Linux Audio

Check our new training course

Loading...
v6.8
   1// SPDX-License-Identifier: GPL-2.0-only
   2/* n2_core.c: Niagara2 Stream Processing Unit (SPU) crypto support.
   3 *
   4 * Copyright (C) 2010, 2011 David S. Miller <davem@davemloft.net>
   5 */
   6
   7#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
   8
   9#include <linux/kernel.h>
  10#include <linux/module.h>
  11#include <linux/of.h>
  12#include <linux/of_address.h>
  13#include <linux/platform_device.h>
  14#include <linux/cpumask.h>
  15#include <linux/slab.h>
  16#include <linux/interrupt.h>
  17#include <linux/crypto.h>
  18#include <crypto/md5.h>
  19#include <crypto/sha1.h>
  20#include <crypto/sha2.h>
  21#include <crypto/aes.h>
  22#include <crypto/internal/des.h>
  23#include <linux/mutex.h>
  24#include <linux/delay.h>
  25#include <linux/sched.h>
  26
  27#include <crypto/internal/hash.h>
  28#include <crypto/internal/skcipher.h>
  29#include <crypto/scatterwalk.h>
  30#include <crypto/algapi.h>
  31
  32#include <asm/hypervisor.h>
  33#include <asm/mdesc.h>
  34
  35#include "n2_core.h"
  36
  37#define DRV_MODULE_NAME		"n2_crypto"
  38#define DRV_MODULE_VERSION	"0.2"
  39#define DRV_MODULE_RELDATE	"July 28, 2011"
  40
  41static const char version[] =
  42	DRV_MODULE_NAME ".c:v" DRV_MODULE_VERSION " (" DRV_MODULE_RELDATE ")\n";
  43
  44MODULE_AUTHOR("David S. Miller (davem@davemloft.net)");
  45MODULE_DESCRIPTION("Niagara2 Crypto driver");
  46MODULE_LICENSE("GPL");
  47MODULE_VERSION(DRV_MODULE_VERSION);
  48
  49#define N2_CRA_PRIORITY		200
  50
  51static DEFINE_MUTEX(spu_lock);
  52
  53struct spu_queue {
  54	cpumask_t		sharing;
  55	unsigned long		qhandle;
  56
  57	spinlock_t		lock;
  58	u8			q_type;
  59	void			*q;
  60	unsigned long		head;
  61	unsigned long		tail;
  62	struct list_head	jobs;
  63
  64	unsigned long		devino;
  65
  66	char			irq_name[32];
  67	unsigned int		irq;
  68
  69	struct list_head	list;
  70};
  71
  72struct spu_qreg {
  73	struct spu_queue	*queue;
  74	unsigned long		type;
  75};
  76
  77static struct spu_queue **cpu_to_cwq;
  78static struct spu_queue **cpu_to_mau;
  79
  80static unsigned long spu_next_offset(struct spu_queue *q, unsigned long off)
  81{
  82	if (q->q_type == HV_NCS_QTYPE_MAU) {
  83		off += MAU_ENTRY_SIZE;
  84		if (off == (MAU_ENTRY_SIZE * MAU_NUM_ENTRIES))
  85			off = 0;
  86	} else {
  87		off += CWQ_ENTRY_SIZE;
  88		if (off == (CWQ_ENTRY_SIZE * CWQ_NUM_ENTRIES))
  89			off = 0;
  90	}
  91	return off;
  92}
  93
  94struct n2_request_common {
  95	struct list_head	entry;
  96	unsigned int		offset;
  97};
  98#define OFFSET_NOT_RUNNING	(~(unsigned int)0)
  99
 100/* An async job request records the final tail value it used in
 101 * n2_request_common->offset, test to see if that offset is in
 102 * the range old_head, new_head, inclusive.
