1 // SPDX-License-Identifier: GPL-2.0-or-later
3 * Copyright 2005, Paul Mackerras, IBM Corporation.
4 * Copyright 2009, Benjamin Herrenschmidt, IBM Corporation.
5 * Copyright 2015-2016, Aneesh Kumar K.V, IBM Corporation.
8 #include <linux/sched.h>
9 #include <linux/mm_types.h>
12 #include <asm/pgalloc.h>
13 #include <asm/sections.h>
17 #include <mm/mmu_decl.h>
19 #define CREATE_TRACE_POINTS
20 #include <trace/events/thp.h>
22 #if H_PGTABLE_RANGE > (USER_VSID_RANGE * (TASK_SIZE_USER64 / TASK_CONTEXT_SIZE))
23 #warning Limited user VSID range means pagetable space is wasted
26 #ifdef CONFIG_SPARSEMEM_VMEMMAP
28 * vmemmap is the starting address of the virtual address space where
29 * struct pages are allocated for all possible PFNs present on the system
30 * including holes and bad memory (hence sparse). These virtual struct
31 * pages are stored in sequence in this virtual address space irrespective
32 * of the fact whether the corresponding PFN is valid or not. This achieves
33 * constant relationship between address of struct page and its PFN.
35 * During boot or memory hotplug operation when a new memory section is
36 * added, physical memory allocation (including hash table bolting) will
37 * be performed for the set of struct pages which are part of the memory
38 * section. This saves memory by not allocating struct pages for PFNs
39 * which are not valid.
41 * ----------------------------------------------
42 * | PHYSICAL ALLOCATION OF VIRTUAL STRUCT PAGES|
43 * ----------------------------------------------
45 * f000000000000000 c000000000000000
46 * vmemmap +--------------+ +--------------+
47 * + | page struct | +--------------> | page struct |
48 * | +--------------+ +--------------+
49 * | | page struct | +--------------> | page struct |
50 * | +--------------+ | +--------------+
51 * | | page struct | + +------> | page struct |
52 * | +--------------+ | +--------------+
53 * | | page struct | | +--> | page struct |
54 * | +--------------+ | | +--------------+
55 * | | page struct | | |
56 * | +--------------+ | |
57 * | | page struct | | |
58 * | +--------------+ | |
59 * | | page struct | | |
60 * | +--------------+ | |
61 * | | page struct | | |
62 * | +--------------+ | |
63 * | | page struct | +-------+ |
64 * | +--------------+ |
65 * | | page struct | +-----------+
67 * | | page struct | No mapping
69 * | | page struct | No mapping
72 * -----------------------------------------
73 * | RELATION BETWEEN STRUCT PAGES AND PFNS|
74 * -----------------------------------------
76 * vmemmap +--------------+ +---------------+
77 * + | page struct | +-------------> | PFN |
78 * | +--------------+ +---------------+
79 * | | page struct | +-------------> | PFN |
80 * | +--------------+ +---------------+
81 * | | page struct | +-------------> | PFN |
82 * | +--------------+ +---------------+
83 * | | page struct | +-------------> | PFN |
84 * | +--------------+ +---------------+
90 * | +--------------+ +---------------+
91 * | | page struct | +-------------> | PFN |
92 * | +--------------+ +---------------+
96 * | +--------------+ +---------------+
97 * | | page struct | +-------------> | PFN |
98 * | +--------------+ +---------------+
99 * | | page struct | +-------------> | PFN |
100 * v +--------------+ +---------------+
103 * On hash-based CPUs, the vmemmap is bolted in the hash table.
106 int __meminit
hash__vmemmap_create_mapping(unsigned long start
,
107 unsigned long page_size
,
112 if ((start
+ page_size
) >= H_VMEMMAP_END
) {
113 pr_warn("Outside the supported range\n");
117 rc
= htab_bolt_mapping(start
, start
+ page_size
, phys
,
118 pgprot_val(PAGE_KERNEL
),
119 mmu_vmemmap_psize
, mmu_kernel_ssize
);
121 int rc2
= htab_remove_mapping(start
, start
+ page_size
,
124 BUG_ON(rc2
&& (rc2
!= -ENOENT
));
129 #ifdef CONFIG_MEMORY_HOTPLUG
130 void hash__vmemmap_remove_mapping(unsigned long start
,
131 unsigned long page_size
)
133 int rc
= htab_remove_mapping(start
, start
+ page_size
,
136 BUG_ON((rc
< 0) && (rc
!= -ENOENT
));
137 WARN_ON(rc
== -ENOENT
);
140 #endif /* CONFIG_SPARSEMEM_VMEMMAP */
143 * map_kernel_page currently only called by __ioremap
144 * map_kernel_page adds an entry to the ioremap page table
145 * and adds an entry to the HPT, possibly bolting it
147 int hash__map_kernel_page(unsigned long ea
, unsigned long pa
, pgprot_t prot
)
155 BUILD_BUG_ON(TASK_SIZE_USER64
> H_PGTABLE_RANGE
);
156 if (slab_is_available()) {
157 pgdp
= pgd_offset_k(ea
);
158 p4dp
= p4d_offset(pgdp
, ea
);
159 pudp
= pud_alloc(&init_mm
, p4dp
, ea
);
162 pmdp
= pmd_alloc(&init_mm
, pudp
, ea
);
165 ptep
= pte_alloc_kernel(pmdp
, ea
);
168 set_pte_at(&init_mm
, ea
, ptep
, pfn_pte(pa
>> PAGE_SHIFT
, prot
));
171 * If the mm subsystem is not fully up, we cannot create a
172 * linux page table entry for this mapping. Simply bolt an
173 * entry in the hardware page table.
