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>
15 #include <asm/pgalloc.h>
20 #define CREATE_TRACE_POINTS
21 #include <trace/events/thp.h>
23 #ifdef CONFIG_SPARSEMEM_VMEMMAP
25 * vmemmap is the starting address of the virtual address space where
26 * struct pages are allocated for all possible PFNs present on the system
27 * including holes and bad memory (hence sparse). These virtual struct
28 * pages are stored in sequence in this virtual address space irrespective
29 * of the fact whether the corresponding PFN is valid or not. This achieves
30 * constant relationship between address of struct page and its PFN.
32 * During boot or memory hotplug operation when a new memory section is
33 * added, physical memory allocation (including hash table bolting) will
34 * be performed for the set of struct pages which are part of the memory
35 * section. This saves memory by not allocating struct pages for PFNs
36 * which are not valid.
38 * ----------------------------------------------
39 * | PHYSICAL ALLOCATION OF VIRTUAL STRUCT PAGES|
40 * ----------------------------------------------
42 * f000000000000000 c000000000000000
43 * vmemmap +--------------+ +--------------+
44 * + | page struct | +--------------> | page struct |
45 * | +--------------+ +--------------+
46 * | | page struct | +--------------> | page struct |
47 * | +--------------+ | +--------------+
48 * | | page struct | + +------> | page struct |
49 * | +--------------+ | +--------------+
50 * | | page struct | | +--> | page struct |
51 * | +--------------+ | | +--------------+
52 * | | page struct | | |
53 * | +--------------+ | |
54 * | | page struct | | |
55 * | +--------------+ | |
56 * | | page struct | | |
57 * | +--------------+ | |
58 * | | page struct | | |
59 * | +--------------+ | |
60 * | | page struct | +-------+ |
61 * | +--------------+ |
62 * | | page struct | +-----------+
64 * | | page struct | No mapping
66 * | | page struct | No mapping
69 * -----------------------------------------
70 * | RELATION BETWEEN STRUCT PAGES AND PFNS|
71 * -----------------------------------------
73 * vmemmap +--------------+ +---------------+
74 * + | page struct | +-------------> | PFN |
75 * | +--------------+ +---------------+
76 * | | page struct | +-------------> | PFN |
77 * | +--------------+ +---------------+
78 * | | page struct | +-------------> | PFN |
79 * | +--------------+ +---------------+
80 * | | page struct | +-------------> | PFN |
81 * | +--------------+ +---------------+
87 * | +--------------+ +---------------+
88 * | | page struct | +-------------> | PFN |
89 * | +--------------+ +---------------+
93 * | +--------------+ +---------------+
94 * | | page struct | +-------------> | PFN |
95 * | +--------------+ +---------------+
96 * | | page struct | +-------------> | PFN |
97 * v +--------------+ +---------------+
100 * On hash-based CPUs, the vmemmap is bolted in the hash table.
103 int __meminit hash__vmemmap_create_mapping(unsigned long start,
104 unsigned long page_size,
107 int rc = htab_bolt_mapping(start, start + page_size, phys,
108 pgprot_val(PAGE_KERNEL),
109 mmu_vmemmap_psize, mmu_kernel_ssize);
111 int rc2 = htab_remove_mapping(start, start + page_size,
114 BUG_ON(rc2 && (rc2 != -ENOENT));
119 #ifdef CONFIG_MEMORY_HOTPLUG
120 void hash__vmemmap_remove_mapping(unsigned long start,
121 unsigned long page_size)
123 int rc = htab_remove_mapping(start, start + page_size,
126 BUG_ON((rc < 0) && (rc != -ENOENT));
127 WARN_ON(rc == -ENOENT);
130 #endif /* CONFIG_SPARSEMEM_VMEMMAP */
133 * map_kernel_page currently only called by __ioremap
134 * map_kernel_page adds an entry to the ioremap page table
135 * and adds an entry to the HPT, possibly bolting it
137 int hash__map_kernel_page(unsigned long ea, unsigned long pa, unsigned long flags)
144 BUILD_BUG_ON(TASK_SIZE_USER64 > H_PGTABLE_RANGE);
145 if (slab_is_available()) {
146 pgdp = pgd_offset_k(ea);
147 pudp = pud_alloc(&init_mm, pgdp, ea);
150 pmdp = pmd_alloc(&init_mm, pudp, ea);
153 ptep = pte_alloc_kernel(pmdp, ea);
156 set_pte_at(&init_mm, ea, ptep, pfn_pte(pa >> PAGE_SHIFT,
160 * If the mm subsystem is not fully up, we cannot create a
161 * linux page table entry for this mapping. Simply bolt an
162 * entry in the hardware page table.