 103 */
 104static inline bool job_finished(struct spu_queue *q, unsigned int offset,
 105				unsigned long old_head, unsigned long new_head)
 106{
 107	if (old_head <= new_head) {
 108		if (offset > old_head && offset <= new_head)
 109			return true;
 110	} else {
 111		if (offset > old_head || offset <= new_head)
 112			return true;
 113	}
 114	return false;
 115}
 116
 117/* When the HEAD marker is unequal to the actual HEAD, we get
 118 * a virtual device INO interrupt.  We should process the
 119 * completed CWQ entries and adjust the HEAD marker to clear
 120 * the IRQ.
 121 */
 122static irqreturn_t cwq_intr(int irq, void *dev_id)
 123{
 124	unsigned long off, new_head, hv_ret;
 125	struct spu_queue *q = dev_id;
 126
 127	pr_err("CPU[%d]: Got CWQ interrupt for qhdl[%lx]\n",
 128	       smp_processor_id(), q->qhandle);
 129
 130	spin_lock(&q->lock);
 131
 132	hv_ret = sun4v_ncs_gethead(q->qhandle, &new_head);
 133
 134	pr_err("CPU[%d]: CWQ gethead[%lx] hv_ret[%lu]\n",
 135	       smp_processor_id(), new_head, hv_ret);
 136
 137	for (off = q->head; off != new_head; off = spu_next_offset(q, off)) {
 138		/* XXX ... XXX */
 139	}
 140
 141	hv_ret = sun4v_ncs_sethead_marker(q->qhandle, new_head);
 142	if (hv_ret == HV_EOK)
 143		q->head = new_head;
 144
 145	spin_unlock(&q->lock);
 146
 147	return IRQ_HANDLED;
 148}
 149
 150static irqreturn_t mau_intr(int irq, void *dev_id)
 151{
 152	struct spu_queue *q = dev_id;
 153	unsigned long head, hv_ret;
 154
 155	spin_lock(&q->lock);
 156
 157	pr_err("CPU[%d]: Got MAU interrupt for qhdl[%lx]\n",
 158	       smp_processor_id(), q->qhandle);
 159
 160	hv_ret = sun4v_ncs_gethead(q->qhandle, &head);
 161
 162	pr_err("CPU[%d]: MAU gethead[%lx] hv_ret[%lu]\n",
 163	       smp_processor_id(), head, hv_ret);
 164
 165	sun4v_ncs_sethead_marker(q->qhandle, head);
 166
 167	spin_unlock(&q->lock);
 168
 169	return IRQ_HANDLED;
 170}
 171
 172static void *spu_queue_next(struct spu_queue *q, void *cur)
 173{
 174	return q->q + spu_next_offset(q, cur - q->q);
 175}
 176
 177static int spu_queue_num_free(struct spu_queue *q)
 178{
 179	unsigned long head = q->head;
 180	unsigned long tail = q->tail;
 181	unsigned long end = (CWQ_ENTRY_SIZE * CWQ_NUM_ENTRIES);
 182	unsigned long diff;
 183
 184	if (head > tail)
 185		diff = head - tail;
 186	else
 187		diff = (end - tail) + head;
 188
 189	return (diff / CWQ_ENTRY_SIZE) - 1;
 190}
 191
 192static void *spu_queue_alloc(struct spu_queue *q, int num_entries)
 193{
 194	int avail = spu_queue_num_free(q);
 195
 196	if (avail >= num_entries)
 197		return q->q + q->tail;
 198
 199	return NULL;
 200}
 201
 202static unsigned long spu_queue_submit(struct spu_queue *q, void *last)
