crypto/async_tx/async_pq.c

   1 // SPDX-License-Identifier: GPL-2.0-or-later
   2 /*
   3  * Copyright(c) 2007 Yuri Tikhonov <yur@emcraft.com>
   4  * Copyright(c) 2009 Intel Corporation
   5  */
   6 #include <linux/kernel.h>
   7 #include <linux/interrupt.h>
   8 #include <linux/module.h>
   9 #include <linux/dma-mapping.h>
  10 #include <linux/raid/pq.h>
  11 #include <linux/async_tx.h>
  12 #include <linux/gfp.h>
  13
  14 /**
  15  * pq_scribble_page - space to hold throwaway P or Q buffer for
  16  * synchronous gen_syndrome
  17  */
  18 static struct page *pq_scribble_page;
  19
  20 /* the struct page *blocks[] parameter passed to async_gen_syndrome()
  21  * and async_syndrome_val() contains the 'P' destination address at
  22  * blocks[disks-2] and the 'Q' destination address at blocks[disks-1]
  23  *
  24  * note: these are macros as they are used as lvalues
  25  */
  26 #define P(b, d) (b[d-2])
  27 #define Q(b, d) (b[d-1])
  28
  29 #define MAX_DISKS 255
  30
  31 /**
  32  * do_async_gen_syndrome - asynchronously calculate P and/or Q
  33  */
  34 static __async_inline struct dma_async_tx_descriptor *
  35 do_async_gen_syndrome(struct dma_chan *chan,
  36                       const unsigned char *scfs, int disks,
  37                       struct dmaengine_unmap_data *unmap,
  38                       enum dma_ctrl_flags dma_flags,
  39                       struct async_submit_ctl *submit)
  40 {
  41         struct dma_async_tx_descriptor *tx = NULL;
  42         struct dma_device *dma = chan->device;
  43         enum async_tx_flags flags_orig = submit->flags;
  44         dma_async_tx_callback cb_fn_orig = submit->cb_fn;
  45         dma_async_tx_callback cb_param_orig = submit->cb_param;
  46         int src_cnt = disks - 2;
  47         unsigned short pq_src_cnt;
  48         dma_addr_t dma_dest[2];
  49         int src_off = 0;
  50
  51         while (src_cnt > 0) {
  52                 submit->flags = flags_orig;
  53                 pq_src_cnt = min(src_cnt, dma_maxpq(dma, dma_flags));
  54                 /* if we are submitting additional pqs, leave the chain open,
  55                  * clear the callback parameters, and leave the destination
  56                  * buffers mapped
  57                  */
  58                 if (src_cnt > pq_src_cnt) {
  59                         submit->flags &= ~ASYNC_TX_ACK;
  60                         submit->flags |= ASYNC_TX_FENCE;
  61                         submit->cb_fn = NULL;
  62                         submit->cb_param = NULL;
  63                 } else {
  64                         submit->cb_fn = cb_fn_orig;
  65                         submit->cb_param = cb_param_orig;
  66                         if (cb_fn_orig)
  67                                 dma_flags |= DMA_PREP_INTERRUPT;
  68                 }
  69                 if (submit->flags & ASYNC_TX_FENCE)
  70                         dma_flags |= DMA_PREP_FENCE;
  71
  72                 /* Drivers force forward progress in case they can not provide
  73                  * a descriptor
  74                  */
  75                 for (;;) {
  76                         dma_dest[0] = unmap->addr[disks - 2];
  77                         dma_dest[1] = unmap->addr[disks - 1];
  78                         tx = dma->device_prep_dma_pq(chan, dma_dest,
  79                                                      &unmap->addr[src_off],
  80                                                      pq_src_cnt,
