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; 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 (unlikely(src_list[i] == dest)) { 55 dma_src[i] = dma_dest; 56 continue; 57 } 58 dma_src[i] = dma_map_page(dma->dev, src_list[i], offset, 59 len, DMA_TO_DEVICE); 60 } 61 62 while (src_cnt) { 63 submit->flags = flags_orig; 64 dma_flags = 0; 65 xor_src_cnt = min(src_cnt, (int)dma->max_xor); 66 /* if we are submitting additional xors, leave the chain open, 67 * clear the callback parameters, and leave the destination 68 * buffer mapped 69 */ 70 if (src_cnt > xor_src_cnt) { 71 submit->flags &= ~ASYNC_TX_ACK; 72 submit->flags |= ASYNC_TX_FENCE; 73 dma_flags = DMA_COMPL_SKIP_DEST_UNMAP; 74 submit->cb_fn = NULL; 75 submit->cb_param = NULL; 76 } else { 77 submit->cb_fn = cb_fn_orig; 78 submit->cb_param = cb_param_orig; 79 } 80 if (submit->cb_fn) 81 dma_flags |= DMA_PREP_INTERRUPT; 82 if (submit->flags & ASYNC_TX_FENCE) 83 dma_flags |= DMA_PREP_FENCE; 84 /* Since we have clobbered the src_list we are committed 85 * to doing this asynchronously. Drivers force forward progress 86 * in case they can not provide a descriptor 87 */ 88 tx = dma->device_prep_dma_xor(chan, dma_dest, &dma_src[src_off], 89 xor_src_cnt, len, dma_flags); 90 91 if (unlikely(!tx)) 92 async_tx_quiesce(&submit->depend_tx); 93 94 /* spin wait for the preceeding transactions to complete */ 95 while (unlikely(!tx)) { 96 dma_async_issue_pending(chan); 97 tx = dma->device_prep_dma_xor(chan, dma_dest, 98 &dma_src[src_off], 99 xor_src_cnt, len, 100 dma_flags); 101 } 102 103 async_tx_submit(chan, tx, submit); 104 submit->depend_tx = tx; 105 106 if (src_cnt > xor_src_cnt) { 107 /* drop completed sources */ 108 src_cnt -= xor_src_cnt; 109 src_off += xor_src_cnt; 110 111 /* use the intermediate result a source */ 112 dma_src[--src_off] = dma_dest; 113 src_cnt++; 114 } else 115 break; 116 } 117 118 return tx; 119 } 120 121 static void 122 do_sync_xor(struct page *dest, struct page **src_list, unsigned int offset, 123 int src_cnt, size_t len, struct async_submit_ctl *submit) 124 { 125 int i; 126 int xor_src_cnt; 127 int src_off = 0; 128 void *dest_buf; 129 void **srcs; 130 131 if (submit->scribble) 132 srcs = submit->scribble; 133 else 134 srcs = (void **) src_list; 135 136 /* convert to buffer pointers */ 137 for (i = 0; i < src_cnt; i++) 138 srcs[i] = page_address(src_list[i]) + offset; 139 140 /* set destination address */ 141 dest_buf = page_address(dest) + offset; 142 143 if (submit->flags & ASYNC_TX_XOR_ZERO_DST) 144 memset(dest_buf, 0, len); 145 146 while (src_cnt > 0) { 147 /* process up to 'MAX_XOR_BLOCKS' sources */ 148 xor_src_cnt = min(src_cnt, MAX_XOR_BLOCKS); 149 xor_blocks(xor_src_cnt, len, dest_buf, &srcs[src_off]); 150 151 /* drop completed sources */ 152 src_cnt -= xor_src_cnt; 153 src_off += xor_src_cnt; 154 } 155 156 async_tx_sync_epilog(submit); 157 } 158 159 /** 160 * async_xor - attempt to xor a set of blocks with a dma engine. 161 * @dest: destination page 162 * @src_list: array of source pages 163 * @offset: common src/dst offset to start transaction 164 * @src_cnt: number of source pages 165 * @len: length in bytes 166 * @submit: submission / completion modifiers 167 * 168 * honored flags: ASYNC_TX_ACK, ASYNC_TX_XOR_ZERO_DST, ASYNC_TX_XOR_DROP_DST 169 * 170 * xor_blocks always uses the dest as a source so the 171 * ASYNC_TX_XOR_ZERO_DST flag must be set to not include dest data in 172 * the calculation. The assumption with dma eninges is that they only 173 * use the destination buffer as a source when it is explicity specified 174 * in the source list. 175 * 176 * src_list note: if the dest is also a source it must be at index zero. 177 * The contents of this array will be overwritten if a scribble region 178 * is not specified. 