block: optionally snapshot page contents to provide stable pages during write
[linux-3.10.git] / mm / bounce.c
1 /* bounce buffer handling for block devices
2  *
3  * - Split from highmem.c
4  */
5
6 #include <linux/mm.h>
7 #include <linux/export.h>
8 #include <linux/swap.h>
9 #include <linux/gfp.h>
10 #include <linux/bio.h>
11 #include <linux/pagemap.h>
12 #include <linux/mempool.h>
13 #include <linux/blkdev.h>
14 #include <linux/init.h>
15 #include <linux/hash.h>
16 #include <linux/highmem.h>
17 #include <linux/bootmem.h>
18 #include <asm/tlbflush.h>
19
20 #include <trace/events/block.h>
21
22 #define POOL_SIZE       64
23 #define ISA_POOL_SIZE   16
24
25 static mempool_t *page_pool, *isa_page_pool;
26
27 #if defined(CONFIG_HIGHMEM) || defined(CONFIG_NEED_BOUNCE_POOL)
28 static __init int init_emergency_pool(void)
29 {
30 #if defined(CONFIG_HIGHMEM) && !defined(CONFIG_MEMORY_HOTPLUG)
31         if (max_pfn <= max_low_pfn)
32                 return 0;
33 #endif
34
35         page_pool = mempool_create_page_pool(POOL_SIZE, 0);
36         BUG_ON(!page_pool);
37         printk("bounce pool size: %d pages\n", POOL_SIZE);
38
39         return 0;
40 }
41
42 __initcall(init_emergency_pool);
43 #endif
44
45 #ifdef CONFIG_HIGHMEM
46 /*
47  * highmem version, map in to vec
48  */
49 static void bounce_copy_vec(struct bio_vec *to, unsigned char *vfrom)
50 {
51         unsigned long flags;
52         unsigned char *vto;
53
54         local_irq_save(flags);
55         vto = kmap_atomic(to->bv_page);
56         memcpy(vto + to->bv_offset, vfrom, to->bv_len);
57         kunmap_atomic(vto);
58         local_irq_restore(flags);
59 }
60
61 #else /* CONFIG_HIGHMEM */
62
63 #define bounce_copy_vec(to, vfrom)      \
64         memcpy(page_address((to)->bv_page) + (to)->bv_offset, vfrom, (to)->bv_len)
65
66 #endif /* CONFIG_HIGHMEM */
67
68 /*
69  * allocate pages in the DMA region for the ISA pool
70  */
71 static void *mempool_alloc_pages_isa(gfp_t gfp_mask, void *data)
72 {
73         return mempool_alloc_pages(gfp_mask | GFP_DMA, data);
74 }
75
76 /*
77  * gets called "every" time someone init's a queue with BLK_BOUNCE_ISA
78  * as the max address, so check if the pool has already been created.
