1// SPDX-License-Identifier: GPL-2.0
  2/*
  3 * Lockless get_user_pages_fast for MIPS
  4 *
  5 * Copyright (C) 2008 Nick Piggin
  6 * Copyright (C) 2008 Novell Inc.
  7 * Copyright (C) 2011 Ralf Baechle
  8 */
  9#include <linux/sched.h>
 10#include <linux/mm.h>
 11#include <linux/vmstat.h>
 12#include <linux/highmem.h>
 13#include <linux/swap.h>
 14#include <linux/hugetlb.h>
 15
 16#include <asm/cpu-features.h>
 17#include <asm/pgtable.h>
 18
 19static inline pte_t gup_get_pte(pte_t *ptep)
 20{
 21#if defined(CONFIG_PHYS_ADDR_T_64BIT) && defined(CONFIG_CPU_MIPS32)
 22	pte_t pte;
 23
 24retry:
 25	pte.pte_low = ptep->pte_low;
 26	smp_rmb();
 27	pte.pte_high = ptep->pte_high;
 28	smp_rmb();
 29	if (unlikely(pte.pte_low != ptep->pte_low))
 30		goto retry;
 31
 32	return pte;
 33#else
 34	return READ_ONCE(*ptep);
 35#endif
 36}
 37
 38static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end,
 39			int write, struct page **pages, int *nr)
 40{
 41	pte_t *ptep = pte_offset_map(&pmd, addr);
 42	do {
 43		pte_t pte = gup_get_pte(ptep);
 44		struct page *page;
 45
 46		if (!pte_present(pte) ||
 47		    pte_special(pte) || (write && !pte_write(pte))) {
 48			pte_unmap(ptep);
 49			return 0;
 50		}
 51		VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
 52		page = pte_page(pte);
 53		get_page(page);
 54		SetPageReferenced(page);
 55		pages[*nr] = page;
 56		(*nr)++;
 57
 58	} while (ptep++, addr += PAGE_SIZE, addr != end);
 59
 60	pte_unmap(ptep - 1);
 61	return 1;
 62}
 63
 64static inline void get_head_page_multiple(struct page *page, int nr)
 65{
 66	VM_BUG_ON(page != compound_head(page));
 67	VM_BUG_ON(page_count(page) == 0);
 68	page_ref_add(page, nr);
 69	SetPageReferenced(page);
 70}
 71
 72static int gup_huge_pmd(pmd_t pmd, unsigned long addr, unsigned long end,
 73			int write, struct page **pages, int *nr)
 74{
 75	pte_t pte = *(pte_t *)&pmd;
 76	struct page *head, *page;
 77	int refs;
 78
 79	if (write && !pte_write(pte))
 80		return 0;
 81	/* hugepages are never "special" */
 82	VM_BUG_ON(pte_special(pte));
 83	VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
 84
 85	refs = 0;
 86	head = pte_page(pte);
 87	page = head + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
 88	do {
 89		VM_BUG_ON(compound_head(page) != head);
 90		pages[*nr] = page;
 91		(*nr)++;
 92		page++;
 93		refs++;
 94	} while (addr += PAGE_SIZE, addr != end);
 95
 96	get_head_page_multiple(head, refs);
 97	return 1;
 98}
 99
100static int gup_pmd_range(pud_t pud, unsigned long addr, unsigned long end,
101			int write, struct page **pages, int *nr)
102{
103	unsigned long next;
104	pmd_t *pmdp;
105
106	pmdp = pmd_offset(&pud, addr);
107	do {
108		pmd_t pmd = *pmdp;
109
110		next = pmd_addr_end(addr, end);
111		if (pmd_none(pmd))
112			return 0;
113		if (unlikely(pmd_huge(pmd))) {
114			if (!gup_huge_pmd(pmd, addr, next, write, pages,nr))
115				return 0;
116		} else {
117			if (!gup_pte_range(pmd, addr, next, write, pages,nr))
118				return 0;
119		}
120	} while (pmdp++, addr = next, addr != end);
121
122	return 1;
123}
124
125static int gup_huge_pud(pud_t pud, unsigned long addr, unsigned long end,
126			int write, struct page **pages, int *nr)
127{
128	pte_t pte = *(pte_t *)&pud;
129	struct page *head, *page;
130	int refs;
131
132	if (write && !pte_write(pte))
133		return 0;
134	/* hugepages are never "special" */
135	VM_BUG_ON(pte_special(pte));
136	VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
137
138	refs = 0;
139	head = pte_page(pte);
140	page = head + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
141	do {
142		VM_BUG_ON(compound_head(page) != head);
143		pages[*nr] = page;
144		(*nr)++;
145		page++;
146		refs++;
147	} while (addr += PAGE_SIZE, addr != end);
148
149	get_head_page_multiple(head, refs);
150	return 1;
151}
152
153static int gup_pud_range(pgd_t pgd, unsigned long addr, unsigned long end,
154			int write, struct page **pages, int *nr)
155{
156	unsigned long next;
157	pud_t *pudp;
158
159	pudp = pud_offset(&pgd, addr);
160	do {
161		pud_t pud = *pudp;
162
163		next = pud_addr_end(addr, end);
164		if (pud_none(pud))
165			return 0;
166		if (unlikely(pud_huge(pud))) {
167			if (!gup_huge_pud(pud, addr, next, write, pages,nr))
168				return 0;
169		} else {
170			if (!gup_pmd_range(pud, addr, next, write, pages,nr))
171				return 0;
172		}
173	} while (pudp++, addr = next, addr != end);
174
175	return 1;
176}
177
178/*
179 * Like get_user_pages_fast() except its IRQ-safe in that it won't fall
180 * back to the regular GUP.
