Merge branch 'merge' of git://git.kernel.org/pub/scm/linux/kernel/git/paulus/powerpc
[linux-2.6] / 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/module.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 <asm/tlbflush.h>
21 #include <asm/io.h>
22
23 /*
24  * We do use our own empty page to avoid interference with other users
25  * of ZERO_PAGE(), such as /dev/zero
26  */
27 static DEFINE_MUTEX(xip_sparse_mutex);
28 static seqcount_t xip_sparse_seq = SEQCNT_ZERO;
29 static struct page *__xip_sparse_page;
30
31 /* called under xip_sparse_mutex */
32 static struct page *xip_sparse_page(void)
33 {
34         if (!__xip_sparse_page) {
35                 struct page *page = alloc_page(GFP_HIGHUSER | __GFP_ZERO);
36
37                 if (page)
38                         __xip_sparse_page = page;
39         }
40         return __xip_sparse_page;
41 }
42
43 /*
44  * This is a file read routine for execute in place files, and uses
45  * the mapping->a_ops->get_xip_mem() function for the actual low-level
46  * stuff.
47  *
48  * Note the struct file* is not used at all.  It may be NULL.
49  */
50 static ssize_t
51 do_xip_mapping_read(struct address_space *mapping,
52                     struct file_ra_state *_ra,
53                     struct file *filp,
54                     char __user *buf,
55                     size_t len,
56                     loff_t *ppos)
57 {
58         struct inode *inode = mapping->host;
59         pgoff_t index, end_index;
60         unsigned long offset;
61         loff_t isize, pos;
62         size_t copied = 0, error = 0;
63
64         BUG_ON(!mapping->a_ops->get_xip_mem);
65
66         pos = *ppos;
67         index = pos >> PAGE_CACHE_SHIFT;
68         offset = pos & ~PAGE_CACHE_MASK;
69
70         isize = i_size_read(inode);
71         if (!isize)
72                 goto out;
73
74         end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
75         do {
76                 unsigned long nr, left;
77                 void *xip_mem;
78                 unsigned long xip_pfn;
79                 int zero = 0;
80
81                 /* nr is the maximum number of bytes to copy from this page */
82                 nr = PAGE_CACHE_SIZE;
83                 if (index >= end_index) {
84                         if (index > end_index)
85                                 goto out;
86                         nr = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
87                         if (nr <= offset) {
88                                 goto out;
89                         }
90                 }
91                 nr = nr - offset;
92                 if (nr > len - copied)
93                         nr = len - copied;
94
95                 error = mapping->a_ops->get_xip_mem(mapping, index, 0,
96                                                         &xip_mem, &xip_pfn);
97                 if (unlikely(error)) {
98                         if (error == -ENODATA) {
99                                 /* sparse */
100                                 zero = 1;
101                         } else
102                                 goto out;
103                 }
104
105                 /* If users can be writing to this page using arbitrary
106                  * virtual addresses, take care about potential aliasing
107                  * before reading the page on the kernel side.
108                  */
109                 if (mapping_writably_mapped(mapping))
110                         /* address based flush */ ;
111
112                 /*
113                  * Ok, we have the mem, so now we can copy it to user space...
114                  *
115                  * The actor routine returns how many bytes were actually used..
116                  * NOTE! This may not be the same as how much of a user buffer
117                  * we filled up (we may be padding etc), so we can only update
118                  * "pos" here (the actor routine has to update the user buffer
119                  * pointers and the remaining count).
120                  */
121                 if (!zero)
122                         left = __copy_to_user(buf+copied, xip_mem+offset, nr);
123                 else
124                         left = __clear_user(buf + copied, nr);
125
126                 if (left) {
127                         error = -EFAULT;
128                         goto out;
129                 }
130
131                 copied += (nr - left);
132                 offset += (nr - left);
133                 index += offset >> PAGE_CACHE_SHIFT;
134                 offset &= ~PAGE_CACHE_MASK;
135         } while (copied < len);
136
137 out:
138         *ppos = pos + copied;
139         if (filp)
140                 file_accessed(filp);
141
142         return (copied ? copied : error);
143 }
144
145 ssize_t
146 xip_file_read(struct file *filp, char __user *buf, size_t len, loff_t *ppos)
147 {
148         if (!access_ok(VERIFY_WRITE, buf, len))
149                 return -EFAULT;
150
151         return do_xip_mapping_read(filp->f_mapping, &filp->f_ra, filp,
152                             buf, len, ppos);
153 }
154 EXPORT_SYMBOL_GPL(xip_file_read);
155
156 /*
157  * __xip_unmap is invoked from xip_unmap and
158  * xip_write
159  *
160  * This function walks all vmas of the address_space and unmaps the
161  * __xip_sparse_page when found at pgoff.
