fb_defio: fix for non-dirty ptes
[linux-2.6.git] / drivers / video / fb_defio.c
1 /*
2  *  linux/drivers/video/fb_defio.c
3  *
4  *  Copyright (C) 2006 Jaya Kumar
5  *
6  * This file is subject to the terms and conditions of the GNU General Public
7  * License. See the file COPYING in the main directory of this archive
8  * for more details.
9  */
10
11 #include <linux/module.h>
12 #include <linux/kernel.h>
13 #include <linux/errno.h>
14 #include <linux/string.h>
15 #include <linux/mm.h>
16 #include <linux/vmalloc.h>
17 #include <linux/delay.h>
18 #include <linux/interrupt.h>
19 #include <linux/fb.h>
20 #include <linux/list.h>
21
22 /* to support deferred IO */
23 #include <linux/rmap.h>
24 #include <linux/pagemap.h>
25
26 struct page *fb_deferred_io_page(struct fb_info *info, unsigned long offs)
27 {
28         void *screen_base = (void __force *) info->screen_base;
29         struct page *page;
30
31         if (is_vmalloc_addr(screen_base + offs))
32                 page = vmalloc_to_page(screen_base + offs);
33         else
34                 page = pfn_to_page((info->fix.smem_start + offs) >> PAGE_SHIFT);
35
36         return page;
37 }
38
39 /* this is to find and return the vmalloc-ed fb pages */
40 static int fb_deferred_io_fault(struct vm_area_struct *vma,
41                                 struct vm_fault *vmf)
42 {
43         unsigned long offset;
44         struct page *page;
45         struct fb_info *info = vma->vm_private_data;
46
47         offset = vmf->pgoff << PAGE_SHIFT;
48         if (offset >= info->fix.smem_len)
49                 return VM_FAULT_SIGBUS;
50
51         page = fb_deferred_io_page(info, offset);
52         if (!page)
53                 return VM_FAULT_SIGBUS;
54
55         get_page(page);
56
57         if (vma->vm_file)
58                 page->mapping = vma->vm_file->f_mapping;
59         else
60                 printk(KERN_ERR "no mapping available\n");
61
62         BUG_ON(!page->mapping);
63         page->index = vmf->pgoff;
64
65         vmf->page = page;
66         return 0;
67 }
68
69 int fb_deferred_io_fsync(struct file *file, struct dentry *dentry, int datasync)
70 {
71         struct fb_info *info = file->private_data;
72
73         /* Skip if deferred io is compiled-in but disabled on this fbdev */
74         if (!info->fbdefio)
75                 return 0;
76
77         /* Kill off the delayed work */
78         cancel_rearming_delayed_work(&info->deferred_work);
79
80         /* Run it immediately */
81         return schedule_delayed_work(&info->deferred_work, 0);
82 }
83 EXPORT_SYMBOL_GPL(fb_deferred_io_fsync);
84
85 /* vm_ops->page_mkwrite handler */
86 static int fb_deferred_io_mkwrite(struct vm_area_struct *vma,
87                                   struct vm_fault *vmf)
88 {
89         struct page *page = vmf->page;
90         struct fb_info *info = vma->vm_private_data;
91         struct fb_deferred_io *fbdefio = info->fbdefio;
92         struct page *cur;
93
94         /* this is a callback we get when userspace first tries to
95         write to the page. we schedule a workqueue. that workqueue
96         will eventually mkclean the touched pages and execute the
97         deferred framebuffer IO. then if userspace touches a page
98         again, we repeat the same scheme */
99
100         /* protect against the workqueue changing the page list */
101         mutex_lock(&fbdefio->lock);
102
103         /* we loop through the pagelist before adding in order
104         to keep the pagelist sorted */
105         list_for_each_entry(cur, &fbdefio->pagelist, lru) {
106                 /* this check is to catch the case where a new
107                 process could start writing to the same page
108                 through a new pte. this new access can cause the
109                 mkwrite even when the original ps's pte is marked
110                 writable */
111                 if (unlikely(cur == page))
112                         goto page_already_added;
113                 else if (cur->index > page->index)
114                         break;
115         }
116
117         list_add_tail(&page->lru, &cur->lru);
118
119 page_already_added:
120         mutex_unlock(&fbdefio->lock);
121
122         /* come back after delay to process the deferred IO */
123         schedule_delayed_work(&info->deferred_work, fbdefio->delay);
124         return 0;
125 }
126
127 static const struct vm_operations_struct fb_deferred_io_vm_ops = {
128         .fault          = fb_deferred_io_fault,
