Merge branch 'x86-acpi-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git...
[kernel/kernel-generic.git] / mm / filemap_xip.c
1 /*
2  *      linux/mm/filemap_xip.c
3  *
4  * Copyright (C) 2005 IBM Corporation
5  * Author: Carsten Otte <cotte@de.ibm.com>
6  *
7  * derived from linux/mm/filemap.c - Copyright (C) Linus Torvalds
8  *
9  */
10
11 #include <linux/fs.h>
12 #include <linux/pagemap.h>
13 #include <linux/export.h>
14 #include <linux/uio.h>
15 #include <linux/rmap.h>
16 #include <linux/mmu_notifier.h>
17 #include <linux/sched.h>
18 #include <linux/seqlock.h>
19 #include <linux/mutex.h>
20 #include <linux/gfp.h>
21 #include <asm/tlbflush.h>
22 #include <asm/io.h>
23
24 /*
25  * We do use our own empty page to avoid interference with other users
26  * of ZERO_PAGE(), such as /dev/zero
27  */
28 static DEFINE_MUTEX(xip_sparse_mutex);
29 static seqcount_t xip_sparse_seq = SEQCNT_ZERO;
30 static struct page *__xip_sparse_page;
31
32 /* called under xip_sparse_mutex */
33 static struct page *xip_sparse_page(void)
34 {
35         if (!__xip_sparse_page) {
36                 struct page *page = alloc_page(GFP_HIGHUSER | __GFP_ZERO);
37
38                 if (page)
39                         __xip_sparse_page = page;
40         }
41         return __xip_sparse_page;
42 }
43
44 /*
45  * This is a file read routine for execute in place files, and uses
46  * the mapping->a_ops->get_xip_mem() function for the actual low-level
47  * stuff.
48  *
49  * Note the struct file* is not used at all.  It may be NULL.
50  */
51 static ssize_t
52 do_xip_mapping_read(struct address_space *mapping,
53                     struct file_ra_state *_ra,
54                     struct file *filp,
55                     char __user *buf,
56                     size_t len,
57                     loff_t *ppos)
58 {
59         struct inode *inode = mapping->host;
60         pgoff_t index, end_index;
61         unsigned long offset;
62         loff_t isize, pos;
63         size_t copied = 0, error = 0;
64
65         BUG_ON(!mapping->a_ops->get_xip_mem);
66
67         pos = *ppos;
68         index = pos >> PAGE_CACHE_SHIFT;
69         offset = pos & ~PAGE_CACHE_MASK;
70
71         isize = i_size_read(inode);
72         if (!isize)
73                 goto out;
74
75         end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
76         do {
77                 unsigned long nr, left;
78                 void *xip_mem;
79                 unsigned long xip_pfn;
80                 int zero = 0;
81
82                 /* nr is the maximum number of bytes to copy from this page */
83                 nr = PAGE_CACHE_SIZE;
84                 if (index >= end_index) {
85                         if (index > end_index)
86                                 goto out;
87                         nr = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
88                         if (nr <= offset) {
89                                 goto out;
90                         }
91                 }
92                 nr = nr - offset;
93                 if (nr > len - copied)
94                         nr = len - copied;
95
96                 error = mapping->a_ops->get_xip_mem(mapping, index, 0,
97                                                         &xip_mem, &xip_pfn);
98                 if (unlikely(error)) {
99                         if (error == -ENODATA) {
100                                 /* sparse */
101                                 zero = 1;
102                         } else
103                                 goto out;
104                 }
105
106                 /* If users can be writing to this page using arbitrary
107                  * virtual addresses, take care about potential aliasing
108                  * before reading the page on the kernel side.
109                  */
110                 if (mapping_writably_mapped(mapping))
111                         /* address based flush */ ;
112
113                 /*
114                  * Ok, we have the mem, so now we can copy it to user space...
115                  *
116                  * The actor routine returns how many bytes were actually used..
117                  * NOTE! This may not be the same as how much of a user buffer
118                  * we filled up (we may be padding etc), so we can only update
119                  * "pos" here (the actor routine has to update the user buffer
120                  * pointers and the remaining count).
