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         dma_flags |= DMA_COMPL_SKIP_SRC_UNMAP | DMA_COMPL_SKIP_DEST_UNMAP;
  66         if (submit->flags & ASYNC_TX_FENCE)
  67                 dma_flags |= DMA_PREP_FENCE;
  68
  69         while (src_cnt > 0) {
  70                 submit->flags = flags_orig;
  71                 pq_src_cnt = min(src_cnt, dma_maxpq(dma, dma_flags));
  72                 /* if we are submitting additional pqs, leave the chain open,
  73                  * clear the callback parameters, and leave the destination
  74                  * buffers mapped
  75                  */
  76                 if (src_cnt > pq_src_cnt) {
  77                         submit->flags &= ~ASYNC_TX_ACK;
  78                         submit->flags |= ASYNC_TX_FENCE;
  79                         submit->cb_fn = NULL;
  80                         submit->cb_param = NULL;
  81                 } else {
  82                         submit->cb_fn = cb_fn_orig;
  83                         submit->cb_param = cb_param_orig;
  84                         if (cb_fn_orig)
  85                                 dma_flags |= DMA_PREP_INTERRUPT;
  86                 }
  87
  88                 /* Drivers force forward progress in case they can not provide
  89                  * a descriptor
  90                  */
  91                 for (;;) {
  92                         dma_dest[0] = unmap->addr[disks - 2];
  93                         dma_dest[1] = unmap->addr[disks - 1];
  94                         tx = dma->device_prep_dma_pq(chan, dma_dest,
  95                                                      &unmap->addr[src_off],
  96                                                      pq_src_cnt,
  97                                                      &scfs[src_off], unmap->len,
  98                                                      dma_flags);
  99                         if (likely(tx))
 100                                 break;
 101                         async_tx_quiesce(&submit->depend_tx);
 102                         dma_async_issue_pending(chan);
 103                 }
 104
 105                 dma_set_unmap(tx, unmap);
 106                 async_tx_submit(chan, tx, submit);
 107                 submit->depend_tx = tx;
 108
 109                 /* drop completed sources */
 110                 src_cnt -= pq_src_cnt;
 111                 src_off += pq_src_cnt;
 112
 113                 dma_flags |= DMA_PREP_CONTINUE;
 114         }
 115
 116         return tx;
 117 }
 118
 119 /**
 120  * do_sync_gen_syndrome - synchronously calculate a raid6 syndrome
 121  */
 122 static void
 123 do_sync_gen_syndrome(struct page **blocks, unsigned int offset, int disks,
 124                      size_t len, struct async_submit_ctl *submit)
 125 {
 126         void **srcs;
 127         int i;
 128
 129         if (submit->scribble)
 130                 srcs = submit->scribble;
 131         else
 132                 srcs = (void **) blocks;
 133
 134         for (i = 0; i < disks; i++) {
 135                 if (blocks[i] == NULL) {
 136                         BUG_ON(i > disks - 3); /* P or Q can't be zero */
 137                         srcs[i] = (void*)raid6_empty_zero_page;
 138                 } else
 139                         srcs[i] = page_address(blocks[i]) + offset;
 140         }
 141         raid6_call.gen_syndrome(disks, len, srcs);
 142         async_tx_sync_epilog(submit);
 143 }
 144
 145 /**
 146  * async_gen_syndrome - asynchronously calculate a raid6 syndrome
 147  * @blocks: source blocks from idx 0..disks-3, P @ disks-2 and Q @ disks-1
 148  * @offset: common offset into each block (src and dest) to start transaction
 149  * @disks: number of blocks (including missing P or Q, see below)
 150  * @len: length of operation in bytes
 151  * @submit: submission/completion modifiers
 152  *
 153  * General note: This routine assumes a field of GF(2^8) with a
 154  * primitive polynomial of 0x11d and a generator of {02}.
