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 * This routine is marked __always_inline so it can be compiled away 35 * when CONFIG_DMA_ENGINE=n 36 */ 37 static __always_inline struct dma_async_tx_descriptor * 38 do_async_xor(struct dma_chan *chan, struct page *dest, struct page **src_list, 39 unsigned int offset, int src_cnt, size_t len, 40 enum async_tx_flags flags, 41 struct dma_async_tx_descriptor *depend_tx, 42 dma_async_tx_callback cb_fn, void *cb_param) 43 { 44 struct dma_device *dma = chan->device; 45 dma_addr_t *dma_src = (dma_addr_t *) src_list; 46 struct dma_async_tx_descriptor *tx = NULL; 47 int src_off = 0; 48 int i; 49 dma_async_tx_callback _cb_fn; 50 void *_cb_param; 51 enum async_tx_flags async_flags; 52 enum dma_ctrl_flags dma_flags; 53 int xor_src_cnt; 54 dma_addr_t dma_dest; 55 56 /* map the dest bidrectional in case it is re-used as a source */ 57 dma_dest = dma_map_page(dma->dev, dest, offset, len, DMA_BIDIRECTIONAL); 58 for (i = 0; i < src_cnt; i++) { 59 /* only map the dest once */ 60 if (unlikely(src_list[i] == dest)) { 61 dma_src[i] = dma_dest; 62 continue; 63 } 64 dma_src[i] = dma_map_page(dma->dev, src_list[i], offset, 65 len, DMA_TO_DEVICE); 66 } 67 68 while (src_cnt) { 69 async_flags = flags; 70 dma_flags = 0; 71 xor_src_cnt = min(src_cnt, dma->max_xor); 72 /* if we are submitting additional xors, leave the chain open, 73 * clear the callback parameters, and leave the destination 74 * buffer mapped 75 */ 76 if (src_cnt > xor_src_cnt) { 77 async_flags &= ~ASYNC_TX_ACK; 78 dma_flags = DMA_COMPL_SKIP_DEST_UNMAP; 79 _cb_fn = NULL; 80 _cb_param = NULL; 81 } else { 82 _cb_fn = cb_fn; 83 _cb_param = cb_param; 84 } 85 if (_cb_fn) 86 dma_flags |= DMA_PREP_INTERRUPT; 87 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(&depend_tx); 97 98 /* spin wait for the preceeding 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, async_flags, depend_tx, _cb_fn, 108 _cb_param); 109 110 depend_tx = tx; 111 flags |= ASYNC_TX_DEP_ACK; 112 113 if (src_cnt > xor_src_cnt) { 114 /* drop completed sources */ 115 src_cnt -= xor_src_cnt; 116 src_off += xor_src_cnt; 117 118 /* use the intermediate result a source */ 119 dma_src[--src_off] = dma_dest; 120 src_cnt++; 121 } else 122 break; 123 } 124 125 return tx; 126 } 127 128 static void 129 do_sync_xor(struct page *dest, struct page **src_list, unsigned int offset, 130 int src_cnt, size_t len, enum async_tx_flags flags, 131 dma_async_tx_callback cb_fn, void *cb_param) 132 { 133 int i; 134 int xor_src_cnt; 135 int src_off = 0; 136 void *dest_buf; 137 void **srcs = (void **) src_list; 138 139 /* reuse the 'src_list' array to convert to buffer pointers */ 140 for (i = 0; i < src_cnt; i++) 141 srcs[i] = page_address(src_list[i]) + offset; 142 143 /* set destination address */ 144 dest_buf = page_address(dest) + offset; 145 146 if (flags & ASYNC_TX_XOR_ZERO_DST) 147 memset(dest_buf, 0, len); 148 149 while (src_cnt > 0) { 150 /* process up to 'MAX_XOR_BLOCKS' sources */ 151 xor_src_cnt = min(src_cnt, MAX_XOR_BLOCKS); 152 xor_blocks(xor_src_cnt, len, dest_buf, &srcs[src_off]); 153 154 /* drop completed sources */ 155 src_cnt -= xor_src_cnt; 156 src_off += xor_src_cnt; 157 } 158 159 async_tx_sync_epilog(cb_fn, cb_param); 160 } 161 162 /** 163 * async_xor - attempt to xor a set of blocks with a dma engine. 164 * xor_blocks always uses the dest as a source so the ASYNC_TX_XOR_ZERO_DST 165 * flag must be set to not include dest data in the calculation. The 166 * assumption with dma eninges is that they only use the destination 167 * buffer as a source when it is explicity specified in the source list. 168 * @dest: destination page 169 * @src_list: array of source pages (if the dest is also a source it must be 170 * at index zero). The contents of this array may be overwritten. 171 * @offset: offset in pages to start transaction 172 * @src_cnt: number of source pages 173 * @len: length in bytes 174 * @flags: ASYNC_TX_XOR_ZERO_DST, ASYNC_TX_XOR_DROP_DEST, 175 * ASYNC_TX_ACK, ASYNC_TX_DEP_ACK 176 * @depend_tx: xor depends on the result of this transaction. 