1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Page table allocation functions
4 *
5 * Copyright IBM Corp. 2016
6 * Author(s): Martin Schwidefsky <schwidefsky@de.ibm.com>
7 */
8
9#include <linux/sysctl.h>
10#include <linux/slab.h>
11#include <linux/mm.h>
12#include <asm/mmu_context.h>
13#include <asm/page-states.h>
14#include <asm/pgalloc.h>
15#include <asm/gmap.h>
16#include <asm/tlb.h>
17#include <asm/tlbflush.h>
18
19#ifdef CONFIG_PGSTE
20
21int page_table_allocate_pgste = 0;
22EXPORT_SYMBOL(page_table_allocate_pgste);
23
24static struct ctl_table page_table_sysctl[] = {
25 {
26 .procname = "allocate_pgste",
27 .data = &page_table_allocate_pgste,
28 .maxlen = sizeof(int),
29 .mode = S_IRUGO | S_IWUSR,
30 .proc_handler = proc_dointvec_minmax,
31 .extra1 = SYSCTL_ZERO,
32 .extra2 = SYSCTL_ONE,
33 },
34};
35
36static int __init page_table_register_sysctl(void)
37{
38 return register_sysctl("vm", page_table_sysctl) ? 0 : -ENOMEM;
39}
40__initcall(page_table_register_sysctl);
41
42#endif /* CONFIG_PGSTE */
43
44unsigned long *crst_table_alloc(struct mm_struct *mm)
45{
46 struct ptdesc *ptdesc = pagetable_alloc(GFP_KERNEL, order: CRST_ALLOC_ORDER);
47 unsigned long *table;
48
49 if (!ptdesc)
50 return NULL;
51 table = ptdesc_to_virt(pt: ptdesc);
52 __arch_set_page_dat(table, 1UL << CRST_ALLOC_ORDER);
53 return table;
54}
55
56void crst_table_free(struct mm_struct *mm, unsigned long *table)
57{
58 pagetable_free(pt: virt_to_ptdesc(x: table));
59}
60
61static void __crst_table_upgrade(void *arg)
62{
63 struct mm_struct *mm = arg;
64
65 /* change all active ASCEs to avoid the creation of new TLBs */
66 if (current->active_mm == mm) {
67 S390_lowcore.user_asce.val = mm->context.asce;
68 local_ctl_load(7, &S390_lowcore.user_asce);
69 }
70 __tlb_flush_local();
71}
72
73int crst_table_upgrade(struct mm_struct *mm, unsigned long end)
74{
75 unsigned long *pgd = NULL, *p4d = NULL, *__pgd;
76 unsigned long asce_limit = mm->context.asce_limit;
77
78 /* upgrade should only happen from 3 to 4, 3 to 5, or 4 to 5 levels */
79 VM_BUG_ON(asce_limit < _REGION2_SIZE);
80
81 if (end <= asce_limit)
82 return 0;
83
84 if (asce_limit == _REGION2_SIZE) {
85 p4d = crst_table_alloc(mm);
86 if (unlikely(!p4d))
87 goto err_p4d;
88 crst_table_init(p4d, _REGION2_ENTRY_EMPTY);
89 }
90 if (end > _REGION1_SIZE) {
91 pgd = crst_table_alloc(mm);
92 if (unlikely(!pgd))
93 goto err_pgd;
94 crst_table_init(pgd, _REGION1_ENTRY_EMPTY);
95 }
96
97 spin_lock_bh(lock: &mm->page_table_lock);
98
99 /*
100 * This routine gets called with mmap_lock lock held and there is
101 * no reason to optimize for the case of otherwise. However, if
102 * that would ever change, the below check will let us know.