129         .page_mkwrite   = fb_deferred_io_mkwrite,
130 };
131
132 static int fb_deferred_io_set_page_dirty(struct page *page)
133 {
134         if (!PageDirty(page))
135                 SetPageDirty(page);
136         return 0;
137 }
138
139 static const struct address_space_operations fb_deferred_io_aops = {
140         .set_page_dirty = fb_deferred_io_set_page_dirty,
141 };
142
143 static int fb_deferred_io_mmap(struct fb_info *info, struct vm_area_struct *vma)
144 {
145         vma->vm_ops = &fb_deferred_io_vm_ops;
146         vma->vm_flags |= ( VM_RESERVED | VM_DONTEXPAND );
147         if (!(info->flags & FBINFO_VIRTFB))
148                 vma->vm_flags |= VM_IO;
149         vma->vm_private_data = info;
150         return 0;
151 }
152
153 /* workqueue callback */
154 static void fb_deferred_io_work(struct work_struct *work)
155 {
156         struct fb_info *info = container_of(work, struct fb_info,
157                                                 deferred_work.work);
158         struct fb_deferred_io *fbdefio = info->fbdefio;
159         struct page *page, *tmp_page;
160         struct list_head *node, *tmp_node;
161         struct list_head non_dirty;
162
163         INIT_LIST_HEAD(&non_dirty);
164
165         /* here we mkclean the pages, then do all deferred IO */
166         mutex_lock(&fbdefio->lock);
167         list_for_each_entry_safe(page, tmp_page, &fbdefio->pagelist, lru) {
168                 lock_page(page);
169                 /*
170                  * The workqueue callback can be triggered after a
171                  * ->page_mkwrite() call but before the PTE has been marked
172                  * dirty. In this case page_mkclean() won't "rearm" the page.
173                  *
174                  * To avoid this, remove those "non-dirty" pages from the
175                  * pagelist before calling the driver's callback, then add
176                  * them back to get processed on the next work iteration.
177                  * At that time, their PTEs will hopefully be dirty for real.
178                  */
179                 if (!page_mkclean(page))
180                         list_move_tail(&page->lru, &non_dirty);
181                 unlock_page(page);
182         }
183
184         /* driver's callback with pagelist */
185         fbdefio->deferred_io(info, &fbdefio->pagelist);
186
187         /* clear the list... */
188         list_for_each_safe(node, tmp_node, &fbdefio->pagelist) {
189                 list_del(node);
190         }
191         /* ... and add back the "non-dirty" pages to the list */
192         list_splice_tail(&non_dirty, &fbdefio->pagelist);
193         mutex_unlock(&fbdefio->lock);
194 }
195
196 void fb_deferred_io_init(struct fb_info *info)
197 {
198         struct fb_deferred_io *fbdefio = info->fbdefio;
199
200         BUG_ON(!fbdefio);
201         mutex_init(&fbdefio->lock);
202         info->fbops->fb_mmap = fb_deferred_io_mmap;
203         INIT_DELAYED_WORK(&info->deferred_work, fb_deferred_io_work);
204         INIT_LIST_HEAD(&fbdefio->pagelist);
205         if (fbdefio->delay == 0) /* set a default of 1 s */
206                 fbdefio->delay = HZ;
207 }
208 EXPORT_SYMBOL_GPL(fb_deferred_io_init);
209
210 void fb_deferred_io_open(struct fb_info *info,
211                          struct inode *inode,
212                          struct file *file)
213 {
214         file->f_mapping->a_ops = &fb_deferred_io_aops;
215 }
216 EXPORT_SYMBOL_GPL(fb_deferred_io_open);
217
218 void fb_deferred_io_cleanup(struct fb_info *info)
219 {
220         struct fb_deferred_io *fbdefio = info->fbdefio;
221         struct list_head *node, *tmp_node;
222         struct page *page;
223         int i;
224
225         BUG_ON(!fbdefio);
226         cancel_delayed_work(&info->deferred_work);
227         flush_scheduled_work();
228
229         /*  the list may have still some non-dirty pages at this point */
230         mutex_lock(&fbdefio->lock);
231         list_for_each_safe(node, tmp_node, &fbdefio->pagelist) {
232                 list_del(node);
233         }
234         mutex_unlock(&fbdefio->lock);
235
236         /* clear out the mapping that we setup */
237         for (i = 0 ; i < info->fix.smem_len; i += PAGE_SIZE) {
238                 page = fb_deferred_io_page(info, i);
239                 page->mapping = NULL;
240         }
241
242         info->fbops->fb_mmap = NULL;
243         mutex_destroy(&fbdefio->lock);
244 }
245 EXPORT_SYMBOL_GPL(fb_deferred_io_cleanup);
246
247 MODULE_LICENSE("GPL");