  81                                                      &scfs[src_off], unmap->len,
  82                                                      dma_flags);
  83                         if (likely(tx))
  84                                 break;
  85                         async_tx_quiesce(&submit->depend_tx);
  86                         dma_async_issue_pending(chan);
  87                 }
  88
  89                 dma_set_unmap(tx, unmap);
  90                 async_tx_submit(chan, tx, submit);
  91                 submit->depend_tx = tx;
  92
  93                 /* drop completed sources */
  94                 src_cnt -= pq_src_cnt;
  95                 src_off += pq_src_cnt;
  96
  97                 dma_flags |= DMA_PREP_CONTINUE;
  98         }
  99
 100         return tx;
 101 }
 102
 103 /**
 104  * do_sync_gen_syndrome - synchronously calculate a raid6 syndrome
 105  */
 106 static void
 107 do_sync_gen_syndrome(struct page **blocks, unsigned int offset, int disks,
 108                      size_t len, struct async_submit_ctl *submit)
 109 {
 110         void **srcs;
 111         int i;
 112         int start = -1, stop = disks - 3;
 113
 114         if (submit->scribble)
 115                 srcs = submit->scribble;
 116         else
 117                 srcs = (void **) blocks;
 118
 119         for (i = 0; i < disks; i++) {
 120                 if (blocks[i] == NULL) {
 121                         BUG_ON(i > disks - 3); /* P or Q can't be zero */
 122                         srcs[i] = (void*)raid6_empty_zero_page;
 123                 } else {
 124                         srcs[i] = page_address(blocks[i]) + offset;
 125                         if (i < disks - 2) {
 126                                 stop = i;
 127                                 if (start == -1)
 128                                         start = i;
 129                         }
 130                 }
 131         }
 132         if (submit->flags & ASYNC_TX_PQ_XOR_DST) {
 133                 BUG_ON(!raid6_call.xor_syndrome);
 134                 if (start >= 0)
 135                         raid6_call.xor_syndrome(disks, start, stop, len, srcs);
 136         } else
 137                 raid6_call.gen_syndrome(disks, len, srcs);
 138         async_tx_sync_epilog(submit);
 139 }
 140
 141 /**
 142  * async_gen_syndrome - asynchronously calculate a raid6 syndrome
 143  * @blocks: source blocks from idx 0..disks-3, P @ disks-2 and Q @ disks-1
 144  * @offset: common offset into each block (src and dest) to start transaction
 145  * @disks: number of blocks (including missing P or Q, see below)
 146  * @len: length of operation in bytes
 147  * @submit: submission/completion modifiers
 148  *
 149  * General note: This routine assumes a field of GF(2^8) with a
 150  * primitive polynomial of 0x11d and a generator of {02}.
 151  *
 152  * 'disks' note: callers can optionally omit either P or Q (but not
 153  * both) from the calculation by setting blocks[disks-2] or
 154  * blocks[disks-1] to NULL.  When P or Q is omitted 'len' must be <=
 155  * PAGE_SIZE as a temporary buffer of this size is used in the
 156  * synchronous path.  'disks' always accounts for both destination
 157  * buffers.  If any source buffers (blocks[i] where i < disks - 2) are
 158  * set to NULL those buffers will be replaced with the raid6_zero_page
 159  * in the synchronous path and omitted in the hardware-asynchronous
 160  * path.