103 */
104 VM_BUG_ON(asce_limit != mm->context.asce_limit);
105
106 if (p4d) {
107 __pgd = (unsigned long *) mm->pgd;
108 p4d_populate(mm, p4d: (p4d_t *) p4d, pud: (pud_t *) __pgd);
109 mm->pgd = (pgd_t *) p4d;
110 mm->context.asce_limit = _REGION1_SIZE;
111 mm->context.asce = __pa(mm->pgd) | _ASCE_TABLE_LENGTH |
112 _ASCE_USER_BITS | _ASCE_TYPE_REGION2;
113 mm_inc_nr_puds(mm);
114 }
115 if (pgd) {
116 __pgd = (unsigned long *) mm->pgd;
117 pgd_populate(mm, pgd: (pgd_t *) pgd, p4d: (p4d_t *) __pgd);
118 mm->pgd = (pgd_t *) pgd;
119 mm->context.asce_limit = TASK_SIZE_MAX;
120 mm->context.asce = __pa(mm->pgd) | _ASCE_TABLE_LENGTH |
121 _ASCE_USER_BITS | _ASCE_TYPE_REGION1;
122 }
123
124 spin_unlock_bh(lock: &mm->page_table_lock);
125
126 on_each_cpu(func: __crst_table_upgrade, info: mm, wait: 0);
127
128 return 0;
129
130err_pgd:
131 crst_table_free(mm, table: p4d);
132err_p4d:
133 return -ENOMEM;
134}
135
136#ifdef CONFIG_PGSTE
137
138struct ptdesc *page_table_alloc_pgste(struct mm_struct *mm)
139{
140 struct ptdesc *ptdesc;
141 u64 *table;
142
143 ptdesc = pagetable_alloc(GFP_KERNEL, 0);
144 if (ptdesc) {
145 table = (u64 *)ptdesc_to_virt(ptdesc);
146 __arch_set_page_dat(table, 1);
147 memset64(table, _PAGE_INVALID, PTRS_PER_PTE);
148 memset64(table + PTRS_PER_PTE, 0, PTRS_PER_PTE);
149 }
150 return ptdesc;
151}
152
153void page_table_free_pgste(struct ptdesc *ptdesc)
154{
155 pagetable_free(ptdesc);
156}
157
158#endif /* CONFIG_PGSTE */
159
160unsigned long *page_table_alloc(struct mm_struct *mm)
161{
162 struct ptdesc *ptdesc;
163 unsigned long *table;
164
165 ptdesc = pagetable_alloc(GFP_KERNEL, order: 0);
166 if (!ptdesc)
167 return NULL;
168 if (!pagetable_pte_ctor(ptdesc)) {
169 pagetable_free(pt: ptdesc);
170 return NULL;
171 }
172 table = ptdesc_to_virt(pt: ptdesc);
173 __arch_set_page_dat(table, 1);
174 /* pt_list is used by gmap only */
175 INIT_LIST_HEAD(list: &ptdesc->pt_list);
176 memset64(s: (u64 *)table, v: _PAGE_INVALID, PTRS_PER_PTE);
177 memset64(s: (u64 *)table + PTRS_PER_PTE, v: 0, PTRS_PER_PTE);
178 return table;
179}
180
181static void pagetable_pte_dtor_free(struct ptdesc *ptdesc)
182{
183 pagetable_pte_dtor(ptdesc);
184 pagetable_free(pt: ptdesc);
185}
186
187void page_table_free(struct mm_struct *mm, unsigned long *table)
188{
189 struct ptdesc *ptdesc = virt_to_ptdesc(x: table);
190
191 pagetable_pte_dtor_free(ptdesc);
192}
193
194void __tlb_remove_table(void *table)
195{
196 struct ptdesc *ptdesc = virt_to_ptdesc(x: table);
197 struct page *page = ptdesc_page(ptdesc);
198
199 if (compound_order(page) == CRST_ALLOC_ORDER) {
200 /* pmd, pud, or p4d */
201 pagetable_free(pt: ptdesc);
202 return;
203 }
204 pagetable_pte_dtor_free(ptdesc);
205}
206
207#ifdef CONFIG_TRANSPARENT_HUGEPAGE
208static void pte_free_now(struct rcu_head *head)
209{
210 struct ptdesc *ptdesc = container_of(head, struct ptdesc, pt_rcu_head);
211
212 pagetable_pte_dtor_free(ptdesc);
213}
214
215void pte_free_defer(struct mm_struct *mm, pgtable_t pgtable)
216{
217 struct ptdesc *ptdesc = virt_to_ptdesc(x: pgtable);
218
219 call_rcu(head: &ptdesc->pt_rcu_head, func: pte_free_now);
220 /*
221 * THPs are not allowed for KVM guests. Warn if pgste ever reaches here.
