1 // SPDX-License-Identifier: GPL-2.0
2 /* Copyright (c) 2019 HiSilicon Limited. */
3 #include <crypto/internal/acompress.h>
4 #include <linux/bitfield.h>
5 #include <linux/bitmap.h>
6 #include <linux/dma-mapping.h>
7 #include <linux/scatterlist.h>
10 /* hisi_zip_sqe dw3 */
11 #define HZIP_BD_STATUS_M GENMASK(7, 0)
12 /* hisi_zip_sqe dw7 */
13 #define HZIP_IN_SGE_DATA_OFFSET_M GENMASK(23, 0)
14 #define HZIP_SQE_TYPE_M GENMASK(31, 28)
15 /* hisi_zip_sqe dw8 */
16 #define HZIP_OUT_SGE_DATA_OFFSET_M GENMASK(23, 0)
17 /* hisi_zip_sqe dw9 */
18 #define HZIP_REQ_TYPE_M GENMASK(7, 0)
19 #define HZIP_ALG_TYPE_DEFLATE 0x01
20 #define HZIP_BUF_TYPE_M GENMASK(11, 8)
23 #define HZIP_ALG_PRIORITY 300
24 #define HZIP_SGL_SGE_NR 10
26 #define HZIP_ALG_DEFLATE GENMASK(5, 4)
28 static DEFINE_MUTEX(zip_algs_lock);
29 static unsigned int zip_available_devs;
31 enum hisi_zip_alg_type {
32 HZIP_ALG_TYPE_COMP = 0,
33 HZIP_ALG_TYPE_DECOMP = 1,
42 #define COMP_NAME_TO_TYPE(alg_name) \
43 (!strcmp((alg_name), "deflate") ? HZIP_ALG_TYPE_DEFLATE : 0)
46 struct acomp_req *req;
47 struct hisi_acc_hw_sgl *hw_src;
48 struct hisi_acc_hw_sgl *hw_dst;
54 struct hisi_zip_req_q {
55 struct hisi_zip_req *q;
56 unsigned long *req_bitmap;
61 struct hisi_zip_qp_ctx {
63 struct hisi_zip_req_q req_q;
64 struct hisi_acc_sgl_pool *sgl_pool;
65 struct hisi_zip *zip_dev;
66 struct hisi_zip_ctx *ctx;
69 struct hisi_zip_sqe_ops {
71 void (*fill_addr)(struct hisi_zip_sqe *sqe, struct hisi_zip_req *req);
72 void (*fill_buf_size)(struct hisi_zip_sqe *sqe, struct hisi_zip_req *req);
73 void (*fill_buf_type)(struct hisi_zip_sqe *sqe, u8 buf_type);
74 void (*fill_req_type)(struct hisi_zip_sqe *sqe, u8 req_type);
75 void (*fill_tag)(struct hisi_zip_sqe *sqe, struct hisi_zip_req *req);
76 void (*fill_sqe_type)(struct hisi_zip_sqe *sqe, u8 sqe_type);
77 u32 (*get_tag)(struct hisi_zip_sqe *sqe);
78 u32 (*get_status)(struct hisi_zip_sqe *sqe);
79 u32 (*get_dstlen)(struct hisi_zip_sqe *sqe);
83 struct hisi_zip_qp_ctx qp_ctx[HZIP_CTX_Q_NUM];
84 const struct hisi_zip_sqe_ops *ops;
87 static int sgl_sge_nr_set(const char *val, const struct kernel_param *kp)
95 ret = kstrtou16(val, 10, &n);
96 if (ret || n == 0 || n > HISI_ACC_SGL_SGE_NR_MAX)
99 return param_set_ushort(val, kp);
102 static const struct kernel_param_ops sgl_sge_nr_ops = {
103 .set = sgl_sge_nr_set,
104 .get = param_get_ushort,
107 static u16 sgl_sge_nr = HZIP_SGL_SGE_NR;
108 module_param_cb(sgl_sge_nr, &sgl_sge_nr_ops, &sgl_sge_nr, 0444);
109 MODULE_PARM_DESC(sgl_sge_nr, "Number of sge in sgl(1-255)");
111 static struct hisi_zip_req *hisi_zip_create_req(struct hisi_zip_qp_ctx *qp_ctx,
112 struct acomp_req *req)
114 struct hisi_zip_req_q *req_q = &qp_ctx->req_q;
115 struct hisi_zip_req *q = req_q->q;
116 struct hisi_zip_req *req_cache;
119 write_lock(&req_q->req_lock);
121 req_id = find_first_zero_bit(req_q->req_bitmap, req_q->size);
122 if (req_id >= req_q->size) {