79  */
80 int init_emergency_isa_pool(void)
81 {
82         if (isa_page_pool)
83                 return 0;
84
85         isa_page_pool = mempool_create(ISA_POOL_SIZE, mempool_alloc_pages_isa,
86                                        mempool_free_pages, (void *) 0);
87         BUG_ON(!isa_page_pool);
88
89         printk("isa bounce pool size: %d pages\n", ISA_POOL_SIZE);
90         return 0;
91 }
92
93 /*
94  * Simple bounce buffer support for highmem pages. Depending on the
95  * queue gfp mask set, *to may or may not be a highmem page. kmap it
96  * always, it will do the Right Thing
97  */
98 static void copy_to_high_bio_irq(struct bio *to, struct bio *from)
99 {
100         unsigned char *vfrom;
101         struct bio_vec *tovec, *fromvec;
102         int i;
103
104         __bio_for_each_segment(tovec, to, i, 0) {
105                 fromvec = from->bi_io_vec + i;
106
107                 /*
108                  * not bounced
109                  */
110                 if (tovec->bv_page == fromvec->bv_page)
111                         continue;
112
113                 /*
114                  * fromvec->bv_offset and fromvec->bv_len might have been
115                  * modified by the block layer, so use the original copy,
116                  * bounce_copy_vec already uses tovec->bv_len
117                  */
118                 vfrom = page_address(fromvec->bv_page) + tovec->bv_offset;
119
120                 bounce_copy_vec(tovec, vfrom);
121                 flush_dcache_page(tovec->bv_page);
122         }
123 }
124
125 static void bounce_end_io(struct bio *bio, mempool_t *pool, int err)
126 {
127         struct bio *bio_orig = bio->bi_private;
128         struct bio_vec *bvec, *org_vec;
129         int i;
130
131         if (test_bit(BIO_EOPNOTSUPP, &bio->bi_flags))
132                 set_bit(BIO_EOPNOTSUPP, &bio_orig->bi_flags);
133
134         /*
135          * free up bounce indirect pages used
136          */
137         __bio_for_each_segment(bvec, bio, i, 0) {
138                 org_vec = bio_orig->bi_io_vec + i;
139                 if (bvec->bv_page == org_vec->bv_page)
140                         continue;
141
142                 dec_zone_page_state(bvec->bv_page, NR_BOUNCE);
143                 mempool_free(bvec->bv_page, pool);
144         }
145
146         bio_endio(bio_orig, err);
147         bio_put(bio);
148 }
149
150 static void bounce_end_io_write(struct bio *bio, int err)
151 {
152         bounce_end_io(bio, page_pool, err);
153 }
154
155 static void bounce_end_io_write_isa(struct bio *bio, int err)
156 {
157
158         bounce_end_io(bio, isa_page_pool, err);
159 }
160
161 static void __bounce_end_io_read(struct bio *bio, mempool_t *pool, int err)
162 {
163         struct bio *bio_orig = bio->bi_private;
164
165         if (test_bit(BIO_UPTODATE, &bio->bi_flags))
166                 copy_to_high_bio_irq(bio_orig, bio);
167
168         bounce_end_io(bio, pool, err);
169 }
170
171 static void bounce_end_io_read(struct bio *bio, int err)
172 {
173         __bounce_end_io_read(bio, page_pool, err);
174 }
175
176 static void bounce_end_io_read_isa(struct bio *bio, int err)
177 {
178         __bounce_end_io_read(bio, isa_page_pool, err);
179 }
180
181 #ifdef CONFIG_NEED_BOUNCE_POOL
182 static int must_snapshot_stable_pages(struct request_queue *q, struct bio *bio)
183 {
184         struct page *page;
185         struct backing_dev_info *bdi;
186         struct address_space *mapping;
187         struct bio_vec *from;
188         int i;
189
190         if (bio_data_dir(bio) != WRITE)
191                 return 0;
192
193         if (!bdi_cap_stable_pages_required(&q->backing_dev_info))
194                 return 0;
195
196         /*
197          * Based on the first page that has a valid mapping, decide whether or
198          * not we have to employ bounce buffering to guarantee stable pages.
199          */
200         bio_for_each_segment(from, bio, i) {
201                 page = from->bv_page;
202                 mapping = page_mapping(page);
203                 if (!mapping)
204                         continue;
205                 bdi = mapping->backing_dev_info;
206                 return mapping->host->i_sb->s_flags & MS_SNAP_STABLE;
207         }
208
209         return 0;
210 }
211 #else
212 static int must_snapshot_stable_pages(struct request_queue *q, struct bio *bio)
213 {
214         return 0;
215 }
216 #endif /* CONFIG_NEED_BOUNCE_POOL */
217
218 static void __blk_queue_bounce(struct request_queue *q, struct bio **bio_orig,
219                                mempool_t *pool, int force)
220 {
221         struct page *page;
222         struct bio *bio = NULL;
223         int i, rw = bio_data_dir(*bio_orig);
224         struct bio_vec *to, *from;
225
226         bio_for_each_segment(from, *bio_orig, i) {
227                 page = from->bv_page;
228
229                 /*
230                  * is destination page below bounce pfn?