 203{
 204	unsigned long hv_ret, new_tail;
 205
 206	new_tail = spu_next_offset(q, last - q->q);
 207
 208	hv_ret = sun4v_ncs_settail(q->qhandle, new_tail);
 209	if (hv_ret == HV_EOK)
 210		q->tail = new_tail;
 211	return hv_ret;
 212}
 213
 214static u64 control_word_base(unsigned int len, unsigned int hmac_key_len,
 215			     int enc_type, int auth_type,
 216			     unsigned int hash_len,
 217			     bool sfas, bool sob, bool eob, bool encrypt,
 218			     int opcode)
 219{
 220	u64 word = (len - 1) & CONTROL_LEN;
 221
 222	word |= ((u64) opcode << CONTROL_OPCODE_SHIFT);
 223	word |= ((u64) enc_type << CONTROL_ENC_TYPE_SHIFT);
 224	word |= ((u64) auth_type << CONTROL_AUTH_TYPE_SHIFT);
 225	if (sfas)
 226		word |= CONTROL_STORE_FINAL_AUTH_STATE;
 227	if (sob)
 228		word |= CONTROL_START_OF_BLOCK;
 229	if (eob)
 230		word |= CONTROL_END_OF_BLOCK;
 231	if (encrypt)
 232		word |= CONTROL_ENCRYPT;
 233	if (hmac_key_len)
 234		word |= ((u64) (hmac_key_len - 1)) << CONTROL_HMAC_KEY_LEN_SHIFT;
 235	if (hash_len)
 236		word |= ((u64) (hash_len - 1)) << CONTROL_HASH_LEN_SHIFT;
 237
 238	return word;
 239}
 240
 241#if 0
 242static inline bool n2_should_run_async(struct spu_queue *qp, int this_len)
 243{
 244	if (this_len >= 64 ||
 245	    qp->head != qp->tail)
 246		return true;
 247	return false;
 248}
 249#endif
 250
 251struct n2_ahash_alg {
 252	struct list_head	entry;
 253	const u8		*hash_zero;
 254	const u8		*hash_init;
 255	u8			hw_op_hashsz;
 256	u8			digest_size;
 257	u8			auth_type;
 258	u8			hmac_type;
 259	struct ahash_alg	alg;
 260};
 261
 262static inline struct n2_ahash_alg *n2_ahash_alg(struct crypto_tfm *tfm)
 263{
 264	struct crypto_alg *alg = tfm->__crt_alg;
 265	struct ahash_alg *ahash_alg;
 266
 267	ahash_alg = container_of(alg, struct ahash_alg, halg.base);
 268
 269	return container_of(ahash_alg, struct n2_ahash_alg, alg);
 270}
 271
 272struct n2_hmac_alg {
 273	const char		*child_alg;
 274	struct n2_ahash_alg	derived;
 275};
 276
 277static inline struct n2_hmac_alg *n2_hmac_alg(struct crypto_tfm *tfm)
 278{
 279	struct crypto_alg *alg = tfm->__crt_alg;
 280	struct ahash_alg *ahash_alg;
 281
 282	ahash_alg = container_of(alg, struct ahash_alg, halg.base);
 283
 284	return container_of(ahash_alg, struct n2_hmac_alg, derived.alg);
 285}
 286
 287struct n2_hash_ctx {
 288	struct crypto_ahash		*fallback_tfm;
 289};
 290
 291#define N2_HASH_KEY_MAX			32 /* HW limit for all HMAC requests */
 292
 293struct n2_hmac_ctx {
 294	struct n2_hash_ctx		base;
 295
 296	struct crypto_shash		*child_shash;
 297
 298	int				hash_key_len;
 299	unsigned char			hash_key[N2_HASH_KEY_MAX];
 300};
 301
 302struct n2_hash_req_ctx {
 303	union {
