2 * xor offload engine api
4 * Copyright © 2006, Intel Corporation.
6 * Dan Williams <dan.j.williams@intel.com>
8 * with architecture considerations by:
9 * Neil Brown <neilb@suse.de>
10 * Jeff Garzik <jeff@garzik.org>
12 * This program is free software; you can redistribute it and/or modify it
13 * under the terms and conditions of the GNU General Public License,
14 * version 2, as published by the Free Software Foundation.
16 * This program is distributed in the hope it will be useful, but WITHOUT
17 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
18 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for
21 * You should have received a copy of the GNU General Public License along with
22 * this program; if not, write to the Free Software Foundation, Inc.,
23 * 51 Franklin St - Fifth Floor, Boston, MA 02110-1301 USA.
26 #include <linux/kernel.h>
27 #include <linux/interrupt.h>
28 #include <linux/module.h>
30 #include <linux/dma-mapping.h>
31 #include <linux/raid/xor.h>
32 #include <linux/async_tx.h>
34 /* do_async_xor - dma map the pages and perform the xor with an engine */
35 static __async_inline struct dma_async_tx_descriptor *
36 do_async_xor(struct dma_chan *chan, struct page *dest, struct page **src_list,
37 unsigned int offset, int src_cnt, size_t len, dma_addr_t *dma_src,
38 struct async_submit_ctl *submit)
40 struct dma_device *dma = chan->device;
41 struct dma_async_tx_descriptor *tx = NULL;
44 dma_async_tx_callback cb_fn_orig = submit->cb_fn;
45 void *cb_param_orig = submit->cb_param;
46 enum async_tx_flags flags_orig = submit->flags;
47 enum dma_ctrl_flags dma_flags;
51 /* map the dest bidrectional in case it is re-used as a source */
52 dma_dest = dma_map_page(dma->dev, dest, offset, len, DMA_BIDIRECTIONAL);
53 for (i = 0; i < src_cnt; i++) {
54 /* only map the dest once */
57 if (unlikely(src_list[i] == dest)) {
58 dma_src[xor_src_cnt++] = dma_dest;
61 dma_src[xor_src_cnt++] = dma_map_page(dma->dev, src_list[i], offset,
64 src_cnt = xor_src_cnt;
67 submit->flags = flags_orig;
69 xor_src_cnt = min(src_cnt, (int)dma->max_xor);
70 /* if we are submitting additional xors, leave the chain open,
71 * clear the callback parameters, and leave the destination
74 if (src_cnt > xor_src_cnt) {
75 submit->flags &= ~ASYNC_TX_ACK;
76 submit->flags |= ASYNC_TX_FENCE;
77 dma_flags = DMA_COMPL_SKIP_DEST_UNMAP;
79 submit->cb_param = NULL;
81 submit->cb_fn = cb_fn_orig;
82 submit->cb_param = cb_param_orig;
85 dma_flags |= DMA_PREP_INTERRUPT;
86 if (submit->flags & ASYNC_TX_FENCE)
87 dma_flags |= DMA_PREP_FENCE;
88 /* Since we have clobbered the src_list we are committed
89 * to doing this asynchronously. Drivers force forward progress
90 * in case they can not provide a descriptor
92 tx = dma->device_prep_dma_xor(chan, dma_dest, &dma_src[src_off],
93 xor_src_cnt, len, dma_flags);
96 async_tx_quiesce(&submit->depend_tx);
98 /* spin wait for the preceding transactions to complete */
99 while (unlikely(!tx)) {
100 dma_async_issue_pending(chan);
101 tx = dma->device_prep_dma_xor(chan, dma_dest,
107 async_tx_submit(chan, tx, submit);
108 submit->depend_tx = tx;
110 if (src_cnt > xor_src_cnt) {
111 /* drop completed sources */
112 src_cnt -= xor_src_cnt;
113 src_off += xor_src_cnt;
115 /* use the intermediate result a source */
116 dma_src[--src_off] = dma_dest;
126 do_sync_xor(struct page *dest, struct page **src_list, unsigned int offset,
127 int src_cnt, size_t len, struct async_submit_ctl *submit)
135 if (submit->scribble)
136 srcs = submit->scribble;
138 srcs = (void **) src_list;
140 /* convert to buffer pointers */
141 for (i = 0; i < src_cnt; i++)
143 srcs[xor_src_cnt++] = page_address(src_list[i]) + offset;
144 src_cnt = xor_src_cnt;
145 /* set destination address */
146 dest_buf = page_address(dest) + offset;
148 if (submit->flags & ASYNC_TX_XOR_ZERO_DST)
149 memset(dest_buf, 0, len);
151 while (src_cnt > 0) {
152 /* process up to 'MAX_XOR_BLOCKS' sources */
153 xor_src_cnt = min(src_cnt, MAX_XOR_BLOCKS);
154 xor_blocks(xor_src_cnt, len, dest_buf, &srcs[src_off]);
156 /* drop completed sources */
157 src_cnt -= xor_src_cnt;
158 src_off += xor_src_cnt;
161 async_tx_sync_epilog(submit);
165 * async_xor - attempt to xor a set of blocks with a dma engine.
166 * @dest: destination page
167 * @src_list: array of source pages
168 * @offset: common src/dst offset to start transaction
169 * @src_cnt: number of source pages
170 * @len: length in bytes
171 * @submit: submission / completion modifiers
173 * honored flags: ASYNC_TX_ACK, ASYNC_TX_XOR_ZERO_DST, ASYNC_TX_XOR_DROP_DST
175 * xor_blocks always uses the dest as a source so the
176 * ASYNC_TX_XOR_ZERO_DST flag must be set to not include dest data in
177 * the calculation. The assumption with dma eninges is that they only
178 * use the destination buffer as a source when it is explicity specified
179 * in the source list.
