1/*
2 * IA-32 Huge TLB Page Support for Kernel.
3 *
4 * Copyright (C) 2002, Rohit Seth <rohit.seth@intel.com>
5 */
6
7#include <linux/init.h>
8#include <linux/fs.h>
9#include <linux/mm.h>
10#include <linux/hugetlb.h>
11#include <linux/pagemap.h>
12#include <linux/slab.h>
13#include <linux/err.h>
14#include <linux/sysctl.h>
15#include <asm/mman.h>
16#include <asm/tlb.h>
17#include <asm/tlbflush.h>
18
19static unsigned long page_table_shareable(struct vm_area_struct *svma,
20				struct vm_area_struct *vma,
21				unsigned long addr, pgoff_t idx)
22{
23	unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
24				svma->vm_start;
25	unsigned long sbase = saddr & PUD_MASK;
26	unsigned long s_end = sbase + PUD_SIZE;
27
28	/*
29	 * match the virtual addresses, permission and the alignment of the
30	 * page table page.
31	 */
32	if (pmd_index(addr) != pmd_index(saddr) ||
33	    vma->vm_flags != svma->vm_flags ||
34	    sbase < svma->vm_start || svma->vm_end < s_end)
35		return 0;
36
37	return saddr;
38}
39
40static int vma_shareable(struct vm_area_struct *vma, unsigned long addr)
41{
42	unsigned long base = addr & PUD_MASK;
43	unsigned long end = base + PUD_SIZE;
44
45	/*
46	 * check on proper vm_flags and page table alignment
47	 */
48	if (vma->vm_flags & VM_MAYSHARE &&
49	    vma->vm_start <= base && end <= vma->vm_end)
50		return 1;
51	return 0;
52}
53
54/*
55 * search for a shareable pmd page for hugetlb.
56 */
57static void huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
58{
59	struct vm_area_struct *vma = find_vma(mm, addr);
60	struct address_space *mapping = vma->vm_file->f_mapping;
61	pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
62			vma->vm_pgoff;
63	struct prio_tree_iter iter;
64	struct vm_area_struct *svma;
65	unsigned long saddr;
66	pte_t *spte = NULL;
67
68	if (!vma_shareable(vma, addr))
69		return;
70
71	spin_lock(&mapping->i_mmap_lock);
72	vma_prio_tree_foreach(svma, &iter, &mapping->i_mmap, idx, idx) {
73		if (svma == vma)
74			continue;
75
76		saddr = page_table_shareable(svma, vma, addr, idx);
77		if (saddr) {
78			spte = huge_pte_offset(svma->vm_mm, saddr);
79			if (spte) {
80				get_page(virt_to_page(spte));
81				break;
82			}
83		}
84	}
85
86	if (!spte)
87		goto out;
88
89	spin_lock(&mm->page_table_lock);
90	if (pud_none(*pud))
91		pud_populate(mm, pud, (unsigned long) spte & PAGE_MASK);
92	else
93		put_page(virt_to_page(spte));
94	spin_unlock(&mm->page_table_lock);
95out:
96	spin_unlock(&mapping->i_mmap_lock);
97}
98
99/*
100 * unmap huge page backed by shared pte.
101 *
102 * Hugetlb pte page is ref counted at the time of mapping.  If pte is shared
103 * indicated by page_count > 1, unmap is achieved by clearing pud and
104 * decrementing the ref count. If count == 1, the pte page is not shared.
105 *
106 * called with vma->vm_mm->page_table_lock held.
