xref: /openbmc/linux/crypto/async_tx/async_xor.c (revision 54525552)
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/mm.h>
29 #include <linux/dma-mapping.h>
30 #include <linux/raid/xor.h>
31 #include <linux/async_tx.h>
32 
33 /* do_async_xor - dma map the pages and perform the xor with an engine */
34 static __async_inline struct dma_async_tx_descriptor *
35 do_async_xor(struct dma_chan *chan, struct page *dest, struct page **src_list,
36 	     unsigned int offset, int src_cnt, size_t len, dma_addr_t *dma_src,
37 	     struct async_submit_ctl *submit)
38 {
39 	struct dma_device *dma = chan->device;
40 	struct dma_async_tx_descriptor *tx = NULL;
41 	int src_off = 0;
42 	int i;
43 	dma_async_tx_callback cb_fn_orig = submit->cb_fn;
44 	void *cb_param_orig = submit->cb_param;
45 	enum async_tx_flags flags_orig = submit->flags;
46 	enum dma_ctrl_flags dma_flags;
47 	int xor_src_cnt = 0;
48 	dma_addr_t dma_dest;
49 
50 	/* map the dest bidrectional in case it is re-used as a source */
51 	dma_dest = dma_map_page(dma->dev, dest, offset, len, DMA_BIDIRECTIONAL);
52 	for (i = 0; i < src_cnt; i++) {
53 		/* only map the dest once */
54 		if (!src_list[i])
55 			continue;
56 		if (unlikely(src_list[i] == dest)) {
57 			dma_src[xor_src_cnt++] = dma_dest;
58 			continue;
59 		}
60 		dma_src[xor_src_cnt++] = dma_map_page(dma->dev, src_list[i], offset,
61 						      len, DMA_TO_DEVICE);
62 	}
63 	src_cnt = xor_src_cnt;
64 
65 	while (src_cnt) {
66 		submit->flags = flags_orig;
67 		dma_flags = 0;
68 		xor_src_cnt = min(src_cnt, (int)dma->max_xor);
69 		/* if we are submitting additional xors, leave the chain open,
70 		 * clear the callback parameters, and leave the destination
71 		 * buffer mapped
72 		 */
73 		if (src_cnt > xor_src_cnt) {
74 			submit->flags &= ~ASYNC_TX_ACK;
75 			submit->flags |= ASYNC_TX_FENCE;
76 			dma_flags = DMA_COMPL_SKIP_DEST_UNMAP;
77 			submit->cb_fn = NULL;
78 			submit->cb_param = NULL;
79 		} else {
80 			submit->cb_fn = cb_fn_orig;
81 			submit->cb_param = cb_param_orig;
82 		}
83 		if (submit->cb_fn)
84 			dma_flags |= DMA_PREP_INTERRUPT;
85 		if (submit->flags & ASYNC_TX_FENCE)
86 			dma_flags |= DMA_PREP_FENCE;
87 		/* Since we have clobbered the src_list we are committed
88 		 * to doing this asynchronously.  Drivers force forward progress
89 		 * in case they can not provide a descriptor
90 		 */
91 		tx = dma->device_prep_dma_xor(chan, dma_dest, &dma_src[src_off],
92 					      xor_src_cnt, len, dma_flags);
93 
94 		if (unlikely(!tx))
95 			async_tx_quiesce(&submit->depend_tx);
96 
97 		/* spin wait for the preceding transactions to complete */
98 		while (unlikely(!tx)) {
99 			dma_async_issue_pending(chan);
100 			tx = dma->device_prep_dma_xor(chan, dma_dest,
101 						      &dma_src[src_off],
102 						      xor_src_cnt, len,
103 						      dma_flags);
104 		}
105 
106 		async_tx_submit(chan, tx, submit);
107 		submit->depend_tx = tx;
108 
109 		if (src_cnt > xor_src_cnt) {
110 			/* drop completed sources */
111 			src_cnt -= xor_src_cnt;
112 			src_off += xor_src_cnt;
113 
114 			/* use the intermediate result a source */
115 			dma_src[--src_off] = dma_dest;
116 			src_cnt++;
117 		} else
118 			break;
119 	}
120 
121 	return tx;
122 }
123 
124 static void
125 do_sync_xor(struct page *dest, struct page **src_list, unsigned int offset,
126 	    int src_cnt, size_t len, struct async_submit_ctl *submit)
127 {
128 	int i;
129 	int xor_src_cnt = 0;
130 	int src_off = 0;
131 	void *dest_buf;
132 	void **srcs;
133 
134 	if (submit->scribble)
135 		srcs = submit->scribble;
136 	else
137 		srcs = (void **) src_list;
138 
139 	/* convert to buffer pointers */
140 	for (i = 0; i < src_cnt; i++)