222 * Turn to the generic pte_free_defer() version once gmap is removed.
223 */
224 WARN_ON_ONCE(mm_has_pgste(mm));
225}
226#endif /* CONFIG_TRANSPARENT_HUGEPAGE */
227
228/*
229 * Base infrastructure required to generate basic asces, region, segment,
230 * and page tables that do not make use of enhanced features like EDAT1.
231 */
232
233static struct kmem_cache *base_pgt_cache;
234
235static unsigned long *base_pgt_alloc(void)
236{
237 unsigned long *table;
238
239 table = kmem_cache_alloc(cachep: base_pgt_cache, GFP_KERNEL);
240 if (table)
241 memset64(s: (u64 *)table, v: _PAGE_INVALID, PTRS_PER_PTE);
242 return table;
243}
244
245static void base_pgt_free(unsigned long *table)
246{
247 kmem_cache_free(s: base_pgt_cache, objp: table);
248}
249
250static unsigned long *base_crst_alloc(unsigned long val)
251{
252 unsigned long *table;
253 struct ptdesc *ptdesc;
254
255 ptdesc = pagetable_alloc(GFP_KERNEL, order: CRST_ALLOC_ORDER);
256 if (!ptdesc)
257 return NULL;
258 table = ptdesc_address(pt: ptdesc);
259 crst_table_init(table, val);
260 return table;
261}
262
263static void base_crst_free(unsigned long *table)
264{
265 pagetable_free(pt: virt_to_ptdesc(x: table));
266}
267
268#define BASE_ADDR_END_FUNC(NAME, SIZE) \
269static inline unsigned long base_##NAME##_addr_end(unsigned long addr, \
270 unsigned long end) \
271{ \
272 unsigned long next = (addr + (SIZE)) & ~((SIZE) - 1); \
273 \
274 return (next - 1) < (end - 1) ? next : end; \
275}
276
277BASE_ADDR_END_FUNC(page, _PAGE_SIZE)
278BASE_ADDR_END_FUNC(segment, _SEGMENT_SIZE)
279BASE_ADDR_END_FUNC(region3, _REGION3_SIZE)
280BASE_ADDR_END_FUNC(region2, _REGION2_SIZE)
281BASE_ADDR_END_FUNC(region1, _REGION1_SIZE)
282
283static inline unsigned long base_lra(unsigned long address)
284{
285 unsigned long real;
286
287 asm volatile(
288 " lra %0,0(%1)\n"
289 : "=d" (real) : "a" (address) : "cc");
290 return real;
291}
292
293static int base_page_walk(unsigned long *origin, unsigned long addr,
294 unsigned long end, int alloc)
295{
296 unsigned long *pte, next;
297
298 if (!alloc)
299 return 0;
300 pte = origin;
301 pte += (addr & _PAGE_INDEX) >> _PAGE_SHIFT;
302 do {
303 next = base_page_addr_end(addr, end);
304 *pte = base_lra(address: addr);
305 } while (pte++, addr = next, addr < end);
306 return 0;
307}
308
309static int base_segment_walk(unsigned long *origin, unsigned long addr,
310 unsigned long end, int alloc)
311{
312 unsigned long *ste, next, *table;
313 int rc;
314
315 ste = origin;
316 ste += (addr & _SEGMENT_INDEX) >> _SEGMENT_SHIFT;
317 do {
318 next = base_segment_addr_end(addr, end);
319 if (*ste & _SEGMENT_ENTRY_INVALID) {
320 if (!alloc)
321 continue;
322 table = base_pgt_alloc();
323 if (!table)
324 return -ENOMEM;
325 *ste = __pa(table) | _SEGMENT_ENTRY;
326 }
