Loading...
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/pgalloc.h>
14#include <asm/gmap.h>
15#include <asm/tlb.h>
16#include <asm/tlbflush.h>
17
18#ifdef CONFIG_PGSTE
19
20int page_table_allocate_pgste = 0;
21EXPORT_SYMBOL(page_table_allocate_pgste);
22
23static struct ctl_table page_table_sysctl[] = {
24 {
25 .procname = "allocate_pgste",
26 .data = &page_table_allocate_pgste,
27 .maxlen = sizeof(int),
28 .mode = S_IRUGO | S_IWUSR,
29 .proc_handler = proc_dointvec_minmax,
30 .extra1 = SYSCTL_ZERO,
31 .extra2 = SYSCTL_ONE,
32 },
33 { }
34};
35
36static struct ctl_table page_table_sysctl_dir[] = {
37 {
38 .procname = "vm",
39 .maxlen = 0,
40 .mode = 0555,
41 .child = page_table_sysctl,
42 },
43 { }
44};
45
46static int __init page_table_register_sysctl(void)
47{
48 return register_sysctl_table(page_table_sysctl_dir) ? 0 : -ENOMEM;
49}
50__initcall(page_table_register_sysctl);
51
52#endif /* CONFIG_PGSTE */
53
54unsigned long *crst_table_alloc(struct mm_struct *mm)
55{
56 struct page *page = alloc_pages(GFP_KERNEL, 2);
57
58 if (!page)
59 return NULL;
60 arch_set_page_dat(page, 2);
61 return (unsigned long *) page_to_phys(page);
62}
63
64void crst_table_free(struct mm_struct *mm, unsigned long *table)
65{
66 free_pages((unsigned long) table, 2);
67}
68
69static void __crst_table_upgrade(void *arg)
70{
71 struct mm_struct *mm = arg;
72
73 /* we must change all active ASCEs to avoid the creation of new TLBs */
74 if (current->active_mm == mm) {
75 S390_lowcore.user_asce = mm->context.asce;
76 if (current->thread.mm_segment == USER_DS) {
77 __ctl_load(S390_lowcore.user_asce, 1, 1);
78 /* Mark user-ASCE present in CR1 */
79 clear_cpu_flag(CIF_ASCE_PRIMARY);
80 }
81 if (current->thread.mm_segment == USER_DS_SACF) {
82 __ctl_load(S390_lowcore.user_asce, 7, 7);
83 /* enable_sacf_uaccess does all or nothing */
84 WARN_ON(!test_cpu_flag(CIF_ASCE_SECONDARY));
85 }
86 }
87 __tlb_flush_local();
88}
89
90int crst_table_upgrade(struct mm_struct *mm, unsigned long end)
91{
92 unsigned long *pgd = NULL, *p4d = NULL, *__pgd;
93 unsigned long asce_limit = mm->context.asce_limit;
94
95 /* upgrade should only happen from 3 to 4, 3 to 5, or 4 to 5 levels */
96 VM_BUG_ON(asce_limit < _REGION2_SIZE);
97
98 if (end <= asce_limit)
99 return 0;
100
101 if (asce_limit == _REGION2_SIZE) {
102 p4d = crst_table_alloc(mm);
103 if (unlikely(!p4d))
104 goto err_p4d;
105 crst_table_init(p4d, _REGION2_ENTRY_EMPTY);
106 }
107 if (end > _REGION1_SIZE) {
108 pgd = crst_table_alloc(mm);
109 if (unlikely(!pgd))
110 goto err_pgd;
111 crst_table_init(pgd, _REGION1_ENTRY_EMPTY);
112 }
113
114 spin_lock_bh(&mm->page_table_lock);
115
116 /*
117 * This routine gets called with mmap_lock lock held and there is
118 * no reason to optimize for the case of otherwise. However, if
119 * that would ever change, the below check will let us know.
120 */
121 VM_BUG_ON(asce_limit != mm->context.asce_limit);
122
123 if (p4d) {
124 __pgd = (unsigned long *) mm->pgd;
125 p4d_populate(mm, (p4d_t *) p4d, (pud_t *) __pgd);
126 mm->pgd = (pgd_t *) p4d;
127 mm->context.asce_limit = _REGION1_SIZE;
128 mm->context.asce = __pa(mm->pgd) | _ASCE_TABLE_LENGTH |
129 _ASCE_USER_BITS | _ASCE_TYPE_REGION2;
130 mm_inc_nr_puds(mm);
131 }
132 if (pgd) {
133 __pgd = (unsigned long *) mm->pgd;
134 pgd_populate(mm, (pgd_t *) pgd, (p4d_t *) __pgd);
135 mm->pgd = (pgd_t *) pgd;
136 mm->context.asce_limit = TASK_SIZE_MAX;
137 mm->context.asce = __pa(mm->pgd) | _ASCE_TABLE_LENGTH |
138 _ASCE_USER_BITS | _ASCE_TYPE_REGION1;
139 }
140
141 spin_unlock_bh(&mm->page_table_lock);
142
143 on_each_cpu(__crst_table_upgrade, mm, 0);
144
145 return 0;
146
147err_pgd:
148 crst_table_free(mm, p4d);
149err_p4d:
150 return -ENOMEM;
151}
152
153static inline unsigned int atomic_xor_bits(atomic_t *v, unsigned int bits)
154{
155 unsigned int old, new;
156
157 do {
158 old = atomic_read(v);
159 new = old ^ bits;
160 } while (atomic_cmpxchg(v, old, new) != old);
161 return new;
162}
163
164#ifdef CONFIG_PGSTE
165
166struct page *page_table_alloc_pgste(struct mm_struct *mm)
167{
168 struct page *page;
169 u64 *table;
170
171 page = alloc_page(GFP_KERNEL);
172 if (page) {
173 table = (u64 *)page_to_phys(page);
174 memset64(table, _PAGE_INVALID, PTRS_PER_PTE);
175 memset64(table + PTRS_PER_PTE, 0, PTRS_PER_PTE);
176 }
177 return page;
178}
179
180void page_table_free_pgste(struct page *page)
181{
182 __free_page(page);
183}
184
185#endif /* CONFIG_PGSTE */
186
187/*
188 * page table entry allocation/free routines.