165 if (htab_bolt_mapping(ea, ea + PAGE_SIZE, pa, flags,
166 mmu_io_psize, mmu_kernel_ssize)) {
167 printk(KERN_ERR "Failed to do bolted mapping IO "
168 "memory at %016lx !\n", pa);
177 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
179 unsigned long hash__pmd_hugepage_update(struct mm_struct *mm, unsigned long addr,
180 pmd_t *pmdp, unsigned long clr,
186 #ifdef CONFIG_DEBUG_VM
187 WARN_ON(!pmd_trans_huge(*pmdp));
188 assert_spin_locked(&mm->page_table_lock);
191 __asm__ __volatile__(
199 : "=&r" (old_be), "=&r" (tmp), "=m" (*pmdp)
200 : "r" (pmdp), "r" (cpu_to_be64(clr)), "m" (*pmdp),
201 "r" (cpu_to_be64(H_PAGE_BUSY)), "r" (cpu_to_be64(set))
204 old = be64_to_cpu(old_be);
206 trace_hugepage_update(addr, old, clr, set);
207 if (old & H_PAGE_HASHPTE)
208 hpte_do_hugepage_flush(mm, addr, pmdp, old);
212 pmd_t hash__pmdp_collapse_flush(struct vm_area_struct *vma, unsigned long address,
217 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
218 VM_BUG_ON(pmd_trans_huge(*pmdp));
223 * Wait for all pending hash_page to finish. This is needed
224 * in case of subpage collapse. When we collapse normal pages
225 * to hugepage, we first clear the pmd, then invalidate all
226 * the PTE entries. The assumption here is that any low level
227 * page fault will see a none pmd and take the slow path that
228 * will wait on mmap_sem. But we could very well be in a
229 * hash_page with local ptep pointer value. Such a hash page
230 * can result in adding new HPTE entries for normal subpages.
231 * That means we could be modifying the page content as we
232 * copy them to a huge page. So wait for parallel hash_page
233 * to finish before invalidating HPTE entries. We can do this
234 * by sending an IPI to all the cpus and executing a dummy
237 kick_all_cpus_sync();
239 * Now invalidate the hpte entries in the range
240 * covered by pmd. This make sure we take a
241 * fault and will find the pmd as none, which will
242 * result in a major fault which takes mmap_sem and
243 * hence wait for collapse to complete. Without this
244 * the __collapse_huge_page_copy can result in copying
247 flush_tlb_pmd_range(vma->vm_mm, &pmd, address);
252 * We want to put the pgtable in pmd and use pgtable for tracking
253 * the base page size hptes
255 void hash__pgtable_trans_huge_deposit(struct mm_struct *mm, pmd_t *pmdp,
258 pgtable_t *pgtable_slot;
259 assert_spin_locked(&mm->page_table_lock);
261 * we store the pgtable in the second half of PMD
263 pgtable_slot = (pgtable_t *)pmdp + PTRS_PER_PMD;
264 *pgtable_slot = pgtable;
266 * expose the deposited pgtable to other cpus.
267 * before we set the hugepage PTE at pmd level
268 * hash fault code looks at the deposted pgtable
269 * to store hash index values.
274 pgtable_t hash__pgtable_trans_huge_withdraw(struct mm_struct *mm, pmd_t *pmdp)
277 pgtable_t *pgtable_slot;
279 assert_spin_locked(&mm->page_table_lock);
280 pgtable_slot = (pgtable_t *)pmdp + PTRS_PER_PMD;
281 pgtable = *pgtable_slot;
283 * Once we withdraw, mark the entry NULL.
285 *pgtable_slot = NULL;
287 * We store HPTE information in the deposited PTE fragment.
288 * zero out the content on withdraw.
