1// SPDX-License-Identifier: GPL-2.0
  2/*
  3 *  Lockless get_user_pages_fast for s390
  4 *
  5 *  Copyright IBM Corp. 2010
  6 *  Author(s): Martin Schwidefsky <schwidefsky@de.ibm.com>
  7 */
  8#include <linux/sched.h>
  9#include <linux/mm.h>
 10#include <linux/hugetlb.h>
 11#include <linux/vmstat.h>
 12#include <linux/pagemap.h>
 13#include <linux/rwsem.h>
 14#include <asm/pgtable.h>
 15
 16/*
 17 * The performance critical leaf functions are made noinline otherwise gcc
 18 * inlines everything into a single function which results in too much
 19 * register pressure.
 20 */
 21static inline int gup_pte_range(pmd_t *pmdp, pmd_t pmd, unsigned long addr,
 22		unsigned long end, int write, struct page **pages, int *nr)
 23{
 24	struct page *head, *page;
 25	unsigned long mask;
 26	pte_t *ptep, pte;
 27
 28	mask = (write ? _PAGE_PROTECT : 0) | _PAGE_INVALID | _PAGE_SPECIAL;
 29
 30	ptep = ((pte_t *) pmd_deref(pmd)) + pte_index(addr);
 31	do {
 32		pte = *ptep;
 33		barrier();
 34		/* Similar to the PMD case, NUMA hinting must take slow path */
 35		if (pte_protnone(pte))
 36			return 0;
 37		if ((pte_val(pte) & mask) != 0)
 38			return 0;
 39		VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
 40		page = pte_page(pte);
 41		head = compound_head(page);
 42		if (!page_cache_get_speculative(head))
 43			return 0;
 44		if (unlikely(pte_val(pte) != pte_val(*ptep))) {
 45			put_page(head);
 46			return 0;
 47		}
 48		VM_BUG_ON_PAGE(compound_head(page) != head, page);
 49		pages[*nr] = page;
 50		(*nr)++;
 51
 52	} while (ptep++, addr += PAGE_SIZE, addr != end);
 53
 54	return 1;
 55}
 56
 57static inline int gup_huge_pmd(pmd_t *pmdp, pmd_t pmd, unsigned long addr,
 58		unsigned long end, int write, struct page **pages, int *nr)
 59{
 60	struct page *head, *page;
 61	unsigned long mask;
 62	int refs;
 63
 64	mask = (write ? _SEGMENT_ENTRY_PROTECT : 0) | _SEGMENT_ENTRY_INVALID;
 65	if ((pmd_val(pmd) & mask) != 0)
 66		return 0;
 67	VM_BUG_ON(!pfn_valid(pmd_val(pmd) >> PAGE_SHIFT));
 68
 69	refs = 0;
 70	head = pmd_page(pmd);
 71	page = head + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
 72	do {
 73		VM_BUG_ON(compound_head(page) != head);
 74		pages[*nr] = page;
 75		(*nr)++;
 76		page++;
 77		refs++;
 78	} while (addr += PAGE_SIZE, addr != end);
 79
 80	if (!page_cache_add_speculative(head, refs)) {
 81		*nr -= refs;
 82		return 0;
 83	}
 84
 85	if (unlikely(pmd_val(pmd) != pmd_val(*pmdp))) {
 86		*nr -= refs;
 87		while (refs--)
 88			put_page(head);
 89		return 0;
 90	}
 91
 92	return 1;
 93}
 94
 95
 96static inline int gup_pmd_range(pud_t *pudp, pud_t pud, unsigned long addr,
 97		unsigned long end, int write, struct page **pages, int *nr)
 98{
 99	unsigned long next;
100	pmd_t *pmdp, pmd;
101
102	pmdp = (pmd_t *) pudp;
103	if ((pud_val(pud) & _REGION_ENTRY_TYPE_MASK) == _REGION_ENTRY_TYPE_R3)
104		pmdp = (pmd_t *) pud_deref(pud);
105	pmdp += pmd_index(addr);
106	do {
107		pmd = *pmdp;
108		barrier();
109		next = pmd_addr_end(addr, end);
110		if (pmd_none(pmd))
111			return 0;
112		if (unlikely(pmd_large(pmd))) {
113			/*
114			 * NUMA hinting faults need to be handled in the GUP
115			 * slowpath for accounting purposes and so that they
116			 * can be serialised against THP migration.