189 */
190unsigned long *page_table_alloc(struct mm_struct *mm)
191{
192 unsigned long *table;
193 struct page *page;
194 unsigned int mask, bit;
195
196 /* Try to get a fragment of a 4K page as a 2K page table */
197 if (!mm_alloc_pgste(mm)) {
198 table = NULL;
199 spin_lock_bh(&mm->context.lock);
200 if (!list_empty(&mm->context.pgtable_list)) {
201 page = list_first_entry(&mm->context.pgtable_list,
202 struct page, lru);
203 mask = atomic_read(&page->_refcount) >> 24;
204 mask = (mask | (mask >> 4)) & 3;
205 if (mask != 3) {
206 table = (unsigned long *) page_to_phys(page);
207 bit = mask & 1; /* =1 -> second 2K */
208 if (bit)
209 table += PTRS_PER_PTE;
210 atomic_xor_bits(&page->_refcount,
211 1U << (bit + 24));
212 list_del(&page->lru);
213 }
214 }
215 spin_unlock_bh(&mm->context.lock);
216 if (table)
217 return table;
218 }
219 /* Allocate a fresh page */
220 page = alloc_page(GFP_KERNEL);
221 if (!page)
222 return NULL;
223 if (!pgtable_pte_page_ctor(page)) {
224 __free_page(page);
225 return NULL;
226 }
227 arch_set_page_dat(page, 0);
228 /* Initialize page table */
229 table = (unsigned long *) page_to_phys(page);
230 if (mm_alloc_pgste(mm)) {
231 /* Return 4K page table with PGSTEs */
232 atomic_xor_bits(&page->_refcount, 3 << 24);
233 memset64((u64 *)table, _PAGE_INVALID, PTRS_PER_PTE);
234 memset64((u64 *)table + PTRS_PER_PTE, 0, PTRS_PER_PTE);
235 } else {
236 /* Return the first 2K fragment of the page */
237 atomic_xor_bits(&page->_refcount, 1 << 24);
238 memset64((u64 *)table, _PAGE_INVALID, 2 * PTRS_PER_PTE);
239 spin_lock_bh(&mm->context.lock);
240 list_add(&page->lru, &mm->context.pgtable_list);
241 spin_unlock_bh(&mm->context.lock);
242 }
243 return table;
244}
245
246void page_table_free(struct mm_struct *mm, unsigned long *table)
247{
248 struct page *page;
249 unsigned int bit, mask;
250
251 page = pfn_to_page(__pa(table) >> PAGE_SHIFT);
252 if (!mm_alloc_pgste(mm)) {
253 /* Free 2K page table fragment of a 4K page */
254 bit = (__pa(table) & ~PAGE_MASK)/(PTRS_PER_PTE*sizeof(pte_t));
255 spin_lock_bh(&mm->context.lock);
256 mask = atomic_xor_bits(&page->_refcount, 1U << (bit + 24));
257 mask >>= 24;
258 if (mask & 3)
259 list_add(&page->lru, &mm->context.pgtable_list);
260 else
261 list_del(&page->lru);
262 spin_unlock_bh(&mm->context.lock);
263 if (mask != 0)
264 return;
265 } else {
266 atomic_xor_bits(&page->_refcount, 3U << 24);
267 }
268
269 pgtable_pte_page_dtor(page);
270 __free_page(page);
271}
272
273void page_table_free_rcu(struct mmu_gather *tlb, unsigned long *table,
274 unsigned long vmaddr)
275{
276 struct mm_struct *mm;
277 struct page *page;
278 unsigned int bit, mask;
279
280 mm = tlb->mm;
281 page = pfn_to_page(__pa(table) >> PAGE_SHIFT);
282 if (mm_alloc_pgste(mm)) {
283 gmap_unlink(mm, table, vmaddr);
284 table = (unsigned long *) (__pa(table) | 3);
285 tlb_remove_table(tlb, table);
286 return;
287 }
288 bit = (__pa(table) & ~PAGE_MASK) / (PTRS_PER_PTE*sizeof(pte_t));
289 spin_lock_bh(&mm->context.lock);
290 mask = atomic_xor_bits(&page->_refcount, 0x11U << (bit + 24));
291 mask >>= 24;
292 if (mask & 3)
293 list_add_tail(&page->lru, &mm->context.pgtable_list);
294 else
295 list_del(&page->lru);
296 spin_unlock_bh(&mm->context.lock);
297 table = (unsigned long *) (__pa(table) | (1U << bit));
298 tlb_remove_table(tlb, table);
299}
300
301void __tlb_remove_table(void *_table)
302{
303 unsigned int mask = (unsigned long) _table & 3;
304 void *table = (void *)((unsigned long) _table ^ mask);
305 struct page *page = pfn_to_page(__pa(table) >> PAGE_SHIFT);
306
307 switch (mask) {
308 case 0: /* pmd, pud, or p4d */
309 free_pages((unsigned long) table, 2);
310 break;
311 case 1: /* lower 2K of a 4K page table */
312 case 2: /* higher 2K of a 4K page table */
313 mask = atomic_xor_bits(&page->_refcount, mask << (4 + 24));
314 mask >>= 24;
315 if (mask != 0)
316 break;
317 fallthrough;
318 case 3: /* 4K page table with pgstes */
319 if (mask & 3)
320 atomic_xor_bits(&page->_refcount, 3 << 24);
321 pgtable_pte_page_dtor(page);
322 __free_page(page);
323 break;
324 }
325}
326
327/*
328 * Base infrastructure required to generate basic asces, region, segment,
329 * and page tables that do not make use of enhanced features like EDAT1.