290 memset(pgtable, 0, PTE_FRAG_SIZE);
294 void hash__pmdp_huge_split_prepare(struct vm_area_struct *vma,
295 unsigned long address, pmd_t *pmdp)
297 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
298 VM_BUG_ON(REGION_ID(address) != USER_REGION_ID);
301 * We can't mark the pmd none here, because that will cause a race
302 * against exit_mmap. We need to continue mark pmd TRANS HUGE, while
303 * we spilt, but at the same time we wan't rest of the ppc64 code
304 * not to insert hash pte on this, because we will be modifying
305 * the deposited pgtable in the caller of this function. Hence
306 * clear the _PAGE_USER so that we move the fault handling to
307 * higher level function and that will serialize against ptl.
308 * We need to flush existing hash pte entries here even though,
309 * the translation is still valid, because we will withdraw
310 * pgtable_t after this.
312 pmd_hugepage_update(vma->vm_mm, address, pmdp, 0, _PAGE_PRIVILEGED);
316 * A linux hugepage PMD was changed and the corresponding hash table entries
317 * neesd to be flushed.
319 void hpte_do_hugepage_flush(struct mm_struct *mm, unsigned long addr,
320 pmd_t *pmdp, unsigned long old_pmd)
325 unsigned long flags = 0;
326 const struct cpumask *tmp;
328 /* get the base page size,vsid and segment size */
329 #ifdef CONFIG_DEBUG_VM
330 psize = get_slice_psize(mm, addr);
331 BUG_ON(psize == MMU_PAGE_16M);
333 if (old_pmd & H_PAGE_COMBO)
336 psize = MMU_PAGE_64K;
338 if (!is_kernel_addr(addr)) {
339 ssize = user_segment_size(addr);
340 vsid = get_vsid(mm->context.id, addr, ssize);
343 vsid = get_kernel_vsid(addr, mmu_kernel_ssize);
344 ssize = mmu_kernel_ssize;
347 tmp = cpumask_of(smp_processor_id());
348 if (cpumask_equal(mm_cpumask(mm), tmp))
349 flags |= HPTE_LOCAL_UPDATE;
351 return flush_hash_hugepage(vsid, addr, pmdp, psize, ssize, flags);
354 pmd_t hash__pmdp_huge_get_and_clear(struct mm_struct *mm,
355 unsigned long addr, pmd_t *pmdp)
360 pgtable_t *pgtable_slot;
362 old = pmd_hugepage_update(mm, addr, pmdp, ~0UL, 0);
363 old_pmd = __pmd(old);
365 * We have pmd == none and we are holding page_table_lock.
366 * So we can safely go and clear the pgtable hash
369 pgtable_slot = (pgtable_t *)pmdp + PTRS_PER_PMD;
370 pgtable = *pgtable_slot;
372 * Let's zero out old valid and hash index details
373 * hash fault look at them.
375 memset(pgtable, 0, PTE_FRAG_SIZE);
377 * Serialize against find_linux_pte_or_hugepte which does lock-less
378 * lookup in page tables with local interrupts disabled. For huge pages
379 * it casts pmd_t to pte_t. Since format of pte_t is different from
380 * pmd_t we want to prevent transit from pmd pointing to page table
381 * to pmd pointing to huge page (and back) while interrupts are disabled.
382 * We clear pmd to possibly replace it with page table pointer in
383 * different code paths. So make sure we wait for the parallel
384 * find_linux_pte_or_hugepage to finish.
386 kick_all_cpus_sync();
390 int hash__has_transparent_hugepage(void)
393 if (!mmu_has_feature(MMU_FTR_16M_PAGE))
396 * We support THP only if PMD_SIZE is 16MB.
398 if (mmu_psize_defs[MMU_PAGE_16M].shift != PMD_SHIFT)
401 * We need to make sure that we support 16MB hugepage in a segement
402 * with base page size 64K or 4K. We only enable THP with a PAGE_SIZE
406 * If we have 64K HPTE, we will be using that by default
408 if (mmu_psize_defs[MMU_PAGE_64K].shift &&
409 (mmu_psize_defs[MMU_PAGE_64K].penc[MMU_PAGE_16M] == -1))
412 * Ok we only have 4K HPTE
414 if (mmu_psize_defs[MMU_PAGE_4K].penc[MMU_PAGE_16M] == -1)
419 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */