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 *
do_async_xor(struct dma_chan * chan,struct dmaengine_unmap_data * unmap,struct async_submit_ctl * submit)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 
82 		if (unlikely(!tx))
83 			async_tx_quiesce(&submit->depend_tx);
84 
85 		/* spin wait for the preceding transactions to complete */
86 		while (unlikely(!tx)) {
87 			dma_async_issue_pending(chan);
88 			tx = dma->device_prep_dma_xor(chan, dma_dest,
89 						      src_list,
90 						      xor_src_cnt, unmap->len,
91 						      dma_flags);
92 		}
93 		src_list[0] = tmp;
94 
95 		dma_set_unmap(tx, unmap);
96 		async_tx_submit(chan, tx, submit);
97 		submit->depend_tx = tx;
98 
99 		if (src_cnt > xor_src_cnt) {
100 			/* drop completed sources */
101 			src_cnt -= xor_src_cnt;
102 			/* use the intermediate result a source */
103 			src_cnt++;
104 			src_list += xor_src_cnt - 1;
105 		} else
106 			break;
107 	}
108 
109 	return tx;
110 }
111 
112 static void
do_sync_xor(struct page * dest,struct page ** src_list,unsigned int offset,int src_cnt,size_t len,struct async_submit_ctl * submit)113 do_sync_xor(struct page *dest, struct page **src_list, unsigned int offset,
114 	    int src_cnt, size_t len, struct async_submit_ctl *submit)
115 {
116 	int i;
117 	int xor_src_cnt = 0;
118 	int src_off = 0;
119 	void *dest_buf;
120 	void **srcs;
121 
122 	if (submit->scribble)
123 		srcs = submit->scribble;
124 	else
125 		srcs = (void **) src_list;
126 
127 	/* convert to buffer pointers */
128 	for (i = 0; i < src_cnt; i++)
129 		if (src_list[i])
130 			srcs[xor_src_cnt++] = page_address(src_list[i]) + offset;
131 	src_cnt = xor_src_cnt;
132 	/* set destination address */
133 	dest_buf = page_address(dest) + offset;
134 
135 	if (submit->flags & ASYNC_TX_XOR_ZERO_DST)
136 		memset(dest_buf, 0, len);
137 
138 	while (src_cnt > 0) {
139 		/* process up to 'MAX_XOR_BLOCKS' sources */
140 		xor_src_cnt = min(src_cnt, MAX_XOR_BLOCKS);
141 		xor_blocks(xor_src_cnt, len, dest_buf, &srcs[src_off]);
142 
143 		/* drop completed sources */
144 		src_cnt -= xor_src_cnt;
145 		src_off += xor_src_cnt;
146 	}
147 
148 	async_tx_sync_epilog(submit);
149 }
150 
151 /**
152  * async_xor - attempt to xor a set of blocks with a dma engine.
153  * @dest: destination page
154  * @src_list: array of source pages
155  * @offset: common src/dst offset to start transaction
156  * @src_cnt: number of source pages
157  * @len: length in bytes
158  * @submit: submission / completion modifiers
159  *
160  * honored flags: ASYNC_TX_ACK, ASYNC_TX_XOR_ZERO_DST, ASYNC_TX_XOR_DROP_DST
161  *
162  * xor_blocks always uses the dest as a source so the
163  * ASYNC_TX_XOR_ZERO_DST flag must be set to not include dest data in
164  * the calculation.  The assumption with dma eninges is that they only
165  * use the destination buffer as a source when it is explicity specified
166  * in the source list.
167  *
168  * src_list note: if the dest is also a source it must be at index zero.
169  * The contents of this array will be overwritten if a scribble region
170  * is not specified.
171  */
172 struct dma_async_tx_descriptor *
async_xor(struct page * dest,struct page ** src_list,unsigned int offset,int src_cnt,size_t len,struct async_submit_ctl * submit)173 async_xor(struct page *dest, struct page **src_list, unsigned int offset,
174 	  int src_cnt, size_t len, struct async_submit_ctl *submit)
175 {
176 	struct dma_chan *chan = async_tx_find_channel(submit, DMA_XOR,
177 						      &dest, 1, src_list,
178 						      src_cnt, len);
179 	struct dma_device *device = chan ? chan->device : NULL;
180 	struct dmaengine_unmap_data *unmap = NULL;
181 
182 	BUG_ON(src_cnt <= 1);
183 
184 	if (device)
185 		unmap = dmaengine_get_unmap_data(device->dev, src_cnt+1, GFP_NOWAIT);
186 
187 	if (unmap && is_dma_xor_aligned(device, offset, 0, len)) {
188 		struct dma_async_tx_descriptor *tx;
189 		int i, j;
190 
191 		/* run the xor asynchronously */
192 		pr_debug("%s (async): len: %zu\n", __func__, len);
193 
194 		unmap->len = len;
195 		for (i = 0, j = 0; i < src_cnt; i++) {
196 			if (!src_list[i])
197 				continue;
198 			unmap->to_cnt++;
199 			unmap->addr[j++] = dma_map_page(device->dev, src_list[i],
200 							offset, len, DMA_TO_DEVICE);
201 		}
202 
203 		/* map it bidirectional as it may be re-used as a source */
204 		unmap->addr[j] = dma_map_page(device->dev, dest, offset, len,
205 					      DMA_BIDIRECTIONAL);
206 		unmap->bidi_cnt = 1;
207 
208 		tx = do_async_xor(chan, unmap, submit);
209 		dmaengine_unmap_put(unmap);
210 		return tx;
211 	} else {
212 		dmaengine_unmap_put(unmap);
213 		/* run the xor synchronously */
214 		pr_debug("%s (sync): len: %zu\n", __func__, len);
215 		WARN_ONCE(chan, "%s: no space for dma address conversion\n",
216 			  __func__);
217 
218 		/* in the sync case the dest is an implied source
219 		 * (assumes the dest is the first source)
220 		 */
221 		if (submit->flags & ASYNC_TX_XOR_DROP_DST) {
222 			src_cnt--;
223 			src_list++;
224 		}
225 
226 		/* wait for any prerequisite operations */
227 		async_tx_quiesce(&submit->depend_tx);
228 
229 		do_sync_xor(dest, src_list, offset, src_cnt, len, submit);
230 
231 		return NULL;
232 	}
233 }
234 EXPORT_SYMBOL_GPL(async_xor);
235 
page_is_zero(struct page * p,unsigned int offset,size_t len)236 static int page_is_zero(struct page *p, unsigned int offset, size_t len)
237 {
238 	return !memchr_inv(page_address(p) + offset, 0, len);
239 }
240 
241 static inline struct dma_chan *
xor_val_chan(struct async_submit_ctl * submit,struct page * dest,struct page ** src_list,int src_cnt,size_t len)242 xor_val_chan(struct async_submit_ctl *submit, struct page *dest,
243 		 struct page **src_list, int src_cnt, size_t len)
244 {
245 	#ifdef CONFIG_ASYNC_TX_DISABLE_XOR_VAL_DMA
246 	return NULL;
247 	#endif
248 	return async_tx_find_channel(submit, DMA_XOR_VAL, &dest, 1, src_list,
249 				     src_cnt, len);
250 }
251 
252 /**
253  * async_xor_val - attempt a xor parity check with a dma engine.
