xref: /openbmc/linux/arch/s390/mm/pgalloc.c (revision 65417d9f)
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/mm.h>
10 #include <linux/sysctl.h>
11 #include <asm/mmu_context.h>
12 #include <asm/pgalloc.h>
13 #include <asm/gmap.h>
14 #include <asm/tlb.h>
15 #include <asm/tlbflush.h>
16 
17 #ifdef CONFIG_PGSTE
18 
19 static int page_table_allocate_pgste_min = 0;
20 static int page_table_allocate_pgste_max = 1;
21 int page_table_allocate_pgste = 0;
22 EXPORT_SYMBOL(page_table_allocate_pgste);
23 
24 static 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,
31 		.extra1		= &page_table_allocate_pgste_min,
32 		.extra2		= &page_table_allocate_pgste_max,
33 	},
34 	{ }
35 };
36 
37 static struct ctl_table page_table_sysctl_dir[] = {
38 	{
39 		.procname	= "vm",
40 		.maxlen		= 0,
41 		.mode		= 0555,
42 		.child		= page_table_sysctl,
43 	},
44 	{ }
45 };
46 
47 static int __init page_table_register_sysctl(void)
48 {
49 	return register_sysctl_table(page_table_sysctl_dir) ? 0 : -ENOMEM;
50 }
51 __initcall(page_table_register_sysctl);
52 
53 #endif /* CONFIG_PGSTE */
54 
55 unsigned long *crst_table_alloc(struct mm_struct *mm)
56 {
57 	struct page *page = alloc_pages(GFP_KERNEL, 2);
58 
59 	if (!page)
60 		return NULL;
61 	arch_set_page_dat(page, 2);
62 	return (unsigned long *) page_to_phys(page);
63 }
64 
65 void crst_table_free(struct mm_struct *mm, unsigned long *table)
66 {
67 	free_pages((unsigned long) table, 2);
68 }
69 
70 static void __crst_table_upgrade(void *arg)
71 {
72 	struct mm_struct *mm = arg;
73 
74 	if (current->active_mm == mm)
75 		set_user_asce(mm);
76 	__tlb_flush_local();
77 }
78 
79 int crst_table_upgrade(struct mm_struct *mm, unsigned long end)
80 {
81 	unsigned long *table, *pgd;
82 	int rc, notify;
83 
84 	/* upgrade should only happen from 3 to 4, 3 to 5, or 4 to 5 levels */
85 	VM_BUG_ON(mm->context.asce_limit < _REGION2_SIZE);
86 	if (end >= TASK_SIZE_MAX)
87 		return -ENOMEM;
88 	rc = 0;
89 	notify = 0;
90 	while (mm->context.asce_limit < end) {
91 		table = crst_table_alloc(mm);
92 		if (!table) {
93 			rc = -ENOMEM;
94 			break;
95 		}
96 		spin_lock_bh(&mm->page_table_lock);
97 		pgd = (unsigned long *) mm->pgd;
98 		if (mm->context.asce_limit == _REGION2_SIZE) {
99 			crst_table_init(table, _REGION2_ENTRY_EMPTY);
100 			p4d_populate(mm, (p4d_t *) table, (pud_t *) pgd);
101 			mm->pgd = (pgd_t *) table;
102 			mm->context.asce_limit = _REGION1_SIZE;
103 			mm->context.asce = __pa(mm->pgd) | _ASCE_TABLE_LENGTH |
104 				_ASCE_USER_BITS | _ASCE_TYPE_REGION2;
105 		} else {
106 			crst_table_init(table, _REGION1_ENTRY_EMPTY);
107 			pgd_populate(mm, (pgd_t *) table, (p4d_t *) pgd);
108 			mm->pgd = (pgd_t *) table;
109 			mm->context.asce_limit = -PAGE_SIZE;
110 			mm->context.asce = __pa(mm->pgd) | _ASCE_TABLE_LENGTH |
111 				_ASCE_USER_BITS | _ASCE_TYPE_REGION1;
112 		}
113 		notify = 1;
114 		spin_unlock_bh(&mm->page_table_lock);
115 	}
116 	if (notify)
117 		on_each_cpu(__crst_table_upgrade, mm, 0);
118 	return rc;
119 }
120 
121 void crst_table_downgrade(struct mm_struct *mm)
122 {
123 	pgd_t *pgd;
124 
125 	/* downgrade should only happen from 3 to 2 levels (compat only) */
126 	VM_BUG_ON(mm->context.asce_limit != _REGION2_SIZE);
127 
128 	if (current->active_mm == mm) {
129 		clear_user_asce();
130 		__tlb_flush_mm(mm);
131 	}
132 
133 	pgd = mm->pgd;
134 	mm->pgd = (pgd_t *) (pgd_val(*pgd) & _REGION_ENTRY_ORIGIN);
