crypto/async_tx/async_pq.c

   1 /*
   2  * Copyright(c) 2007 Yuri Tikhonov <yur@emcraft.com>
   3  * Copyright(c) 2009 Intel Corporation
   4  *
   5  * This program is free software; you can redistribute it and/or modify it
   6  * under the terms of the GNU General Public License as published by the Free
   7  * Software Foundation; either version 2 of the License, or (at your option)
   8  * any later version.
   9  *
  10  * This program is distributed in the hope that it will be useful, but WITHOUT
  11  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  12  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
  13  * more details.
  14  *
  15  * You should have received a copy of the GNU General Public License along with
  16  * this program; if not, write to the Free Software Foundation, Inc., 59
  17  * Temple Place - Suite 330, Boston, MA  02111-1307, USA.
  18  *
  19  * The full GNU General Public License is included in this distribution in the
  20  * file called COPYING.
  21  */
  22 #include <linux/kernel.h>
  23 #include <linux/interrupt.h>
  24 #include <linux/module.h>
  25 #include <linux/dma-mapping.h>
  26 #include <linux/raid/pq.h>
  27 #include <linux/async_tx.h>
  28 #include <linux/gfp.h>
  29
  30 /**
  31  * pq_scribble_page - space to hold throwaway P or Q buffer for
  32  * synchronous gen_syndrome
  33  */
  34 static struct page *pq_scribble_page;
  35
  36 /* the struct page *blocks[] parameter passed to async_gen_syndrome()
  37  * and async_syndrome_val() contains the 'P' destination address at
  38  * blocks[disks-2] and the 'Q' destination address at blocks[disks-1]
  39  *
  40  * note: these are macros as they are used as lvalues
  41  */
  42 #define P(b, d) (b[d-2])
  43 #define Q(b, d) (b[d-1])
  44
  45 /**
  46  * do_async_gen_syndrome - asynchronously calculate P and/or Q
  47  */
  48 static __async_inline struct dma_async_tx_descriptor *
  49 do_async_gen_syndrome(struct dma_chan *chan,
  50                       const unsigned char *scfs, int disks,
  51                       struct dmaengine_unmap_data *unmap,
  52                       enum dma_ctrl_flags dma_flags,
  53                       struct async_submit_ctl *submit)
  54 {
  55         struct dma_async_tx_descriptor *tx = NULL;
  56         struct dma_device *dma = chan->device;
  57         enum async_tx_flags flags_orig = submit->flags;
  58         dma_async_tx_callback cb_fn_orig = submit->cb_fn;
  59         dma_async_tx_callback cb_param_orig = submit->cb_param;
  60         int src_cnt = disks - 2;
  61         unsigned short pq_src_cnt;
  62         dma_addr_t dma_dest[2];
  63         int src_off = 0;
  64
  65         while (src_cnt > 0) {
  66                 submit->flags = flags_orig;
  67                 pq_src_cnt = min(src_cnt, dma_maxpq(dma, dma_flags));
  68                 /* if we are submitting additional pqs, leave the chain open,
  69                  * clear the callback parameters, and leave the destination
  70                  * buffers mapped
  71                  */
  72                 if (src_cnt > pq_src_cnt) {
  73                         submit->flags &= ~ASYNC_TX_ACK;
  74                         submit->flags |= ASYNC_TX_FENCE;
  75                         submit->cb_fn = NULL;
  76                         submit->cb_param = NULL;
  77                 } else {
  78                         submit->cb_fn = cb_fn_orig;
  79                         submit->cb_param = cb_param_orig;
  80                         if (cb_fn_orig)
  81                                 dma_flags |= DMA_PREP_INTERRUPT;
  82                 }
  83                 if (submit->flags & ASYNC_TX_FENCE)
  84                         dma_flags |= DMA_PREP_FENCE;
  85
  86                 /* Drivers force forward progress in case they can not provide
  87                  * a descriptor