 161  */
 162 struct dma_async_tx_descriptor *
 163 async_gen_syndrome(struct page **blocks, unsigned int offset, int disks,
 164                    size_t len, struct async_submit_ctl *submit)
 165 {
 166         int src_cnt = disks - 2;
 167         struct dma_chan *chan = async_tx_find_channel(submit, DMA_PQ,
 168                                                       &P(blocks, disks), 2,
 169                                                       blocks, src_cnt, len);
 170         struct dma_device *device = chan ? chan->device : NULL;
 171         struct dmaengine_unmap_data *unmap = NULL;
 172
 173         BUG_ON(disks > MAX_DISKS || !(P(blocks, disks) || Q(blocks, disks)));
 174
 175         if (device)
 176                 unmap = dmaengine_get_unmap_data(device->dev, disks, GFP_NOWAIT);
 177
 178         /* XORing P/Q is only implemented in software */
 179         if (unmap && !(submit->flags & ASYNC_TX_PQ_XOR_DST) &&
 180             (src_cnt <= dma_maxpq(device, 0) ||
 181              dma_maxpq(device, DMA_PREP_CONTINUE) > 0) &&
 182             is_dma_pq_aligned(device, offset, 0, len)) {
 183                 struct dma_async_tx_descriptor *tx;
 184                 enum dma_ctrl_flags dma_flags = 0;
 185                 unsigned char coefs[MAX_DISKS];
 186                 int i, j;
 187
 188                 /* run the p+q asynchronously */
 189                 pr_debug("%s: (async) disks: %d len: %zu\n",
 190                          __func__, disks, len);
 191
 192                 /* convert source addresses being careful to collapse 'empty'
 193                  * sources and update the coefficients accordingly
 194                  */
 195                 unmap->len = len;
 196                 for (i = 0, j = 0; i < src_cnt; i++) {
 197                         if (blocks[i] == NULL)
 198                                 continue;
 199                         unmap->addr[j] = dma_map_page(device->dev, blocks[i], offset,
 200                                                       len, DMA_TO_DEVICE);
 201                         coefs[j] = raid6_gfexp[i];
 202                         unmap->to_cnt++;
 203                         j++;
 204                 }
 205
 206                 /*
 207                  * DMAs use destinations as sources,
 208                  * so use BIDIRECTIONAL mapping
 209                  */
 210                 unmap->bidi_cnt++;
 211                 if (P(blocks, disks))
 212                         unmap->addr[j++] = dma_map_page(device->dev, P(blocks, disks),
 213                                                         offset, len, DMA_BIDIRECTIONAL);
 214                 else {
 215                         unmap->addr[j++] = 0;
 216                         dma_flags |= DMA_PREP_PQ_DISABLE_P;
 217                 }
 218
 219                 unmap->bidi_cnt++;
 220                 if (Q(blocks, disks))
 221                         unmap->addr[j++] = dma_map_page(device->dev, Q(blocks, disks),
 222                                                        offset, len, DMA_BIDIRECTIONAL);
 223                 else {
 224                         unmap->addr[j++] = 0;
 225                         dma_flags |= DMA_PREP_PQ_DISABLE_Q;
 226                 }
 227
 228                 tx = do_async_gen_syndrome(chan, coefs, j, unmap, dma_flags, submit);
 229                 dmaengine_unmap_put(unmap);
 230                 return tx;
 231         }
 232
 233         dmaengine_unmap_put(unmap);
 234
 235         /* run the pq synchronously */
 236         pr_debug("%s: (sync) disks: %d len: %zu\n", __func__, disks, len);
 237
 238         /* wait for any prerequisite operations */
 239         async_tx_quiesce(&submit->depend_tx);
 240
 241         if (!P(blocks, disks)) {
 242                 P(blocks, disks) = pq_scribble_page;
 243                 BUG_ON(len + offset > PAGE_SIZE);
 244         }
 245         if (!Q(blocks, disks)) {
 246                 Q(blocks, disks) = pq_scribble_page;
 247                 BUG_ON(len + offset > PAGE_SIZE);
 248         }
 249         do_sync_gen_syndrome(blocks, offset, disks, len, submit);
 250
 251         return NULL;
 252 }
 253 EXPORT_SYMBOL_GPL(async_gen_syndrome);
 254
 255 static inline struct dma_chan *
 256 pq_val_chan(struct async_submit_ctl *submit, struct page **blocks, int disks, size_t len)
 257 {
 258         #ifdef CONFIG_ASYNC_TX_DISABLE_PQ_VAL_DMA
 259         return NULL;
 260         #endif
 261         return async_tx_find_channel(submit, DMA_PQ_VAL, NULL, 0,  blocks,
 262                                      disks, len);
 263 }
 264
 265 /**
 266  * async_syndrome_val - asynchronously validate a raid6 syndrome
 267  * @blocks: source blocks from idx 0..disks-3, P @ disks-2 and Q @ disks-1
 268  * @offset: common offset into each block (src and dest) to start transaction
 269  * @disks: number of blocks (including missing P or Q, see below)
 270  * @len: length of operation in bytes
 271  * @pqres: on val failure SUM_CHECK_P_RESULT and/or SUM_CHECK_Q_RESULT are set
 272  * @spare: temporary result buffer for the synchronous case
 273  * @submit: submission / completion modifiers
 274  *
 275  * The same notes from async_gen_syndrome apply to the 'blocks',
 276  * and 'disks' parameters of this routine.  The synchronous path
 277  * requires a temporary result buffer and submit->scribble to be
 278  * specified.