181 * Note a difference with get_user_pages_fast: this always returns the
182 * number of pages pinned, 0 if no pages were pinned.
183 */
184int __get_user_pages_fast(unsigned long start, int nr_pages, int write,
185			  struct page **pages)
186{
187	struct mm_struct *mm = current->mm;
188	unsigned long addr, len, end;
189	unsigned long next;
190	unsigned long flags;
191	pgd_t *pgdp;
192	int nr = 0;
193
194	start &= PAGE_MASK;
195	addr = start;
196	len = (unsigned long) nr_pages << PAGE_SHIFT;
197	end = start + len;
198	if (unlikely(!access_ok(write ? VERIFY_WRITE : VERIFY_READ,
199					(void __user *)start, len)))
200		return 0;
201
202	/*
203	 * XXX: batch / limit 'nr', to avoid large irq off latency
204	 * needs some instrumenting to determine the common sizes used by
205	 * important workloads (eg. DB2), and whether limiting the batch
206	 * size will decrease performance.
207	 *
208	 * It seems like we're in the clear for the moment. Direct-IO is
209	 * the main guy that batches up lots of get_user_pages, and even
210	 * they are limited to 64-at-a-time which is not so many.
211	 */
212	/*
213	 * This doesn't prevent pagetable teardown, but does prevent
214	 * the pagetables and pages from being freed.
215	 *
216	 * So long as we atomically load page table pointers versus teardown,
217	 * we can follow the address down to the page and take a ref on it.
218	 */
219	local_irq_save(flags);
220	pgdp = pgd_offset(mm, addr);
221	do {
222		pgd_t pgd = *pgdp;
223
224		next = pgd_addr_end(addr, end);
225		if (pgd_none(pgd))
226			break;
227		if (!gup_pud_range(pgd, addr, next, write, pages, &nr))
228			break;
229	} while (pgdp++, addr = next, addr != end);
230	local_irq_restore(flags);
231
232	return nr;
233}
234
235/**
236 * get_user_pages_fast() - pin user pages in memory
237 * @start:	starting user address
238 * @nr_pages:	number of pages from start to pin
239 * @write:	whether pages will be written to
240 * @pages:	array that receives pointers to the pages pinned.
241 *		Should be at least nr_pages long.
242 *
243 * Attempt to pin user pages in memory without taking mm->mmap_sem.
244 * If not successful, it will fall back to taking the lock and
245 * calling get_user_pages().
246 *
247 * Returns number of pages pinned. This may be fewer than the number
248 * requested. If nr_pages is 0 or negative, returns 0. If no pages
249 * were pinned, returns -errno.
250 */
251int get_user_pages_fast(unsigned long start, int nr_pages, int write,
252			struct page **pages)
253{
254	struct mm_struct *mm = current->mm;
255	unsigned long addr, len, end;
256	unsigned long next;
257	pgd_t *pgdp;
258	int ret, nr = 0;
259
260	start &= PAGE_MASK;
261	addr = start;
262	len = (unsigned long) nr_pages << PAGE_SHIFT;
263
264	end = start + len;
265	if (end < start || cpu_has_dc_aliases)
266		goto slow_irqon;
267
268	/* XXX: batch / limit 'nr' */
269	local_irq_disable();
270	pgdp = pgd_offset(mm, addr);
271	do {
272		pgd_t pgd = *pgdp;
273
274		next = pgd_addr_end(addr, end);
275		if (pgd_none(pgd))
276			goto slow;
277		if (!gup_pud_range(pgd, addr, next, write, pages, &nr))
278			goto slow;
279	} while (pgdp++, addr = next, addr != end);
280	local_irq_enable();
281
282	VM_BUG_ON(nr != (end - start) >> PAGE_SHIFT);
283	return nr;
284slow:
285	local_irq_enable();
286
287slow_irqon:
288	/* Try to get the remaining pages with get_user_pages */
289	start += nr << PAGE_SHIFT;
290	pages += nr;
291
292	ret = get_user_pages_unlocked(start, (end - start) >> PAGE_SHIFT,
293				      pages, write ? FOLL_WRITE : 0);
294
295	/* Have to be a bit careful with return values */
296	if (nr > 0) {
297		if (ret < 0)
298			ret = nr;
299		else
300			ret += nr;
301	}
302	return ret;
303}