330 */
331
332static struct kmem_cache *base_pgt_cache;
333
334static unsigned long base_pgt_alloc(void)
335{
336 u64 *table;
337
338 table = kmem_cache_alloc(base_pgt_cache, GFP_KERNEL);
339 if (table)
340 memset64(table, _PAGE_INVALID, PTRS_PER_PTE);
341 return (unsigned long) table;
342}
343
344static void base_pgt_free(unsigned long table)
345{
346 kmem_cache_free(base_pgt_cache, (void *) table);
347}
348
349static unsigned long base_crst_alloc(unsigned long val)
350{
351 unsigned long table;
352
353 table = __get_free_pages(GFP_KERNEL, CRST_ALLOC_ORDER);
354 if (table)
355 crst_table_init((unsigned long *)table, val);
356 return table;
357}
358
359static void base_crst_free(unsigned long table)
360{
361 free_pages(table, CRST_ALLOC_ORDER);
362}
363
364#define BASE_ADDR_END_FUNC(NAME, SIZE) \
365static inline unsigned long base_##NAME##_addr_end(unsigned long addr, \
366 unsigned long end) \
367{ \
368 unsigned long next = (addr + (SIZE)) & ~((SIZE) - 1); \
369 \
370 return (next - 1) < (end - 1) ? next : end; \
371}
372
373BASE_ADDR_END_FUNC(page, _PAGE_SIZE)
374BASE_ADDR_END_FUNC(segment, _SEGMENT_SIZE)
375BASE_ADDR_END_FUNC(region3, _REGION3_SIZE)
376BASE_ADDR_END_FUNC(region2, _REGION2_SIZE)
377BASE_ADDR_END_FUNC(region1, _REGION1_SIZE)
378
379static inline unsigned long base_lra(unsigned long address)
380{
381 unsigned long real;
382
383 asm volatile(
384 " lra %0,0(%1)\n"
385 : "=d" (real) : "a" (address) : "cc");
386 return real;
387}
388
389static int base_page_walk(unsigned long origin, unsigned long addr,
390 unsigned long end, int alloc)
391{
392 unsigned long *pte, next;
393
394 if (!alloc)
395 return 0;
396 pte = (unsigned long *) origin;
397 pte += (addr & _PAGE_INDEX) >> _PAGE_SHIFT;
398 do {
399 next = base_page_addr_end(addr, end);
400 *pte = base_lra(addr);
401 } while (pte++, addr = next, addr < end);
402 return 0;
403}
404
405static int base_segment_walk(unsigned long origin, unsigned long addr,
406 unsigned long end, int alloc)
407{
408 unsigned long *ste, next, table;
409 int rc;
410
411 ste = (unsigned long *) origin;
412 ste += (addr & _SEGMENT_INDEX) >> _SEGMENT_SHIFT;
413 do {
414 next = base_segment_addr_end(addr, end);
415 if (*ste & _SEGMENT_ENTRY_INVALID) {
416 if (!alloc)
417 continue;
418 table = base_pgt_alloc();
419 if (!table)
420 return -ENOMEM;
421 *ste = table | _SEGMENT_ENTRY;
422 }
423 table = *ste & _SEGMENT_ENTRY_ORIGIN;
424 rc = base_page_walk(table, addr, next, alloc);
425 if (rc)
426 return rc;
427 if (!alloc)
428 base_pgt_free(table);
429 cond_resched();
430 } while (ste++, addr = next, addr < end);
431 return 0;
432}
433
434static int base_region3_walk(unsigned long origin, unsigned long addr,
435 unsigned long end, int alloc)
436{
437 unsigned long *rtte, next, table;
438 int rc;
439
440 rtte = (unsigned long *) origin;
441 rtte += (addr & _REGION3_INDEX) >> _REGION3_SHIFT;
442 do {
443 next = base_region3_addr_end(addr, end);
444 if (*rtte & _REGION_ENTRY_INVALID) {
445 if (!alloc)
446 continue;
447 table = base_crst_alloc(_SEGMENT_ENTRY_EMPTY);
448 if (!table)
449 return -ENOMEM;
450 *rtte = table | _REGION3_ENTRY;
451 }
452 table = *rtte & _REGION_ENTRY_ORIGIN;
453 rc = base_segment_walk(table, addr, next, alloc);
454 if (rc)
455 return rc;
456 if (!alloc)
457 base_crst_free(table);
458 } while (rtte++, addr = next, addr < end);
459 return 0;
460}
461
462static int base_region2_walk(unsigned long origin, unsigned long addr,
463 unsigned long end, int alloc)
464{
465 unsigned long *rste, next, table;
466 int rc;
467
468 rste = (unsigned long *) origin;
469 rste += (addr & _REGION2_INDEX) >> _REGION2_SHIFT;
470 do {
471 next = base_region2_addr_end(addr, end);
472 if (*rste & _REGION_ENTRY_INVALID) {
473 if (!alloc)
474 continue;
475 table = base_crst_alloc(_REGION3_ENTRY_EMPTY);
476 if (!table)
477 return -ENOMEM;
478 *rste = table | _REGION2_ENTRY;
479 }
480 table = *rste & _REGION_ENTRY_ORIGIN;
481 rc = base_region3_walk(table, addr, next, alloc);
482 if (rc)
483 return rc;
484 if (!alloc)
485 base_crst_free(table);
486 } while (rste++, addr = next, addr < end);
487 return 0;
488}
489
490static int base_region1_walk(unsigned long origin, unsigned long addr,
491 unsigned long end, int alloc)
492{
493 unsigned long *rfte, next, table;
494 int rc;
495
496 rfte = (unsigned long *) origin;
497 rfte += (addr & _REGION1_INDEX) >> _REGION1_SHIFT;
498 do {
499 next = base_region1_addr_end(addr, end);
500 if (*rfte & _REGION_ENTRY_INVALID) {
501 if (!alloc)
502 continue;
503 table = base_crst_alloc(_REGION2_ENTRY_EMPTY);
504 if (!table)
505 return -ENOMEM;
506 *rfte = table | _REGION1_ENTRY;
507 }
508 table = *rfte & _REGION_ENTRY_ORIGIN;
509 rc = base_region2_walk(table, addr, next, alloc);
510 if (rc)
511 return rc;
512 if (!alloc)
513 base_crst_free(table);
514 } while (rfte++, addr = next, addr < end);
515 return 0;
516}
517
518/**
519 * base_asce_free - free asce and tables returned from base_asce_alloc()
520 * @asce: asce to be freed
521 *
522 * Frees all region, segment, and page tables that were allocated with a
523 * corresponding base_asce_alloc() call.
