crypto/async_tx/async_pq.c

0001 // SPDX-License-Identifier: GPL-2.0-or-later
0002 /*
0003  * Copyright(c) 2007 Yuri Tikhonov <yur@emcraft.com>
0004  * Copyright(c) 2009 Intel Corporation
0005  */
0006 #include <linux/kernel.h>
0007 #include <linux/interrupt.h>
0008 #include <linux/module.h>
0009 #include <linux/dma-mapping.h>
0010 #include <linux/raid/pq.h>
0011 #include <linux/async_tx.h>
0012 #include <linux/gfp.h>
0013
0014 /**
0015  * pq_scribble_page - space to hold throwaway P or Q buffer for
0016  * synchronous gen_syndrome
0017  */
0018 static struct page *pq_scribble_page;
0019
0020 /* the struct page *blocks[] parameter passed to async_gen_syndrome()
0021  * and async_syndrome_val() contains the 'P' destination address at
0022  * blocks[disks-2] and the 'Q' destination address at blocks[disks-1]
0023  *
0024  * note: these are macros as they are used as lvalues
0025  */
0026 #define P(b, d) (b[d-2])
0027 #define Q(b, d) (b[d-1])
0028
0029 #define MAX_DISKS 255
0030
0031 /**
0032  * do_async_gen_syndrome - asynchronously calculate P and/or Q
0033  */
0034 static __async_inline struct dma_async_tx_descriptor *
0035 do_async_gen_syndrome(struct dma_chan *chan,
0036               const unsigned char *scfs, int disks,
0037               struct dmaengine_unmap_data *unmap,
0038               enum dma_ctrl_flags dma_flags,
0039               struct async_submit_ctl *submit)
0040 {
0041     struct dma_async_tx_descriptor *tx = NULL;
0042     struct dma_device *dma = chan->device;
0043     enum async_tx_flags flags_orig = submit->flags;
0044     dma_async_tx_callback cb_fn_orig = submit->cb_fn;
0045     dma_async_tx_callback cb_param_orig = submit->cb_param;
0046     int src_cnt = disks - 2;
0047     unsigned short pq_src_cnt;
0048     dma_addr_t dma_dest[2];
0049     int src_off = 0;
0050
0051     while (src_cnt > 0) {
0052         submit->flags = flags_orig;
0053         pq_src_cnt = min(src_cnt, dma_maxpq(dma, dma_flags));
0054         /* if we are submitting additional pqs, leave the chain open,
0055          * clear the callback parameters, and leave the destination
0056          * buffers mapped
0057          */
0058         if (src_cnt > pq_src_cnt) {
0059             submit->flags &= ~ASYNC_TX_ACK;
0060             submit->flags |= ASYNC_TX_FENCE;
0061             submit->cb_fn = NULL;
0062             submit->cb_param = NULL;
0063         } else {
0064             submit->cb_fn = cb_fn_orig;
0065             submit->cb_param = cb_param_orig;
0066             if (cb_fn_orig)
0067                 dma_flags |= DMA_PREP_INTERRUPT;
0068         }
0069         if (submit->flags & ASYNC_TX_FENCE)
0070             dma_flags |= DMA_PREP_FENCE;
0071
0072         /* Drivers force forward progress in case they can not provide
0073          * a descriptor
0074          */
0075         for (;;) {
0076             dma_dest[0] = unmap->addr[disks - 2];
0077             dma_dest[1] = unmap->addr[disks - 1];
0078             tx = dma->device_prep_dma_pq(chan, dma_dest,
0079                              &unmap->addr[src_off],
0080                              pq_src_cnt,
0081                              &scfs[src_off], unmap->len,
0082                              dma_flags);
0083             if (likely(tx))