254  * @dest: destination page used if the xor is performed synchronously
255  * @src_list: array of source pages
256  * @offset: offset in pages to start transaction
257  * @src_cnt: number of source pages
258  * @len: length in bytes
259  * @result: 0 if sum == 0 else non-zero
260  * @submit: submission / completion modifiers
261  *
262  * honored flags: ASYNC_TX_ACK
263  *
264  * src_list note: if the dest is also a source it must be at index zero.
265  * The contents of this array will be overwritten if a scribble region
266  * is not specified.
267  */
268 struct dma_async_tx_descriptor *
async_xor_val(struct page * dest,struct page ** src_list,unsigned int offset,int src_cnt,size_t len,enum sum_check_flags * result,struct async_submit_ctl * submit)269 async_xor_val(struct page *dest, struct page **src_list, unsigned int offset,
270 	      int src_cnt, size_t len, enum sum_check_flags *result,
271 	      struct async_submit_ctl *submit)
272 {
273 	struct dma_chan *chan = xor_val_chan(submit, dest, src_list, src_cnt, len);
274 	struct dma_device *device = chan ? chan->device : NULL;
275 	struct dma_async_tx_descriptor *tx = NULL;
276 	struct dmaengine_unmap_data *unmap = NULL;
277 
278 	BUG_ON(src_cnt <= 1);
279 
280 	if (device)
281 		unmap = dmaengine_get_unmap_data(device->dev, src_cnt, GFP_NOWAIT);
282 
283 	if (unmap && src_cnt <= device->max_xor &&
284 	    is_dma_xor_aligned(device, offset, 0, len)) {
285 		unsigned long dma_prep_flags = 0;
286 		int i;
287 
288 		pr_debug("%s: (async) len: %zu\n", __func__, len);
289 
290 		if (submit->cb_fn)
291 			dma_prep_flags |= DMA_PREP_INTERRUPT;
292 		if (submit->flags & ASYNC_TX_FENCE)
293 			dma_prep_flags |= DMA_PREP_FENCE;
294 
295 		for (i = 0; i < src_cnt; i++) {
296 			unmap->addr[i] = dma_map_page(device->dev, src_list[i],
297 						      offset, len, DMA_TO_DEVICE);
298 			unmap->to_cnt++;
299 		}
300 		unmap->len = len;
301 
302 		tx = device->device_prep_dma_xor_val(chan, unmap->addr, src_cnt,
303 						     len, result,
304 						     dma_prep_flags);
305 		if (unlikely(!tx)) {
306 			async_tx_quiesce(&submit->depend_tx);
307 
308 			while (!tx) {
309 				dma_async_issue_pending(chan);
310 				tx = device->device_prep_dma_xor_val(chan,
311 					unmap->addr, src_cnt, len, result,
312 					dma_prep_flags);
313 			}
314 		}
315 		dma_set_unmap(tx, unmap);
316 		async_tx_submit(chan, tx, submit);
317 	} else {
318 		enum async_tx_flags flags_orig = submit->flags;
319 
320 		pr_debug("%s: (sync) len: %zu\n", __func__, len);
321 		WARN_ONCE(device && src_cnt <= device->max_xor,
322 			  "%s: no space for dma address conversion\n",
323 			  __func__);
324 
325 		submit->flags |= ASYNC_TX_XOR_DROP_DST;
326 		submit->flags &= ~ASYNC_TX_ACK;
327 
328 		tx = async_xor(dest, src_list, offset, src_cnt, len, submit);
329 
330 		async_tx_quiesce(&tx);
331 
332 		*result = !page_is_zero(dest, offset, len) << SUM_CHECK_P;
333 
334 		async_tx_sync_epilog(submit);
335 		submit->flags = flags_orig;
336 	}
337 	dmaengine_unmap_put(unmap);
338 
339 	return tx;
340 }
341 EXPORT_SYMBOL_GPL(async_xor_val);
342 
343 MODULE_AUTHOR("Intel Corporation");
344 MODULE_DESCRIPTION("asynchronous xor/xor-zero-sum api");
345 MODULE_LICENSE("GPL");
346