0x%x\n",
425 start, end, newpp, step);
427 for (idx = start; idx < end; idx += step)
428 /* Not sure if we can do much with the return value */
429 mmu_hash_ops.hpte_updateboltedpp(newpp, idx, mmu_linear_psize,
435 void hash__mark_rodata_ro(void)
437 unsigned long start, end;
439 start = (unsigned long)_stext;
440 end = (unsigned long)__init_begin;
442 WARN_ON(!hash__change_memory_range(start, end, PP_RXXX));
445 void hash__mark_initmem_nx(void)
447 unsigned long start, end, pp;
449 start = (unsigned long)__init_begin;
450 end = (unsigned long)__init_end;
452 pp = htab_convert_pte_flags(pgprot_val(PAGE_KERNEL));
454 WARN_ON(!hash__change_memory_range(start, end, pp));