 304		struct md5_state	md5;
 305		struct sha1_state	sha1;
 306		struct sha256_state	sha256;
 307	} u;
 308
 309	struct ahash_request		fallback_req;
 310};
 311
 312static int n2_hash_async_init(struct ahash_request *req)
 313{
 314	struct n2_hash_req_ctx *rctx = ahash_request_ctx(req);
 315	struct crypto_ahash *tfm = crypto_ahash_reqtfm(req);
 316	struct n2_hash_ctx *ctx = crypto_ahash_ctx(tfm);
 317
 318	ahash_request_set_tfm(&rctx->fallback_req, ctx->fallback_tfm);
 319	rctx->fallback_req.base.flags = req->base.flags & CRYPTO_TFM_REQ_MAY_SLEEP;
 320
 321	return crypto_ahash_init(&rctx->fallback_req);
 322}
 323
 324static int n2_hash_async_update(struct ahash_request *req)
 325{
 326	struct n2_hash_req_ctx *rctx = ahash_request_ctx(req);
 327	struct crypto_ahash *tfm = crypto_ahash_reqtfm(req);
 328	struct n2_hash_ctx *ctx = crypto_ahash_ctx(tfm);
 329
 330	ahash_request_set_tfm(&rctx->fallback_req, ctx->fallback_tfm);
 331	rctx->fallback_req.base.flags = req->base.flags & CRYPTO_TFM_REQ_MAY_SLEEP;
 332	rctx->fallback_req.nbytes = req->nbytes;
 333	rctx->fallback_req.src = req->src;
 334
 335	return crypto_ahash_update(&rctx->fallback_req);
 336}
 337
 338static int n2_hash_async_final(struct ahash_request *req)
 339{
 340	struct n2_hash_req_ctx *rctx = ahash_request_ctx(req);
 341	struct crypto_ahash *tfm = crypto_ahash_reqtfm(req);
 342	struct n2_hash_ctx *ctx = crypto_ahash_ctx(tfm);
 343
 344	ahash_request_set_tfm(&rctx->fallback_req, ctx->fallback_tfm);
 345	rctx->fallback_req.base.flags = req->base.flags & CRYPTO_TFM_REQ_MAY_SLEEP;
 346	rctx->fallback_req.result = req->result;
 347
 348	return crypto_ahash_final(&rctx->fallback_req);
 349}
 350
 351static int n2_hash_async_finup(struct ahash_request *req)
 352{
 353	struct n2_hash_req_ctx *rctx = ahash_request_ctx(req);
 354	struct crypto_ahash *tfm = crypto_ahash_reqtfm(req);
 355	struct n2_hash_ctx *ctx = crypto_ahash_ctx(tfm);
 356
 357	ahash_request_set_tfm(&rctx->fallback_req, ctx->fallback_tfm);
 358	rctx->fallback_req.base.flags = req->base.flags & CRYPTO_TFM_REQ_MAY_SLEEP;
 359	rctx->fallback_req.nbytes = req->nbytes;
 360	rctx->fallback_req.src = req->src;
 361	rctx->fallback_req.result = req->result;
 362
 363	return crypto_ahash_finup(&rctx->fallback_req);
 364}
 365
 366static int n2_hash_async_noimport(struct ahash_request *req, const void *in)
 367{
 368	return -ENOSYS;
 369}
 370
 371static int n2_hash_async_noexport(struct ahash_request *req, void *out)
 372{
 373	return -ENOSYS;
 374}
 375
 376static int n2_hash_cra_init(struct crypto_tfm *tfm)
 377{
 378	const char *fallback_driver_name = crypto_tfm_alg_name(tfm);
 379	struct crypto_ahash *ahash = __crypto_ahash_cast(tfm);