177 * @cb_fn: function to call when the xor completes 178 * @cb_param: parameter to pass to the callback routine 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, enum async_tx_flags flags, 183 struct dma_async_tx_descriptor *depend_tx, 184 dma_async_tx_callback cb_fn, void *cb_param) 185 { 186 struct dma_chan *chan = async_tx_find_channel(depend_tx, DMA_XOR, 187 &dest, 1, src_list, 188 src_cnt, len); 189 BUG_ON(src_cnt <= 1); 190 191 if (chan) { 192 /* run the xor asynchronously */ 193 pr_debug("%s (async): len: %zu\n", __func__, len); 194 195 return do_async_xor(chan, dest, src_list, offset, src_cnt, len, 196 flags, depend_tx, cb_fn, cb_param); 197 } else { 198 /* run the xor synchronously */ 199 pr_debug("%s (sync): len: %zu\n", __func__, len); 200 201 /* in the sync case the dest is an implied source 202 * (assumes the dest is the first source) 203 */ 204 if (flags & ASYNC_TX_XOR_DROP_DST) { 205 src_cnt--; 206 src_list++; 207 } 208 209 /* wait for any prerequisite operations */ 210 async_tx_quiesce(&depend_tx); 211 212 do_sync_xor(dest, src_list, offset, src_cnt, len, 213 flags, cb_fn, cb_param); 214 215 return NULL; 216 } 217 } 218 EXPORT_SYMBOL_GPL(async_xor); 219 220 static int page_is_zero(struct page *p, unsigned int offset, size_t len) 221 { 222 char *a = page_address(p) + offset; 223 return ((*(u32 *) a) == 0 && 224 memcmp(a, a + 4, len - 4) == 0); 225 } 226 227 /** 228 * async_xor_zero_sum - attempt a xor parity check with a dma engine. 229 * @dest: destination page used if the xor is performed synchronously 230 * @src_list: array of source pages. The dest page must be listed as a source 231 * at index zero. The contents of this array may be overwritten. 232 * @offset: offset in pages to start transaction 233 * @src_cnt: number of source pages 234 * @len: length in bytes 235 * @result: 0 if sum == 0 else non-zero 236 * @flags: ASYNC_TX_ACK, ASYNC_TX_DEP_ACK 237 * @depend_tx: xor depends on the result of this transaction. 238 * @cb_fn: function to call when the xor completes 239 * @cb_param: parameter to pass to the callback routine 240 */ 241 struct dma_async_tx_descriptor * 242 async_xor_zero_sum(struct page *dest, struct page **src_list, 243 unsigned int offset, int src_cnt, size_t len, 244 u32 *result, enum async_tx_flags flags, 245 struct dma_async_tx_descriptor *depend_tx, 246 dma_async_tx_callback cb_fn, void *cb_param) 247 { 248 struct dma_chan *chan = async_tx_find_channel(depend_tx, DMA_ZERO_SUM, 249 &dest, 1, src_list, 250 src_cnt, len); 251 struct dma_device *device = chan ? chan->device : NULL; 252 struct dma_async_tx_descriptor *tx = NULL; 253 254 BUG_ON(src_cnt <= 1); 255 256 if (device && src_cnt <= device->max_xor) { 257 dma_addr_t *dma_src = (dma_addr_t *) src_list; 258 unsigned long dma_prep_flags = cb_fn ? DMA_PREP_INTERRUPT : 0; 259 int i; 260 261 pr_debug("%s: (async) len: %zu\n", __func__, len); 262 263 for (i = 0; i < src_cnt; i++) 264 dma_src[i] = dma_map_page(device->dev, src_list[i], 265 offset, len, DMA_TO_DEVICE); 266 267 tx = device->device_prep_dma_zero_sum(chan, dma_src, src_cnt, 268 len, result, 269 dma_prep_flags); 270 if (unlikely(!tx)) { 271 async_tx_quiesce(&depend_tx); 272 273 while (!tx) { 274 dma_async_issue_pending(chan); 275 tx = device->device_prep_dma_zero_sum(chan, 276 dma_src, src_cnt, len, result, 277 dma_prep_flags); 278 } 279 } 280 281 async_tx_submit(chan, tx, flags, depend_tx, cb_fn, cb_param); 282 } else { 283 unsigned long xor_flags = flags; 284 285 pr_debug("%s: (sync) len: %zu\n", __func__, len); 286 287 xor_flags |= ASYNC_TX_XOR_DROP_DST; 288 xor_flags &= ~ASYNC_TX_ACK; 289 290 tx = async_xor(dest, src_list, offset, src_cnt, len, xor_flags, 291 depend_tx, NULL, NULL); 292 293 async_tx_quiesce(&tx); 294 295 *result = page_is_zero(dest, offset, len) ? 0 : 1; 296 297 async_tx_sync_epilog(cb_fn, cb_param); 298 } 299 300 return tx; 301 } 302 EXPORT_SYMBOL_GPL(async_xor_zero_sum); 303 304 static int __init async_xor_init(void) 305 { 306 #ifdef CONFIG_DMA_ENGINE 307 /* To conserve stack space the input src_list (array of page pointers) 308 * is reused to hold the array of dma addresses passed to the driver. 309 * This conversion is only possible when dma_addr_t is less than the 310 * the size of a pointer. HIGHMEM64G is known to violate this 311 * assumption. 312 */ 313 BUILD_BUG_ON(sizeof(dma_addr_t) > sizeof(struct page *)); 314 #endif 315 316 return 0; 317 } 318 319 static void __exit async_xor_exit(void) 320 { 321 do { } while (0); 322 } 323 324 module_init(async_xor_init); 325 module_exit(async_xor_exit); 326 327 MODULE_AUTHOR("Intel Corporation"); 328 MODULE_DESCRIPTION("asynchronous xor/xor-zero-sum api"); 329 MODULE_LICENSE("GPL"); 330