 279  */
 280 struct dma_async_tx_descriptor *
 281 async_syndrome_val(struct page **blocks, unsigned int offset, int disks,
 282                    size_t len, enum sum_check_flags *pqres, struct page *spare,
 283                    struct async_submit_ctl *submit)
 284 {
 285         struct dma_chan *chan = pq_val_chan(submit, blocks, disks, len);
 286         struct dma_device *device = chan ? chan->device : NULL;
 287         struct dma_async_tx_descriptor *tx;
 288         unsigned char coefs[MAX_DISKS];
 289         enum dma_ctrl_flags dma_flags = submit->cb_fn ? DMA_PREP_INTERRUPT : 0;
 290         struct dmaengine_unmap_data *unmap = NULL;
 291
 292         BUG_ON(disks < 4 || disks > MAX_DISKS);
 293
 294         if (device)
 295                 unmap = dmaengine_get_unmap_data(device->dev, disks, GFP_NOWAIT);
 296
 297         if (unmap && disks <= dma_maxpq(device, 0) &&
 298             is_dma_pq_aligned(device, offset, 0, len)) {
 299                 struct device *dev = device->dev;
 300                 dma_addr_t pq[2];
 301                 int i, j = 0, src_cnt = 0;
 302
 303                 pr_debug("%s: (async) disks: %d len: %zu\n",
 304                          __func__, disks, len);
 305
 306                 unmap->len = len;
 307                 for (i = 0; i < disks-2; i++)
 308                         if (likely(blocks[i])) {
 309                                 unmap->addr[j] = dma_map_page(dev, blocks[i],
 310                                                               offset, len,
 311                                                               DMA_TO_DEVICE);
 312                                 coefs[j] = raid6_gfexp[i];
 313                                 unmap->to_cnt++;
 314                                 src_cnt++;
 315                                 j++;
 316                         }
 317
 318                 if (!P(blocks, disks)) {
 319                         pq[0] = 0;
 320                         dma_flags |= DMA_PREP_PQ_DISABLE_P;
 321                 } else {
 322                         pq[0] = dma_map_page(dev, P(blocks, disks),
 323                                              offset, len,
 324                                              DMA_TO_DEVICE);
 325                         unmap->addr[j++] = pq[0];
 326                         unmap->to_cnt++;
 327                 }
 328                 if (!Q(blocks, disks)) {
 329                         pq[1] = 0;
 330                         dma_flags |= DMA_PREP_PQ_DISABLE_Q;
 331                 } else {
 332                         pq[1] = dma_map_page(dev, Q(blocks, disks),
 333                                              offset, len,
 334                                              DMA_TO_DEVICE);
 335                         unmap->addr[j++] = pq[1];
 336                         unmap->to_cnt++;
 337                 }
 338
 339                 if (submit->flags & ASYNC_TX_FENCE)
 340                         dma_flags |= DMA_PREP_FENCE;
 341                 for (;;) {
 342                         tx = device->device_prep_dma_pq_val(chan, pq,
 343                                                             unmap->addr,
 344                                                             src_cnt,
 345                                                             coefs,
 346                                                             len, pqres,
 347                                                             dma_flags);
 348                         if (likely(tx))
 349                                 break;
 350                         async_tx_quiesce(&submit->depend_tx);