524 */
525void base_asce_free(unsigned long asce)
526{
527 unsigned long table = asce & _ASCE_ORIGIN;
528
529 if (!asce)
530 return;
531 switch (asce & _ASCE_TYPE_MASK) {
532 case _ASCE_TYPE_SEGMENT:
533 base_segment_walk(table, 0, _REGION3_SIZE, 0);
534 break;
535 case _ASCE_TYPE_REGION3:
536 base_region3_walk(table, 0, _REGION2_SIZE, 0);
537 break;
538 case _ASCE_TYPE_REGION2:
539 base_region2_walk(table, 0, _REGION1_SIZE, 0);
540 break;
541 case _ASCE_TYPE_REGION1:
542 base_region1_walk(table, 0, TASK_SIZE_MAX, 0);
543 break;
544 }
545 base_crst_free(table);
546}
547
548static int base_pgt_cache_init(void)
549{
550 static DEFINE_MUTEX(base_pgt_cache_mutex);
551 unsigned long sz = _PAGE_TABLE_SIZE;
552
553 if (base_pgt_cache)
554 return 0;
555 mutex_lock(&base_pgt_cache_mutex);
556 if (!base_pgt_cache)
557 base_pgt_cache = kmem_cache_create("base_pgt", sz, sz, 0, NULL);
558 mutex_unlock(&base_pgt_cache_mutex);
559 return base_pgt_cache ? 0 : -ENOMEM;
560}
561
562/**
563 * base_asce_alloc - create kernel mapping without enhanced DAT features
564 * @addr: virtual start address of kernel mapping
565 * @num_pages: number of consecutive pages
566 *
567 * Generate an asce, including all required region, segment and page tables,
568 * that can be used to access the virtual kernel mapping. The difference is
569 * that the returned asce does not make use of any enhanced DAT features like
570 * e.g. large pages. This is required for some I/O functions that pass an
571 * asce, like e.g. some service call requests.
572 *
573 * Note: the returned asce may NEVER be attached to any cpu. It may only be
574 * used for I/O requests. tlb entries that might result because the
575 * asce was attached to a cpu won't be cleared.
576 */
577unsigned long base_asce_alloc(unsigned long addr, unsigned long num_pages)
578{
579 unsigned long asce, table, end;
580 int rc;
581
582 if (base_pgt_cache_init())
583 return 0;
584 end = addr + num_pages * PAGE_SIZE;
585 if (end <= _REGION3_SIZE) {
586 table = base_crst_alloc(_SEGMENT_ENTRY_EMPTY);
587 if (!table)
588 return 0;
589 rc = base_segment_walk(table, addr, end, 1);
590 asce = table | _ASCE_TYPE_SEGMENT | _ASCE_TABLE_LENGTH;
591 } else if (end <= _REGION2_SIZE) {
592 table = base_crst_alloc(_REGION3_ENTRY_EMPTY);
593 if (!table)
594 return 0;
595 rc = base_region3_walk(table, addr, end, 1);
596 asce = table | _ASCE_TYPE_REGION3 | _ASCE_TABLE_LENGTH;
597 } else if (end <= _REGION1_SIZE) {
598 table = base_crst_alloc(_REGION2_ENTRY_EMPTY);
599 if (!table)
600 return 0;
601 rc = base_region2_walk(table, addr, end, 1);
602 asce = table | _ASCE_TYPE_REGION2 | _ASCE_TABLE_LENGTH;
603 } else {
604 table = base_crst_alloc(_REGION1_ENTRY_EMPTY);
605 if (!table)
606 return 0;
607 rc = base_region1_walk(table, addr, end, 1);
608 asce = table | _ASCE_TYPE_REGION1 | _ASCE_TABLE_LENGTH;
609 }
610 if (rc) {
611 base_asce_free(asce);
612 asce = 0;
613 }
614 return asce;
615}
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/pgalloc.h>
14#include <asm/gmap.h>
15#include <asm/tlb.h>
16#include <asm/tlbflush.h>
17
18#ifdef CONFIG_PGSTE
19
20static int page_table_allocate_pgste_min = 0;
21static int page_table_allocate_pgste_max = 1;
22int page_table_allocate_pgste = 0;
23EXPORT_SYMBOL(page_table_allocate_pgste);
24
25static struct ctl_table page_table_sysctl[] = {
26 {
27 .procname = "allocate_pgste",
28 .data = &page_table_allocate_pgste,
29 .maxlen = sizeof(int),
30 .mode = S_IRUGO | S_IWUSR,
31 .proc_handler = proc_dointvec,
32 .extra1 = &page_table_allocate_pgste_min,
33 .extra2 = &page_table_allocate_pgste_max,