121                  */
122                 if (!zero)
123                         left = __copy_to_user(buf+copied, xip_mem+offset, nr);
124                 else
125                         left = __clear_user(buf + copied, nr);
126
127                 if (left) {
128                         error = -EFAULT;
129                         goto out;
130                 }
131
132                 copied += (nr - left);
133                 offset += (nr - left);
134                 index += offset >> PAGE_CACHE_SHIFT;
135                 offset &= ~PAGE_CACHE_MASK;
136         } while (copied < len);
137
138 out:
139         *ppos = pos + copied;
140         if (filp)
141                 file_accessed(filp);
142
143         return (copied ? copied : error);
144 }
145
146 ssize_t
147 xip_file_read(struct file *filp, char __user *buf, size_t len, loff_t *ppos)
148 {
149         if (!access_ok(VERIFY_WRITE, buf, len))
150                 return -EFAULT;
151
152         return do_xip_mapping_read(filp->f_mapping, &filp->f_ra, filp,
153                             buf, len, ppos);
154 }
155 EXPORT_SYMBOL_GPL(xip_file_read);
156
157 /*
158  * __xip_unmap is invoked from xip_unmap and
159  * xip_write
160  *
161  * This function walks all vmas of the address_space and unmaps the
162  * __xip_sparse_page when found at pgoff.
163  */
164 static void
165 __xip_unmap (struct address_space * mapping,
166                      unsigned long pgoff)
167 {
168         struct vm_area_struct *vma;
169         struct mm_struct *mm;
170         unsigned long address;
171         pte_t *pte;
172         pte_t pteval;
173         spinlock_t *ptl;
174         struct page *page;
175         unsigned count;
176         int locked = 0;
177
178         count = read_seqcount_begin(&xip_sparse_seq);
179
180         page = __xip_sparse_page;
181         if (!page)
182                 return;
183
184 retry:
185         mutex_lock(&mapping->i_mmap_mutex);
186         vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) {
187                 mm = vma->vm_mm;
188                 address = vma->vm_start +
189                         ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
190                 BUG_ON(address < vma->vm_start || address >= vma->vm_end);
191                 pte = page_check_address(page, mm, address, &ptl, 1);
192                 if (pte) {
193                         /* Nuke the page table entry. */
194                         flush_cache_page(vma, address, pte_pfn(*pte));
195                         pteval = ptep_clear_flush(vma, address, pte);
196                         page_remove_rmap(page);
197                         dec_mm_counter(mm, MM_FILEPAGES);
198                         BUG_ON(pte_dirty(pteval));
199                         pte_unmap_unlock(pte, ptl);
200                         /* must invalidate_page _before_ freeing the page */
201                         mmu_notifier_invalidate_page(mm, address);
202                         page_cache_release(page);
203                 }
204         }
205         mutex_unlock(&mapping->i_mmap_mutex);
206
207         if (locked) {
208                 mutex_unlock(&xip_sparse_mutex);
209         } else if (read_seqcount_retry(&xip_sparse_seq, count)) {
210                 mutex_lock(&xip_sparse_mutex);
211                 locked = 1;
212                 goto retry;
213         }
214 }
215
216 /*
217  * xip_fault() is invoked via the vma operations vector for a
218  * mapped memory region to read in file data during a page fault.