135 	mm->context.asce_limit = _REGION3_SIZE;
136 	mm->context.asce = __pa(mm->pgd) | _ASCE_TABLE_LENGTH |
137 			   _ASCE_USER_BITS | _ASCE_TYPE_SEGMENT;
138 	crst_table_free(mm, (unsigned long *) pgd);
139 
140 	if (current->active_mm == mm)
141 		set_user_asce(mm);
142 }
143 
144 static inline unsigned int atomic_xor_bits(atomic_t *v, unsigned int bits)
145 {
146 	unsigned int old, new;
147 
148 	do {
149 		old = atomic_read(v);
150 		new = old ^ bits;
151 	} while (atomic_cmpxchg(v, old, new) != old);
152 	return new;
153 }
154 
155 #ifdef CONFIG_PGSTE
156 
157 struct page *page_table_alloc_pgste(struct mm_struct *mm)
158 {
159 	struct page *page;
160 	u64 *table;
161 
162 	page = alloc_page(GFP_KERNEL);
163 	if (page) {
164 		table = (u64 *)page_to_phys(page);
165 		memset64(table, _PAGE_INVALID, PTRS_PER_PTE);
166 		memset64(table + PTRS_PER_PTE, 0, PTRS_PER_PTE);
167 	}
168 	return page;
169 }
170 
171 void page_table_free_pgste(struct page *page)
172 {
173 	__free_page(page);
174 }
175 
176 #endif /* CONFIG_PGSTE */
177 
178 /*
179  * page table entry allocation/free routines.
180  */
181 unsigned long *page_table_alloc(struct mm_struct *mm)
182 {
183 	unsigned long *table;
184 	struct page *page;
185 	unsigned int mask, bit;
186 
187 	/* Try to get a fragment of a 4K page as a 2K page table */
188 	if (!mm_alloc_pgste(mm)) {
189 		table = NULL;
190 		spin_lock_bh(&mm->context.lock);
191 		if (!list_empty(&mm->context.pgtable_list)) {
192 			page = list_first_entry(&mm->context.pgtable_list,
193 						struct page, lru);
194 			mask = atomic_read(&page->_mapcount);
195 			mask = (mask | (mask >> 4)) & 3;
196 			if (mask != 3) {
197 				table = (unsigned long *) page_to_phys(page);
198 				bit = mask & 1;		/* =1 -> second 2K */
199 				if (bit)
200 					table += PTRS_PER_PTE;
201 				atomic_xor_bits(&page->_mapcount, 1U << bit);
202 				list_del(&page->lru);
203 			}
204 		}
205 		spin_unlock_bh(&mm->context.lock);
206 		if (table)
207 			return table;
208 	}
209 	/* Allocate a fresh page */
210 	page = alloc_page(GFP_KERNEL);
211 	if (!page)
212 		return NULL;
213 	if (!pgtable_page_ctor(page)) {
214 		__free_page(page);
215 		return NULL;
216 	}
217 	arch_set_page_dat(page, 0);
218 	/* Initialize page table */
219 	table = (unsigned long *) page_to_phys(page);
220 	if (mm_alloc_pgste(mm)) {
221 		/* Return 4K page table with PGSTEs */
222 		atomic_set(&page->_mapcount, 3);
223 		memset64((u64 *)table, _PAGE_INVALID, PTRS_PER_PTE);
224 		memset64((u64 *)table + PTRS_PER_PTE, 0, PTRS_PER_PTE);
225 	} else {
226 		/* Return the first 2K fragment of the page */
227 		atomic_set(&page->_mapcount, 1);
228 		memset64((u64 *)table, _PAGE_INVALID, 2 * PTRS_PER_PTE);
229 		spin_lock_bh(&mm->context.lock);
230 		list_add(&page->lru, &mm->context.pgtable_list);
231 		spin_unlock_bh(&mm->context.lock);
232 	}
233 	return table;
234 }
235 
236 void page_table_free(struct mm_struct *mm, unsigned long *table)
237 {
238 	struct page *page;
239 	unsigned int bit, mask;
240 
241 	page = pfn_to_page(__pa(table) >> PAGE_SHIFT);
242 	if (!mm_alloc_pgste(mm)) {
243 		/* Free 2K page table fragment of a 4K page */
244 		bit = (__pa(table) & ~PAGE_MASK)/(PTRS_PER_PTE*sizeof(pte_t));
245 		spin_lock_bh(&mm->context.lock);
246 		mask = atomic_xor_bits(&page->_mapcount, 1U << bit);
247 		if (mask & 3)
248 			list_add(&page->lru, &mm->context.pgtable_list);
249 		else
250 			list_del(&page->lru);