107 *
108 * returns: 1 successfully unmapped a shared pte page
109 *	    0 the underlying pte page is not shared, or it is the last user
110 */
111int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
112{
113	pgd_t *pgd = pgd_offset(mm, *addr);
114	pud_t *pud = pud_offset(pgd, *addr);
115
116	BUG_ON(page_count(virt_to_page(ptep)) == 0);
117	if (page_count(virt_to_page(ptep)) == 1)
118		return 0;
119
120	pud_clear(pud);
121	put_page(virt_to_page(ptep));
122	*addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE;
123	return 1;
124}
125
126pte_t *huge_pte_alloc(struct mm_struct *mm, unsigned long addr)
127{
128	pgd_t *pgd;
129	pud_t *pud;
130	pte_t *pte = NULL;
131
132	pgd = pgd_offset(mm, addr);
133	pud = pud_alloc(mm, pgd, addr);
134	if (pud) {
135		if (pud_none(*pud))
136			huge_pmd_share(mm, addr, pud);
137		pte = (pte_t *) pmd_alloc(mm, pud, addr);
138	}
139	BUG_ON(pte && !pte_none(*pte) && !pte_huge(*pte));
140
141	return pte;
142}
143
144pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr)
145{
146	pgd_t *pgd;
147	pud_t *pud;
148	pmd_t *pmd = NULL;
149
150	pgd = pgd_offset(mm, addr);
151	if (pgd_present(*pgd)) {
152		pud = pud_offset(pgd, addr);
153		if (pud_present(*pud))
154			pmd = pmd_offset(pud, addr);
155	}
156	return (pte_t *) pmd;
157}
158
159
160struct page *
161follow_huge_addr(struct mm_struct *mm, unsigned long address, int write)
162{
163	return ERR_PTR(-EINVAL);
164}
165
166int pmd_huge(pmd_t pmd)
167{
168	return !!(pmd_val(pmd) & _PAGE_PSE);
169}
170
171struct page *
172follow_huge_pmd(struct mm_struct *mm, unsigned long address,
173		pmd_t *pmd, int write)
174{
175	struct page *page;
176
177	page = pte_page(*(pte_t *)pmd);
178	if (page)
179		page += ((address & ~HPAGE_MASK) >> PAGE_SHIFT);
180	return page;
181}
182
183/* x86_64 also uses this file */
184
185#ifdef HAVE_ARCH_HUGETLB_UNMAPPED_AREA
186static unsigned long hugetlb_get_unmapped_area_bottomup(struct file *file,
187		unsigned long addr, unsigned long len,
188		unsigned long pgoff, unsigned long flags)
189{
190	struct mm_struct *mm = current->mm;
191	struct vm_area_struct *vma;
192	unsigned long start_addr;
193
194	if (len > mm->cached_hole_size) {
195	        start_addr = mm->free_area_cache;
196	} else {
197	        start_addr = TASK_UNMAPPED_BASE;
198	        mm->cached_hole_size = 0;
199	}
200
201full_search:
202	addr = ALIGN(start_addr, HPAGE_SIZE);
203
204	for (vma = find_vma(mm, addr); ; vma = vma->vm_next) {
205		/* At this point:  (!vma || addr < vma->vm_end). */
206		if (TASK_SIZE - len < addr) {
207			/*
208			 * Start a new search - just in case we missed
209			 * some holes.
