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>
29 #include <linux/dma-mapping.h>
30 #include <linux/raid/xor.h>
31 #include <linux/async_tx.h>
34 * async_tx_quiesce - ensure tx is complete and freeable upon return
35 * @tx - transaction to quiesce
37 static void async_tx_quiesce(struct dma_async_tx_descriptor **tx)
40 /* if ack is already set then we cannot be sure
41 * we are referring to the correct operation
43 BUG_ON(async_tx_test_ack(*tx));
44 if (dma_wait_for_async_tx(*tx) == DMA_ERROR)
45 panic("DMA_ERROR waiting for transaction\n");
51 /* do_async_xor - dma map the pages and perform the xor with an engine.
52 * This routine is marked __always_inline so it can be compiled away
53 * when CONFIG_DMA_ENGINE=n
55 static __always_inline struct dma_async_tx_descriptor *
56 do_async_xor(struct dma_chan *chan, struct page *dest, struct page **src_list,
57 unsigned int offset, int src_cnt, size_t len,
58 enum async_tx_flags flags,
59 struct dma_async_tx_descriptor *depend_tx,
60 dma_async_tx_callback cb_fn, void *cb_param)
62 struct dma_device *dma = chan->device;
63 dma_addr_t *dma_src = (dma_addr_t *) src_list;
64 struct dma_async_tx_descriptor *tx = NULL;
67 dma_async_tx_callback _cb_fn;
69 enum async_tx_flags async_flags;
70 enum dma_ctrl_flags dma_flags;
74 dma_dest = dma_map_page(dma->dev, dest, offset, len, DMA_FROM_DEVICE);
75 for (i = 0; i < src_cnt; i++)
76 dma_src[i] = dma_map_page(dma->dev, src_list[i], offset,
82 xor_src_cnt = min(src_cnt, dma->max_xor);
83 /* if we are submitting additional xors, leave the chain open,
84 * clear the callback parameters, and leave the destination
87 if (src_cnt > xor_src_cnt) {
88 async_flags &= ~ASYNC_TX_ACK;
89 dma_flags = DMA_COMPL_SKIP_DEST_UNMAP;
97 dma_flags |= DMA_PREP_INTERRUPT;
99 /* Since we have clobbered the src_list we are committed
100 * to doing this asynchronously. Drivers force forward progress
101 * in case they can not provide a descriptor
103 tx = dma->device_prep_dma_xor(chan, dma_dest, &dma_src[src_off],
104 xor_src_cnt, len, dma_flags);
107 async_tx_quiesce(&depend_tx);
109 /* spin wait for the preceeding transactions to complete */
110 while (unlikely(!tx)) {
111 dma_async_issue_pending(chan);
112 tx = dma->device_prep_dma_xor(chan, dma_dest,
118 async_tx_submit(chan, tx, async_flags, depend_tx, _cb_fn,
122 flags |= ASYNC_TX_DEP_ACK;
124 if (src_cnt > xor_src_cnt) {
125 /* drop completed sources */
126 src_cnt -= xor_src_cnt;
127 src_off += xor_src_cnt;
129 /* use the intermediate result a source */
130 dma_src[--src_off] = dma_dest;
140 do_sync_xor(struct page *dest, struct page **src_list, unsigned int offset,
141 int src_cnt, size_t len, enum async_tx_flags flags,
142 struct dma_async_tx_descriptor *depend_tx,
143 dma_async_tx_callback cb_fn, void *cb_param)
149 void **srcs = (void **) src_list;
151 /* reuse the 'src_list' array to convert to buffer pointers */
152 for (i = 0; i < src_cnt; i++)
153 srcs[i] = page_address(src_list[i]) + offset;
155 /* set destination address */
156 dest_buf = page_address(dest) + offset;
158 if (flags & ASYNC_TX_XOR_ZERO_DST)
159 memset(dest_buf, 0, len);
161 while (src_cnt > 0) {
162 /* process up to 'MAX_XOR_BLOCKS' sources */
163 xor_src_cnt = min(src_cnt, MAX_XOR_BLOCKS);
164 xor_blocks(xor_src_cnt, len, dest_buf, &srcs[src_off]);
166 /* drop completed sources */
167 src_cnt -= xor_src_cnt;
168 src_off += xor_src_cnt;
171 async_tx_sync_epilog(flags, depend_tx, cb_fn, cb_param);
175 * async_xor - attempt to xor a set of blocks with a dma engine.
176 * xor_blocks always uses the dest as a source so the ASYNC_TX_XOR_ZERO_DST
177 * flag must be set to not include dest data in the calculation. The
178 * assumption with dma eninges is that they only use the destination
179 * buffer as a source when it is explicity specified in the source list.
180 * @dest: destination page
181 * @src_list: array of source pages (if the dest is also a source it must be
182 * at index zero). The contents of this array may be overwritten.
183 * @offset: offset in pages to start transaction
184 * @src_cnt: number of source pages
185 * @len: length in bytes
186 * @flags: ASYNC_TX_XOR_ZERO_DST, ASYNC_TX_XOR_DROP_DEST,
187 * ASYNC_TX_ACK, ASYNC_TX_DEP_ACK
188 * @depend_tx: xor depends on the result of this transaction.