123 write_unlock(&req_q->req_lock);
124 dev_dbg(&qp_ctx->qp->qm->pdev->dev, "req cache is full!\n");
125 return ERR_PTR(-EAGAIN);
127 set_bit(req_id, req_q->req_bitmap);
129 write_unlock(&req_q->req_lock);
131 req_cache = q + req_id;
132 req_cache->req_id = req_id;
133 req_cache->req = req;
138 static void hisi_zip_remove_req(struct hisi_zip_qp_ctx *qp_ctx,
139 struct hisi_zip_req *req)
141 struct hisi_zip_req_q *req_q = &qp_ctx->req_q;
143 write_lock(&req_q->req_lock);
144 clear_bit(req->req_id, req_q->req_bitmap);
145 write_unlock(&req_q->req_lock);
148 static void hisi_zip_fill_addr(struct hisi_zip_sqe *sqe, struct hisi_zip_req *req)
150 sqe->source_addr_l = lower_32_bits(req->dma_src);
151 sqe->source_addr_h = upper_32_bits(req->dma_src);
152 sqe->dest_addr_l = lower_32_bits(req->dma_dst);
153 sqe->dest_addr_h = upper_32_bits(req->dma_dst);
156 static void hisi_zip_fill_buf_size(struct hisi_zip_sqe *sqe, struct hisi_zip_req *req)
158 struct acomp_req *a_req = req->req;
160 sqe->input_data_length = a_req->slen;
161 sqe->dest_avail_out = a_req->dlen;
164 static void hisi_zip_fill_buf_type(struct hisi_zip_sqe *sqe, u8 buf_type)
168 val = sqe->dw9 & ~HZIP_BUF_TYPE_M;
169 val |= FIELD_PREP(HZIP_BUF_TYPE_M, buf_type);
173 static void hisi_zip_fill_req_type(struct hisi_zip_sqe *sqe, u8 req_type)
177 val = sqe->dw9 & ~HZIP_REQ_TYPE_M;
178 val |= FIELD_PREP(HZIP_REQ_TYPE_M, req_type);
182 static void hisi_zip_fill_tag(struct hisi_zip_sqe *sqe, struct hisi_zip_req *req)
184 sqe->dw26 = req->req_id;
187 static void hisi_zip_fill_sqe_type(struct hisi_zip_sqe *sqe, u8 sqe_type)
191 val = sqe->dw7 & ~HZIP_SQE_TYPE_M;
192 val |= FIELD_PREP(HZIP_SQE_TYPE_M, sqe_type);
196 static void hisi_zip_fill_sqe(struct hisi_zip_ctx *ctx, struct hisi_zip_sqe *sqe,
197 u8 req_type, struct hisi_zip_req *req)
199 const struct hisi_zip_sqe_ops *ops = ctx->ops;
201 memset(sqe, 0, sizeof(struct hisi_zip_sqe));
203 ops->fill_addr(sqe, req);
204 ops->fill_buf_size(sqe, req);
205 ops->fill_buf_type(sqe, HZIP_SGL);
206 ops->fill_req_type(sqe, req_type);
207 ops->fill_tag(sqe, req);
208 ops->fill_sqe_type(sqe, ops->sqe_type);
211 static int hisi_zip_do_work(struct hisi_zip_qp_ctx *qp_ctx,
212 struct hisi_zip_req *req)
214 struct hisi_acc_sgl_pool *pool = qp_ctx->sgl_pool;
215 struct hisi_zip_dfx *dfx = &qp_ctx->zip_dev->dfx;
216 struct acomp_req *a_req = req->req;
217 struct hisi_qp *qp = qp_ctx->qp;
218 struct device *dev = &qp->qm->pdev->dev;
219 struct hisi_zip_sqe zip_sqe;
222 if (unlikely(!a_req->src || !a_req->slen || !a_req->dst || !a_req->dlen))
225 req->hw_src = hisi_acc_sg_buf_map_to_hw_sgl(dev, a_req->src, pool,
226 req->req_id << 1, &req->dma_src);
227 if (IS_ERR(req->hw_src)) {
228 dev_err(dev, "failed to map the src buffer to hw sgl (%ld)!\n",
229 PTR_ERR(req->hw_src));
230 return PTR_ERR(req->hw_src);
233 req->hw_dst = hisi_acc_sg_buf_map_to_hw_sgl(dev, a_req->dst, pool,
234 (req->req_id << 1) + 1,