162  */
163 static void
164 __xip_unmap (struct address_space * mapping,
165                      unsigned long pgoff)
166 {
167         struct vm_area_struct *vma;
168         struct mm_struct *mm;
169         struct prio_tree_iter iter;
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         spin_lock(&mapping->i_mmap_lock);
186         vma_prio_tree_foreach(vma, &iter, &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_notify(vma, address, pte);
196                         page_remove_rmap(page);
197                         dec_mm_counter(mm, file_rss);
198                         BUG_ON(pte_dirty(pteval));
199                         pte_unmap_unlock(pte, ptl);
200                         page_cache_release(page);
201                 }
202         }
203         spin_unlock(&mapping->i_mmap_lock);
204
205         if (locked) {
206                 mutex_unlock(&xip_sparse_mutex);
207         } else if (read_seqcount_retry(&xip_sparse_seq, count)) {
208                 mutex_lock(&xip_sparse_mutex);
209                 locked = 1;
210                 goto retry;
211         }
212 }
213
214 /*
215  * xip_fault() is invoked via the vma operations vector for a
216  * mapped memory region to read in file data during a page fault.
217  *
218  * This function is derived from filemap_fault, but used for execute in place
219  */
220 static int xip_file_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
221 {
222         struct file *file = vma->vm_file;
223         struct address_space *mapping = file->f_mapping;
224         struct inode *inode = mapping->host;
225         pgoff_t size;
226         void *xip_mem;
227         unsigned long xip_pfn;
228         struct page *page;
229         int error;
230
231         /* XXX: are VM_FAULT_ codes OK? */
232 again:
233         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
234         if (vmf->pgoff >= size)
235                 return VM_FAULT_SIGBUS;
236
237         error = mapping->a_ops->get_xip_mem(mapping, vmf->pgoff, 0,
238                                                 &xip_mem, &xip_pfn);
239         if (likely(!error))
240                 goto found;
241         if (error != -ENODATA)
242                 return VM_FAULT_OOM;
243
244         /* sparse block */
245         if ((vma->vm_flags & (VM_WRITE | VM_MAYWRITE)) &&
246             (vma->vm_flags & (VM_SHARED | VM_MAYSHARE)) &&
247             (!(mapping->host->i_sb->s_flags & MS_RDONLY))) {
248                 int err;
249
250                 /* maybe shared writable, allocate new block */
251                 mutex_lock(&xip_sparse_mutex);
252                 error = mapping->a_ops->get_xip_mem(mapping, vmf->pgoff, 1,
253                                                         &xip_mem, &xip_pfn);
254                 mutex_unlock(&xip_sparse_mutex);
255                 if (error)
256                         return VM_FAULT_SIGBUS;
257                 /* unmap sparse mappings at pgoff from all other vmas */
258                 __xip_unmap(mapping, vmf->pgoff);
259
260 found:
261                 err = vm_insert_mixed(vma, (unsigned long)vmf->virtual_address,
262                                                         xip_pfn);
263                 if (err == -ENOMEM)
264                         return VM_FAULT_OOM;
265                 BUG_ON(err);
266                 return VM_FAULT_NOPAGE;
267         } else {
268                 int err, ret = VM_FAULT_OOM;
269
270                 mutex_lock(&xip_sparse_mutex);
271                 write_seqcount_begin(&xip_sparse_seq);
272                 error = mapping->a_ops->get_xip_mem(mapping, vmf->pgoff, 0,
273                                                         &xip_mem, &xip_pfn);
274                 if (unlikely(!error)) {
275                         write_seqcount_end(&xip_sparse_seq);
276                         mutex_unlock(&xip_sparse_mutex);
277                         goto again;
278                 }
279                 if (error != -ENODATA)
280                         goto out;
281                 /* not shared and writable, use xip_sparse_page() */
282                 page = xip_sparse_page();
283                 if (!page)
284                         goto out;
285                 err = vm_insert_page(vma, (unsigned long)vmf->virtual_address,
286                                                         page);
287                 if (err == -ENOMEM)
288                         goto out;
289
290                 ret = VM_FAULT_NOPAGE;
291 out:
292                 write_seqcount_end(&xip_sparse_seq);
293                 mutex_unlock(&xip_sparse_mutex);
294
295                 return ret;
296         }
297 }
298
299 static struct vm_operations_struct xip_file_vm_ops = {
300         .fault  = xip_file_fault,
301 };
302
303 int xip_file_mmap(struct file * file, struct vm_area_struct * vma)
304 {
305         BUG_ON(!file->f_mapping->a_ops->get_xip_mem);
306
307         file_accessed(file);
308         vma->vm_ops = &xip_file_vm_ops;
309         vma->vm_flags |= VM_CAN_NONLINEAR | VM_MIXEDMAP;
310         return 0;
311 }
312 EXPORT_SYMBOL_GPL(xip_file_mmap);
313
314 static ssize_t
315 __xip_file_write(struct file *filp, const char __user *buf,
316                   size_t count, loff_t pos, loff_t *ppos)
317 {
318         struct address_space * mapping = filp->f_mapping;
319         const struct address_space_operations *a_ops = mapping->a_ops;