 380	struct n2_hash_ctx *ctx = crypto_ahash_ctx(ahash);
 381	struct crypto_ahash *fallback_tfm;
 382	int err;
 383
 384	fallback_tfm = crypto_alloc_ahash(fallback_driver_name, 0,
 385					  CRYPTO_ALG_NEED_FALLBACK);
 386	if (IS_ERR(fallback_tfm)) {
 387		pr_warn("Fallback driver '%s' could not be loaded!\n",
 388			fallback_driver_name);
 389		err = PTR_ERR(fallback_tfm);
 390		goto out;
 391	}
 392
 393	crypto_ahash_set_reqsize(ahash, (sizeof(struct n2_hash_req_ctx) +
 394					 crypto_ahash_reqsize(fallback_tfm)));
 395
 396	ctx->fallback_tfm = fallback_tfm;
 397	return 0;
 398
 399out:
 400	return err;
 401}
 402
 403static void n2_hash_cra_exit(struct crypto_tfm *tfm)
 404{
 405	struct crypto_ahash *ahash = __crypto_ahash_cast(tfm);
 406	struct n2_hash_ctx *ctx = crypto_ahash_ctx(ahash);
 407
 408	crypto_free_ahash(ctx->fallback_tfm);
 409}
 410
 411static int n2_hmac_cra_init(struct crypto_tfm *tfm)
 412{
 413	const char *fallback_driver_name = crypto_tfm_alg_name(tfm);
 414	struct crypto_ahash *ahash = __crypto_ahash_cast(tfm);
 415	struct n2_hmac_ctx *ctx = crypto_ahash_ctx(ahash);
 416	struct n2_hmac_alg *n2alg = n2_hmac_alg(tfm);
 417	struct crypto_ahash *fallback_tfm;
 418	struct crypto_shash *child_shash;
 419	int err;
 420
 421	fallback_tfm = crypto_alloc_ahash(fallback_driver_name, 0,
 422					  CRYPTO_ALG_NEED_FALLBACK);
 423	if (IS_ERR(fallback_tfm)) {
 424		pr_warn("Fallback driver '%s' could not be loaded!\n",
 425			fallback_driver_name);
 426		err = PTR_ERR(fallback_tfm);
 427		goto out;
 428	}
 429
 430	child_shash = crypto_alloc_shash(n2alg->child_alg, 0, 0);
 431	if (IS_ERR(child_shash)) {
 432		pr_warn("Child shash '%s' could not be loaded!\n",
 433			n2alg->child_alg);
 434		err = PTR_ERR(child_shash);
 435		goto out_free_fallback;
 436	}
 437
 438	crypto_ahash_set_reqsize(ahash, (sizeof(struct n2_hash_req_ctx) +
 439					 crypto_ahash_reqsize(fallback_tfm)));
 440
 441	ctx->child_shash = child_shash;
 442	ctx->base.fallback_tfm = fallback_tfm;
 443	return 0;
 444
 445out_free_fallback:
 446	crypto_free_ahash(fallback_tfm);
 447
 448out:
 449	return err;
 450}
 451
 452static void n2_hmac_cra_exit(struct crypto_tfm *tfm)
 453{
 454	struct crypto_ahash *ahash = __crypto_ahash_cast(tfm);
 455	struct n2_hmac_ctx *ctx = crypto_ahash_ctx(ahash);
 456
 457	crypto_free_ahash(ctx->base.fallback_tfm);
 458	crypto_free_shash(ctx->child_shash);
 459}
 460
 461static int n2_hmac_async_setkey(struct crypto_ahash *tfm, const u8 *key,
 462				unsigned int keylen)
 463{
 464	struct n2_hmac_ctx *ctx = crypto_ahash_ctx(tfm);
 465	struct crypto_shash *child_shash = ctx->child_shash;
 466	struct crypto_ahash *fallback_tfm;
