1 /*
2 * Copyright 2005, Paul Mackerras, IBM Corporation.
3 * Copyright 2009, Benjamin Herrenschmidt, IBM Corporation.
4 * Copyright 2015-2016, Aneesh Kumar K.V, IBM Corporation.
5 *
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.
10 */
11
12 #include <linux/sched.h>
13 #include <linux/mm_types.h>
14 #include <linux/mm.h>
15
16 #include <asm/pgalloc.h>
17 #include <asm/pgtable.h>
18 #include <asm/sections.h>
19 #include <asm/mmu.h>
20 #include <asm/tlb.h>
21
22 #include "mmu_decl.h"
23
24 #define CREATE_TRACE_POINTS
25 #include <trace/events/thp.h>
26
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
29 #endif
30
31 #ifdef CONFIG_SPARSEMEM_VMEMMAP
32 /*
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.
39 *
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.
45 *
46 * ----------------------------------------------
47 * | PHYSICAL ALLOCATION OF VIRTUAL STRUCT PAGES|
48 * ----------------------------------------------
49 *
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 | +-----------+
71 * | +--------------+
72 * | | page struct | No mapping
73 * | +--------------+
74 * | | page struct | No mapping
75 * v +--------------+
76 *
77 * -----------------------------------------
78 * | RELATION BETWEEN STRUCT PAGES AND PFNS|
79 * -----------------------------------------
80 *
81 * vmemmap +--------------+ +---------------+
82 * + | page struct | +-------------> | PFN |
83 * | +--------------+ +---------------+
84 * | | page struct | +-------------> | PFN |
85 * | +--------------+ +---------------+
86 * | | page struct | +-------------> | PFN |
87 * | +--------------+ +---------------+
88 * | | page struct | +-------------> | PFN |
89 * | +--------------+ +---------------+
90 * | | |
91 * | +--------------+
92 * | | |
93 * | +--------------+
94 * | | |
95 * | +--------------+ +---------------+
96 * | | page struct | +-------------> | PFN |
97 * | +--------------+ +---------------+
98 * | | |
99 * | +--------------+
100 * | | |
101 * | +--------------+ +---------------+
102 * | | page struct | +-------------> | PFN |
103 * | +--------------+ +---------------+
104 * | | page struct | +-------------> | PFN |
105 * v +--------------+ +---------------+
106 */
107 /*
108 * On hash-based CPUs, the vmemmap is bolted in the hash table.
109 *
110 */
hash__vmemmap_create_mapping(unsigned long start,unsigned long page_size,unsigned long phys)111 int __meminit hash__vmemmap_create_mapping(unsigned long start,
112 unsigned long page_size,
113 unsigned long phys)
114 {
115 int rc = htab_bolt_mapping(start, start + page_size, phys,
116 pgprot_val(PAGE_KERNEL),
117 mmu_vmemmap_psize, mmu_kernel_ssize);
118 if (rc < 0) {
119 int rc2 = htab_remove_mapping(start, start + page_size,
120 mmu_vmemmap_psize,
121 mmu_kernel_ssize);
122 BUG_ON(rc2 && (rc2 != -ENOENT));
123 }
124 return rc;
125 }
126
127 #ifdef CONFIG_MEMORY_HOTPLUG
hash__vmemmap_remove_mapping(unsigned long start,unsigned long page_size)128 void hash__vmemmap_remove_mapping(unsigned long start,
129 unsigned long page_size)
130 {
131 int rc = htab_remove_mapping(start, start + page_size,
132 mmu_vmemmap_psize,
133 mmu_kernel_ssize);
134 BUG_ON((rc < 0) && (rc != -ENOENT));
135 WARN_ON(rc == -ENOENT);
136 }
137 #endif
138 #endif /* CONFIG_SPARSEMEM_VMEMMAP */
139
140 /*
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
144 */
hash__map_kernel_page(unsigned long ea,unsigned long pa,unsigned long flags)145 int hash__map_kernel_page(unsigned long ea, unsigned long pa, unsigned long flags)
146 {
147 pgd_t *pgdp;
148 pud_t *pudp;
149 pmd_t *pmdp;
150 pte_t *ptep;
151
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);
156 if (!pudp)
157 return -ENOMEM;
158 pmdp = pmd_alloc(&init_mm, pudp, ea);
159 if (!pmdp)
160 return -ENOMEM;
161 ptep = pte_alloc_kernel(pmdp, ea);
162 if (!ptep)
163 return -ENOMEM;
164 set_pte_at(&init_mm, ea, ptep, pfn_pte(pa >> PAGE_SHIFT,
165 __pgprot(flags)));
166 } else {
167 /*
168 * If the mm subsystem is not fully up, we cannot create a
169 * linux page table entry for this mapping. Simply bolt an
170 * entry in the hardware page table.