  88                  */
  89                 for (;;) {
  90                         dma_dest[0] = unmap->addr[disks - 2];
  91                         dma_dest[1] = unmap->addr[disks - 1];
  92                         tx = dma->device_prep_dma_pq(chan, dma_dest,
  93                                                      &unmap->addr[src_off],
  94                                                      pq_src_cnt,
  95                                                      &scfs[src_off], unmap->len,
  96                                                      dma_flags);
  97                         if (likely(tx))
  98                                 break;
  99                         async_tx_quiesce(&submit->depend_tx);
 100                         dma_async_issue_pending(chan);
 101                 }
 102
 103                 dma_set_unmap(tx, unmap);
 104                 async_tx_submit(chan, tx, submit);
 105                 submit->depend_tx = tx;
 106
 107                 /* drop completed sources */
 108                 src_cnt -= pq_src_cnt;
 109                 src_off += pq_src_cnt;
 110
 111                 dma_flags |= DMA_PREP_CONTINUE;
 112         }
 113
 114         return tx;
 115 }
 116
 117 /**
 118  * do_sync_gen_syndrome - synchronously calculate a raid6 syndrome
 119  */
 120 static void
 121 do_sync_gen_syndrome(struct page **blocks, unsigned int offset, int disks,
 122                      size_t len, struct async_submit_ctl *submit)
 123 {
 124         void **srcs;
 125         int i;
 126         int start = -1, stop = disks - 3;
 127
 128         if (submit->scribble)
 129                 srcs = submit->scribble;
 130         else
 131                 srcs = (void **) blocks;
 132
 133         for (i = 0; i < disks; i++) {
 134                 if (blocks[i] == NULL) {
 135                         BUG_ON(i > disks - 3); /* P or Q can't be zero */
 136                         srcs[i] = (void*)raid6_empty_zero_page;
 137                 } else {
 138                         srcs[i] = page_address(blocks[i]) + offset;
 139                         if (i < disks - 2) {
 140                                 stop = i;
 141                                 if (start == -1)
 142                                         start = i;
 143                         }
 144                 }
 145         }
 146         if (submit->flags & ASYNC_TX_PQ_XOR_DST) {
 147                 BUG_ON(!raid6_call.xor_syndrome);
 148                 if (start >= 0)
 149                         raid6_call.xor_syndrome(disks, start, stop, len, srcs);
 150         } else
 151                 raid6_call.gen_syndrome(disks, len, srcs);
 152         async_tx_sync_epilog(submit);
 153 }
 154
 155 /**
 156  * async_gen_syndrome - asynchronously calculate a raid6 syndrome
 157  * @blocks: source blocks from idx 0..disks-3, P @ disks-2 and Q @ disks-1
 158  * @offset: common offset into each block (src and dest) to start transaction
 159  * @disks: number of blocks (including missing P or Q, see below)
 160  * @len: length of operation in bytes
 161  * @submit: submission/completion modifiers
 162  *
 163  * General note: This routine assumes a field of GF(2^8) with a
 164  * primitive polynomial of 0x11d and a generator of {02}.
 165  *
 166  * 'disks' note: callers can optionally omit either P or Q (but not
 167  * both) from the calculation by setting blocks[disks-2] or
 168  * blocks[disks-1] to NULL.  When P or Q is omitted 'len' must be <=
 169  * PAGE_SIZE as a temporary buffer of this size is used in the
 170  * synchronous path.  'disks' always accounts for both destination
 171  * buffers.  If any source buffers (blocks[i] where i < disks - 2) are
 172  * set to NULL those buffers will be replaced with the raid6_zero_page
 173  * in the synchronous path and omitted in the hardware-asynchronous
 174  * path.
