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 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) 39 { 40 struct dma_device *dma = chan->device; 41 struct dma_async_tx_descriptor *tx = NULL; 42 int src_off = 0; 43 int i; 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; 48 int xor_src_cnt = 0; 49 dma_addr_t dma_dest; 50 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 */ 55 if (!src_list[i]) 56 continue; 57 if (unlikely(src_list[i] == dest)) { 58 dma_src[xor_src_cnt++] = dma_dest; 59 continue; 60 } 61 dma_src[xor_src_cnt++] = dma_map_page(dma->dev, src_list[i], offset, 62 len, DMA_TO_DEVICE); 63 } 64 src_cnt = xor_src_cnt; 65 66 while (src_cnt) { 67 submit->flags = flags_orig; 68 dma_flags = 0; 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 72 * buffer mapped 73 */ 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; 78 submit->cb_fn = NULL; 79 submit->cb_param = NULL; 80 } else { 81 submit->cb_fn = cb_fn_orig; 82 submit->cb_param = cb_param_orig; 83 } 84 if (submit->cb_fn) 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 91 */ 92 tx = dma->device_prep_dma_xor(chan, dma_dest, &dma_src[src_off], 93 xor_src_cnt, len, dma_flags); 94 95 if (unlikely(!tx)) 96 async_tx_quiesce(&submit->depend_tx); 97 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, 102 &dma_src[src_off], 103 xor_src_cnt, len, 104 dma_flags); 105 } 106 107 async_tx_submit(chan, tx, submit); 108 submit->depend_tx = tx; 109 110 if (src_cnt > xor_src_cnt) { 111 /* drop completed sources */ 112 src_cnt -= xor_src_cnt; 113 src_off += xor_src_cnt; 114 115 /* use the intermediate result a source */ 116 dma_src[--src_off] = dma_dest; 117 src_cnt++; 118 } else 119 break; 120 } 121 122 return tx; 123 } 124 125 static void 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) 128 { 129 int i; 130 int xor_src_cnt = 0; 131 int src_off = 0; 132 void *dest_buf; 133 void **srcs; 134 135 if (submit->scribble) 136 srcs = submit->scribble; 137 else 138 srcs = (void **) src_list; 139 140 /* convert to buffer pointers */ 141 for (i = 0; i < src_cnt; i++) 142 if (src_list[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; 147 148 if (submit->flags & ASYNC_TX_XOR_ZERO_DST) 149 memset(dest_buf, 0, len); 150 151 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]); 155 156 /* drop completed sources */ 157 src_cnt -= xor_src_cnt; 158 src_off += xor_src_cnt; 159 } 160 161 async_tx_sync_epilog(submit); 162 } 163 164 /** 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 172 * 173 * honored flags: ASYNC_TX_ACK, ASYNC_TX_XOR_ZERO_DST, ASYNC_TX_XOR_DROP_DST 174 * 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. 180 * 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 183 * is not specified. 184 */ 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) 188 { 189 struct dma_chan *chan = async_tx_find_channel(submit, DMA_XOR, 190 &dest, 1, src_list, 191 src_cnt, len); 192 dma_addr_t *dma_src = NULL; 193 194 BUG_ON(src_cnt <= 1); 195 196 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; 200 201 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); 204 205 return do_async_xor(chan, dest, src_list, offset, src_cnt, len, 206 dma_src, submit); 207 } else { 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", 211 __func__); 212 213 /* in the sync case the dest is an implied source 214 * (assumes the dest is the first source) 215 */ 216 if (submit->flags & ASYNC_TX_XOR_DROP_DST) { 217 src_cnt--; 218 src_list++; 219 } 220 221 /* wait for any