 467	int err, bs, ds;
 468
 469	fallback_tfm = ctx->base.fallback_tfm;
 470	err = crypto_ahash_setkey(fallback_tfm, key, keylen);
 471	if (err)
 472		return err;
 473
 474	bs = crypto_shash_blocksize(child_shash);
 475	ds = crypto_shash_digestsize(child_shash);
 476	BUG_ON(ds > N2_HASH_KEY_MAX);
 477	if (keylen > bs) {
 478		err = crypto_shash_tfm_digest(child_shash, key, keylen,
 479					      ctx->hash_key);
 480		if (err)
 481			return err;
 482		keylen = ds;
 483	} else if (keylen <= N2_HASH_KEY_MAX)
 484		memcpy(ctx->hash_key, key, keylen);
 485
 486	ctx->hash_key_len = keylen;
 487
 488	return err;
 489}
 490
 491static unsigned long wait_for_tail(struct spu_queue *qp)
 492{
 493	unsigned long head, hv_ret;
 494
 495	do {
 496		hv_ret = sun4v_ncs_gethead(qp->qhandle, &head);
 497		if (hv_ret != HV_EOK) {
 498			pr_err("Hypervisor error on gethead\n");
 499			break;
 500		}
 501		if (head == qp->tail) {
 502			qp->head = head;
 503			break;
 504		}
 505	} while (1);
 506	return hv_ret;
 507}
 508
 509static unsigned long submit_and_wait_for_tail(struct spu_queue *qp,
 510					      struct cwq_initial_entry *ent)
 511{
 512	unsigned long hv_ret = spu_queue_submit(qp, ent);
 513
 514	if (hv_ret == HV_EOK)
 515		hv_ret = wait_for_tail(qp);
 516
 517	return hv_ret;
 518}
 519
 520static int n2_do_async_digest(struct ahash_request *req,
 521			      unsigned int auth_type, unsigned int digest_size,
 522			      unsigned int result_size, void *hash_loc,
 523			      unsigned long auth_key, unsigned int auth_key_len)
 524{
 525	struct crypto_ahash *tfm = crypto_ahash_reqtfm(req);
 526	struct cwq_initial_entry *ent;
 527	struct crypto_hash_walk walk;
 528	struct spu_queue *qp;
 529	unsigned long flags;
 530	int err = -ENODEV;
 531	int nbytes, cpu;
 532
 533	/* The total effective length of the operation may not
 534	 * exceed 2^16.
 535	 */
 536	if (unlikely(req->nbytes > (1 << 16))) {
 537		struct n2_hash_req_ctx *rctx = ahash_request_ctx(req);
 538		struct n2_hash_ctx *ctx = crypto_ahash_ctx(tfm);
 539
 540		ahash_request_set_tfm(&rctx->fallback_req, ctx->fallback_tfm);
 541		rctx->fallback_req.base.flags =
 542			req->base.flags & CRYPTO_TFM_REQ_MAY_SLEEP;
 543		rctx->fallback_req.nbytes = req->nbytes;
 544		rctx->fallback_req.src = req->src;
 545		rctx->fallback_req.result = req->result;
 546
 547		return crypto_ahash_digest(&rctx->fallback_req);
 548	}
 549
 550	nbytes = crypto_hash_walk_first(req, &walk);
 551
 552	cpu = get_cpu();
 553	qp = cpu_to_cwq[cpu];
 554	if (!qp)
 555		goto out;
 556
 557	spin_lock_irqsave(&qp->lock, flags);
 558
 559	/* XXX can do better, improve this later by doing a by-hand scatterlist
 560	 * XXX walk, etc.