 175  */
 176 struct dma_async_tx_descriptor *
 177 async_gen_syndrome(struct page **blocks, unsigned int offset, int disks,
 178                    size_t len, struct async_submit_ctl *submit)
 179 {
 180         int src_cnt = disks - 2;
 181         struct dma_chan *chan = async_tx_find_channel(submit, DMA_PQ,
 182                                                       &P(blocks, disks), 2,
 183                                                       blocks, src_cnt, len);
 184         struct dma_device *device = chan ? chan->device : NULL;
 185         struct dmaengine_unmap_data *unmap = NULL;
 186
 187         BUG_ON(disks > 255 || !(P(blocks, disks) || Q(blocks, disks)));
 188
 189         if (device)
 190                 unmap = dmaengine_get_unmap_data(device->dev, disks, GFP_NOWAIT);
 191
 192         /* XORing P/Q is only implemented in software */
 193         if (unmap && !(submit->flags & ASYNC_TX_PQ_XOR_DST) &&
 194             (src_cnt <= dma_maxpq(device, 0) ||
 195              dma_maxpq(device, DMA_PREP_CONTINUE) > 0) &&
 196             is_dma_pq_aligned(device, offset, 0, len)) {
 197                 struct dma_async_tx_descriptor *tx;
 198                 enum dma_ctrl_flags dma_flags = 0;
 199                 unsigned char coefs[src_cnt];
 200                 int i, j;
 201
 202                 /* run the p+q asynchronously */
 203                 pr_debug("%s: (async) disks: %d len: %zu\n",
 204                          __func__, disks, len);
 205
 206                 /* convert source addresses being careful to collapse 'empty'
 207                  * sources and update the coefficients accordingly
 208                  */
 209                 unmap->len = len;
 210                 for (i = 0, j = 0; i < src_cnt; i++) {
 211                         if (blocks[i] == NULL)
 212                                 continue;
 213                         unmap->addr[j] = dma_map_page(device->dev, blocks[i], offset,
 214                                                       len, DMA_TO_DEVICE);
 215                         coefs[j] = raid6_gfexp[i];
 216                         unmap->to_cnt++;
 217                         j++;
 218                 }
 219
 220                 /*
 221                  * DMAs use destinations as sources,
 222                  * so use BIDIRECTIONAL mapping
 223                  */
 224                 unmap->bidi_cnt++;
 225                 if (P(blocks, disks))
 226                         unmap->addr[j++] = dma_map_page(device->dev, P(blocks, disks),
 227                                                         offset, len, DMA_BIDIRECTIONAL);
 228                 else {
 229                         unmap->addr[j++] = 0;
 230                         dma_flags |= DMA_PREP_PQ_DISABLE_P;
 231                 }
 232
 233                 unmap->bidi_cnt++;
 234                 if (Q(blocks, disks))
 235                         unmap->addr[j++] = dma_map_page(device->dev, Q(blocks, disks),
 236                                                        offset, len, DMA_BIDIRECTIONAL);
 237                 else {
 238                         unmap->addr[j++] = 0;
 239                         dma_flags |= DMA_PREP_PQ_DISABLE_Q;
 240                 }
 241
 242                 tx = do_async_gen_syndrome(chan, coefs, j, unmap, dma_flags, submit);
 243                 dmaengine_unmap_put(unmap);
 244                 return tx;
 245         }
 246
 247         dmaengine_unmap_put(unmap);
 248
 249         /* run the pq synchronously */
 250         pr_debug("%s: (sync) disks: %d len: %zu\n", __func__, disks, len);
 251
 252         /* wait for any prerequisite operations */
 253         async_tx_quiesce(&submit->depend_tx);
 254
 255         if (!P(blocks, disks)) {
 256                 P(blocks, disks) = pq_scribble_page;
 257                 BUG_ON(len + offset > PAGE_SIZE);
 258         }
 259         if (!Q(blocks, disks)) {
 260                 Q(blocks, disks) = pq_scribble_page;
 261                 BUG_ON(len + offset > PAGE_SIZE);
 262         }
 263         do_sync_gen_syndrome(blocks, offset, disks, len, submit);
 264
 265         return NULL;
 266 }
 267 EXPORT_SYMBOL_GPL(async_gen_syndrome);