219  *
220  * This function is derived from filemap_fault, but used for execute in place
221  */
222 static int xip_file_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
223 {
224         struct file *file = vma->vm_file;
225         struct address_space *mapping = file->f_mapping;
226         struct inode *inode = mapping->host;
227         pgoff_t size;
228         void *xip_mem;
229         unsigned long xip_pfn;
230         struct page *page;
231         int error;
232
233         /* XXX: are VM_FAULT_ codes OK? */
234 again:
235         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
236         if (vmf->pgoff >= size)
237                 return VM_FAULT_SIGBUS;
238
239         error = mapping->a_ops->get_xip_mem(mapping, vmf->pgoff, 0,
240                                                 &xip_mem, &xip_pfn);
241         if (likely(!error))
242                 goto found;
243         if (error != -ENODATA)
244                 return VM_FAULT_OOM;
245
246         /* sparse block */
247         if ((vma->vm_flags & (VM_WRITE | VM_MAYWRITE)) &&
248             (vma->vm_flags & (VM_SHARED | VM_MAYSHARE)) &&
249             (!(mapping->host->i_sb->s_flags & MS_RDONLY))) {
250                 int err;
251
252                 /* maybe shared writable, allocate new block */
253                 mutex_lock(&xip_sparse_mutex);
254                 error = mapping->a_ops->get_xip_mem(mapping, vmf->pgoff, 1,
255                                                         &xip_mem, &xip_pfn);
256                 mutex_unlock(&xip_sparse_mutex);
257                 if (error)
258                         return VM_FAULT_SIGBUS;
259                 /* unmap sparse mappings at pgoff from all other vmas */
260                 __xip_unmap(mapping, vmf->pgoff);
261
262 found:
263                 err = vm_insert_mixed(vma, (unsigned long)vmf->virtual_address,
264                                                         xip_pfn);
265                 if (err == -ENOMEM)
266                         return VM_FAULT_OOM;
267                 /*
268                  * err == -EBUSY is fine, we've raced against another thread
269                  * that faulted-in the same page
270                  */
271                 if (err != -EBUSY)
272                         BUG_ON(err);
273                 return VM_FAULT_NOPAGE;
274         } else {
275                 int err, ret = VM_FAULT_OOM;
276
277                 mutex_lock(&xip_sparse_mutex);
278                 write_seqcount_begin(&xip_sparse_seq);
279                 error = mapping->a_ops->get_xip_mem(mapping, vmf->pgoff, 0,
280                                                         &xip_mem, &xip_pfn);
281                 if (unlikely(!error)) {
282                         write_seqcount_end(&xip_sparse_seq);
283                         mutex_unlock(&xip_sparse_mutex);
284                         goto again;
285                 }
286                 if (error != -ENODATA)
287                         goto out;
288                 /* not shared and writable, use xip_sparse_page() */
289                 page = xip_sparse_page();
290                 if (!page)
291                         goto out;
292                 err = vm_insert_page(vma, (unsigned long)vmf->virtual_address,
293                                                         page);
294                 if (err == -ENOMEM)
295                         goto out;
296
297                 ret = VM_FAULT_NOPAGE;
298 out:
299                 write_seqcount_end(&xip_sparse_seq);
300                 mutex_unlock(&xip_sparse_mutex);
301
302                 return ret;
303         }
304 }
305
306 static const struct vm_operations_struct xip_file_vm_ops = {
307         .fault  = xip_file_fault,
308         .page_mkwrite   = filemap_page_mkwrite,
309         .remap_pages = generic_file_remap_pages,
310 };
311
312 int xip_file_mmap(struct file * file, struct vm_area_struct * vma)
313 {
314         BUG_ON(!file->f_mapping->a_ops->get_xip_mem);
315
316         file_accessed(file);
317         vma->vm_ops = &xip_file_vm_ops;
318         vma->vm_flags |= VM_MIXEDMAP;
319         return 0;
320 }
321 EXPORT_SYMBOL_GPL(xip_file_mmap);
322
323 static ssize_t
324 __xip_file_write(struct file *filp, const char __user *buf,
325                   size_t count, loff_t pos, loff_t *ppos)
326 {
327         struct address_space * mapping = filp->f_mapping;
328         const struct address_space_operations *a_ops = mapping->a_ops;
329         struct inode    *inode = mapping->host;
330         long            status = 0;
331         size_t          bytes;
332         ssize_t         written = 0;
333
334         BUG_ON(!mapping->a_ops->get_xip_mem);
335
336         do {
337                 unsigned long index;
338                 unsigned long offset;
339                 size_t copied;
340                 void *xip_mem;
341                 unsigned long xip_pfn;
342
343                 offset = (pos & (PAGE_CACHE_SIZE -1)); /* Within page */
344                 index = pos >> PAGE_CACHE_SHIFT;
345                 bytes = PAGE_CACHE_SIZE - offset;
346                 if (bytes > count)
347                         bytes = count;
348
349                 status = a_ops->get_xip_mem(mapping, index, 0,
350                                                 &xip_mem, &xip_pfn);
351                 if (status == -ENODATA) {
352                         /* we allocate a new page unmap it */
353                         mutex_lock(&xip_sparse_mutex);
354                         status = a_ops->get_xip_mem(mapping, index, 1,
355                                                         &xip_mem, &xip_pfn);
356                         mutex_unlock(&xip_sparse_mutex);
357                         if (!status)
358                                 /* unmap page at pgoff from all other vmas */
359                                 __xip_unmap(mapping, index);
360                 }
361
362                 if (status)
363                         break;
364
365                 copied = bytes -
366                         __copy_from_user_nocache(xip_mem + offset, buf, bytes);
367
368                 if (likely(copied > 0)) {
369                         status = copied;
370
371                         if (status >= 0) {
372                                 written += status;
373                                 count -= status;
374                                 pos += status;
375                                 buf += status;
376                         }
377                 }
378                 if (unlikely(copied != bytes))
379                         if (status >= 0)
380                                 status = -EFAULT;
381                 if (status < 0)
382                         break;
383         } while (count);
384         *ppos = pos;
385         /*
386          * No need to use i_size_read() here, the i_size
387          * cannot change under us because we hold i_mutex.