251 		spin_unlock_bh(&mm->context.lock);
252 		if (mask != 0)
253 			return;
254 	}
255 
256 	pgtable_page_dtor(page);
257 	atomic_set(&page->_mapcount, -1);
258 	__free_page(page);
259 }
260 
261 void page_table_free_rcu(struct mmu_gather *tlb, unsigned long *table,
262 			 unsigned long vmaddr)
263 {
264 	struct mm_struct *mm;
265 	struct page *page;
266 	unsigned int bit, mask;
267 
268 	mm = tlb->mm;
269 	page = pfn_to_page(__pa(table) >> PAGE_SHIFT);
270 	if (mm_alloc_pgste(mm)) {
271 		gmap_unlink(mm, table, vmaddr);
272 		table = (unsigned long *) (__pa(table) | 3);
273 		tlb_remove_table(tlb, table);
274 		return;
275 	}
276 	bit = (__pa(table) & ~PAGE_MASK) / (PTRS_PER_PTE*sizeof(pte_t));
277 	spin_lock_bh(&mm->context.lock);
278 	mask = atomic_xor_bits(&page->_mapcount, 0x11U << bit);
279 	if (mask & 3)
280 		list_add_tail(&page->lru, &mm->context.pgtable_list);
281 	else
282 		list_del(&page->lru);
283 	spin_unlock_bh(&mm->context.lock);
284 	table = (unsigned long *) (__pa(table) | (1U << bit));
285 	tlb_remove_table(tlb, table);
286 }
287 
288 static void __tlb_remove_table(void *_table)
289 {
290 	unsigned int mask = (unsigned long) _table & 3;
291 	void *table = (void *)((unsigned long) _table ^ mask);
292 	struct page *page = pfn_to_page(__pa(table) >> PAGE_SHIFT);
293 
294 	switch (mask) {
295 	case 0:		/* pmd, pud, or p4d */
296 		free_pages((unsigned long) table, 2);
297 		break;
298 	case 1:		/* lower 2K of a 4K page table */
299 	case 2:		/* higher 2K of a 4K page table */
300 		if (atomic_xor_bits(&page->_mapcount, mask << 4) != 0)
301 			break;
302 		/* fallthrough */
303 	case 3:		/* 4K page table with pgstes */
304 		pgtable_page_dtor(page);
305 		atomic_set(&page->_mapcount, -1);
306 		__free_page(page);
307 		break;
308 	}
309 }
310 
311 static void tlb_remove_table_smp_sync(void *arg)
312 {
313 	/* Simply deliver the interrupt */
314 }
315 
316 static void tlb_remove_table_one(void *table)
317 {
318 	/*
319 	 * This isn't an RCU grace period and hence the page-tables cannot be
320 	 * assumed to be actually RCU-freed.
321 	 *
322 	 * It is however sufficient for software page-table walkers that rely
323 	 * on IRQ disabling. See the comment near struct mmu_table_batch.
324 	 */
325 	smp_call_function(tlb_remove_table_smp_sync, NULL, 1);
326 	__tlb_remove_table(table);
327 }
328 
329 static void tlb_remove_table_rcu(struct rcu_head *head)
330 {
331 	struct mmu_table_batch *batch;
332 	int i;
333 
334 	batch = container_of(head, struct mmu_table_batch, rcu);
335 
336 	for (i = 0; i < batch->nr; i++)
337 		__tlb_remove_table(batch->tables[i]);
338 
339 	free_page((unsigned long)batch);
340 }
341 
342 void tlb_table_flush(struct mmu_gather *tlb)
343 {
344 	struct mmu_table_batch **batch = &tlb->batch;
345 
346 	if (*batch) {
347 		call_rcu_sched(&(*batch)->rcu, tlb_remove_table_rcu);
348 		*batch = NULL;
349 	}
350 }
351 
352 void tlb_remove_table(struct mmu_gather *tlb, void *table)
353 {
354 	struct mmu_table_batch **batch = &tlb->batch;
355 
356 	tlb->mm->context.flush_mm = 1;
357 	if (*batch == NULL) {
358 		*batch = (struct mmu_table_batch *)
359 			__get_free_page(GFP_NOWAIT | __GFP_NOWARN);
360 		if (*batch == NULL) {
361 			__tlb_flush_mm_lazy(tlb->mm);
362 			tlb_remove_table_one(table);
363 			return;
364 		}
365 		(*batch)->nr = 0;
366 	}
367 	(*batch)->tables[(*batch)->nr++] = table;
368 	if ((*batch)->nr == MAX_TABLE_BATCH)
369 		tlb_flush_mmu(tlb);
370 }
371