 155  *
 156  * 'disks' note: callers can optionally omit either P or Q (but not
 157  * both) from the calculation by setting blocks[disks-2] or
 158  * blocks[disks-1] to NULL.  When P or Q is omitted 'len' must be <=
 159  * PAGE_SIZE as a temporary buffer of this size is used in the
 160  * synchronous path.  'disks' always accounts for both destination
 161  * buffers.  If any source buffers (blocks[i] where i < disks - 2) are
 162  * set to NULL those buffers will be replaced with the raid6_zero_page
 163  * in the synchronous path and omitted in the hardware-asynchronous
 164  * path.
 165  */
 166 struct dma_async_tx_descriptor *
 167 async_gen_syndrome(struct page **blocks, unsigned int offset, int disks,
 168                    size_t len, struct async_submit_ctl *submit)
 169 {
 170         int src_cnt = disks - 2;
 171         struct dma_chan *chan = async_tx_find_channel(submit, DMA_PQ,
 172                                                       &P(blocks, disks), 2,
 173                                                       blocks, src_cnt, len);
 174         struct dma_device *device = chan ? chan->device : NULL;
 175         struct dmaengine_unmap_data *unmap = NULL;
 176
 177         BUG_ON(disks > 255 || !(P(blocks, disks) || Q(blocks, disks)));
 178
 179         if (device)
 180                 unmap = dmaengine_get_unmap_data(device->dev, disks, GFP_NOIO);
 181
 182         if (unmap &&
 183             (src_cnt <= dma_maxpq(device, 0) ||
 184              dma_maxpq(device, DMA_PREP_CONTINUE) > 0) &&
 185             is_dma_pq_aligned(device, offset, 0, len)) {
 186                 struct dma_async_tx_descriptor *tx;
 187                 enum dma_ctrl_flags dma_flags = 0;
 188                 unsigned char coefs[src_cnt];
 189                 int i, j;
 190
 191                 /* run the p+q asynchronously */
 192                 pr_debug("%s: (async) disks: %d len: %zu\n",
 193                          __func__, disks, len);
 194
 195                 /* convert source addresses being careful to collapse 'empty'
 196                  * sources and update the coefficients accordingly
 197                  */
 198                 unmap->len = len;
 199                 for (i = 0, j = 0; i < src_cnt; i++) {
 200                         if (blocks[i] == NULL)
 201                                 continue;
 202                         unmap->addr[j] = dma_map_page(device->dev, blocks[i], offset,
 203                                                       len, DMA_TO_DEVICE);
 204                         coefs[j] = raid6_gfexp[i];
 205                         unmap->to_cnt++;
 206                         j++;
 207                 }
 208
 209                 /*
 210                  * DMAs use destinations as sources,
 211                  * so use BIDIRECTIONAL mapping
 212                  */
 213                 unmap->bidi_cnt++;
 214                 if (P(blocks, disks))
 215                         unmap->addr[j++] = dma_map_page(device->dev, P(blocks, disks),
 216                                                         offset, len, DMA_BIDIRECTIONAL);
 217                 else {
 218                         unmap->addr[j++] = 0;
 219                         dma_flags |= DMA_PREP_PQ_DISABLE_P;
 220                 }
 221
 222                 unmap->bidi_cnt++;
 223                 if (Q(blocks, disks))
 224                         unmap->addr[j++] = dma_map_page(device->dev, Q(blocks, disks),
 225                                                        offset, len, DMA_BIDIRECTIONAL);
 226                 else {
 227                         unmap->addr[j++] = 0;
 228                         dma_flags |= DMA_PREP_PQ_DISABLE_Q;
 229                 }
 230
 231                 tx = do_async_gen_syndrome(chan, coefs, j, unmap, dma_flags, submit);
 232                 dmaengine_unmap_put(unmap);
 233                 return tx;
 234         }
 235
 236         dmaengine_unmap_put(unmap);
 237
 238         /* run the pq synchronously */
 239         pr_debug("%s: (sync) disks: %d len: %zu\n", __func__, disks, len);
 240
 241         /* wait for any prerequisite operations */
 242         async_tx_quiesce(&submit->depend_tx);