171 *
172 */
173 if (htab_bolt_mapping(ea, ea + PAGE_SIZE, pa, flags,
174 mmu_io_psize, mmu_kernel_ssize)) {
175 printk(KERN_ERR "Failed to do bolted mapping IO "
176 "memory at %016lx !\n", pa);
177 return -ENOMEM;
178 }
179 }
180
181 smp_wmb();
182 return 0;
183 }
184
185 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
186
hash__pmd_hugepage_update(struct mm_struct * mm,unsigned long addr,pmd_t * pmdp,unsigned long clr,unsigned long set)187 unsigned long hash__pmd_hugepage_update(struct mm_struct *mm, unsigned long addr,
188 pmd_t *pmdp, unsigned long clr,
189 unsigned long set)
190 {
191 __be64 old_be, tmp;
192 unsigned long old;
193
194 #ifdef CONFIG_DEBUG_VM
195 WARN_ON(!hash__pmd_trans_huge(*pmdp) && !pmd_devmap(*pmdp));
196 assert_spin_locked(pmd_lockptr(mm, pmdp));
197 #endif
198
199 __asm__ __volatile__(
200 "1: ldarx %0,0,%3\n\
201 and. %1,%0,%6\n\
202 bne- 1b \n\
203 andc %1,%0,%4 \n\
204 or %1,%1,%7\n\
205 stdcx. %1,0,%3 \n\
206 bne- 1b"
207 : "=&r" (old_be), "=&r" (tmp), "=m" (*pmdp)
208 : "r" (pmdp), "r" (cpu_to_be64(clr)), "m" (*pmdp),
209 "r" (cpu_to_be64(H_PAGE_BUSY)), "r" (cpu_to_be64(set))
210 : "cc" );
211
212 old = be64_to_cpu(old_be);
213
214 trace_hugepage_update(addr, old, clr, set);
215 if (old & H_PAGE_HASHPTE)
216 hpte_do_hugepage_flush(mm, addr, pmdp, old);
217 return old;
218 }
219
hash__pmdp_collapse_flush(struct vm_area_struct * vma,unsigned long address,pmd_t * pmdp)220 pmd_t hash__pmdp_collapse_flush(struct vm_area_struct *vma, unsigned long address,
221 pmd_t *pmdp)
222 {
223 pmd_t pmd;
224
225 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
226 VM_BUG_ON(pmd_trans_huge(*pmdp));
227 VM_BUG_ON(pmd_devmap(*pmdp));
228
229 pmd = *pmdp;
230 pmd_clear(pmdp);
231 /*
232 * Wait for all pending hash_page to finish. This is needed
233 * in case of subpage collapse. When we collapse normal pages
234 * to hugepage, we first clear the pmd, then invalidate all
235 * the PTE entries. The assumption here is that any low level
236 * page fault will see a none pmd and take the slow path that
237 * will wait on mmap_sem. But we could very well be in a
238 * hash_page with local ptep pointer value. Such a hash page
239 * can result in adding new HPTE entries for normal subpages.
240 * That means we could be modifying the page content as we
241 * copy them to a huge page. So wait for parallel hash_page
242 * to finish before invalidating HPTE entries. We can do this
243 * by sending an IPI to all the cpus and executing a dummy
244 * function there.