236 if (IS_ERR(req->hw_dst)) {
237 ret = PTR_ERR(req->hw_dst);
238 dev_err(dev, "failed to map the dst buffer to hw slg (%d)!\n",
240 goto err_unmap_input;
243 hisi_zip_fill_sqe(qp_ctx->ctx, &zip_sqe, qp->req_type, req);
245 /* send command to start a task */
246 atomic64_inc(&dfx->send_cnt);
247 ret = hisi_qp_send(qp, &zip_sqe);
248 if (unlikely(ret < 0)) {
249 atomic64_inc(&dfx->send_busy_cnt);
251 dev_dbg_ratelimited(dev, "failed to send request!\n");
252 goto err_unmap_output;
258 hisi_acc_sg_buf_unmap(dev, a_req->dst, req->hw_dst);
260 hisi_acc_sg_buf_unmap(dev, a_req->src, req->hw_src);
264 static u32 hisi_zip_get_tag(struct hisi_zip_sqe *sqe)
269 static u32 hisi_zip_get_status(struct hisi_zip_sqe *sqe)
271 return sqe->dw3 & HZIP_BD_STATUS_M;
274 static u32 hisi_zip_get_dstlen(struct hisi_zip_sqe *sqe)
276 return sqe->produced;
279 static void hisi_zip_acomp_cb(struct hisi_qp *qp, void *data)
281 struct hisi_zip_qp_ctx *qp_ctx = qp->qp_ctx;
282 const struct hisi_zip_sqe_ops *ops = qp_ctx->ctx->ops;
283 struct hisi_zip_dfx *dfx = &qp_ctx->zip_dev->dfx;
284 struct hisi_zip_req_q *req_q = &qp_ctx->req_q;
285 struct device *dev = &qp->qm->pdev->dev;
286 struct hisi_zip_sqe *sqe = data;
287 u32 tag = ops->get_tag(sqe);
288 struct hisi_zip_req *req = req_q->q + tag;
289 struct acomp_req *acomp_req = req->req;
293 atomic64_inc(&dfx->recv_cnt);
294 status = ops->get_status(sqe);
295 if (unlikely(status != 0 && status != HZIP_NC_ERR)) {
296 dev_err(dev, "%scompress fail in qp%u: %u, output: %u\n",
297 (qp->alg_type == 0) ? "" : "de", qp->qp_id, status,
299 atomic64_inc(&dfx->err_bd_cnt);
303 hisi_acc_sg_buf_unmap(dev, acomp_req->src, req->hw_src);
304 hisi_acc_sg_buf_unmap(dev, acomp_req->dst, req->hw_dst);
306 acomp_req->dlen = ops->get_dstlen(sqe);
308 if (acomp_req->base.complete)
309 acomp_request_complete(acomp_req, err);
311 hisi_zip_remove_req(qp_ctx, req);
314 static int hisi_zip_acompress(struct acomp_req *acomp_req)
316 struct hisi_zip_ctx *ctx = crypto_tfm_ctx(acomp_req->base.tfm);
317 struct hisi_zip_qp_ctx *qp_ctx = &ctx->qp_ctx[HZIP_QPC_COMP];
318 struct device *dev = &qp_ctx->qp->qm->pdev->dev;
319 struct hisi_zip_req *req;
322 req = hisi_zip_create_req(qp_ctx, acomp_req);
326 ret = hisi_zip_do_work(qp_ctx, req);
327 if (unlikely(ret != -EINPROGRESS)) {
328 dev_info_ratelimited(dev, "failed to do compress (%d)!\n", ret);
329 hisi_zip_remove_req(qp_ctx, req);
335 static int hisi_zip_adecompress(struct acomp_req *acomp_req)
337 struct hisi_zip_ctx *ctx = crypto_tfm_ctx(acomp_req->base.tfm);
338 struct hisi_zip_qp_ctx *qp_ctx = &ctx->qp_ctx[HZIP_QPC_DECOMP];
339 struct device *dev = &qp_ctx->qp->qm->pdev->dev;
340 struct hisi_zip_req *req;
343 req = hisi_zip_create_req(qp_ctx, acomp_req);
347 ret = hisi_zip_do_work(qp_ctx, req);
348 if (unlikely(ret != -EINPROGRESS)) {
349 dev_info_ratelimited(dev, "failed to do decompress (%d)!\n",
351 hisi_zip_remove_req(qp_ctx, req);