0084                 break;
0085             async_tx_quiesce(&submit->depend_tx);
0086             dma_async_issue_pending(chan);
0087         }
0088
0089         dma_set_unmap(tx, unmap);
0090         async_tx_submit(chan, tx, submit);
0091         submit->depend_tx = tx;
0092
0093         /* drop completed sources */
0094         src_cnt -= pq_src_cnt;
0095         src_off += pq_src_cnt;
0096
0097         dma_flags |= DMA_PREP_CONTINUE;
0098     }
0099
0100     return tx;
0101 }
0102
0103 /**
0104  * do_sync_gen_syndrome - synchronously calculate a raid6 syndrome
0105  */
0106 static void
0107 do_sync_gen_syndrome(struct page **blocks, unsigned int *offsets, int disks,
0108              size_t len, struct async_submit_ctl *submit)
0109 {
0110     void **srcs;
0111     int i;
0112     int start = -1, stop = disks - 3;
0113
0114     if (submit->scribble)
0115         srcs = submit->scribble;
0116     else
0117         srcs = (void **) blocks;
0118
0119     for (i = 0; i < disks; i++) {
0120         if (blocks[i] == NULL) {
0121             BUG_ON(i > disks - 3); /* P or Q can't be zero */
0122             srcs[i] = (void*)raid6_empty_zero_page;
0123         } else {
0124             srcs[i] = page_address(blocks[i]) + offsets[i];
0125
0126             if (i < disks - 2) {
0127                 stop = i;
0128                 if (start == -1)
0129                     start = i;
0130             }
0131         }
0132     }
0133     if (submit->flags & ASYNC_TX_PQ_XOR_DST) {
0134         BUG_ON(!raid6_call.xor_syndrome);
0135         if (start >= 0)
0136             raid6_call.xor_syndrome(disks, start, stop, len, srcs);
0137     } else
0138         raid6_call.gen_syndrome(disks, len, srcs);
0139     async_tx_sync_epilog(submit);
0140 }
0141
0142 static inline bool
0143 is_dma_pq_aligned_offs(struct dma_device *dev, unsigned int *offs,
0144                      int src_cnt, size_t len)
0145 {
0146     int i;
0147
0148     for (i = 0; i < src_cnt; i++) {
0149         if (!is_dma_pq_aligned(dev, offs[i], 0, len))
0150             return false;
0151     }
0152     return true;
0153 }
0154
0155 /**
0156  * async_gen_syndrome - asynchronously calculate a raid6 syndrome
0157  * @blocks: source blocks from idx 0..disks-3, P @ disks-2 and Q @ disks-1
0158  * @offsets: offset array into each block (src and dest) to start transaction
0159  * @disks: number of blocks (including missing P or Q, see below)
0160  * @len: length of operation in bytes
0161  * @submit: submission/completion modifiers
0162  *
0163  * General note: This routine assumes a field of GF(2^8) with a
0164  * primitive polynomial of 0x11d and a generator of {02}.
0165  *
0166  * 'disks' note: callers can optionally omit either P or Q (but not
0167  * both) from the calculation by setting blocks[disks-2] or
0168  * blocks[disks-1] to NULL.  When P or Q is omitted 'len' must be <=
0169  * PAGE_SIZE as a temporary buffer of this size is used in the
0170  * synchronous path.  'disks' always accounts for both destination
0171  * buffers.  If any source buffers (blocks[i] where i < disks - 2) are
0172  * set to NULL those buffers will be replaced with the raid6_zero_page
0173  * in the synchronous path and omitted in the hardware-asynchronous
0174  * path.
