xref: /openbmc/linux/crypto/async_tx/async_xor.c (revision 77d84ff8)
1 /*
2  * xor offload engine api
3  *
4  * Copyright © 2006, Intel Corporation.
5  *
6  *      Dan Williams <dan.j.williams@intel.com>
7  *
8  *      with architecture considerations by:
9  *      Neil Brown <neilb@suse.de>
10  *      Jeff Garzik <jeff@garzik.org>
11  *
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.
15  *
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
19  * more details.
20  *
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.
24  *
25  */
26 #include <linux/kernel.h>
27 #include <linux/interrupt.h>
28 #include <linux/module.h>
29 #include <linux/mm.h>
30 #include <linux/dma-mapping.h>
31 #include <linux/raid/xor.h>
32 #include <linux/async_tx.h>
33 
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 dmaengine_unmap_data *unmap,
37 	     struct async_submit_ctl *submit)
38 {
39 	struct dma_device *dma = chan->device;
40 	struct dma_async_tx_descriptor *tx = NULL;
41 	dma_async_tx_callback cb_fn_orig = submit->cb_fn;
42 	void *cb_param_orig = submit->cb_param;
43 	enum async_tx_flags flags_orig = submit->flags;
44 	enum dma_ctrl_flags dma_flags = 0;
45 	int src_cnt = unmap->to_cnt;
46 	int xor_src_cnt;
47 	dma_addr_t dma_dest = unmap->addr[unmap->to_cnt];
48 	dma_addr_t *src_list = unmap->addr;
49 
50 	while (src_cnt) {
51 		dma_addr_t tmp;
52 
53 		submit->flags = flags_orig;
54 		xor_src_cnt = min(src_cnt, (int)dma->max_xor);
55 		/* if we are submitting additional xors, leave the chain open
56 		 * and clear the callback parameters
57 		 */
58 		if (src_cnt > xor_src_cnt) {
59 			submit->flags &= ~ASYNC_TX_ACK;
60 			submit->flags |= ASYNC_TX_FENCE;
61 			submit->cb_fn = NULL;
62 			submit->cb_param = NULL;
63 		} else {
64 			submit->cb_fn = cb_fn_orig;
65 			submit->cb_param = cb_param_orig;
66 		}
67 		if (submit->cb_fn)
68 			dma_flags |= DMA_PREP_INTERRUPT;
69 		if (submit->flags & ASYNC_TX_FENCE)
70 			dma_flags |= DMA_PREP_FENCE;
71 
72 		/* Drivers force forward progress in case they can not provide a
73 		 * descriptor
74 		 */
75 		tmp = src_list[0];
76 		if (src_list > unmap->addr)
77 			src_list[0] = dma_dest;
78 		tx = dma->device_prep_dma_xor(chan, dma_dest, src_list,
79 					      xor_src_cnt, unmap->len,
80 					      dma_flags);
81 		src_list[0] = tmp;
82 
83 
84 		if (unlikely(!tx))
85 			async_tx_quiesce(&submit->depend_tx);
86 
87 		/* spin wait for the preceding transactions to complete */
88 		while (unlikely(!tx)) {
89 			dma_async_issue_pending(chan);
90 			tx = dma->device_prep_dma_xor(chan, dma_dest,
91 						      src_list,
92 						      xor_src_cnt, unmap->len,
93 						      dma_flags);
94 		}
95 
96 		dma_set_unmap(tx, unmap);
97 		async_tx_submit(chan, tx, submit);
98 		submit->depend_tx = tx;
99 
100 		if (src_cnt > xor_src_cnt) {
101 			/* drop completed sources */
102 			src_cnt -= xor_src_cnt;
103 			/* use the intermediate result a source */
104 			src_cnt++;
105 			src_list += xor_src_cnt - 1;
106 		} else
107 			break;
108 	}
109 
110 	return tx;
111 }
112 
113 static void
114 do_sync_xor(struct page *dest, struct page **src_list, unsigned int offset,
115 	    int src_cnt, size_t len, struct async_submit_ctl *submit)
116 {
117 	int i;
118 	int xor_src_cnt = 0;
119 	int src_off = 0;
120 	void *dest_buf;
121 	void **srcs;
122 
123 	if (submit->scribble)
124 		srcs = submit->scribble;
125 	else
126 		srcs = (void **) src_list;
127 
128 	/* convert to buffer pointers */
129 	for (i = 0; i < src_cnt; i++)
130 		if (src_list[i])
131 			srcs[xor_src_cnt++] = page_address(src_list[i]) + offset;
132 	src_cnt = xor_src_cnt;
133 	/* set destination address */
134 	dest_buf = page_address(dest) + offset;
135 
136 	if (submit->flags & ASYNC_TX_XOR_ZERO_DST)
137 		memset(dest_buf, 0, len);
138 
139 	while (src_cnt > 0) {
140 		/* process up to 'MAX_XOR_BLOCKS' sources */
141 		xor_src_cnt = min(src_cnt, MAX_XOR_BLOCKS);
142 		xor_blocks(xor_src_cnt, len, dest_buf, &srcs[src_off]);
143 
144 		/* drop completed sources */
145 		src_cnt -= xor_src_cnt;
146 		src_off += xor_src_cnt;
147 	}
148 
149 	async_tx_sync_epilog(submit);
150 }
151 
152 /**
153  * async_xor - attempt to xor a set of blocks with a dma engine.