 243
 244         if (!P(blocks, disks)) {
 245                 P(blocks, disks) = pq_scribble_page;
 246                 BUG_ON(len + offset > PAGE_SIZE);
 247         }
 248         if (!Q(blocks, disks)) {
 249                 Q(blocks, disks) = pq_scribble_page;
 250                 BUG_ON(len + offset > PAGE_SIZE);
 251         }
 252         do_sync_gen_syndrome(blocks, offset, disks, len, submit);
 253
 254         return NULL;
 255 }
 256 EXPORT_SYMBOL_GPL(async_gen_syndrome);
 257
 258 static inline struct dma_chan *
 259 pq_val_chan(struct async_submit_ctl *submit, struct page **blocks, int disks, size_t len)
 260 {
 261         #ifdef CONFIG_ASYNC_TX_DISABLE_PQ_VAL_DMA
 262         return NULL;
 263         #endif
 264         return async_tx_find_channel(submit, DMA_PQ_VAL, NULL, 0,  blocks,
 265                                      disks, len);
 266 }
 267
 268 /**
 269  * async_syndrome_val - asynchronously validate a raid6 syndrome
 270  * @blocks: source blocks from idx 0..disks-3, P @ disks-2 and Q @ disks-1
 271  * @offset: common offset into each block (src and dest) to start transaction
 272  * @disks: number of blocks (including missing P or Q, see below)
 273  * @len: length of operation in bytes
 274  * @pqres: on val failure SUM_CHECK_P_RESULT and/or SUM_CHECK_Q_RESULT are set
 275  * @spare: temporary result buffer for the synchronous case
 276  * @submit: submission / completion modifiers
 277  *
 278  * The same notes from async_gen_syndrome apply to the 'blocks',
 279  * and 'disks' parameters of this routine.  The synchronous path
 280  * requires a temporary result buffer and submit->scribble to be
 281  * specified.
 282  */
 283 struct dma_async_tx_descriptor *
 284 async_syndrome_val(struct page **blocks, unsigned int offset, int disks,
 285                    size_t len, enum sum_check_flags *pqres, struct page *spare,
 286                    struct async_submit_ctl *submit)
 287 {
 288         struct dma_chan *chan = pq_val_chan(submit, blocks, disks, len);
 289         struct dma_device *device = chan ? chan->device : NULL;
 290         struct dma_async_tx_descriptor *tx;
 291         unsigned char coefs[disks-2];
 292         enum dma_ctrl_flags dma_flags = submit->cb_fn ? DMA_PREP_INTERRUPT : 0;
 293         dma_addr_t *dma_src = NULL;
 294         int src_cnt = 0;
 295
 296         BUG_ON(disks < 4);
 297
 298         if (submit->scribble)
 299                 dma_src = submit->scribble;
 300         else if (sizeof(dma_addr_t) <= sizeof(struct page *))
 301                 dma_src = (dma_addr_t *) blocks;
 302
 303         if (dma_src && device && disks <= dma_maxpq(device, 0) &&
 304             is_dma_pq_aligned(device, offset, 0, len)) {
 305                 struct device *dev = device->dev;
 306                 dma_addr_t *pq = &dma_src[disks-2];
 307                 int i;
 308
 309                 pr_debug("%s: (async) disks: %d len: %zu\n",
 310                          __func__, disks, len);
 311                 if (!P(blocks, disks))
 312                         dma_flags |= DMA_PREP_PQ_DISABLE_P;
 313                 else
 314                         pq[0] = dma_map_page(dev, P(blocks, disks),
 315                                              offset, len,
 316                                              DMA_TO_DEVICE);
 317                 if (!Q(blocks, disks))
 318                         dma_flags |= DMA_PREP_PQ_DISABLE_Q;
 319                 else
 320                         pq[1] = dma_map_page(dev, Q(blocks, disks),
 321                                              offset, len,
 322                                              DMA_TO_DEVICE);
 323
 324                 if (submit->flags & ASYNC_TX_FENCE)
 325                         dma_flags |= DMA_PREP_FENCE;
 326                 for (i = 0; i < disks-2; i++)
 327                         if (likely(blocks[i])) {
 328                                 dma_src[src_cnt] = dma_map_page(dev, blocks[i],
 329                                                                 offset, len,
 330                                                                 DMA_TO_DEVICE);
 331                                 coefs[src_cnt] = raid6_gfexp[i];