327 table = __va(*ste & _SEGMENT_ENTRY_ORIGIN);
328 rc = base_page_walk(origin: table, addr, end: next, alloc);
329 if (rc)
330 return rc;
331 if (!alloc)
332 base_pgt_free(table);
333 cond_resched();
334 } while (ste++, addr = next, addr < end);
335 return 0;
336}
337
338static int base_region3_walk(unsigned long *origin, unsigned long addr,
339 unsigned long end, int alloc)
340{
341 unsigned long *rtte, next, *table;
342 int rc;
343
344 rtte = origin;
345 rtte += (addr & _REGION3_INDEX) >> _REGION3_SHIFT;
346 do {
347 next = base_region3_addr_end(addr, end);
348 if (*rtte & _REGION_ENTRY_INVALID) {
349 if (!alloc)
350 continue;
351 table = base_crst_alloc(_SEGMENT_ENTRY_EMPTY);
352 if (!table)
353 return -ENOMEM;
354 *rtte = __pa(table) | _REGION3_ENTRY;
355 }
356 table = __va(*rtte & _REGION_ENTRY_ORIGIN);
357 rc = base_segment_walk(origin: table, addr, end: next, alloc);
358 if (rc)
359 return rc;
360 if (!alloc)
361 base_crst_free(table);
362 } while (rtte++, addr = next, addr < end);
363 return 0;
364}
365
366static int base_region2_walk(unsigned long *origin, unsigned long addr,
367 unsigned long end, int alloc)
368{
369 unsigned long *rste, next, *table;
370 int rc;
371
372 rste = origin;
373 rste += (addr & _REGION2_INDEX) >> _REGION2_SHIFT;
374 do {
375 next = base_region2_addr_end(addr, end);
376 if (*rste & _REGION_ENTRY_INVALID) {
377 if (!alloc)
378 continue;
379 table = base_crst_alloc(_REGION3_ENTRY_EMPTY);
380 if (!table)
381 return -ENOMEM;
382 *rste = __pa(table) | _REGION2_ENTRY;
383 }
384 table = __va(*rste & _REGION_ENTRY_ORIGIN);
385 rc = base_region3_walk(origin: table, addr, end: next, alloc);
386 if (rc)
387 return rc;
388 if (!alloc)
389 base_crst_free(table);
390 } while (rste++, addr = next, addr < end);
391 return 0;
392}
393
394static int base_region1_walk(unsigned long *origin, unsigned long addr,
395 unsigned long end, int alloc)
396{
397 unsigned long *rfte, next, *table;
398 int rc;
399
400 rfte = origin;
401 rfte += (addr & _REGION1_INDEX) >> _REGION1_SHIFT;
402 do {
403 next = base_region1_addr_end(addr, end);
404 if (*rfte & _REGION_ENTRY_INVALID) {
405 if (!alloc)
406 continue;
407 table = base_crst_alloc(_REGION2_ENTRY_EMPTY);
408 if (!table)
409 return -ENOMEM;
410 *rfte = __pa(table) | _REGION1_ENTRY;
411 }
412 table = __va(*rfte & _REGION_ENTRY_ORIGIN);
413 rc = base_region2_walk(origin: table, addr, end: next, alloc);
414 if (rc)
415 return rc;
416 if (!alloc)
417 base_crst_free(table);
418 } while (rfte++, addr = next, addr < end);
419 return 0;
420}
421
422/**
423 * base_asce_free - free asce and tables returned from base_asce_alloc()
424 * @asce: asce to be freed
425 *
426 * Frees all region, segment, and page tables that were allocated with a
427 * corresponding base_asce_alloc() call.