154  * @dest: destination page
155  * @src_list: array of source pages
156  * @offset: common src/dst offset to start transaction
157  * @src_cnt: number of source pages
158  * @len: length in bytes
159  * @submit: submission / completion modifiers
160  *
161  * honored flags: ASYNC_TX_ACK, ASYNC_TX_XOR_ZERO_DST, ASYNC_TX_XOR_DROP_DST
162  *
163  * xor_blocks always uses the dest as a source so the
164  * ASYNC_TX_XOR_ZERO_DST flag must be set to not include dest data in
165  * the calculation.  The assumption with dma eninges is that they only
166  * use the destination buffer as a source when it is explicity specified
167  * in the source list.
168  *
169  * src_list note: if the dest is also a source it must be at index zero.
170  * The contents of this array will be overwritten if a scribble region
171  * is not specified.
172  */
173 struct dma_async_tx_descriptor *
174 async_xor(struct page *dest, struct page **src_list, unsigned int offset,
175 	  int src_cnt, size_t len, struct async_submit_ctl *submit)
176 {
177 	struct dma_chan *chan = async_tx_find_channel(submit, DMA_XOR,
178 						      &dest, 1, src_list,
179 						      src_cnt, len);
180 	struct dma_device *device = chan ? chan->device : NULL;
181 	struct dmaengine_unmap_data *unmap = NULL;
182 
183 	BUG_ON(src_cnt <= 1);
184 
185 	if (device)
186 		unmap = dmaengine_get_unmap_data(device->dev, src_cnt+1, GFP_NOIO);
187 
188 	if (unmap && is_dma_xor_aligned(device, offset, 0, len)) {
189 		struct dma_async_tx_descriptor *tx;
190 		int i, j;
191 
192 		/* run the xor asynchronously */
193 		pr_debug("%s (async): len: %zu\n", __func__, len);
194 
195 		unmap->len = len;
196 		for (i = 0, j = 0; i < src_cnt; i++) {
197 			if (!src_list[i])
198 				continue;
199 			unmap->to_cnt++;
200 			unmap->addr[j++] = dma_map_page(device->dev, src_list[i],
201 							offset, len, DMA_TO_DEVICE);
202 		}
203 
204 		/* map it bidirectional as it may be re-used as a source */
205 		unmap->addr[j] = dma_map_page(device->dev, dest, offset, len,
206 					      DMA_BIDIRECTIONAL);
207 		unmap->bidi_cnt = 1;
208 
209 		tx = do_async_xor(chan, unmap, submit);
210 		dmaengine_unmap_put(unmap);
211 		return tx;
212 	} else {
213 		dmaengine_unmap_put(unmap);
214 		/* run the xor synchronously */
215 		pr_debug("%s (sync): len: %zu\n", __func__, len);
216 		WARN_ONCE(chan, "%s: no space for dma address conversion\n",
217 			  __func__);
218 
219 		/* in the sync case the dest is an implied source
220 		 * (assumes the dest is the first source)
221 		 */
222 		if (submit->flags & ASYNC_TX_XOR_DROP_DST) {
223 			src_cnt--;
224 			src_list++;
225 		}
226 
227 		/* wait for any prerequisite operations */
228 		async_tx_quiesce(&submit->depend_tx);
229 
230 		do_sync_xor(dest, src_list, offset, src_cnt, len, submit);
231 
232 		return NULL;
233 	}
234 }
235 EXPORT_SYMBOL_GPL(async_xor);
236 
237 static int page_is_zero(struct page *p, unsigned int offset, size_t len)
238 {
239 	return !memchr_inv(page_address(p) + offset, 0, len);
240 }
241 
242 static inline struct dma_chan *
243 xor_val_chan(struct async_submit_ctl *submit, struct page *dest,
244 		 struct page **src_list, int src_cnt, size_t len)
245 {
246 	#ifdef CONFIG_ASYNC_TX_DISABLE_XOR_VAL_DMA
247 	return NULL;
248 	#endif
249 	return async_tx_find_channel(submit, DMA_XOR_VAL, &dest, 1, src_list,
250 				     src_cnt, len);
251 }
252 
253 /**
254  * async_xor_val - attempt a xor parity check with a dma engine.