210			 */
211			if (start_addr != TASK_UNMAPPED_BASE) {
212				start_addr = TASK_UNMAPPED_BASE;
213				mm->cached_hole_size = 0;
214				goto full_search;
215			}
216			return -ENOMEM;
217		}
218		if (!vma || addr + len <= vma->vm_start) {
219			mm->free_area_cache = addr + len;
220			return addr;
221		}
222		if (addr + mm->cached_hole_size < vma->vm_start)
223		        mm->cached_hole_size = vma->vm_start - addr;
224		addr = ALIGN(vma->vm_end, HPAGE_SIZE);
225	}
226}
227
228static unsigned long hugetlb_get_unmapped_area_topdown(struct file *file,
229		unsigned long addr0, unsigned long len,
230		unsigned long pgoff, unsigned long flags)
231{
232	struct mm_struct *mm = current->mm;
233	struct vm_area_struct *vma, *prev_vma;
234	unsigned long base = mm->mmap_base, addr = addr0;
235	unsigned long largest_hole = mm->cached_hole_size;
236	int first_time = 1;
237
238	/* don't allow allocations above current base */
239	if (mm->free_area_cache > base)
240		mm->free_area_cache = base;
241
242	if (len <= largest_hole) {
243	        largest_hole = 0;
244		mm->free_area_cache  = base;
245	}
246try_again:
247	/* make sure it can fit in the remaining address space */
248	if (mm->free_area_cache < len)
249		goto fail;
250
251	/* either no address requested or cant fit in requested address hole */
252	addr = (mm->free_area_cache - len) & HPAGE_MASK;
253	do {
254		/*
255		 * Lookup failure means no vma is above this address,
256		 * i.e. return with success:
257		 */
258		if (!(vma = find_vma_prev(mm, addr, &prev_vma)))
259			return addr;
260
261		/*
262		 * new region fits between prev_vma->vm_end and
263		 * vma->vm_start, use it:
264		 */
265		if (addr + len <= vma->vm_start &&
266		            (!prev_vma || (addr >= prev_vma->vm_end))) {
267			/* remember the address as a hint for next time */
268		        mm->cached_hole_size = largest_hole;
269		        return (mm->free_area_cache = addr);
270		} else {
271			/* pull free_area_cache down to the first hole */
272		        if (mm->free_area_cache == vma->vm_end) {
273				mm->free_area_cache = vma->vm_start;
274				mm->cached_hole_size = largest_hole;
275			}
276		}
277
278		/* remember the largest hole we saw so far */
279		if (addr + largest_hole < vma->vm_start)
280		        largest_hole = vma->vm_start - addr;
281
282		/* try just below the current vma->vm_start */
283		addr = (vma->vm_start - len) & HPAGE_MASK;
284	} while (len <= vma->vm_start);
285
286fail:
287	/*
288	 * if hint left us with no space for the requested
289	 * mapping then try again:
290	 */
291	if (first_time) {
292		mm->free_area_cache = base;
293		largest_hole = 0;
294		first_time = 0;
295		goto try_again;
296	}
297	/*
298	 * A failed mmap() very likely causes application failure,
299	 * so fall back to the bottom-up function here. This scenario
300	 * can happen with large stack limits and large mmap()
301	 * allocations.
302	 */
303	mm->free_area_cache = TASK_UNMAPPED_BASE;
304	mm->cached_hole_size = ~0UL;
305	addr = hugetlb_get_unmapped_area_bottomup(file, addr0,
306			len, pgoff, flags);
307
308	/*
309	 * Restore the topdown base:
310	 */
311	mm->free_area_cache = base;
312	mm->cached_hole_size = ~0UL;
313
314	return addr;
315}
316
317unsigned long
318hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
319		unsigned long len, unsigned long pgoff, unsigned long flags)
320{
321	struct mm_struct *mm = current->mm;
322	struct vm_area_struct *vma;
323
324	if (len & ~HPAGE_MASK)
325		return -EINVAL;
326	if (len > TASK_SIZE)
327		return -ENOMEM;
328
329	if (flags & MAP_FIXED) {
330		if (prepare_hugepage_range(addr, len, pgoff))
331			return -EINVAL;
332		return addr;
333	}
334
335	if (addr) {
336		addr = ALIGN(addr, HPAGE_SIZE);
337		vma = find_vma(mm, addr);
338		if (TASK_SIZE - len >= addr &&
339		    (!vma || addr + len <= vma->vm_start))
340			return addr;
341	}
342	if (mm->get_unmapped_area == arch_get_unmapped_area)
343		return hugetlb_get_unmapped_area_bottomup(file, addr, len,
344				pgoff, flags);
345	else
346		return hugetlb_get_unmapped_area_topdown(file, addr, len,
347				pgoff, flags);
348}
349
350#endif /*HAVE_ARCH_HUGETLB_UNMAPPED_AREA*/
351