 351                         dma_async_issue_pending(chan);
 352                 }
 353
 354                 dma_set_unmap(tx, unmap);
 355                 async_tx_submit(chan, tx, submit);
 356         } else {
 357                 struct page *p_src = P(blocks, disks);
 358                 struct page *q_src = Q(blocks, disks);
 359                 enum async_tx_flags flags_orig = submit->flags;
 360                 dma_async_tx_callback cb_fn_orig = submit->cb_fn;
 361                 void *scribble = submit->scribble;
 362                 void *cb_param_orig = submit->cb_param;
 363                 void *p, *q, *s;
 364
 365                 pr_debug("%s: (sync) disks: %d len: %zu\n",
 366                          __func__, disks, len);
 367
 368                 /* caller must provide a temporary result buffer and
 369                  * allow the input parameters to be preserved
 370                  */
 371                 BUG_ON(!spare || !scribble);
 372
 373                 /* wait for any prerequisite operations */
 374                 async_tx_quiesce(&submit->depend_tx);
 375
 376                 /* recompute p and/or q into the temporary buffer and then
 377                  * check to see the result matches the current value
 378                  */
 379                 tx = NULL;
 380                 *pqres = 0;
 381                 if (p_src) {
 382                         init_async_submit(submit, ASYNC_TX_XOR_ZERO_DST, NULL,
 383                                           NULL, NULL, scribble);
 384                         tx = async_xor(spare, blocks, offset, disks-2, len, submit);
 385                         async_tx_quiesce(&tx);
 386                         p = page_address(p_src) + offset;
 387                         s = page_address(spare) + offset;
 388                         *pqres |= !!memcmp(p, s, len) << SUM_CHECK_P;
 389                 }
 390
 391                 if (q_src) {
 392                         P(blocks, disks) = NULL;
 393                         Q(blocks, disks) = spare;
 394                         init_async_submit(submit, 0, NULL, NULL, NULL, scribble);
 395                         tx = async_gen_syndrome(blocks, offset, disks, len, submit);
 396                         async_tx_quiesce(&tx);
 397                         q = page_address(q_src) + offset;
 398                         s = page_address(spare) + offset;
 399                         *pqres |= !!memcmp(q, s, len) << SUM_CHECK_Q;
 400                 }
 401
 402                 /* restore P, Q and submit */
 403                 P(blocks, disks) = p_src;
 404                 Q(blocks, disks) = q_src;
 405
 406                 submit->cb_fn = cb_fn_orig;
 407                 submit->cb_param = cb_param_orig;
 408                 submit->flags = flags_orig;
 409                 async_tx_sync_epilog(submit);
 410                 tx = NULL;
 411         }
 412         dmaengine_unmap_put(unmap);
 413
 414         return tx;
 415 }
 416 EXPORT_SYMBOL_GPL(async_syndrome_val);
 417
 418 static int __init async_pq_init(void)
 419 {
 420         pq_scribble_page = alloc_page(GFP_KERNEL);
 421
 422         if (pq_scribble_page)
 423                 return 0;
 424
 425         pr_err("%s: failed to allocate required spare page\n", __func__);
 426
 427         return -ENOMEM;
 428 }
 429
 430 static void __exit async_pq_exit(void)
 431 {
 432         __free_page(pq_scribble_page);
 433 }
 434
 435 module_init(async_pq_init);
 436 module_exit(async_pq_exit);
 437
 438 MODULE_DESCRIPTION("asynchronous raid6 syndrome generation/validation");
 439 MODULE_LICENSE("GPL");