0175  */
0176 struct dma_async_tx_descriptor *
0177 async_gen_syndrome(struct page **blocks, unsigned int *offsets, int disks,
0178            size_t len, struct async_submit_ctl *submit)
0179 {
0180     int src_cnt = disks - 2;
0181     struct dma_chan *chan = async_tx_find_channel(submit, DMA_PQ,
0182                               &P(blocks, disks), 2,
0183                               blocks, src_cnt, len);
0184     struct dma_device *device = chan ? chan->device : NULL;
0185     struct dmaengine_unmap_data *unmap = NULL;
0186
0187     BUG_ON(disks > MAX_DISKS || !(P(blocks, disks) || Q(blocks, disks)));
0188
0189     if (device)
0190         unmap = dmaengine_get_unmap_data(device->dev, disks, GFP_NOWAIT);
0191
0192     /* XORing P/Q is only implemented in software */
0193     if (unmap && !(submit->flags & ASYNC_TX_PQ_XOR_DST) &&
0194         (src_cnt <= dma_maxpq(device, 0) ||
0195          dma_maxpq(device, DMA_PREP_CONTINUE) > 0) &&
0196         is_dma_pq_aligned_offs(device, offsets, disks, len)) {
0197         struct dma_async_tx_descriptor *tx;
0198         enum dma_ctrl_flags dma_flags = 0;
0199         unsigned char coefs[MAX_DISKS];
0200         int i, j;
0201
0202         /* run the p+q asynchronously */
0203         pr_debug("%s: (async) disks: %d len: %zu\n",
0204              __func__, disks, len);
0205
0206         /* convert source addresses being careful to collapse 'empty'
0207          * sources and update the coefficients accordingly
0208          */
0209         unmap->len = len;
0210         for (i = 0, j = 0; i < src_cnt; i++) {
0211             if (blocks[i] == NULL)
0212                 continue;
0213             unmap->addr[j] = dma_map_page(device->dev, blocks[i],
0214                         offsets[i], len, DMA_TO_DEVICE);
0215             coefs[j] = raid6_gfexp[i];
0216             unmap->to_cnt++;
0217             j++;
0218         }
0219
0220         /*
0221          * DMAs use destinations as sources,
0222          * so use BIDIRECTIONAL mapping
0223          */
0224         unmap->bidi_cnt++;
0225         if (P(blocks, disks))
0226             unmap->addr[j++] = dma_map_page(device->dev, P(blocks, disks),
0227                             P(offsets, disks),
0228                             len, DMA_BIDIRECTIONAL);
0229         else {
0230             unmap->addr[j++] = 0;
0231             dma_flags |= DMA_PREP_PQ_DISABLE_P;
0232         }
0233
0234         unmap->bidi_cnt++;
0235         if (Q(blocks, disks))
0236             unmap->addr[j++] = dma_map_page(device->dev, Q(blocks, disks),
0237                             Q(offsets, disks),
0238                             len, DMA_BIDIRECTIONAL);
0239         else {
0240             unmap->addr[j++] = 0;
0241             dma_flags |= DMA_PREP_PQ_DISABLE_Q;
0242         }
0243
0244         tx = do_async_gen_syndrome(chan, coefs, j, unmap, dma_flags, submit);
0245         dmaengine_unmap_put(unmap);
0246         return tx;
0247     }
0248
0249     dmaengine_unmap_put(unmap);
0250
0251     /* run the pq synchronously */
0252     pr_debug("%s: (sync) disks: %d len: %zu\n", __func__, disks, len);
0253
0254     /* wait for any prerequisite operations */
0255     async_tx_quiesce(&submit->depend_tx);
0256
0257     if (!P(blocks, disks)) {
0258         P(blocks, disks) = pq_scribble_page;
0259         P(offsets, disks) = 0;
0260     }
0261     if (!Q(blocks, disks)) {
0262         Q(blocks, disks) = pq_scribble_page;
0263         Q(offsets, disks) = 0;
0264     }
0265     do_sync_gen_syndrome(blocks, offsets, disks, len, submit);
0266
0267     return NULL;
0268 }
0269 EXPORT_SYMBOL_GPL(async_gen_syndrome);
0270
0271 static inline struct dma_chan *
0272 pq_val_chan(struct async_submit_ctl *submit, struct page **blocks, int disks, size_t len)
0273 {
0274     #ifdef CONFIG_ASYNC_TX_DISABLE_PQ_VAL_DMA
0275     return NULL;
0276     #endif
0277     return async_tx_find_channel(submit, DMA_PQ_VAL, NULL, 0,  blocks,
0278                      disks, len);
0279 }
0280
0281 /**
0282  * async_syndrome_val - asynchronously validate a raid6 syndrome
0283  * @blocks: source blocks from idx 0..disks-3, P @ disks-2 and Q @ disks-1
0284  * @offset: common offset into each block (src and dest) to start transaction
0285  * @disks: number of blocks (including missing P or Q, see below)
0286  * @len: length of operation in bytes
0287  * @pqres: on val failure SUM_CHECK_P_RESULT and/or SUM_CHECK_Q_RESULT are set
0288  * @spare: temporary result buffer for the synchronous case
0289  * @s_off: spare buffer page offset
0290  * @submit: submission / completion modifiers
0291  *
0292  * The same notes from async_gen_syndrome apply to the 'blocks',
0293  * and 'disks' parameters of this routine.  The synchronous path
0294  * requires a temporary result buffer and submit->scribble to be
0295  * specified.