117			 */
118			if (pmd_protnone(pmd))
119				return 0;
120			if (!gup_huge_pmd(pmdp, pmd, addr, next,
121					  write, pages, nr))
122				return 0;
123		} else if (!gup_pte_range(pmdp, pmd, addr, next,
124					  write, pages, nr))
125			return 0;
126	} while (pmdp++, addr = next, addr != end);
127
128	return 1;
129}
130
131static int gup_huge_pud(pud_t *pudp, pud_t pud, unsigned long addr,
132		unsigned long end, int write, struct page **pages, int *nr)
133{
134	struct page *head, *page;
135	unsigned long mask;
136	int refs;
137
138	mask = (write ? _REGION_ENTRY_PROTECT : 0) | _REGION_ENTRY_INVALID;
139	if ((pud_val(pud) & mask) != 0)
140		return 0;
141	VM_BUG_ON(!pfn_valid(pud_pfn(pud)));
142
143	refs = 0;
144	head = pud_page(pud);
145	page = head + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
146	do {
147		VM_BUG_ON_PAGE(compound_head(page) != head, page);
148		pages[*nr] = page;
149		(*nr)++;
150		page++;
151		refs++;
152	} while (addr += PAGE_SIZE, addr != end);
153
154	if (!page_cache_add_speculative(head, refs)) {
155		*nr -= refs;
156		return 0;
157	}
158
159	if (unlikely(pud_val(pud) != pud_val(*pudp))) {
160		*nr -= refs;
161		while (refs--)
162			put_page(head);
163		return 0;
164	}
165
166	return 1;
167}
168
169static inline int gup_pud_range(p4d_t *p4dp, p4d_t p4d, unsigned long addr,
170		unsigned long end, int write, struct page **pages, int *nr)
171{
172	unsigned long next;
173	pud_t *pudp, pud;
174
175	pudp = (pud_t *) p4dp;
176	if ((p4d_val(p4d) & _REGION_ENTRY_TYPE_MASK) == _REGION_ENTRY_TYPE_R2)
177		pudp = (pud_t *) p4d_deref(p4d);
178	pudp += pud_index(addr);
179	do {
180		pud = *pudp;
181		barrier();
182		next = pud_addr_end(addr, end);
183		if (pud_none(pud))
184			return 0;
185		if (unlikely(pud_large(pud))) {
186			if (!gup_huge_pud(pudp, pud, addr, next, write, pages,
187					  nr))
188				return 0;
189		} else if (!gup_pmd_range(pudp, pud, addr, next, write, pages,
190					  nr))
191			return 0;
192	} while (pudp++, addr = next, addr != end);
193
194	return 1;
195}
196
197static inline int gup_p4d_range(pgd_t *pgdp, pgd_t pgd, unsigned long addr,
198		unsigned long end, int write, struct page **pages, int *nr)
199{
200	unsigned long next;
201	p4d_t *p4dp, p4d;
202
203	p4dp = (p4d_t *) pgdp;
204	if ((pgd_val(pgd) & _REGION_ENTRY_TYPE_MASK) == _REGION_ENTRY_TYPE_R1)
205		p4dp = (p4d_t *) pgd_deref(pgd);
206	p4dp += p4d_index(addr);
207	do {
208		p4d = *p4dp;
209		barrier();
210		next = p4d_addr_end(addr, end);
211		if (p4d_none(p4d))
212			return 0;
213		if (!gup_pud_range(p4dp, p4d, addr, next, write, pages, nr))
214			return 0;
215	} while (p4dp++, addr = next, addr != end);
216
217	return 1;
218}
219
220/*
221 * Like get_user_pages_fast() except its IRQ-safe in that it won't fall
222 * back to the regular GUP.
223 * Note a difference with get_user_pages_fast: this always returns the
224 * number of pages pinned, 0 if no pages were pinned.
225 */
226int __get_user_pages_fast(unsigned long start, int nr_pages, int write,
227			  struct page **pages)
228{
229	struct mm_struct *mm = current->mm;
230	unsigned long addr, len, end;
231	unsigned long next, flags;
232	pgd_t *pgdp, pgd;
233	int nr = 0;
234
235	start &= PAGE_MASK;
236	addr = start;
237	len = (unsigned long) nr_pages << PAGE_SHIFT;
238	end = start + len;
239	if ((end <= start) || (end > mm->context.asce_limit))
240		return 0;
241	/*
242	 * local_irq_save() doesn't prevent pagetable teardown, but does
243	 * prevent the pagetables from being freed on s390.
244	 *
245	 * So long as we atomically load page table pointers versus teardown,
246	 * we can follow the address down to the the page and take a ref on it.
247	 */
248	local_irq_save(flags);
249	pgdp = pgd_offset(mm, addr);
250	do {
251		pgd = *pgdp;
252		barrier();
253		next = pgd_addr_end(addr, end);
254		if (pgd_none(pgd))
255			break;
256		if (!gup_p4d_range(pgdp, pgd, addr, next, write, pages, &nr))
257			break;
258	} while (pgdp++, addr = next, addr != end);
259	local_irq_restore(flags);
260
261	return nr;
262}
263
264/**
265 * get_user_pages_fast() - pin user pages in memory
266 * @start:	starting user address
267 * @nr_pages:	number of pages from start to pin
268 * @write:	whether pages will be written to
269 * @pages:	array that receives pointers to the pages pinned.
270 *		Should be at least nr_pages long.
271 *
272 * Attempt to pin user pages in memory without taking mm->mmap_sem.
273 * If not successful, it will fall back to taking the lock and
274 * calling get_user_pages().
275 *
276 * Returns number of pages pinned. This may be fewer than the number
277 * requested. If nr_pages is 0 or negative, returns 0. If no pages
278 * were pinned, returns -errno.
279 */
280int get_user_pages_fast(unsigned long start, int nr_pages, int write,
281			struct page **pages)
282{
283	int nr, ret;
284
285	might_sleep();
286	start &= PAGE_MASK;
287	nr = __get_user_pages_fast(start, nr_pages, write, pages);
288	if (nr == nr_pages)
289		return nr;
290
291	/* Try to get the remaining pages with get_user_pages */
292	start += nr << PAGE_SHIFT;
293	pages += nr;
294	ret = get_user_pages_unlocked(start, nr_pages - nr, pages,
295				      write ? FOLL_WRITE : 0);
296	/* Have to be a bit careful with return values */
297	if (nr > 0)
298		ret = (ret < 0) ? nr : ret + nr;
299	return ret;
300}