34 },
35 { }
36};
37
38static struct ctl_table page_table_sysctl_dir[] = {
39 {
40 .procname = "vm",
41 .maxlen = 0,
42 .mode = 0555,
43 .child = page_table_sysctl,
44 },
45 { }
46};
47
48static int __init page_table_register_sysctl(void)
49{
50 return register_sysctl_table(page_table_sysctl_dir) ? 0 : -ENOMEM;
51}
52__initcall(page_table_register_sysctl);
53
54#endif /* CONFIG_PGSTE */
55
56unsigned long *crst_table_alloc(struct mm_struct *mm)
57{
58 struct page *page = alloc_pages(GFP_KERNEL, 2);
59
60 if (!page)
61 return NULL;
62 arch_set_page_dat(page, 2);
63 return (unsigned long *) page_to_phys(page);
64}
65
66void crst_table_free(struct mm_struct *mm, unsigned long *table)
67{
68 free_pages((unsigned long) table, 2);
69}
70
71static void __crst_table_upgrade(void *arg)
72{
73 struct mm_struct *mm = arg;
74
75 if (current->active_mm == mm)
76 set_user_asce(mm);
77 __tlb_flush_local();
78}
79
80int crst_table_upgrade(struct mm_struct *mm, unsigned long end)
81{
82 unsigned long *table, *pgd;
83 int rc, notify;
84
85 /* upgrade should only happen from 3 to 4, 3 to 5, or 4 to 5 levels */
86 VM_BUG_ON(mm->context.asce_limit < _REGION2_SIZE);
87 rc = 0;
88 notify = 0;
89 while (mm->context.asce_limit < end) {
90 table = crst_table_alloc(mm);
91 if (!table) {
92 rc = -ENOMEM;
93 break;
94 }
95 spin_lock_bh(&mm->page_table_lock);
96 pgd = (unsigned long *) mm->pgd;
97 if (mm->context.asce_limit == _REGION2_SIZE) {
98 crst_table_init(table, _REGION2_ENTRY_EMPTY);
99 p4d_populate(mm, (p4d_t *) table, (pud_t *) pgd);
100 mm->pgd = (pgd_t *) table;
101 mm->context.asce_limit = _REGION1_SIZE;
102 mm->context.asce = __pa(mm->pgd) | _ASCE_TABLE_LENGTH |
103 _ASCE_USER_BITS | _ASCE_TYPE_REGION2;
104 } else {
105 crst_table_init(table, _REGION1_ENTRY_EMPTY);
106 pgd_populate(mm, (pgd_t *) table, (p4d_t *) pgd);
107 mm->pgd = (pgd_t *) table;
108 mm->context.asce_limit = -PAGE_SIZE;
109 mm->context.asce = __pa(mm->pgd) | _ASCE_TABLE_LENGTH |
110 _ASCE_USER_BITS | _ASCE_TYPE_REGION1;
111 }
112 notify = 1;
113 spin_unlock_bh(&mm->page_table_lock);
114 }
115 if (notify)
116 on_each_cpu(__crst_table_upgrade, mm, 0);
117 return rc;
118}
119
120void crst_table_downgrade(struct mm_struct *mm)
121{
122 pgd_t *pgd;
123
124 /* downgrade should only happen from 3 to 2 levels (compat only) */
125 VM_BUG_ON(mm->context.asce_limit != _REGION2_SIZE);
126
127 if (current->active_mm == mm) {
128 clear_user_asce();
129 __tlb_flush_mm(mm);
130 }
131
132 pgd = mm->pgd;
133 mm->pgd = (pgd_t *) (pgd_val(*pgd) & _REGION_ENTRY_ORIGIN);
134 mm->context.asce_limit = _REGION3_SIZE;
135 mm->context.asce = __pa(mm->pgd) | _ASCE_TABLE_LENGTH |
136 _ASCE_USER_BITS | _ASCE_TYPE_SEGMENT;
137 crst_table_free(mm, (unsigned long *) pgd);
138
139 if (current->active_mm == mm)
140 set_user_asce(mm);
141}
142
143static inline unsigned int atomic_xor_bits(atomic_t *v, unsigned int bits)
144{
145 unsigned int old, new;
146
147 do {
148 old = atomic_read(v);
149 new = old ^ bits;
150 } while (atomic_cmpxchg(v, old, new) != old);
151 return new;
152}
153
154#ifdef CONFIG_PGSTE
155
156struct page *page_table_alloc_pgste(struct mm_struct *mm)
157{
158 struct page *page;
159 u64 *table;
160
161 page = alloc_page(GFP_KERNEL);
162 if (page) {
163 table = (u64 *)page_to_phys(page);
164 memset64(table, _PAGE_INVALID, PTRS_PER_PTE);
165 memset64(table + PTRS_PER_PTE, 0, PTRS_PER_PTE);
166 }
167 return page;
168}
169
170void page_table_free_pgste(struct page *page)
171{
172 __free_page(page);
173}
174
175#endif /* CONFIG_PGSTE */
176
177/*
178 * page table entry allocation/free routines.