245 */
246 serialize_against_pte_lookup(vma->vm_mm);
247 /*
248 * Now invalidate the hpte entries in the range
249 * covered by pmd. This make sure we take a
250 * fault and will find the pmd as none, which will
251 * result in a major fault which takes mmap_sem and
252 * hence wait for collapse to complete. Without this
253 * the __collapse_huge_page_copy can result in copying
254 * the old content.
255 */
256 flush_tlb_pmd_range(vma->vm_mm, &pmd, address);
257 return pmd;
258 }
259
260 /*
261 * We want to put the pgtable in pmd and use pgtable for tracking
262 * the base page size hptes
263 */
hash__pgtable_trans_huge_deposit(struct mm_struct * mm,pmd_t * pmdp,pgtable_t pgtable)264 void hash__pgtable_trans_huge_deposit(struct mm_struct *mm, pmd_t *pmdp,
265 pgtable_t pgtable)
266 {
267 pgtable_t *pgtable_slot;
268
269 assert_spin_locked(pmd_lockptr(mm, pmdp));
270 /*
271 * we store the pgtable in the second half of PMD
272 */
273 pgtable_slot = (pgtable_t *)pmdp + PTRS_PER_PMD;
274 *pgtable_slot = pgtable;
275 /*
276 * expose the deposited pgtable to other cpus.
277 * before we set the hugepage PTE at pmd level
278 * hash fault code looks at the deposted pgtable
279 * to store hash index values.
280 */
281 smp_wmb();
282 }
283
hash__pgtable_trans_huge_withdraw(struct mm_struct * mm,pmd_t * pmdp)284 pgtable_t hash__pgtable_trans_huge_withdraw(struct mm_struct *mm, pmd_t *pmdp)
285 {
286 pgtable_t pgtable;
287 pgtable_t *pgtable_slot;
288
289 assert_spin_locked(pmd_lockptr(mm, pmdp));
290
291 pgtable_slot = (pgtable_t *)pmdp + PTRS_PER_PMD;
292 pgtable = *pgtable_slot;
293 /*
294 * Once we withdraw, mark the entry NULL.
295 */
296 *pgtable_slot = NULL;
297 /*
298 * We store HPTE information in the deposited PTE fragment.
299 * zero out the content on withdraw.
300 */
301 memset(pgtable, 0, PTE_FRAG_SIZE);
302 return pgtable;
303 }
304
305 /*
306 * A linux hugepage PMD was changed and the corresponding hash table entries
307 * neesd to be flushed.
308 */
hpte_do_hugepage_flush(struct mm_struct * mm,unsigned long addr,pmd_t * pmdp,unsigned long old_pmd)309 void hpte_do_hugepage_flush(struct mm_struct *mm, unsigned long addr,
310 pmd_t *pmdp, unsigned long old_pmd)
311 {
312 int ssize;
313 unsigned int psize;
314 unsigned long vsid;
315 unsigned long flags = 0;
316
317 /* get the base page size,vsid and segment size */
318 #ifdef CONFIG_DEBUG_VM
319 psize = get_slice_psize(mm, addr);
320 BUG_ON(psize == MMU_PAGE_16M);
321 #endif
322 if (old_pmd & H_PAGE_COMBO)
323 psize = MMU_PAGE_4K;
324 else
325 psize = MMU_PAGE_64K;
326
327 if (!is_kernel_addr(addr)) {
328 ssize = user_segment_size(addr);
329 vsid = get_user_vsid(&mm->context, addr, ssize);
330 WARN_ON(vsid == 0);
331 } else {
332 vsid = get_kernel_vsid(addr, mmu_kernel_ssize);
333 ssize = mmu_kernel_ssize;
334 }
335
336 if (mm_is_thread_local(mm))
337 flags |= HPTE_LOCAL_UPDATE;
338
339 return flush_hash_hugepage(vsid, addr, pmdp, psize, ssize, flags);
340 }
341
hash__pmdp_huge_get_and_clear(struct mm_struct * mm,unsigned long addr,pmd_t * pmdp)342 pmd_t hash__pmdp_huge_get_and_clear(struct mm_struct *mm,
343 unsigned long addr, pmd_t *pmdp)
344 {
345 pmd_t old_pmd;
346 pgtable_t pgtable;
347 unsigned long old;
348 pgtable_t *pgtable_slot;
349
350 old = pmd_hugepage_update(mm, addr, pmdp, ~0UL, 0);
351 old_pmd = __pmd(old);
352 /*
353 * We have pmd == none and we are holding page_table_lock.