179 */ 180 struct dma_async_tx_descriptor * 181 async_xor(struct page *dest, struct page **src_list, unsigned int offset, 182 int src_cnt, size_t len, struct async_submit_ctl *submit) 183 { 184 struct dma_chan *chan = async_tx_find_channel(submit, DMA_XOR, 185 &dest, 1, src_list, 186 src_cnt, len); 187 dma_addr_t *dma_src = NULL; 188 189 BUG_ON(src_cnt <= 1); 190 191 if (submit->scribble) 192 dma_src = submit->scribble; 193 else if (sizeof(dma_addr_t) <= sizeof(struct page *)) 194 dma_src = (dma_addr_t *) src_list; 195 196 if (dma_src && chan && is_dma_xor_aligned(chan->device, offset, 0, len)) { 197 /* run the xor asynchronously */ 198 pr_debug("%s (async): len: %zu\n", __func__, len); 199 200 return do_async_xor(chan, dest, src_list, offset, src_cnt, len, 201 dma_src, submit); 202 } else { 203 /* run the xor synchronously */ 204 pr_debug("%s (sync): len: %zu\n", __func__, len); 205 WARN_ONCE(chan, "%s: no space for dma address conversion\n", 206 __func__); 207 208 /* in the sync case the dest is an implied source 209 * (assumes the dest is the first source) 210 */ 211 if (submit->flags & ASYNC_TX_XOR_DROP_DST) { 212 src_cnt--; 213 src_list++; 214 } 215 216 /* wait for any prerequisite operations */ 217 async_tx_quiesce(&submit->depend_tx); 218 219 do_sync_xor(dest, src_list, offset, src_cnt, len, submit); 220 221 return NULL; 222 } 223 } 224 EXPORT_SYMBOL_GPL(async_xor); 225 226 static int page_is_zero(struct page *p, unsigned int offset, size_t len) 227 { 228 char *a = page_address(p) + offset; 229 return ((*(u32 *) a) == 0 && 230 memcmp(a, a + 4, len - 4) == 0); 231 } 232 233 /** 234 * async_xor_val - attempt a xor parity check with a dma engine. 235 * @dest: destination page used if the xor is performed synchronously 236 * @src_list: array of source pages 237 * @offset: offset in pages to start transaction 238 * @src_cnt: number of source pages 239 * @len: length in bytes 240 * @result: 0 if sum == 0 else non-zero 241 * @submit: submission / completion modifiers 242 * 243 * honored flags: ASYNC_TX_ACK 244 * 245 * src_list note: if the dest is also a source it must be at index zero. 246 * The contents of this array will be overwritten if a scribble region 247 * is not specified. 248 */ 249 struct dma_async_tx_descriptor * 250 async_xor_val(struct page *dest, struct page **src_list, unsigned int offset, 251 int src_cnt, size_t len, enum sum_check_flags *result, 252 struct async_submit_ctl *submit) 253 { 254 struct dma_chan *chan = async_tx_find_channel(submit, DMA_XOR_VAL, 255 &dest, 1, src_list, 256 src_cnt, len); 257 struct dma_device *device = chan ? chan->device : NULL; 258 struct dma_async_tx_descriptor *tx = NULL; 259 dma_addr_t *dma_src = NULL; 260 261 BUG_ON(src_cnt <= 1); 262 263 if (submit->scribble) 264 dma_src = submit->scribble; 265 else if (sizeof(dma_addr_t) <= sizeof(struct page *)) 266 dma_src = (dma_addr_t *) src_list; 267 268 if (dma_src && device && src_cnt <= device->max_xor && 269 is_dma_xor_aligned(device, offset, 0, len)) { 270 unsigned long dma_prep_flags = 0; 271 int i; 272 273 pr_debug("%s: (async) len: %zu\n", __func__, len); 274 275 if (submit->cb_fn) 276 dma_prep_flags |= DMA_PREP_INTERRUPT; 277 if (submit->flags & ASYNC_TX_FENCE) 278 dma_prep_flags |= DMA_PREP_FENCE; 279 for (i = 0; i < src_cnt; i++) 280 dma_src[i] = dma_map_page(device->dev, src_list[i], 281 offset, len, DMA_TO_DEVICE); 282 283 tx = device->device_prep_dma_xor_val(chan, dma_src, src_cnt, 284 len, result, 285 dma_prep_flags); 286 if (unlikely(!tx)) { 287 async_tx_quiesce(&submit->depend_tx); 288 289 while (!tx) { 290 dma_async_issue_pending(chan); 291 tx = device->device_prep_dma_xor_val(chan, 292 dma_src, src_cnt, len, result, 293 dma_prep_flags); 294 } 295 } 296 297 async_tx_submit(chan, tx, submit); 298 } else { 299 enum async_tx_flags flags_orig = submit->flags; 300 301 pr_debug("%s: (sync) len: %zu\n", __func__, len); 302 WARN_ONCE(device && src_cnt <= device->max_xor, 303 "%s: no space for dma address conversion\n", 304 __func__); 305 306 submit->flags |= ASYNC_TX_XOR_DROP_DST; 307 submit->flags &= ~ASYNC_TX_ACK; 308 309 tx = async_xor(dest, src_list, offset, src_cnt, len, submit); 310 311 async_tx_quiesce(&tx); 312 313 *result = !page_is_zero(dest, offset, len) << SUM_CHECK_P; 314 315 async_tx_sync_epilog(submit); 316 submit->flags = flags_orig; 317 } 318 319 return tx; 320 } 321 EXPORT_SYMBOL_GPL(async_xor_val); 322 323 MODULE_AUTHOR("Intel Corporation"); 324 MODULE_DESCRIPTION("asynchronous xor/xor-zero-sum api"); 325 MODULE_LICENSE("GPL"); 326