179 */
180unsigned long *page_table_alloc(struct mm_struct *mm)
181{
182 unsigned long *table;
183 struct page *page;
184 unsigned int mask, bit;
185
186 /* Try to get a fragment of a 4K page as a 2K page table */
187 if (!mm_alloc_pgste(mm)) {
188 table = NULL;
189 spin_lock_bh(&mm->context.lock);
190 if (!list_empty(&mm->context.pgtable_list)) {
191 page = list_first_entry(&mm->context.pgtable_list,
192 struct page, lru);
193 mask = atomic_read(&page->_mapcount);
194 mask = (mask | (mask >> 4)) & 3;
195 if (mask != 3) {
196 table = (unsigned long *) page_to_phys(page);
197 bit = mask & 1; /* =1 -> second 2K */
198 if (bit)
199 table += PTRS_PER_PTE;
200 atomic_xor_bits(&page->_mapcount, 1U << bit);
201 list_del(&page->lru);
202 }
203 }
204 spin_unlock_bh(&mm->context.lock);
205 if (table)
206 return table;
207 }
208 /* Allocate a fresh page */
209 page = alloc_page(GFP_KERNEL);
210 if (!page)
211 return NULL;
212 if (!pgtable_page_ctor(page)) {
213 __free_page(page);
214 return NULL;
215 }
216 arch_set_page_dat(page, 0);
217 /* Initialize page table */
218 table = (unsigned long *) page_to_phys(page);
219 if (mm_alloc_pgste(mm)) {
220 /* Return 4K page table with PGSTEs */
221 atomic_set(&page->_mapcount, 3);
222 memset64((u64 *)table, _PAGE_INVALID, PTRS_PER_PTE);
223 memset64((u64 *)table + PTRS_PER_PTE, 0, PTRS_PER_PTE);
224 } else {
225 /* Return the first 2K fragment of the page */
226 atomic_set(&page->_mapcount, 1);
227 memset64((u64 *)table, _PAGE_INVALID, 2 * PTRS_PER_PTE);
228 spin_lock_bh(&mm->context.lock);
229 list_add(&page->lru, &mm->context.pgtable_list);
230 spin_unlock_bh(&mm->context.lock);
231 }
232 return table;
233}
234
235void page_table_free(struct mm_struct *mm, unsigned long *table)
236{
237 struct page *page;
238 unsigned int bit, mask;
239
240 page = pfn_to_page(__pa(table) >> PAGE_SHIFT);
241 if (!mm_alloc_pgste(mm)) {
242 /* Free 2K page table fragment of a 4K page */
243 bit = (__pa(table) & ~PAGE_MASK)/(PTRS_PER_PTE*sizeof(pte_t));
244 spin_lock_bh(&mm->context.lock);
245 mask = atomic_xor_bits(&page->_mapcount, 1U << bit);
246 if (mask & 3)
247 list_add(&page->lru, &mm->context.pgtable_list);
248 else
249 list_del(&page->lru);
250 spin_unlock_bh(&mm->context.lock);
251 if (mask != 0)
252 return;
253 }
254
255 pgtable_page_dtor(page);
256 atomic_set(&page->_mapcount, -1);
257 __free_page(page);
258}
259
260void page_table_free_rcu(struct mmu_gather *tlb, unsigned long *table,
261 unsigned long vmaddr)
262{
263 struct mm_struct *mm;
264 struct page *page;
265 unsigned int bit, mask;
266
267 mm = tlb->mm;
268 page = pfn_to_page(__pa(table) >> PAGE_SHIFT);
269 if (mm_alloc_pgste(mm)) {
270 gmap_unlink(mm, table, vmaddr);
271 table = (unsigned long *) (__pa(table) | 3);
272 tlb_remove_table(tlb, table);
273 return;
274 }
275 bit = (__pa(table) & ~PAGE_MASK) / (PTRS_PER_PTE*sizeof(pte_t));
276 spin_lock_bh(&mm->context.lock);
277 mask = atomic_xor_bits(&page->_mapcount, 0x11U << bit);
278 if (mask & 3)
279 list_add_tail(&page->lru, &mm->context.pgtable_list);
280 else
281 list_del(&page->lru);
282 spin_unlock_bh(&mm->context.lock);
283 table = (unsigned long *) (__pa(table) | (1U << bit));
284 tlb_remove_table(tlb, table);
285}
286
287static void __tlb_remove_table(void *_table)
288{
289 unsigned int mask = (unsigned long) _table & 3;
290 void *table = (void *)((unsigned long) _table ^ mask);
291 struct page *page = pfn_to_page(__pa(table) >> PAGE_SHIFT);
292
293 switch (mask) {
294 case 0: /* pmd, pud, or p4d */
295 free_pages((unsigned long) table, 2);
296 break;
297 case 1: /* lower 2K of a 4K page table */
298 case 2: /* higher 2K of a 4K page table */
299 if (atomic_xor_bits(&page->_mapcount, mask << 4) != 0)
300 break;
301 /* fallthrough */
302 case 3: /* 4K page table with pgstes */
303 pgtable_page_dtor(page);
304 atomic_set(&page->_mapcount, -1);
305 __free_page(page);
306 break;
307 }
308}
309
310static void tlb_remove_table_smp_sync(void *arg)
311{
312 /* Simply deliver the interrupt */
313}
314
315static void tlb_remove_table_one(void *table)
316{
317 /*
318 * This isn't an RCU grace period and hence the page-tables cannot be
319 * assumed to be actually RCU-freed.
320 *
321 * It is however sufficient for software page-table walkers that rely
322 * on IRQ disabling. See the comment near struct mmu_table_batch.