428 */
429void base_asce_free(unsigned long asce)
430{
431 unsigned long *table = __va(asce & _ASCE_ORIGIN);
432
433 if (!asce)
434 return;
435 switch (asce & _ASCE_TYPE_MASK) {
436 case _ASCE_TYPE_SEGMENT:
437 base_segment_walk(table, 0, _REGION3_SIZE, 0);
438 break;
439 case _ASCE_TYPE_REGION3:
440 base_region3_walk(table, 0, _REGION2_SIZE, 0);
441 break;
442 case _ASCE_TYPE_REGION2:
443 base_region2_walk(table, 0, _REGION1_SIZE, 0);
444 break;
445 case _ASCE_TYPE_REGION1:
446 base_region1_walk(origin: table, addr: 0, TASK_SIZE_MAX, alloc: 0);
447 break;
448 }
449 base_crst_free(table);
450}
451
452static int base_pgt_cache_init(void)
453{
454 static DEFINE_MUTEX(base_pgt_cache_mutex);
455 unsigned long sz = _PAGE_TABLE_SIZE;
456
457 if (base_pgt_cache)
458 return 0;
459 mutex_lock(&base_pgt_cache_mutex);
460 if (!base_pgt_cache)
461 base_pgt_cache = kmem_cache_create(name: "base_pgt", size: sz, align: sz, flags: 0, NULL);
462 mutex_unlock(lock: &base_pgt_cache_mutex);
463 return base_pgt_cache ? 0 : -ENOMEM;
464}
465
466/**
467 * base_asce_alloc - create kernel mapping without enhanced DAT features
468 * @addr: virtual start address of kernel mapping
469 * @num_pages: number of consecutive pages
470 *
471 * Generate an asce, including all required region, segment and page tables,
472 * that can be used to access the virtual kernel mapping. The difference is
473 * that the returned asce does not make use of any enhanced DAT features like
474 * e.g. large pages. This is required for some I/O functions that pass an
475 * asce, like e.g. some service call requests.
476 *
477 * Note: the returned asce may NEVER be attached to any cpu. It may only be
478 * used for I/O requests. tlb entries that might result because the
479 * asce was attached to a cpu won't be cleared.
480 */
481unsigned long base_asce_alloc(unsigned long addr, unsigned long num_pages)
482{
483 unsigned long asce, *table, end;
484 int rc;
485
486 if (base_pgt_cache_init())
487 return 0;
488 end = addr + num_pages * PAGE_SIZE;
489 if (end <= _REGION3_SIZE) {
490 table = base_crst_alloc(_SEGMENT_ENTRY_EMPTY);
491 if (!table)
492 return 0;
493 rc = base_segment_walk(origin: table, addr, end, alloc: 1);
494 asce = __pa(table) | _ASCE_TYPE_SEGMENT | _ASCE_TABLE_LENGTH;
495 } else if (end <= _REGION2_SIZE) {
496 table = base_crst_alloc(_REGION3_ENTRY_EMPTY);
497 if (!table)
498 return 0;
499 rc = base_region3_walk(origin: table, addr, end, alloc: 1);
500 asce = __pa(table) | _ASCE_TYPE_REGION3 | _ASCE_TABLE_LENGTH;
501 } else if (end <= _REGION1_SIZE) {
502 table = base_crst_alloc(_REGION2_ENTRY_EMPTY);
503 if (!table)
504 return 0;
505 rc = base_region2_walk(origin: table, addr, end, alloc: 1);
506 asce = __pa(table) | _ASCE_TYPE_REGION2 | _ASCE_TABLE_LENGTH;
507 } else {
508 table = base_crst_alloc(_REGION1_ENTRY_EMPTY);
509 if (!table)
510 return 0;
511 rc = base_region1_walk(origin: table, addr, end, alloc: 1);
512 asce = __pa(table) | _ASCE_TYPE_REGION1 | _ASCE_TABLE_LENGTH;
513 }
514 if (rc) {
515 base_asce_free(asce);
516 asce = 0;
517 }
518 return asce;
519}
520

source code of linux/arch/s390/mm/pgalloc.c