 332                                 src_cnt++;
 333                         }
 334
 335                 for (;;) {
 336                         tx = device->device_prep_dma_pq_val(chan, pq, dma_src,
 337                                                             src_cnt,
 338                                                             coefs,
 339                                                             len, pqres,
 340                                                             dma_flags);
 341                         if (likely(tx))
 342                                 break;
 343                         async_tx_quiesce(&submit->depend_tx);
 344                         dma_async_issue_pending(chan);
 345                 }
 346                 async_tx_submit(chan, tx, submit);
 347
 348                 return tx;
 349         } else {
 350                 struct page *p_src = P(blocks, disks);
 351                 struct page *q_src = Q(blocks, disks);
 352                 enum async_tx_flags flags_orig = submit->flags;
 353                 dma_async_tx_callback cb_fn_orig = submit->cb_fn;
 354                 void *scribble = submit->scribble;
 355                 void *cb_param_orig = submit->cb_param;
 356                 void *p, *q, *s;
 357
 358                 pr_debug("%s: (sync) disks: %d len: %zu\n",
 359                          __func__, disks, len);
 360
 361                 /* caller must provide a temporary result buffer and
 362                  * allow the input parameters to be preserved
 363                  */
 364                 BUG_ON(!spare || !scribble);
 365
 366                 /* wait for any prerequisite operations */
 367                 async_tx_quiesce(&submit->depend_tx);
 368
 369                 /* recompute p and/or q into the temporary buffer and then
 370                  * check to see the result matches the current value
 371                  */
 372                 tx = NULL;
 373                 *pqres = 0;
 374                 if (p_src) {
 375                         init_async_submit(submit, ASYNC_TX_XOR_ZERO_DST, NULL,
 376                                           NULL, NULL, scribble);
 377                         tx = async_xor(spare, blocks, offset, disks-2, len, submit);
 378                         async_tx_quiesce(&tx);
 379                         p = page_address(p_src) + offset;
 380                         s = page_address(spare) + offset;
 381                         *pqres |= !!memcmp(p, s, len) << SUM_CHECK_P;
 382                 }
 383
 384                 if (q_src) {
 385                         P(blocks, disks) = NULL;
 386                         Q(blocks, disks) = spare;
 387                         init_async_submit(submit, 0, NULL, NULL, NULL, scribble);
 388                         tx = async_gen_syndrome(blocks, offset, disks, len, submit);
 389                         async_tx_quiesce(&tx);
 390                         q = page_address(q_src) + offset;
 391                         s = page_address(spare) + offset;
 392                         *pqres |= !!memcmp(q, s, len) << SUM_CHECK_Q;
 393                 }
 394
 395                 /* restore P, Q and submit */
 396                 P(blocks, disks) = p_src;
 397                 Q(blocks, disks) = q_src;
 398
 399                 submit->cb_fn = cb_fn_orig;
 400                 submit->cb_param = cb_param_orig;
 401                 submit->flags = flags_orig;
 402                 async_tx_sync_epilog(submit);
 403
 404                 return NULL;
 405         }
 406 }
 407 EXPORT_SYMBOL_GPL(async_syndrome_val);
 408
 409 static int __init async_pq_init(void)
 410 {
 411         pq_scribble_page = alloc_page(GFP_KERNEL);
 412
 413         if (pq_scribble_page)
 414                 return 0;
 415
 416         pr_err("%s: failed to allocate required spare page\n", __func__);
 417
 418         return -ENOMEM;
 419 }
 420
 421 static void __exit async_pq_exit(void)
 422 {
 423         put_page(pq_scribble_page);
 424 }
 425
 426 module_init(async_pq_init);
 427 module_exit(async_pq_exit);
 428
 429 MODULE_DESCRIPTION("asynchronous raid6 syndrome generation/validation");
 430 MODULE_LICENSE("GPL");