388          */
389         if (pos > inode->i_size) {
390                 i_size_write(inode, pos);
391                 mark_inode_dirty(inode);
392         }
393
394         return written ? written : status;
395 }
396
397 ssize_t
398 xip_file_write(struct file *filp, const char __user *buf, size_t len,
399                loff_t *ppos)
400 {
401         struct address_space *mapping = filp->f_mapping;
402         struct inode *inode = mapping->host;
403         size_t count;
404         loff_t pos;
405         ssize_t ret;
406
407         sb_start_write(inode->i_sb);
408
409         mutex_lock(&inode->i_mutex);
410
411         if (!access_ok(VERIFY_READ, buf, len)) {
412                 ret=-EFAULT;
413                 goto out_up;
414         }
415
416         pos = *ppos;
417         count = len;
418
419         /* We can write back this queue in page reclaim */
420         current->backing_dev_info = mapping->backing_dev_info;
421
422         ret = generic_write_checks(filp, &pos, &count, S_ISBLK(inode->i_mode));
423         if (ret)
424                 goto out_backing;
425         if (count == 0)
426                 goto out_backing;
427
428         ret = file_remove_suid(filp);
429         if (ret)
430                 goto out_backing;
431
432         ret = file_update_time(filp);
433         if (ret)
434                 goto out_backing;
435
436         ret = __xip_file_write (filp, buf, count, pos, ppos);
437
438  out_backing:
439         current->backing_dev_info = NULL;
440  out_up:
441         mutex_unlock(&inode->i_mutex);
442         sb_end_write(inode->i_sb);
443         return ret;
444 }
445 EXPORT_SYMBOL_GPL(xip_file_write);
446
447 /*
448  * truncate a page used for execute in place
449  * functionality is analog to block_truncate_page but does use get_xip_mem
450  * to get the page instead of page cache
451  */
452 int
453 xip_truncate_page(struct address_space *mapping, loff_t from)
454 {
455         pgoff_t index = from >> PAGE_CACHE_SHIFT;
456         unsigned offset = from & (PAGE_CACHE_SIZE-1);
457         unsigned blocksize;
458         unsigned length;
459         void *xip_mem;
460         unsigned long xip_pfn;
461         int err;
462
463         BUG_ON(!mapping->a_ops->get_xip_mem);
464
465         blocksize = 1 << mapping->host->i_blkbits;
466         length = offset & (blocksize - 1);
467
468         /* Block boundary? Nothing to do */
469         if (!length)
470                 return 0;
471
472         length = blocksize - length;
473
474         err = mapping->a_ops->get_xip_mem(mapping, index, 0,
475                                                 &xip_mem, &xip_pfn);
476         if (unlikely(err)) {
477                 if (err == -ENODATA)
478                         /* Hole? No need to truncate */
479                         return 0;
480                 else
481                         return err;
482         }
483         memset(xip_mem + offset, 0, length);
484         return 0;
485 }
486 EXPORT_SYMBOL_GPL(xip_truncate_page);