 268
 269 static inline struct dma_chan *
 270 pq_val_chan(struct async_submit_ctl *submit, struct page **blocks, int disks, size_t len)
 271 {
 272         #ifdef CONFIG_ASYNC_TX_DISABLE_PQ_VAL_DMA
 273         return NULL;
 274         #endif
 275         return async_tx_find_channel(submit, DMA_PQ_VAL, NULL, 0,  blocks,
 276                                      disks, len);
 277 }
 278
 279 /**
 280  * async_syndrome_val - asynchronously validate a raid6 syndrome
 281  * @blocks: source blocks from idx 0..disks-3, P @ disks-2 and Q @ disks-1
 282  * @offset: common offset into each block (src and dest) to start transaction
 283  * @disks: number of blocks (including missing P or Q, see below)
 284  * @len: length of operation in bytes
 285  * @pqres: on val failure SUM_CHECK_P_RESULT and/or SUM_CHECK_Q_RESULT are set
 286  * @spare: temporary result buffer for the synchronous case
 287  * @submit: submission / completion modifiers
 288  *
 289  * The same notes from async_gen_syndrome apply to the 'blocks',
 290  * and 'disks' parameters of this routine.  The synchronous path
 291  * requires a temporary result buffer and submit->scribble to be
 292  * specified.
 293  */
 294 struct dma_async_tx_descriptor *
 295 async_syndrome_val(struct page **blocks, unsigned int offset, int disks,
 296                    size_t len, enum sum_check_flags *pqres, struct page *spare,
 297                    struct async_submit_ctl *submit)
 298 {
 299         struct dma_chan *chan = pq_val_chan(submit, blocks, disks, len);
 300         struct dma_device *device = chan ? chan->device : NULL;
 301         struct dma_async_tx_descriptor *tx;
 302         unsigned char coefs[disks-2];
 303         enum dma_ctrl_flags dma_flags = submit->cb_fn ? DMA_PREP_INTERRUPT : 0;
 304         struct dmaengine_unmap_data *unmap = NULL;
 305
 306         BUG_ON(disks < 4);
 307
 308         if (device)
 309                 unmap = dmaengine_get_unmap_data(device->dev, disks, GFP_NOWAIT);
 310
 311         if (unmap && disks <= dma_maxpq(device, 0) &&
 312             is_dma_pq_aligned(device, offset, 0, len)) {
 313                 struct device *dev = device->dev;
 314                 dma_addr_t pq[2];
 315                 int i, j = 0, src_cnt = 0;
 316
 317                 pr_debug("%s: (async) disks: %d len: %zu\n",
 318                          __func__, disks, len);
 319
 320                 unmap->len = len;
 321                 for (i = 0; i < disks-2; i++)
 322                         if (likely(blocks[i])) {
 323                                 unmap->addr[j] = dma_map_page(dev, blocks[i],
 324                                                               offset, len,
 325                                                               DMA_TO_DEVICE);
 326                                 coefs[j] = raid6_gfexp[i];
 327                                 unmap->to_cnt++;
 328                                 src_cnt++;
 329                                 j++;
 330                         }
 331
 332                 if (!P(blocks, disks)) {
 333                         pq[0] = 0;
 334                         dma_flags |= DMA_PREP_PQ_DISABLE_P;
 335                 } else {
 336                         pq[0] = dma_map_page(dev, P(blocks, disks),
 337                                              offset, len,
 338                                              DMA_TO_DEVICE);
 339                         unmap->addr[j++] = pq[0];
 340                         unmap->to_cnt++;
 341                 }
 342                 if (!Q(blocks, disks)) {
 343                         pq[1] = 0;
 344                         dma_flags |= DMA_PREP_PQ_DISABLE_Q;
 345                 } else {
 346                         pq[1] = dma_map_page(dev, Q(blocks, disks),
 347                                              offset, len,
 348                                              DMA_TO_DEVICE);
 349                         unmap->addr[j++] = pq[1];
 350                         unmap->to_cnt++;
 351                 }
 352
 353                 if (submit->flags & ASYNC_TX_FENCE)
 354                         dma_flags |= DMA_PREP_FENCE;
 355                 for (;;) {