0296  */
0297 struct dma_async_tx_descriptor *
0298 async_syndrome_val(struct page **blocks, unsigned int *offsets, int disks,
0299            size_t len, enum sum_check_flags *pqres, struct page *spare,
0300            unsigned int s_off, struct async_submit_ctl *submit)
0301 {
0302     struct dma_chan *chan = pq_val_chan(submit, blocks, disks, len);
0303     struct dma_device *device = chan ? chan->device : NULL;
0304     struct dma_async_tx_descriptor *tx;
0305     unsigned char coefs[MAX_DISKS];
0306     enum dma_ctrl_flags dma_flags = submit->cb_fn ? DMA_PREP_INTERRUPT : 0;
0307     struct dmaengine_unmap_data *unmap = NULL;
0308
0309     BUG_ON(disks < 4 || disks > MAX_DISKS);
0310
0311     if (device)
0312         unmap = dmaengine_get_unmap_data(device->dev, disks, GFP_NOWAIT);
0313
0314     if (unmap && disks <= dma_maxpq(device, 0) &&
0315         is_dma_pq_aligned_offs(device, offsets, disks, len)) {
0316         struct device *dev = device->dev;
0317         dma_addr_t pq[2];
0318         int i, j = 0, src_cnt = 0;
0319
0320         pr_debug("%s: (async) disks: %d len: %zu\n",
0321              __func__, disks, len);
0322
0323         unmap->len = len;
0324         for (i = 0; i < disks-2; i++)
0325             if (likely(blocks[i])) {
0326                 unmap->addr[j] = dma_map_page(dev, blocks[i],
0327                                   offsets[i], len,
0328                                   DMA_TO_DEVICE);
0329                 coefs[j] = raid6_gfexp[i];
0330                 unmap->to_cnt++;
0331                 src_cnt++;
0332                 j++;
0333             }
0334
0335         if (!P(blocks, disks)) {
0336             pq[0] = 0;
0337             dma_flags |= DMA_PREP_PQ_DISABLE_P;
0338         } else {
0339             pq[0] = dma_map_page(dev, P(blocks, disks),
0340                          P(offsets, disks), len,
0341                          DMA_TO_DEVICE);
0342             unmap->addr[j++] = pq[0];
0343             unmap->to_cnt++;
0344         }
0345         if (!Q(blocks, disks)) {
0346             pq[1] = 0;
0347             dma_flags |= DMA_PREP_PQ_DISABLE_Q;
0348         } else {
0349             pq[1] = dma_map_page(dev, Q(blocks, disks),
0350                          Q(offsets, disks), len,
0351                          DMA_TO_DEVICE);
0352             unmap->addr[j++] = pq[1];
0353             unmap->to_cnt++;
0354         }
0355
0356         if (submit->flags & ASYNC_TX_FENCE)
0357             dma_flags |= DMA_PREP_FENCE;
0358         for (;;) {
0359             tx = device->device_prep_dma_pq_val(chan, pq,
0360                                 unmap->addr,
0361                                 src_cnt,
0362                                 coefs,
0363                                 len, pqres,
0364                                 dma_flags);
0365             if (likely(tx))
0366                 break;
0367             async_tx_quiesce(&submit->depend_tx);
0368             dma_async_issue_pending(chan);
0369         }
0370
0371         dma_set_unmap(tx, unmap);
0372         async_tx_submit(chan, tx, submit);
0373     } else {
0374         struct page *p_src = P(blocks, disks);