141 		if (src_list[i])
142 			srcs[xor_src_cnt++] = page_address(src_list[i]) + offset;
143 	src_cnt = xor_src_cnt;
144 	/* set destination address */
145 	dest_buf = page_address(dest) + offset;
146 
147 	if (submit->flags & ASYNC_TX_XOR_ZERO_DST)
148 		memset(dest_buf, 0, len);
149 
150 	while (src_cnt > 0) {
151 		/* process up to 'MAX_XOR_BLOCKS' sources */
152 		xor_src_cnt = min(src_cnt, MAX_XOR_BLOCKS);
153 		xor_blocks(xor_src_cnt, len, dest_buf, &srcs[src_off]);
154 
155 		/* drop completed sources */
156 		src_cnt -= xor_src_cnt;
157 		src_off += xor_src_cnt;
158 	}
159 
160 	async_tx_sync_epilog(submit);
161 }
162 
163 /**
164  * async_xor - attempt to xor a set of blocks with a dma engine.
165  * @dest: destination page
166  * @src_list: array of source pages
167  * @offset: common src/dst offset to start transaction
168  * @src_cnt: number of source pages
169  * @len: length in bytes
170  * @submit: submission / completion modifiers
171  *
172  * honored flags: ASYNC_TX_ACK, ASYNC_TX_XOR_ZERO_DST, ASYNC_TX_XOR_DROP_DST
173  *
174  * xor_blocks always uses the dest as a source so the
175  * ASYNC_TX_XOR_ZERO_DST flag must be set to not include dest data in
176  * the calculation.  The assumption with dma eninges is that they only
177  * use the destination buffer as a source when it is explicity specified
178  * in the source list.
179  *
180  * src_list note: if the dest is also a source it must be at index zero.
181  * The contents of this array will be overwritten if a scribble region
182  * is not specified.
183  */
184 struct dma_async_tx_descriptor *
185 async_xor(struct page *dest, struct page **src_list, unsigned int offset,
186 	  int src_cnt, size_t len, struct async_submit_ctl *submit)
187 {
188 	struct dma_chan *chan = async_tx_find_channel(submit, DMA_XOR,
189 						      &dest, 1, src_list,
190 						      src_cnt, len);
191 	dma_addr_t *dma_src = NULL;
192 
193 	BUG_ON(src_cnt <= 1);
194 
195 	if (submit->scribble)
196 		dma_src = submit->scribble;
197 	else if (sizeof(dma_addr_t) <= sizeof(struct page *))
198 		dma_src = (dma_addr_t *) src_list;
199 
200 	if (dma_src && chan && is_dma_xor_aligned(chan->device, offset, 0, len)) {
201 		/* run the xor asynchronously */
202 		pr_debug("%s (async): len: %zu\n", __func__, len);
203 
204 		return do_async_xor(chan, dest, src_list, offset, src_cnt, len,
205 				    dma_src, submit);
206 	} else {
207 		/* run the xor synchronously */
208 		pr_debug("%s (sync): len: %zu\n", __func__, len);
209 		WARN_ONCE(chan, "%s: no space for dma address conversion\n",
210 			  __func__);
211 
212 		/* in the sync case the dest is an implied source
213 		 * (assumes the dest is the first source)
214 		 */
215 		if (submit->flags & ASYNC_TX_XOR_DROP_DST) {
216 			src_cnt--;
217 			src_list++;
218 		}
219 
220 		/* wait for any prerequisite operations */
221 		async_tx_quiesce(&submit->depend_tx);
222 
223 		do_sync_xor(dest, src_list, offset, src_cnt, len, submit);
224 
225 		return NULL;
226 	}
227 }
228 EXPORT_SYMBOL_GPL(async_xor);
229 
230 static int page_is_zero(struct page *p, unsigned int offset, size_t len)
231 {
232 	char *a = page_address(p) + offset;
233 	return ((*(u32 *) a) == 0 &&
234 		memcmp(a, a + 4, len - 4) == 0);
235 }
236 
237 static inline struct dma_chan *
238 xor_val_chan(struct async_submit_ctl *submit, struct page *dest,
239 		 struct page **src_list, int src_cnt, size_t len)
240 {
241 	#ifdef CONFIG_ASYNC_TX_DISABLE_XOR_VAL_DMA
242 	return NULL;
243 	#endif
244 	return async_tx_find_channel(submit, DMA_XOR_VAL, &dest, 1, src_list,
245 				     src_cnt, len);
246 }
247 
248 /**
249  * async_xor_val - attempt a xor parity check with a dma engine.