176 if (htab_bolt_mapping(ea
, ea
+ PAGE_SIZE
, pa
, pgprot_val(prot
),
177 mmu_io_psize
, mmu_kernel_ssize
)) {
178 printk(KERN_ERR
"Failed to do bolted mapping IO "
179 "memory at %016lx !\n", pa
);
188 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
190 unsigned long hash__pmd_hugepage_update(struct mm_struct
*mm
, unsigned long addr
,
191 pmd_t
*pmdp
, unsigned long clr
,
197 #ifdef CONFIG_DEBUG_VM
198 WARN_ON(!hash__pmd_trans_huge(*pmdp
) && !pmd_devmap(*pmdp
));
199 assert_spin_locked(pmd_lockptr(mm
, pmdp
));
202 __asm__
__volatile__(
210 : "=&r" (old_be
), "=&r" (tmp
), "=m" (*pmdp
)
211 : "r" (pmdp
), "r" (cpu_to_be64(clr
)), "m" (*pmdp
),
212 "r" (cpu_to_be64(H_PAGE_BUSY
)), "r" (cpu_to_be64(set
))
215 old
= be64_to_cpu(old_be
);
217 trace_hugepage_update(addr
, old
, clr
, set
);
218 if (old
& H_PAGE_HASHPTE
)
219 hpte_do_hugepage_flush(mm
, addr
, pmdp
, old
);
223 pmd_t
hash__pmdp_collapse_flush(struct vm_area_struct
*vma
, unsigned long address
,
228 VM_BUG_ON(address
& ~HPAGE_PMD_MASK
);
229 VM_BUG_ON(pmd_trans_huge(*pmdp
));
230 VM_BUG_ON(pmd_devmap(*pmdp
));
235 * Wait for all pending hash_page to finish. This is needed
236 * in case of subpage collapse. When we collapse normal pages
237 * to hugepage, we first clear the pmd, then invalidate all
238 * the PTE entries. The assumption here is that any low level
239 * page fault will see a none pmd and take the slow path that
240 * will wait on mmap_sem. But we could very well be in a
241 * hash_page with local ptep pointer value. Such a hash page
242 * can result in adding new HPTE entries for normal subpages.
243 * That means we could be modifying the page content as we
244 * copy them to a huge page. So wait for parallel hash_page
245 * to finish before invalidating HPTE entries. We can do this
246 * by sending an IPI to all the cpus and executing a dummy
249 serialize_against_pte_lookup(vma
->vm_mm
);
251 * Now invalidate the hpte entries in the range
252 * covered by pmd. This make sure we take a
253 * fault and will find the pmd as none, which will
254 * result in a major fault which takes mmap_sem and
255 * hence wait for collapse to complete. Without this
256 * the __collapse_huge_page_copy can result in copying
259 flush_tlb_pmd_range(vma
->vm_mm
, &pmd
, address
);
264 * We want to put the pgtable in pmd and use pgtable for tracking
265 * the base page size hptes
267 void hash__pgtable_trans_huge_deposit(struct mm_struct
*mm
, pmd_t
*pmdp
,
270 pgtable_t
*pgtable_slot
;
272 assert_spin_locked(pmd_lockptr(mm
, pmdp
));
274 * we store the pgtable in the second half of PMD
276 pgtable_slot
= (pgtable_t
*)pmdp
+ PTRS_PER_PMD
;
277 *pgtable_slot
= pgtable
;
279 * expose the deposited pgtable to other cpus.
280 * before we set the hugepage PTE at pmd level
281 * hash fault code looks at the deposted pgtable
282 * to store hash index values.
287 pgtable_t
hash__pgtable_trans_huge_withdraw(struct mm_struct
*mm
, pmd_t
*pmdp
)
290 pgtable_t
*pgtable_slot
;
292 assert_spin_locked(pmd_lockptr(mm
, pmdp
));
294 pgtable_slot
= (pgtable_t
*)pmdp
+ PTRS_PER_PMD
;
295 pgtable
= *pgtable_slot
;
297 * Once we withdraw, mark the entry NULL.