357 static int hisi_zip_start_qp(struct hisi_qp *qp, struct hisi_zip_qp_ctx *qp_ctx,
358 int alg_type, int req_type)
360 struct device *dev = &qp->qm->pdev->dev;
363 qp->req_type = req_type;
364 qp->alg_type = alg_type;
367 ret = hisi_qm_start_qp(qp, 0);
369 dev_err(dev, "failed to start qp (%d)!\n", ret);
378 static void hisi_zip_release_qp(struct hisi_zip_qp_ctx *qp_ctx)
380 hisi_qm_stop_qp(qp_ctx->qp);
381 hisi_qm_free_qps(&qp_ctx->qp, 1);
384 static const struct hisi_zip_sqe_ops hisi_zip_ops = {
386 .fill_addr = hisi_zip_fill_addr,
387 .fill_buf_size = hisi_zip_fill_buf_size,
388 .fill_buf_type = hisi_zip_fill_buf_type,
389 .fill_req_type = hisi_zip_fill_req_type,
390 .fill_tag = hisi_zip_fill_tag,
391 .fill_sqe_type = hisi_zip_fill_sqe_type,
392 .get_tag = hisi_zip_get_tag,
393 .get_status = hisi_zip_get_status,
394 .get_dstlen = hisi_zip_get_dstlen,
397 static int hisi_zip_ctx_init(struct hisi_zip_ctx *hisi_zip_ctx, u8 req_type, int node)
399 struct hisi_qp *qps[HZIP_CTX_Q_NUM] = { NULL };
400 struct hisi_zip_qp_ctx *qp_ctx;
401 struct hisi_zip *hisi_zip;
404 ret = zip_create_qps(qps, HZIP_CTX_Q_NUM, node);
406 pr_err("failed to create zip qps (%d)!\n", ret);
410 hisi_zip = container_of(qps[0]->qm, struct hisi_zip, qm);
412 for (i = 0; i < HZIP_CTX_Q_NUM; i++) {
413 /* alg_type = 0 for compress, 1 for decompress in hw sqe */
414 qp_ctx = &hisi_zip_ctx->qp_ctx[i];
415 qp_ctx->ctx = hisi_zip_ctx;
416 ret = hisi_zip_start_qp(qps[i], qp_ctx, i, req_type);
418 for (j = i - 1; j >= 0; j--)
419 hisi_qm_stop_qp(hisi_zip_ctx->qp_ctx[j].qp);
421 hisi_qm_free_qps(qps, HZIP_CTX_Q_NUM);
425 qp_ctx->zip_dev = hisi_zip;
428 hisi_zip_ctx->ops = &hisi_zip_ops;
433 static void hisi_zip_ctx_exit(struct hisi_zip_ctx *hisi_zip_ctx)
437 for (i = 0; i < HZIP_CTX_Q_NUM; i++)
438 hisi_zip_release_qp(&hisi_zip_ctx->qp_ctx[i]);
441 static int hisi_zip_create_req_q(struct hisi_zip_ctx *ctx)
443 u16 q_depth = ctx->qp_ctx[0].qp->sq_depth;
444 struct hisi_zip_req_q *req_q;
447 for (i = 0; i < HZIP_CTX_Q_NUM; i++) {
448 req_q = &ctx->qp_ctx[i].req_q;
449 req_q->size = q_depth;
451 req_q->req_bitmap = bitmap_zalloc(req_q->size, GFP_KERNEL);
452 if (!req_q->req_bitmap) {
457 goto err_free_comp_q;
459 rwlock_init(&req_q->req_lock);
461 req_q->q = kcalloc(req_q->size, sizeof(struct hisi_zip_req),
466 goto err_free_comp_bitmap;
468 goto err_free_decomp_bitmap;
474 err_free_decomp_bitmap:
475 bitmap_free(ctx->qp_ctx[HZIP_QPC_DECOMP].req_q.req_bitmap);
477 kfree(ctx->qp_ctx[HZIP_QPC_COMP].req_q.q);
478 err_free_comp_bitmap:
479 bitmap_free(ctx->qp_ctx[HZIP_QPC_COMP].req_q.req_bitmap);
483 static void hisi_zip_release_req_q(struct hisi_zip_ctx *ctx)
487 for (i = 0; i < HZIP_CTX_Q_NUM; i++) {
488 kfree(ctx->qp_ctx[i].req_q.q);
489 bitmap_free(ctx->qp_ctx[i].req_q.req_bitmap);
493 static int hisi_zip_create_sgl_pool(struct hisi_zip_ctx *ctx)
495 u16 q_depth = ctx->qp_ctx[0].qp->sq_depth;
496 struct hisi_zip_qp_ctx *tmp;