231                  */
232                 if (page_to_pfn(page) <= queue_bounce_pfn(q) && !force)
233                         continue;
234
235                 /*
236                  * irk, bounce it
237                  */
238                 if (!bio) {
239                         unsigned int cnt = (*bio_orig)->bi_vcnt;
240
241                         bio = bio_alloc(GFP_NOIO, cnt);
242                         memset(bio->bi_io_vec, 0, cnt * sizeof(struct bio_vec));
243                 }
244                         
245
246                 to = bio->bi_io_vec + i;
247
248                 to->bv_page = mempool_alloc(pool, q->bounce_gfp);
249                 to->bv_len = from->bv_len;
250                 to->bv_offset = from->bv_offset;
251                 inc_zone_page_state(to->bv_page, NR_BOUNCE);
252
253                 if (rw == WRITE) {
254                         char *vto, *vfrom;
255
256                         flush_dcache_page(from->bv_page);
257                         vto = page_address(to->bv_page) + to->bv_offset;
258                         vfrom = kmap(from->bv_page) + from->bv_offset;
259                         memcpy(vto, vfrom, to->bv_len);
260                         kunmap(from->bv_page);
261                 }
262         }
263
264         /*
265          * no pages bounced
266          */
267         if (!bio)
268                 return;
269
270         trace_block_bio_bounce(q, *bio_orig);
271
272         /*
273          * at least one page was bounced, fill in possible non-highmem
274          * pages
275          */
276         __bio_for_each_segment(from, *bio_orig, i, 0) {
277                 to = bio_iovec_idx(bio, i);
278                 if (!to->bv_page) {
279                         to->bv_page = from->bv_page;
280                         to->bv_len = from->bv_len;
281                         to->bv_offset = from->bv_offset;
282                 }
283         }
284
285         bio->bi_bdev = (*bio_orig)->bi_bdev;
286         bio->bi_flags |= (1 << BIO_BOUNCED);
287         bio->bi_sector = (*bio_orig)->bi_sector;
288         bio->bi_rw = (*bio_orig)->bi_rw;
289
290         bio->bi_vcnt = (*bio_orig)->bi_vcnt;
291         bio->bi_idx = (*bio_orig)->bi_idx;
292         bio->bi_size = (*bio_orig)->bi_size;
293
294         if (pool == page_pool) {
295                 bio->bi_end_io = bounce_end_io_write;
296                 if (rw == READ)
297                         bio->bi_end_io = bounce_end_io_read;
298         } else {
299                 bio->bi_end_io = bounce_end_io_write_isa;
300                 if (rw == READ)
301                         bio->bi_end_io = bounce_end_io_read_isa;
302         }
303
304         bio->bi_private = *bio_orig;
305         *bio_orig = bio;
306 }
307
308 void blk_queue_bounce(struct request_queue *q, struct bio **bio_orig)
309 {
310         int must_bounce;
311         mempool_t *pool;
312
313         /*
314          * Data-less bio, nothing to bounce
315          */
316         if (!bio_has_data(*bio_orig))
317                 return;
318
319         must_bounce = must_snapshot_stable_pages(q, *bio_orig);
320
321         /*
322          * for non-isa bounce case, just check if the bounce pfn is equal
323          * to or bigger than the highest pfn in the system -- in that case,
324          * don't waste time iterating over bio segments
325          */
326         if (!(q->bounce_gfp & GFP_DMA)) {
327                 if (queue_bounce_pfn(q) >= blk_max_pfn && !must_bounce)
328                         return;
329                 pool = page_pool;
330         } else {
331                 BUG_ON(!isa_page_pool);
332                 pool = isa_page_pool;
333         }
334
335         /*
336          * slow path
337          */
338         __blk_queue_bounce(q, bio_orig, pool, must_bounce);
339 }
340
341 EXPORT_SYMBOL(blk_queue_bounce);