181 * src_list note: if the dest is also a source it must be at index zero.
182 * The contents of this array will be overwritten if a scribble region
185 struct dma_async_tx_descriptor *
186 async_xor(struct page *dest, struct page **src_list, unsigned int offset,
187 int src_cnt, size_t len, struct async_submit_ctl *submit)
189 struct dma_chan *chan = async_tx_find_channel(submit, DMA_XOR,
192 dma_addr_t *dma_src = NULL;
194 BUG_ON(src_cnt <= 1);
196 if (submit->scribble)
197 dma_src = submit->scribble;
198 else if (sizeof(dma_addr_t) <= sizeof(struct page *))
199 dma_src = (dma_addr_t *) src_list;
201 if (dma_src && chan && is_dma_xor_aligned(chan->device, offset, 0, len)) {
202 /* run the xor asynchronously */
203 pr_debug("%s (async): len: %zu\n", __func__, len);
205 return do_async_xor(chan, dest, src_list, offset, src_cnt, len,
208 /* run the xor synchronously */
209 pr_debug("%s (sync): len: %zu\n", __func__, len);
210 WARN_ONCE(chan, "%s: no space for dma address conversion\n",
213 /* in the sync case the dest is an implied source
214 * (assumes the dest is the first source)
216 if (submit->flags & ASYNC_TX_XOR_DROP_DST) {
221 /* wait for any prerequisite operations */
222 async_tx_quiesce(&submit->depend_tx);
224 do_sync_xor(dest, src_list, offset, src_cnt, len, submit);
229 EXPORT_SYMBOL_GPL(async_xor);
231 static int page_is_zero(struct page *p, unsigned int offset, size_t len)
233 char *a = page_address(p) + offset;
234 return ((*(u32 *) a) == 0 &&
235 memcmp(a, a + 4, len - 4) == 0);
238 static inline struct dma_chan *
239 xor_val_chan(struct async_submit_ctl *submit, struct page *dest,
240 struct page **src_list, int src_cnt, size_t len)
242 #ifdef CONFIG_ASYNC_TX_DISABLE_XOR_VAL_DMA
245 return async_tx_find_channel(submit, DMA_XOR_VAL, &dest, 1, src_list,
250 * async_xor_val - attempt a xor parity check with a dma engine.
251 * @dest: destination page used if the xor is performed synchronously
252 * @src_list: array of source pages
253 * @offset: offset in pages to start transaction
254 * @src_cnt: number of source pages
255 * @len: length in bytes
256 * @result: 0 if sum == 0 else non-zero
257 * @submit: submission / completion modifiers
259 * honored flags: ASYNC_TX_ACK
261 * src_list note: if the dest is also a source it must be at index zero.
262 * The contents of this array will be overwritten if a scribble region
265 struct dma_async_tx_descriptor *
266 async_xor_val(struct page *dest, struct page **src_list, unsigned int offset,
267 int src_cnt, size_t len, enum sum_check_flags *result,
268 struct async_submit_ctl *submit)
270 struct dma_chan *chan = xor_val_chan(submit, dest, src_list, src_cnt, len);
271 struct dma_device *device = chan ? chan->device : NULL;
272 struct dma_async_tx_descriptor *tx = NULL;
273 dma_addr_t *dma_src = NULL;
275 BUG_ON(src_cnt <= 1);
277 if (submit->scribble)
278 dma_src = submit->scribble;
279 else if (sizeof(dma_addr_t) <= sizeof(struct page *))
280 dma_src = (dma_addr_t *) src_list;
282 if (dma_src && device && src_cnt <= device->max_xor &&
283 is_dma_xor_aligned(device, offset, 0, len)) {
284 unsigned long dma_prep_flags = 0;
287 pr_debug("%s: (async) len: %zu\n", __func__, len);
290 dma_prep_flags |= DMA_PREP_INTERRUPT;
291 if (submit->flags & ASYNC_TX_FENCE)
292 dma_prep_flags |= DMA_PREP_FENCE;
293 for (i = 0; i < src_cnt; i++)
294 dma_src[i] = dma_map_page(device->dev, src_list[i],
295 offset, len, DMA_TO_DEVICE);
297 tx = device->device_prep_dma_xor_val(chan, dma_src, src_cnt,
301 async_tx_quiesce(&submit->depend_tx);
304 dma_async_issue_pending(chan);
305 tx = device->device_prep_dma_xor_val(chan,
306 dma_src, src_cnt, len, result,
311 async_tx_submit(chan, tx, submit);
313 enum async_tx_flags flags_orig = submit->flags;
315 pr_debug("%s: (sync) len: %zu\n", __func__, len);
316 WARN_ONCE(device && src_cnt <= device->max_xor,
317 "%s: no space for dma address conversion\n",
320 submit->flags |= ASYNC_TX_XOR_DROP_DST;
321 submit->flags &= ~ASYNC_TX_ACK;
323 tx = async_xor(dest, src_list, offset, src_cnt, len, submit);
325 async_tx_quiesce(&tx);
327 *result = !page_is_zero(dest, offset, len) << SUM_CHECK_P;
329 async_tx_sync_epilog(submit);
330 submit->flags = flags_orig;
335 EXPORT_SYMBOL_GPL(async_xor_val);
337 MODULE_AUTHOR("Intel Corporation");
338 MODULE_DESCRIPTION("asynchronous xor/xor-zero-sum api");
339 MODULE_LICENSE("GPL");