255  * @dest: destination page used if the xor is performed synchronously
256  * @src_list: array of source pages
257  * @offset: offset in pages to start transaction
258  * @src_cnt: number of source pages
259  * @len: length in bytes
260  * @result: 0 if sum == 0 else non-zero
261  * @submit: submission / completion modifiers
262  *
263  * honored flags: ASYNC_TX_ACK
264  *
265  * src_list note: if the dest is also a source it must be at index zero.
266  * The contents of this array will be overwritten if a scribble region
267  * is not specified.
268  */
269 struct dma_async_tx_descriptor *
270 async_xor_val(struct page *dest, struct page **src_list, unsigned int offset,
271 	      int src_cnt, size_t len, enum sum_check_flags *result,
272 	      struct async_submit_ctl *submit)
273 {
274 	struct dma_chan *chan = xor_val_chan(submit, dest, src_list, src_cnt, len);
275 	struct dma_device *device = chan ? chan->device : NULL;
276 	struct dma_async_tx_descriptor *tx = NULL;
277 	struct dmaengine_unmap_data *unmap = NULL;
278 
279 	BUG_ON(src_cnt <= 1);
280 
281 	if (device)
282 		unmap = dmaengine_get_unmap_data(device->dev, src_cnt, GFP_NOIO);
283 
284 	if (unmap && src_cnt <= device->max_xor &&
285 	    is_dma_xor_aligned(device, offset, 0, len)) {
286 		unsigned long dma_prep_flags = 0;
287 		int i;
288 
289 		pr_debug("%s: (async) len: %zu\n", __func__, len);
290 
291 		if (submit->cb_fn)
292 			dma_prep_flags |= DMA_PREP_INTERRUPT;
293 		if (submit->flags & ASYNC_TX_FENCE)
294 			dma_prep_flags |= DMA_PREP_FENCE;
295 
296 		for (i = 0; i < src_cnt; i++) {
297 			unmap->addr[i] = dma_map_page(device->dev, src_list[i],
298 						      offset, len, DMA_TO_DEVICE);
299 			unmap->to_cnt++;
300 		}
301 		unmap->len = len;
302 
303 		tx = device->device_prep_dma_xor_val(chan, unmap->addr, src_cnt,
304 						     len, result,
305 						     dma_prep_flags);
306 		if (unlikely(!tx)) {
307 			async_tx_quiesce(&submit->depend_tx);
308 
309 			while (!tx) {
310 				dma_async_issue_pending(chan);
311 				tx = device->device_prep_dma_xor_val(chan,
312 					unmap->addr, src_cnt, len, result,
313 					dma_prep_flags);
314 			}
315 		}
316 		dma_set_unmap(tx, unmap);
317 		async_tx_submit(chan, tx, submit);
318 	} else {
319 		enum async_tx_flags flags_orig = submit->flags;
320 
321 		pr_debug("%s: (sync) len: %zu\n", __func__, len);
322 		WARN_ONCE(device && src_cnt <= device->max_xor,
323 			  "%s: no space for dma address conversion\n",
324 			  __func__);
325 
326 		submit->flags |= ASYNC_TX_XOR_DROP_DST;
327 		submit->flags &= ~ASYNC_TX_ACK;
328 
329 		tx = async_xor(dest, src_list, offset, src_cnt, len, submit);
330 
331 		async_tx_quiesce(&tx);
332 
333 		*result = !page_is_zero(dest, offset, len) << SUM_CHECK_P;
334 
335 		async_tx_sync_epilog(submit);
336 		submit->flags = flags_orig;
337 	}
338 	dmaengine_unmap_put(unmap);
339 
340 	return tx;
341 }
342 EXPORT_SYMBOL_GPL(async_xor_val);
343 
344 MODULE_AUTHOR("Intel Corporation");
345 MODULE_DESCRIPTION("asynchronous xor/xor-zero-sum api");
346 MODULE_LICENSE("GPL");
347