189 * @cb_fn: function to call when the xor completes
190 * @cb_param: parameter to pass to the callback routine
192 struct dma_async_tx_descriptor *
193 async_xor(struct page *dest, struct page **src_list, unsigned int offset,
194 int src_cnt, size_t len, enum async_tx_flags flags,
195 struct dma_async_tx_descriptor *depend_tx,
196 dma_async_tx_callback cb_fn, void *cb_param)
198 struct dma_chan *chan = async_tx_find_channel(depend_tx, DMA_XOR,
201 BUG_ON(src_cnt <= 1);
204 /* run the xor asynchronously */
205 pr_debug("%s (async): len: %zu\n", __func__, len);
207 return do_async_xor(chan, dest, src_list, offset, src_cnt, len,
208 flags, depend_tx, cb_fn, cb_param);
210 /* run the xor synchronously */
211 pr_debug("%s (sync): len: %zu\n", __func__, len);
213 /* in the sync case the dest is an implied source
214 * (assumes the dest is the first source)
216 if (flags & ASYNC_TX_XOR_DROP_DST) {
221 /* wait for any prerequisite operations */
223 /* if ack is already set then we cannot be sure
224 * we are referring to the correct operation
226 BUG_ON(async_tx_test_ack(depend_tx));
227 if (dma_wait_for_async_tx(depend_tx) == DMA_ERROR)
228 panic("%s: DMA_ERROR waiting for depend_tx\n",
232 do_sync_xor(dest, src_list, offset, src_cnt, len,
233 flags, depend_tx, cb_fn, cb_param);
238 EXPORT_SYMBOL_GPL(async_xor);
240 static int page_is_zero(struct page *p, unsigned int offset, size_t len)
242 char *a = page_address(p) + offset;
243 return ((*(u32 *) a) == 0 &&
244 memcmp(a, a + 4, len - 4) == 0);
248 * async_xor_zero_sum - attempt a xor parity check with a dma engine.
249 * @dest: destination page used if the xor is performed synchronously
250 * @src_list: array of source pages. The dest page must be listed as a source
251 * at index zero. The contents of this array may be overwritten.
252 * @offset: offset in pages to start transaction
253 * @src_cnt: number of source pages
254 * @len: length in bytes
255 * @result: 0 if sum == 0 else non-zero
256 * @flags: ASYNC_TX_ACK, ASYNC_TX_DEP_ACK
257 * @depend_tx: xor depends on the result of this transaction.
258 * @cb_fn: function to call when the xor completes
259 * @cb_param: parameter to pass to the callback routine
261 struct dma_async_tx_descriptor *
262 async_xor_zero_sum(struct page *dest, struct page **src_list,
263 unsigned int offset, int src_cnt, size_t len,
264 u32 *result, enum async_tx_flags flags,
265 struct dma_async_tx_descriptor *depend_tx,
266 dma_async_tx_callback cb_fn, void *cb_param)
268 struct dma_chan *chan = async_tx_find_channel(depend_tx, DMA_ZERO_SUM,
271 struct dma_device *device = chan ? chan->device : NULL;
272 struct dma_async_tx_descriptor *tx = NULL;
274 BUG_ON(src_cnt <= 1);
276 if (device && src_cnt <= device->max_xor) {
277 dma_addr_t *dma_src = (dma_addr_t *) src_list;
278 unsigned long dma_prep_flags = cb_fn ? DMA_PREP_INTERRUPT : 0;
281 pr_debug("%s: (async) len: %zu\n", __func__, len);
283 for (i = 0; i < src_cnt; i++)
284 dma_src[i] = dma_map_page(device->dev, src_list[i],
285 offset, len, DMA_TO_DEVICE);
287 tx = device->device_prep_dma_zero_sum(chan, dma_src, src_cnt,
291 async_tx_quiesce(&depend_tx);
294 dma_async_issue_pending(chan);
295 tx = device->device_prep_dma_zero_sum(chan,
296 dma_src, src_cnt, len, result,
300 async_tx_submit(chan, tx, flags, depend_tx, cb_fn, cb_param);
302 unsigned long xor_flags = flags;
304 pr_debug("%s: (sync) len: %zu\n", __func__, len);
306 xor_flags |= ASYNC_TX_XOR_DROP_DST;
307 xor_flags &= ~ASYNC_TX_ACK;
309 tx = async_xor(dest, src_list, offset, src_cnt, len, xor_flags,
310 depend_tx, NULL, NULL);
313 if (dma_wait_for_async_tx(tx) == DMA_ERROR)
314 panic("%s: DMA_ERROR waiting for tx\n",
319 *result = page_is_zero(dest, offset, len) ? 0 : 1;
323 async_tx_sync_epilog(flags, depend_tx, cb_fn, cb_param);
328 EXPORT_SYMBOL_GPL(async_xor_zero_sum);
330 static int __init async_xor_init(void)
332 #ifdef CONFIG_DMA_ENGINE
333 /* To conserve stack space the input src_list (array of page pointers)
334 * is reused to hold the array of dma addresses passed to the driver.
335 * This conversion is only possible when dma_addr_t is less than the
336 * the size of a pointer. HIGHMEM64G is known to violate this
339 BUILD_BUG_ON(sizeof(dma_addr_t) > sizeof(struct page *));
345 static void __exit async_xor_exit(void)
350 module_init(async_xor_init);
351 module_exit(async_xor_exit);
353 MODULE_AUTHOR("Intel Corporation");
354 MODULE_DESCRIPTION("asynchronous xor/xor-zero-sum api");
355 MODULE_LICENSE("GPL");