500 for (i = 0; i < HZIP_CTX_Q_NUM; i++) {
501 tmp = &ctx->qp_ctx[i];
502 dev = &tmp->qp->qm->pdev->dev;
503 tmp->sgl_pool = hisi_acc_create_sgl_pool(dev, q_depth << 1,
505 if (IS_ERR(tmp->sgl_pool)) {
507 goto err_free_sgl_pool0;
515 hisi_acc_free_sgl_pool(&ctx->qp_ctx[HZIP_QPC_COMP].qp->qm->pdev->dev,
516 ctx->qp_ctx[HZIP_QPC_COMP].sgl_pool);
520 static void hisi_zip_release_sgl_pool(struct hisi_zip_ctx *ctx)
524 for (i = 0; i < HZIP_CTX_Q_NUM; i++)
525 hisi_acc_free_sgl_pool(&ctx->qp_ctx[i].qp->qm->pdev->dev,
526 ctx->qp_ctx[i].sgl_pool);
529 static void hisi_zip_set_acomp_cb(struct hisi_zip_ctx *ctx,
530 void (*fn)(struct hisi_qp *, void *))
534 for (i = 0; i < HZIP_CTX_Q_NUM; i++)
535 ctx->qp_ctx[i].qp->req_cb = fn;
538 static int hisi_zip_acomp_init(struct crypto_acomp *tfm)
540 const char *alg_name = crypto_tfm_alg_name(&tfm->base);
541 struct hisi_zip_ctx *ctx = crypto_tfm_ctx(&tfm->base);
545 ret = hisi_zip_ctx_init(ctx, COMP_NAME_TO_TYPE(alg_name), tfm->base.node);
547 pr_err("failed to init ctx (%d)!\n", ret);
551 dev = &ctx->qp_ctx[0].qp->qm->pdev->dev;
553 ret = hisi_zip_create_req_q(ctx);
555 dev_err(dev, "failed to create request queue (%d)!\n", ret);
559 ret = hisi_zip_create_sgl_pool(ctx);
561 dev_err(dev, "failed to create sgl pool (%d)!\n", ret);
562 goto err_release_req_q;
565 hisi_zip_set_acomp_cb(ctx, hisi_zip_acomp_cb);
570 hisi_zip_release_req_q(ctx);
572 hisi_zip_ctx_exit(ctx);
576 static void hisi_zip_acomp_exit(struct crypto_acomp *tfm)
578 struct hisi_zip_ctx *ctx = crypto_tfm_ctx(&tfm->base);
580 hisi_zip_set_acomp_cb(ctx, NULL);
581 hisi_zip_release_sgl_pool(ctx);
582 hisi_zip_release_req_q(ctx);
583 hisi_zip_ctx_exit(ctx);
586 static struct acomp_alg hisi_zip_acomp_deflate = {
587 .init = hisi_zip_acomp_init,
588 .exit = hisi_zip_acomp_exit,
589 .compress = hisi_zip_acompress,
590 .decompress = hisi_zip_adecompress,
592 .cra_name = "deflate",
593 .cra_driver_name = "hisi-deflate-acomp",
594 .cra_flags = CRYPTO_ALG_ASYNC,
595 .cra_module = THIS_MODULE,
596 .cra_priority = HZIP_ALG_PRIORITY,
597 .cra_ctxsize = sizeof(struct hisi_zip_ctx),
601 static int hisi_zip_register_deflate(struct hisi_qm *qm)
605 if (!hisi_zip_alg_support(qm, HZIP_ALG_DEFLATE))
608 ret = crypto_register_acomp(&hisi_zip_acomp_deflate);
610 dev_err(&qm->pdev->dev, "failed to register to deflate (%d)!\n", ret);
615 static void hisi_zip_unregister_deflate(struct hisi_qm *qm)
617 if (!hisi_zip_alg_support(qm, HZIP_ALG_DEFLATE))
620 crypto_unregister_acomp(&hisi_zip_acomp_deflate);
623 int hisi_zip_register_to_crypto(struct hisi_qm *qm)
627 mutex_lock(&zip_algs_lock);
628 if (zip_available_devs++)
631 ret = hisi_zip_register_deflate(qm);
633 zip_available_devs--;
636 mutex_unlock(&zip_algs_lock);
640 void hisi_zip_unregister_from_crypto(struct hisi_qm *qm)
642 mutex_lock(&zip_algs_lock);
643 if (--zip_available_devs)
646 hisi_zip_unregister_deflate(qm);
649 mutex_unlock(&zip_algs_lock);