299 *pgtable_slot
= NULL
;
301 * We store HPTE information in the deposited PTE fragment.
302 * zero out the content on withdraw.
304 memset(pgtable
, 0, PTE_FRAG_SIZE
);
309 * A linux hugepage PMD was changed and the corresponding hash table entries
310 * neesd to be flushed.
312 void hpte_do_hugepage_flush(struct mm_struct
*mm
, unsigned long addr
,
313 pmd_t
*pmdp
, unsigned long old_pmd
)
318 unsigned long flags
= 0;
320 /* get the base page size,vsid and segment size */
321 #ifdef CONFIG_DEBUG_VM
322 psize
= get_slice_psize(mm
, addr
);
323 BUG_ON(psize
== MMU_PAGE_16M
);
325 if (old_pmd
& H_PAGE_COMBO
)
328 psize
= MMU_PAGE_64K
;
330 if (!is_kernel_addr(addr
)) {
331 ssize
= user_segment_size(addr
);
332 vsid
= get_user_vsid(&mm
->context
, addr
, ssize
);
335 vsid
= get_kernel_vsid(addr
, mmu_kernel_ssize
);
336 ssize
= mmu_kernel_ssize
;
339 if (mm_is_thread_local(mm
))
340 flags
|= HPTE_LOCAL_UPDATE
;
342 return flush_hash_hugepage(vsid
, addr
, pmdp
, psize
, ssize
, flags
);
345 pmd_t
hash__pmdp_huge_get_and_clear(struct mm_struct
*mm
,
346 unsigned long addr
, pmd_t
*pmdp
)
351 pgtable_t
*pgtable_slot
;
353 old
= pmd_hugepage_update(mm
, addr
, pmdp
, ~0UL, 0);
354 old_pmd
= __pmd(old
);
356 * We have pmd == none and we are holding page_table_lock.
357 * So we can safely go and clear the pgtable hash
360 pgtable_slot
= (pgtable_t
*)pmdp
+ PTRS_PER_PMD
;
361 pgtable
= *pgtable_slot
;
363 * Let's zero out old valid and hash index details
364 * hash fault look at them.
366 memset(pgtable
, 0, PTE_FRAG_SIZE
);
370 int hash__has_transparent_hugepage(void)
373 if (!mmu_has_feature(MMU_FTR_16M_PAGE
))
376 * We support THP only if PMD_SIZE is 16MB.
378 if (mmu_psize_defs
[MMU_PAGE_16M
].shift
!= PMD_SHIFT
)
381 * We need to make sure that we support 16MB hugepage in a segement
382 * with base page size 64K or 4K. We only enable THP with a PAGE_SIZE
386 * If we have 64K HPTE, we will be using that by default
388 if (mmu_psize_defs
[MMU_PAGE_64K
].shift
&&
389 (mmu_psize_defs
[MMU_PAGE_64K
].penc
[MMU_PAGE_16M
] == -1))
392 * Ok we only have 4K HPTE
394 if (mmu_psize_defs
[MMU_PAGE_4K
].penc
[MMU_PAGE_16M
] == -1)
399 EXPORT_SYMBOL_GPL(hash__has_transparent_hugepage
);
401 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
403 #ifdef CONFIG_STRICT_KERNEL_RWX
404 static bool hash__change_memory_range(unsigned long start
, unsigned long end
,
408 unsigned int step
, shift
;
410 shift
= mmu_psize_defs
[mmu_linear_psize
].shift
;
413 start
= ALIGN_DOWN(start
, step
);
414 end
= ALIGN(end
, step
); // aligns up
419 pr_debug("Changing page protection on range 0x%lx-0x%lx, to 0x%lx, step 0x%x\n",
420 start
, end
, newpp
, step
);
422 for (idx
= start
; idx
< end
; idx
+= step
)
423 /* Not sure if we can do much with the return value */
424 mmu_hash_ops
.hpte_updateboltedpp(newpp
, idx
, mmu_linear_psize
,
430 void hash__mark_rodata_ro(void)
432 unsigned long start
, end
;
434 start
= (unsigned long)_stext
;
435 end
= (unsigned long)__init_begin
;
437 WARN_ON(!hash__change_memory_range(start
, end
, PP_RXXX
));
440 void hash__mark_initmem_nx(void)
442 unsigned long start
, end
, pp
;
444 start
= (unsigned long)__init_begin
;
445 end
= (unsigned long)__init_end
;
447 pp
= htab_convert_pte_flags(pgprot_val(PAGE_KERNEL
));
449 WARN_ON(!hash__change_memory_range(start
, end
, pp
));