323 */
324 smp_call_function(tlb_remove_table_smp_sync, NULL, 1);
325 __tlb_remove_table(table);
326}
327
328static void tlb_remove_table_rcu(struct rcu_head *head)
329{
330 struct mmu_table_batch *batch;
331 int i;
332
333 batch = container_of(head, struct mmu_table_batch, rcu);
334
335 for (i = 0; i < batch->nr; i++)
336 __tlb_remove_table(batch->tables[i]);
337
338 free_page((unsigned long)batch);
339}
340
341void tlb_table_flush(struct mmu_gather *tlb)
342{
343 struct mmu_table_batch **batch = &tlb->batch;
344
345 if (*batch) {
346 call_rcu_sched(&(*batch)->rcu, tlb_remove_table_rcu);
347 *batch = NULL;
348 }
349}
350
351void tlb_remove_table(struct mmu_gather *tlb, void *table)
352{
353 struct mmu_table_batch **batch = &tlb->batch;
354
355 tlb->mm->context.flush_mm = 1;
356 if (*batch == NULL) {
357 *batch = (struct mmu_table_batch *)
358 __get_free_page(GFP_NOWAIT | __GFP_NOWARN);
359 if (*batch == NULL) {
360 __tlb_flush_mm_lazy(tlb->mm);
361 tlb_remove_table_one(table);
362 return;
363 }
364 (*batch)->nr = 0;
365 }
366 (*batch)->tables[(*batch)->nr++] = table;
367 if ((*batch)->nr == MAX_TABLE_BATCH)
368 tlb_flush_mmu(tlb);
369}
370
371/*
372 * Base infrastructure required to generate basic asces, region, segment,
373 * and page tables that do not make use of enhanced features like EDAT1.
374 */
375
376static struct kmem_cache *base_pgt_cache;
377
378static unsigned long base_pgt_alloc(void)
379{
380 u64 *table;
381
382 table = kmem_cache_alloc(base_pgt_cache, GFP_KERNEL);
383 if (table)
384 memset64(table, _PAGE_INVALID, PTRS_PER_PTE);
385 return (unsigned long) table;
386}
387
388static void base_pgt_free(unsigned long table)
389{
390 kmem_cache_free(base_pgt_cache, (void *) table);
391}
392
393static unsigned long base_crst_alloc(unsigned long val)
394{
395 unsigned long table;
396
397 table = __get_free_pages(GFP_KERNEL, CRST_ALLOC_ORDER);
398 if (table)
399 crst_table_init((unsigned long *)table, val);
400 return table;
401}
402
403static void base_crst_free(unsigned long table)
404{
405 free_pages(table, CRST_ALLOC_ORDER);
406}
407
408#define BASE_ADDR_END_FUNC(NAME, SIZE) \
409static inline unsigned long base_##NAME##_addr_end(unsigned long addr, \
410 unsigned long end) \
411{ \
412 unsigned long next = (addr + (SIZE)) & ~((SIZE) - 1); \
413 \
414 return (next - 1) < (end - 1) ? next : end; \
415}
416
417BASE_ADDR_END_FUNC(page, _PAGE_SIZE)
418BASE_ADDR_END_FUNC(segment, _SEGMENT_SIZE)
419BASE_ADDR_END_FUNC(region3, _REGION3_SIZE)
420BASE_ADDR_END_FUNC(region2, _REGION2_SIZE)
421BASE_ADDR_END_FUNC(region1, _REGION1_SIZE)
422
423static inline unsigned long base_lra(unsigned long address)
424{
425 unsigned long real;
426
427 asm volatile(
428 " lra %0,0(%1)\n"
429 : "=d" (real) : "a" (address) : "cc");
430 return real;
431}
432
433static int base_page_walk(unsigned long origin, unsigned long addr,
434 unsigned long end, int alloc)
435{
436 unsigned long *pte, next;
437
438 if (!alloc)
439 return 0;
440 pte = (unsigned long *) origin;
441 pte += (addr & _PAGE_INDEX) >> _PAGE_SHIFT;
442 do {
443 next = base_page_addr_end(addr, end);
444 *pte = base_lra(addr);
445 } while (pte++, addr = next, addr < end);
446 return 0;
447}
448
449static int base_segment_walk(unsigned long origin, unsigned long addr,
450 unsigned long end, int alloc)
451{
452 unsigned long *ste, next, table;
453 int rc;
454
455 ste = (unsigned long *) origin;
456 ste += (addr & _SEGMENT_INDEX) >> _SEGMENT_SHIFT;
457 do {
458 next = base_segment_addr_end(addr, end);
459 if (*ste & _SEGMENT_ENTRY_INVALID) {
460 if (!alloc)
461 continue;
462 table = base_pgt_alloc();
463 if (!table)
464 return -ENOMEM;
465 *ste = table | _SEGMENT_ENTRY;
466 }
467 table = *ste & _SEGMENT_ENTRY_ORIGIN;
468 rc = base_page_walk(table, addr, next, alloc);
469 if (rc)
470 return rc;
471 if (!alloc)
472 base_pgt_free(table);
473 cond_resched();
474 } while (ste++, addr = next, addr < end);
475 return 0;
476}
477
478static int base_region3_walk(unsigned long origin, unsigned long addr,
479 unsigned long end, int alloc)
480{
481 unsigned long *rtte, next, table;
482 int rc;
483
484 rtte = (unsigned long *) origin;
485 rtte += (addr & _REGION3_INDEX) >> _REGION3_SHIFT;
486 do {
487 next = base_region3_addr_end(addr, end);
488 if (*rtte & _REGION_ENTRY_INVALID) {
489 if (!alloc)
490 continue;
491 table = base_crst_alloc(_SEGMENT_ENTRY_EMPTY);
492 if (!table)
493 return -ENOMEM;
494 *rtte = table | _REGION3_ENTRY;
495 }
496 table = *rtte & _REGION_ENTRY_ORIGIN;
497 rc = base_segment_walk(table, addr, next, alloc);
498 if (rc)
499 return rc;
500 if (!alloc)
501 base_crst_free(table);
502 } while (rtte++, addr = next, addr < end);
503 return 0;