0375         unsigned int p_off = P(offsets, disks);
0376         struct page *q_src = Q(blocks, disks);
0377         unsigned int q_off = Q(offsets, disks);
0378         enum async_tx_flags flags_orig = submit->flags;
0379         dma_async_tx_callback cb_fn_orig = submit->cb_fn;
0380         void *scribble = submit->scribble;
0381         void *cb_param_orig = submit->cb_param;
0382         void *p, *q, *s;
0383
0384         pr_debug("%s: (sync) disks: %d len: %zu\n",
0385              __func__, disks, len);
0386
0387         /* caller must provide a temporary result buffer and
0388          * allow the input parameters to be preserved
0389          */
0390         BUG_ON(!spare || !scribble);
0391
0392         /* wait for any prerequisite operations */
0393         async_tx_quiesce(&submit->depend_tx);
0394
0395         /* recompute p and/or q into the temporary buffer and then
0396          * check to see the result matches the current value
0397          */
0398         tx = NULL;
0399         *pqres = 0;
0400         if (p_src) {
0401             init_async_submit(submit, ASYNC_TX_XOR_ZERO_DST, NULL,
0402                       NULL, NULL, scribble);
0403             tx = async_xor_offs(spare, s_off,
0404                     blocks, offsets, disks-2, len, submit);
0405             async_tx_quiesce(&tx);
0406             p = page_address(p_src) + p_off;
0407             s = page_address(spare) + s_off;
0408             *pqres |= !!memcmp(p, s, len) << SUM_CHECK_P;
0409         }
0410
0411         if (q_src) {
0412             P(blocks, disks) = NULL;
0413             Q(blocks, disks) = spare;
0414             Q(offsets, disks) = s_off;
0415             init_async_submit(submit, 0, NULL, NULL, NULL, scribble);
0416             tx = async_gen_syndrome(blocks, offsets, disks,
0417                     len, submit);
0418             async_tx_quiesce(&tx);
0419             q = page_address(q_src) + q_off;
0420             s = page_address(spare) + s_off;
0421             *pqres |= !!memcmp(q, s, len) << SUM_CHECK_Q;
0422         }
0423
0424         /* restore P, Q and submit */
0425         P(blocks, disks) = p_src;
0426         P(offsets, disks) = p_off;
0427         Q(blocks, disks) = q_src;
0428         Q(offsets, disks) = q_off;
0429
0430         submit->cb_fn = cb_fn_orig;
0431         submit->cb_param = cb_param_orig;
0432         submit->flags = flags_orig;
0433         async_tx_sync_epilog(submit);
0434         tx = NULL;
0435     }
0436     dmaengine_unmap_put(unmap);
0437
0438     return tx;
0439 }
0440 EXPORT_SYMBOL_GPL(async_syndrome_val);
0441
0442 static int __init async_pq_init(void)
0443 {
0444     pq_scribble_page = alloc_page(GFP_KERNEL);
0445
0446     if (pq_scribble_page)
0447         return 0;
0448
0449     pr_err("%s: failed to allocate required spare page\n", __func__);
0450
0451     return -ENOMEM;
0452 }
0453
0454 static void __exit async_pq_exit(void)
0455 {
0456     __free_page(pq_scribble_page);
0457 }
0458
0459 module_init(async_pq_init);
0460 module_exit(async_pq_exit);
0461
0462 MODULE_DESCRIPTION("asynchronous raid6 syndrome generation/validation");
0463 MODULE_LICENSE("GPL");