 356                         tx = device->device_prep_dma_pq_val(chan, pq,
 357                                                             unmap->addr,
 358                                                             src_cnt,
 359                                                             coefs,
 360                                                             len, pqres,
 361                                                             dma_flags);
 362                         if (likely(tx))
 363                                 break;
 364                         async_tx_quiesce(&submit->depend_tx);
 365                         dma_async_issue_pending(chan);
 366                 }
 367
 368                 dma_set_unmap(tx, unmap);
 369                 async_tx_submit(chan, tx, submit);
 370         } else {
 371                 struct page *p_src = P(blocks, disks);
 372                 struct page *q_src = Q(blocks, disks);
 373                 enum async_tx_flags flags_orig = submit->flags;
 374                 dma_async_tx_callback cb_fn_orig = submit->cb_fn;
 375                 void *scribble = submit->scribble;
 376                 void *cb_param_orig = submit->cb_param;
 377                 void *p, *q, *s;
 378
 379                 pr_debug("%s: (sync) disks: %d len: %zu\n",
 380                          __func__, disks, len);
 381
 382                 /* caller must provide a temporary result buffer and
 383                  * allow the input parameters to be preserved
 384                  */
 385                 BUG_ON(!spare || !scribble);
 386
 387                 /* wait for any prerequisite operations */
 388                 async_tx_quiesce(&submit->depend_tx);
 389
 390                 /* recompute p and/or q into the temporary buffer and then
 391                  * check to see the result matches the current value
 392                  */
 393                 tx = NULL;
 394                 *pqres = 0;
 395                 if (p_src) {
 396                         init_async_submit(submit, ASYNC_TX_XOR_ZERO_DST, NULL,
 397                                           NULL, NULL, scribble);
 398                         tx = async_xor(spare, blocks, offset, disks-2, len, submit);
 399                         async_tx_quiesce(&tx);
 400                         p = page_address(p_src) + offset;
 401                         s = page_address(spare) + offset;
 402                         *pqres |= !!memcmp(p, s, len) << SUM_CHECK_P;
 403                 }
 404
 405                 if (q_src) {
 406                         P(blocks, disks) = NULL;
 407                         Q(blocks, disks) = spare;
 408                         init_async_submit(submit, 0, NULL, NULL, NULL, scribble);
 409                         tx = async_gen_syndrome(blocks, offset, disks, len, submit);
 410                         async_tx_quiesce(&tx);
 411                         q = page_address(q_src) + offset;
 412                         s = page_address(spare) + offset;
 413                         *pqres |= !!memcmp(q, s, len) << SUM_CHECK_Q;
 414                 }
 415
 416                 /* restore P, Q and submit */
 417                 P(blocks, disks) = p_src;
 418                 Q(blocks, disks) = q_src;
 419
 420                 submit->cb_fn = cb_fn_orig;
 421                 submit->cb_param = cb_param_orig;
 422                 submit->flags = flags_orig;
 423                 async_tx_sync_epilog(submit);
 424                 tx = NULL;
 425         }
 426         dmaengine_unmap_put(unmap);
 427
 428         return tx;
 429 }
 430 EXPORT_SYMBOL_GPL(async_syndrome_val);
 431
 432 static int __init async_pq_init(void)
 433 {
 434         pq_scribble_page = alloc_page(GFP_KERNEL);
 435
 436         if (pq_scribble_page)
 437                 return 0;
 438
 439         pr_err("%s: failed to allocate required spare page\n", __func__);
 440
 441         return -ENOMEM;
 442 }
 443
 444 static void __exit async_pq_exit(void)
 445 {
 446         __free_page(pq_scribble_page);
 447 }
 448
 449 module_init(async_pq_init);
 450 module_exit(async_pq_exit);
 451
 452 MODULE_DESCRIPTION("asynchronous raid6 syndrome generation/validation");
 453 MODULE_LICENSE("GPL");