504}
505
506static int base_region2_walk(unsigned long origin, unsigned long addr,
507 unsigned long end, int alloc)
508{
509 unsigned long *rste, next, table;
510 int rc;
511
512 rste = (unsigned long *) origin;
513 rste += (addr & _REGION2_INDEX) >> _REGION2_SHIFT;
514 do {
515 next = base_region2_addr_end(addr, end);
516 if (*rste & _REGION_ENTRY_INVALID) {
517 if (!alloc)
518 continue;
519 table = base_crst_alloc(_REGION3_ENTRY_EMPTY);
520 if (!table)
521 return -ENOMEM;
522 *rste = table | _REGION2_ENTRY;
523 }
524 table = *rste & _REGION_ENTRY_ORIGIN;
525 rc = base_region3_walk(table, addr, next, alloc);
526 if (rc)
527 return rc;
528 if (!alloc)
529 base_crst_free(table);
530 } while (rste++, addr = next, addr < end);
531 return 0;
532}
533
534static int base_region1_walk(unsigned long origin, unsigned long addr,
535 unsigned long end, int alloc)
536{
537 unsigned long *rfte, next, table;
538 int rc;
539
540 rfte = (unsigned long *) origin;
541 rfte += (addr & _REGION1_INDEX) >> _REGION1_SHIFT;
542 do {
543 next = base_region1_addr_end(addr, end);
544 if (*rfte & _REGION_ENTRY_INVALID) {
545 if (!alloc)
546 continue;
547 table = base_crst_alloc(_REGION2_ENTRY_EMPTY);
548 if (!table)
549 return -ENOMEM;
550 *rfte = table | _REGION1_ENTRY;
551 }
552 table = *rfte & _REGION_ENTRY_ORIGIN;
553 rc = base_region2_walk(table, addr, next, alloc);
554 if (rc)
555 return rc;
556 if (!alloc)
557 base_crst_free(table);
558 } while (rfte++, addr = next, addr < end);
559 return 0;
560}
561
562/**
563 * base_asce_free - free asce and tables returned from base_asce_alloc()
564 * @asce: asce to be freed
565 *
566 * Frees all region, segment, and page tables that were allocated with a
567 * corresponding base_asce_alloc() call.
568 */
569void base_asce_free(unsigned long asce)
570{
571 unsigned long table = asce & _ASCE_ORIGIN;
572
573 if (!asce)
574 return;
575 switch (asce & _ASCE_TYPE_MASK) {
576 case _ASCE_TYPE_SEGMENT:
577 base_segment_walk(table, 0, _REGION3_SIZE, 0);
578 break;
579 case _ASCE_TYPE_REGION3:
580 base_region3_walk(table, 0, _REGION2_SIZE, 0);
581 break;
582 case _ASCE_TYPE_REGION2:
583 base_region2_walk(table, 0, _REGION1_SIZE, 0);
584 break;
585 case _ASCE_TYPE_REGION1:
586 base_region1_walk(table, 0, -_PAGE_SIZE, 0);
587 break;
588 }
589 base_crst_free(table);
590}
591
592static int base_pgt_cache_init(void)
593{
594 static DEFINE_MUTEX(base_pgt_cache_mutex);
595 unsigned long sz = _PAGE_TABLE_SIZE;
596
597 if (base_pgt_cache)
598 return 0;
599 mutex_lock(&base_pgt_cache_mutex);
600 if (!base_pgt_cache)
601 base_pgt_cache = kmem_cache_create("base_pgt", sz, sz, 0, NULL);
602 mutex_unlock(&base_pgt_cache_mutex);
603 return base_pgt_cache ? 0 : -ENOMEM;
604}
605
606/**
607 * base_asce_alloc - create kernel mapping without enhanced DAT features
608 * @addr: virtual start address of kernel mapping
609 * @num_pages: number of consecutive pages
610 *
611 * Generate an asce, including all required region, segment and page tables,
612 * that can be used to access the virtual kernel mapping. The difference is
613 * that the returned asce does not make use of any enhanced DAT features like
614 * e.g. large pages. This is required for some I/O functions that pass an
615 * asce, like e.g. some service call requests.
616 *
617 * Note: the returned asce may NEVER be attached to any cpu. It may only be
618 * used for I/O requests. tlb entries that might result because the
619 * asce was attached to a cpu won't be cleared.
620 */
621unsigned long base_asce_alloc(unsigned long addr, unsigned long num_pages)
622{
623 unsigned long asce, table, end;
624 int rc;
625
626 if (base_pgt_cache_init())
627 return 0;
628 end = addr + num_pages * PAGE_SIZE;
629 if (end <= _REGION3_SIZE) {
630 table = base_crst_alloc(_SEGMENT_ENTRY_EMPTY);
631 if (!table)
632 return 0;
633 rc = base_segment_walk(table, addr, end, 1);
634 asce = table | _ASCE_TYPE_SEGMENT | _ASCE_TABLE_LENGTH;
635 } else if (end <= _REGION2_SIZE) {
636 table = base_crst_alloc(_REGION3_ENTRY_EMPTY);
637 if (!table)
638 return 0;
639 rc = base_region3_walk(table, addr, end, 1);
640 asce = table | _ASCE_TYPE_REGION3 | _ASCE_TABLE_LENGTH;
641 } else if (end <= _REGION1_SIZE) {
642 table = base_crst_alloc(_REGION2_ENTRY_EMPTY);
643 if (!table)
644 return 0;
645 rc = base_region2_walk(table, addr, end, 1);
646 asce = table | _ASCE_TYPE_REGION2 | _ASCE_TABLE_LENGTH;
647 } else {
648 table = base_crst_alloc(_REGION1_ENTRY_EMPTY);
649 if (!table)
650 return 0;
651 rc = base_region1_walk(table, addr, end, 1);
652 asce = table | _ASCE_TYPE_REGION1 | _ASCE_TABLE_LENGTH;
653 }
654 if (rc) {
655 base_asce_free(asce);
656 asce = 0;
657 }
658 return asce;
659}