354 * So we can safely go and clear the pgtable hash
355 * index info.
356 */
357 pgtable_slot = (pgtable_t *)pmdp + PTRS_PER_PMD;
358 pgtable = *pgtable_slot;
359 /*
360 * Let's zero out old valid and hash index details
361 * hash fault look at them.
362 */
363 memset(pgtable, 0, PTE_FRAG_SIZE);
364 /*
365 * Serialize against find_current_mm_pte variants which does lock-less
366 * lookup in page tables with local interrupts disabled. For huge pages
367 * it casts pmd_t to pte_t. Since format of pte_t is different from
368 * pmd_t we want to prevent transit from pmd pointing to page table
369 * to pmd pointing to huge page (and back) while interrupts are disabled.
370 * We clear pmd to possibly replace it with page table pointer in
371 * different code paths. So make sure we wait for the parallel
372 * find_curren_mm_pte to finish.
373 */
374 serialize_against_pte_lookup(mm);
375 return old_pmd;
376 }
377
hash__has_transparent_hugepage(void)378 int hash__has_transparent_hugepage(void)
379 {
380
381 if (!mmu_has_feature(MMU_FTR_16M_PAGE))
382 return 0;
383 /*
384 * We support THP only if PMD_SIZE is 16MB.
385 */
386 if (mmu_psize_defs[MMU_PAGE_16M].shift != PMD_SHIFT)
387 return 0;
388 /*
389 * We need to make sure that we support 16MB hugepage in a segement
390 * with base page size 64K or 4K. We only enable THP with a PAGE_SIZE
391 * of 64K.
392 */
393 /*
394 * If we have 64K HPTE, we will be using that by default
395 */
396 if (mmu_psize_defs[MMU_PAGE_64K].shift &&
397 (mmu_psize_defs[MMU_PAGE_64K].penc[MMU_PAGE_16M] == -1))
398 return 0;
399 /*
400 * Ok we only have 4K HPTE
401 */
402 if (mmu_psize_defs[MMU_PAGE_4K].penc[MMU_PAGE_16M] == -1)
403 return 0;
404
405 return 1;
406 }
407 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
408
409 #ifdef CONFIG_STRICT_KERNEL_RWX
hash__change_memory_range(unsigned long start,unsigned long end,unsigned long newpp)410 static bool hash__change_memory_range(unsigned long start, unsigned long end,
411 unsigned long newpp)
412 {
413 unsigned long idx;
414 unsigned int step, shift;
415
416 shift = mmu_psize_defs[mmu_linear_psize].shift;
417 step = 1 << shift;
418
419 start = ALIGN_DOWN(start, step);
420 end = ALIGN(end, step); // aligns up
421
422 if (start >= end)
423 return false;
424
425 pr_debug("Changing page protection on range 0x%lx-0x%lx, to 0x%lx, step 0x%x\n",
426 start, end, newpp, step);
427
428 for (idx = start; idx < end; idx += step)
429 /* Not sure if we can do much with the return value */
430 mmu_hash_ops.hpte_updateboltedpp(newpp, idx, mmu_linear_psize,
431 mmu_kernel_ssize);
432
433 return true;
434 }
435
hash__mark_rodata_ro(void)436 void hash__mark_rodata_ro(void)
437 {
438 unsigned long start, end;
439
440 start = (unsigned long)_stext;
441 end = (unsigned long)__init_begin;
442
443 WARN_ON(!hash__change_memory_range(start, end, PP_RXXX));
444 }
445
hash__mark_initmem_nx(void)446 void hash__mark_initmem_nx(void)
447 {
448 unsigned long start, end, pp;
449
450 start = (unsigned long)__init_begin;
451 end = (unsigned long)__init_end;
452
453 pp = htab_convert_pte_flags(pgprot_val(PAGE_KERNEL));
454
455 WARN_ON(!hash__change_memory_range(start, end, pp));
456 }
457 #endif
458