prerequisite operations */ 222 async_tx_quiesce(&submit->depend_tx); 223 224 do_sync_xor(dest, src_list, offset, src_cnt, len, submit); 225 226 return NULL; 227 } 228 } 229 EXPORT_SYMBOL_GPL(async_xor); 230 231 static int page_is_zero(struct page *p, unsigned int offset, size_t len) 232 { 233 return !memchr_inv(page_address(p) + offset, 0, len); 234 } 235 236 static inline struct dma_chan * 237 xor_val_chan(struct async_submit_ctl *submit, struct page *dest, 238 struct page **src_list, int src_cnt, size_t len) 239 { 240 #ifdef CONFIG_ASYNC_TX_DISABLE_XOR_VAL_DMA 241 return NULL; 242 #endif 243 return async_tx_find_channel(submit, DMA_XOR_VAL, &dest, 1, src_list, 244 src_cnt, len); 245 } 246 247 /** 248 * async_xor_val - 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 251 * @offset: offset in pages to start transaction 252 * @src_cnt: number of source pages 253 * @len: length in bytes 254 * @result: 0 if sum == 0 else non-zero 255 * @submit: submission / completion modifiers 256 * 257 * honored flags: ASYNC_TX_ACK 258 * 259 * src_list note: if the dest is also a source it must be at index zero. 260 * The contents of this array will be overwritten if a scribble region 261 * is not specified. 262 */ 263 struct dma_async_tx_descriptor * 264 async_xor_val(struct page *dest, struct page **src_list, unsigned int offset, 265 int src_cnt, size_t len, enum sum_check_flags *result, 266 struct async_submit_ctl *submit) 267 { 268 struct dma_chan *chan = xor_val_chan(submit, dest, src_list, src_cnt, len); 269 struct dma_device *device = chan ? chan->device : NULL; 270 struct dma_async_tx_descriptor *tx = NULL; 271 dma_addr_t *dma_src = NULL; 272 273 BUG_ON(src_cnt <= 1); 274 275 if (submit->scribble) 276 dma_src = submit->scribble; 277 else if (sizeof(dma_addr_t) <= sizeof(struct page *)) 278 dma_src = (dma_addr_t *) src_list; 279 280 if (dma_src && device && src_cnt <= device->max_xor && 281 is_dma_xor_aligned(device, offset, 0, len)) { 282 unsigned long dma_prep_flags = 0; 283 int i; 284 285 pr_debug("%s: (async) len: %zu\n", __func__, len); 286 287 if (submit->cb_fn) 288 dma_prep_flags |= DMA_PREP_INTERRUPT; 289 if (submit->flags & ASYNC_TX_FENCE) 290 dma_prep_flags |= DMA_PREP_FENCE; 291 for (i = 0; i < src_cnt; i++) 292 dma_src[i] = dma_map_page(device->dev, src_list[i], 293 offset, len, DMA_TO_DEVICE); 294 295 tx = device->device_prep_dma_xor_val(chan, dma_src, src_cnt, 296 len, result, 297 dma_prep_flags); 298 if (unlikely(!tx)) { 299 async_tx_quiesce(&submit->depend_tx); 300 301 while (!tx) { 302 dma_async_issue_pending(chan); 303 tx = device->device_prep_dma_xor_val(chan, 304 dma_src, src_cnt, len, result, 305 dma_prep_flags); 306 } 307 } 308 309 async_tx_submit(chan, tx, submit); 310 } else { 311 enum async_tx_flags flags_orig = submit->flags; 312 313 pr_debug("%s: (sync) len: %zu\n", __func__, len); 314 WARN_ONCE(device && src_cnt <= device->max_xor, 315 "%s: no space for dma address conversion\n", 316 __func__); 317 318 submit->flags |= ASYNC_TX_XOR_DROP_DST; 319 submit->flags &= ~ASYNC_TX_ACK; 320 321 tx = async_xor(dest, src_list, offset, src_cnt, len, submit); 322 323 async_tx_quiesce(&tx); 324 325 *result = !page_is_zero(dest, offset, len) << SUM_CHECK_P; 326 327 async_tx_sync_epilog(submit); 328 submit->flags = flags_orig; 329 } 330 331 return tx; 332 } 333 EXPORT_SYMBOL_GPL(async_xor_val); 334 335 MODULE_AUTHOR("Intel Corporation"); 336 MODULE_DESCRIPTION("asynchronous xor/xor-zero-sum api"); 337 MODULE_LICENSE("GPL"); 338