250  * @dest: destination page used if the xor is performed synchronously
251  * @src_list: array of source pages
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  * @submit: submission / completion modifiers
257  *
258  * honored flags: ASYNC_TX_ACK
259  *
260  * src_list note: if the dest is also a source it must be at index zero.
261  * The contents of this array will be overwritten if a scribble region
262  * is not specified.
263  */
264 struct dma_async_tx_descriptor *
265 async_xor_val(struct page *dest, struct page **src_list, unsigned int offset,
266 	      int src_cnt, size_t len, enum sum_check_flags *result,
267 	      struct async_submit_ctl *submit)
268 {
269 	struct dma_chan *chan = xor_val_chan(submit, dest, src_list, src_cnt, len);
270 	struct dma_device *device = chan ? chan->device : NULL;
271 	struct dma_async_tx_descriptor *tx = NULL;
272 	dma_addr_t *dma_src = NULL;
273 
274 	BUG_ON(src_cnt <= 1);
275 
276 	if (submit->scribble)
277 		dma_src = submit->scribble;
278 	else if (sizeof(dma_addr_t) <= sizeof(struct page *))
279 		dma_src = (dma_addr_t *) src_list;
280 
281 	if (dma_src && device && src_cnt <= device->max_xor &&
282 	    is_dma_xor_aligned(device, offset, 0, len)) {
283 		unsigned long dma_prep_flags = 0;
284 		int i;
285 
286 		pr_debug("%s: (async) len: %zu\n", __func__, len);
287 
288 		if (submit->cb_fn)
289 			dma_prep_flags |= DMA_PREP_INTERRUPT;
290 		if (submit->flags & ASYNC_TX_FENCE)
291 			dma_prep_flags |= DMA_PREP_FENCE;
292 		for (i = 0; i < src_cnt; i++)
293 			dma_src[i] = dma_map_page(device->dev, src_list[i],
294 						  offset, len, DMA_TO_DEVICE);
295 
296 		tx = device->device_prep_dma_xor_val(chan, dma_src, src_cnt,
297 						     len, result,
298 						     dma_prep_flags);
299 		if (unlikely(!tx)) {
300 			async_tx_quiesce(&submit->depend_tx);
301 
302 			while (!tx) {
303 				dma_async_issue_pending(chan);
304 				tx = device->device_prep_dma_xor_val(chan,
305 					dma_src, src_cnt, len, result,
306 					dma_prep_flags);
307 			}
308 		}
309 
310 		async_tx_submit(chan, tx, submit);
311 	} else {
312 		enum async_tx_flags flags_orig = submit->flags;
313 
314 		pr_debug("%s: (sync) len: %zu\n", __func__, len);
315 		WARN_ONCE(device && src_cnt <= device->max_xor,
316 			  "%s: no space for dma address conversion\n",
317 			  __func__);
318 
319 		submit->flags |= ASYNC_TX_XOR_DROP_DST;
320 		submit->flags &= ~ASYNC_TX_ACK;
321 
322 		tx = async_xor(dest, src_list, offset, src_cnt, len, submit);
323 
324 		async_tx_quiesce(&tx);
325 
326 		*result = !page_is_zero(dest, offset, len) << SUM_CHECK_P;
327 
328 		async_tx_sync_epilog(submit);
329 		submit->flags = flags_orig;
330 	}
331 
332 	return tx;
333 }
334 EXPORT_SYMBOL_GPL(async_xor_val);
335 
336 MODULE_AUTHOR("Intel Corporation");
337 MODULE_DESCRIPTION("asynchronous xor/xor-zero-sum api");
338 MODULE_LICENSE("GPL");
339