320         struct inode    *inode = mapping->host;
321         long            status = 0;
322         size_t          bytes;
323         ssize_t         written = 0;
324
325         BUG_ON(!mapping->a_ops->get_xip_mem);
326
327         do {
328                 unsigned long index;
329                 unsigned long offset;
330                 size_t copied;
331                 void *xip_mem;
332                 unsigned long xip_pfn;
333
334                 offset = (pos & (PAGE_CACHE_SIZE -1)); /* Within page */
335                 index = pos >> PAGE_CACHE_SHIFT;
336                 bytes = PAGE_CACHE_SIZE - offset;
337                 if (bytes > count)
338                         bytes = count;
339
340                 status = a_ops->get_xip_mem(mapping, index, 0,
341                                                 &xip_mem, &xip_pfn);
342                 if (status == -ENODATA) {
343                         /* we allocate a new page unmap it */
344                         mutex_lock(&xip_sparse_mutex);
345                         status = a_ops->get_xip_mem(mapping, index, 1,
346                                                         &xip_mem, &xip_pfn);
347                         mutex_unlock(&xip_sparse_mutex);
348                         if (!status)
349                                 /* unmap page at pgoff from all other vmas */
350                                 __xip_unmap(mapping, index);
351                 }
352
353                 if (status)
354                         break;
355
356                 copied = bytes -
357                         __copy_from_user_nocache(xip_mem + offset, buf, bytes);
358
359                 if (likely(copied > 0)) {
360                         status = copied;
361
362                         if (status >= 0) {
363                                 written += status;
364                                 count -= status;
365                                 pos += status;
366                                 buf += status;
367                         }
368                 }
369                 if (unlikely(copied != bytes))
370                         if (status >= 0)
371                                 status = -EFAULT;
372                 if (status < 0)
373                         break;
374         } while (count);
375         *ppos = pos;
376         /*
377          * No need to use i_size_read() here, the i_size
378          * cannot change under us because we hold i_mutex.
379          */
380         if (pos > inode->i_size) {
381                 i_size_write(inode, pos);
382                 mark_inode_dirty(inode);
383         }
384
385         return written ? written : status;
386 }
387
388 ssize_t
389 xip_file_write(struct file *filp, const char __user *buf, size_t len,
390                loff_t *ppos)
391 {
392         struct address_space *mapping = filp->f_mapping;
393         struct inode *inode = mapping->host;
394         size_t count;
395         loff_t pos;
396         ssize_t ret;
397
398         mutex_lock(&inode->i_mutex);
399
400         if (!access_ok(VERIFY_READ, buf, len)) {
401                 ret=-EFAULT;
402                 goto out_up;
403         }
404
405         pos = *ppos;
406         count = len;
407
408         vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
409
410         /* We can write back this queue in page reclaim */
411         current->backing_dev_info = mapping->backing_dev_info;
412
413         ret = generic_write_checks(filp, &pos, &count, S_ISBLK(inode->i_mode));
414         if (ret)
415                 goto out_backing;
416         if (count == 0)
417                 goto out_backing;
418
419         ret = file_remove_suid(filp);
420         if (ret)
421                 goto out_backing;
422
423         file_update_time(filp);
424
425         ret = __xip_file_write (filp, buf, count, pos, ppos);
426
427  out_backing:
428         current->backing_dev_info = NULL;
429  out_up:
430         mutex_unlock(&inode->i_mutex);
431         return ret;
432 }
433 EXPORT_SYMBOL_GPL(xip_file_write);
434
435 /*
436  * truncate a page used for execute in place
437  * functionality is analog to block_truncate_page but does use get_xip_mem
438  * to get the page instead of page cache
439  */
440 int
441 xip_truncate_page(struct address_space *mapping, loff_t from)
442 {
443         pgoff_t index = from >> PAGE_CACHE_SHIFT;
444         unsigned offset = from & (PAGE_CACHE_SIZE-1);
445         unsigned blocksize;
446         unsigned length;
447         void *xip_mem;
448         unsigned long xip_pfn;
449         int err;
450
451         BUG_ON(!mapping->a_ops->get_xip_mem);
452
453         blocksize = 1 << mapping->host->i_blkbits;
454         length = offset & (blocksize - 1);
455
456         /* Block boundary? Nothing to do */
457         if (!length)
458                 return 0;
459
460         length = blocksize - length;
461
462         err = mapping->a_ops->get_xip_mem(mapping, index, 0,
463                                                 &xip_mem, &xip_pfn);
464         if (unlikely(err)) {
465                 if (err == -ENODATA)
466                         /* Hole? No need to truncate */
467                         return 0;
468                 else
469                         return err;
470         }
471         memset(xip_mem + offset, 0, length);
472         return 0;
473 }
474 EXPORT_SYMBOL_GPL(xip_truncate_page);