 561	 */
 562	ent = qp->q + qp->tail;
 563
 564	ent->control = control_word_base(nbytes, auth_key_len, 0,
 565					 auth_type, digest_size,
 566					 false, true, false, false,
 567					 OPCODE_INPLACE_BIT |
 568					 OPCODE_AUTH_MAC);
 569	ent->src_addr = __pa(walk.data);
 570	ent->auth_key_addr = auth_key;
 571	ent->auth_iv_addr = __pa(hash_loc);
 572	ent->final_auth_state_addr = 0UL;
 573	ent->enc_key_addr = 0UL;
 574	ent->enc_iv_addr = 0UL;
 575	ent->dest_addr = __pa(hash_loc);
 576
 577	nbytes = crypto_hash_walk_done(&walk, 0);
 578	while (nbytes > 0) {
 579		ent = spu_queue_next(qp, ent);
 580
 581		ent->control = (nbytes - 1);
 582		ent->src_addr = __pa(walk.data);
 583		ent->auth_key_addr = 0UL;
 584		ent->auth_iv_addr = 0UL;
 585		ent->final_auth_state_addr = 0UL;
 586		ent->enc_key_addr = 0UL;
 587		ent->enc_iv_addr = 0UL;
 588		ent->dest_addr = 0UL;
 589
 590		nbytes = crypto_hash_walk_done(&walk, 0);
 591	}
 592	ent->control |= CONTROL_END_OF_BLOCK;
 593
 594	if (submit_and_wait_for_tail(qp, ent) != HV_EOK)
 595		err = -EINVAL;
 596	else
 597		err = 0;
 598
 599	spin_unlock_irqrestore(&qp->lock, flags);
 600
 601	if (!err)
 602		memcpy(req->result, hash_loc, result_size);
 603out:
 604	put_cpu();
 605
 606	return err;
 607}
 608
 609static int n2_hash_async_digest(struct ahash_request *req)
 610{
 611	struct n2_ahash_alg *n2alg = n2_ahash_alg(req->base.tfm);
 612	struct n2_hash_req_ctx *rctx = ahash_request_ctx(req);
 613	int ds;
 614
 615	ds = n2alg->digest_size;
 616	if (unlikely(req->nbytes == 0)) {
 617		memcpy(req->result, n2alg->hash_zero, ds);
 618		return 0;
 619	}
 620	memcpy(&rctx->u, n2alg->hash_init, n2alg->hw_op_hashsz);
 621
 622	return n2_do_async_digest(req, n2alg->auth_type,
 623				  n2alg->hw_op_hashsz, ds,
 624				  &rctx->u, 0UL, 0);
 625}
 626
 627static int n2_hmac_async_digest(struct ahash_request *req)
 628{
 629	struct n2_hmac_alg *n2alg = n2_hmac_alg(req->base.tfm);
 630	struct n2_hash_req_ctx *rctx = ahash_request_ctx(req);
 631	struct crypto_ahash *tfm = crypto_ahash_reqtfm(req);
 632	struct n2_hmac_ctx *ctx = crypto_ahash_ctx(tfm);
 633	int ds;
 634
 635	ds = n2alg->derived.digest_size;
 636	if (unlikely(req->nbytes == 0) ||
 637	    unlikely(ctx->hash_key_len > N2_HASH_KEY_MAX)) {
 638		struct n2_hash_req_ctx *rctx = ahash_request_ctx(req);
 639		struct n2_hash_ctx *ctx = crypto_ahash_ctx(tfm);
 640
 641		ahash_request_set_tfm(&rctx->fallback_req, ctx->fallback_tfm);
 642		rctx->fallback_req.base.flags =
 643			req->base.flags & CRYPTO_TFM_REQ_MAY_SLEEP;
 644		rctx->fallback_req.nbytes = req->nbytes;
 645		rctx->fallback_req.src = req->src;
 646		rctx->fallback_req.result = req->result;
 647
 648		return crypto_ahash_digest(&rctx->fallback_req);
 649	}
 650	memcpy(&rctx->u, n2alg->derived.hash_init,
 651	       n2alg->derived.hw_op_hashsz);
 652
 653	return n2_do_async_digest(req, n2alg->derived.hmac_type,
 654				  n2alg->derived.hw_op_hashsz, ds,
 655				  &rctx->u,
 656				  __pa(&ctx->hash_key),
 657				  ctx->hash_key_len);
 658}
 659
 660struct n2_skcipher_context {
 661	int			key_len;
 662	int			enc_type;
 663	union {
 664		u8		aes[AES_MAX_KEY_SIZE];
 665		u8		des[DES_KEY_SIZE];
 666		u8		des3[3 * DES_KEY_SIZE];
 667	} key;
 668};
 669
 670#define N2_CHUNK_ARR_LEN	16
 671
 672struct n2_crypto_chunk {
 673	struct list_head	entry;
 674	unsigned long		iv_paddr : 44;
 675	unsigned long		arr_len : 20;
 676	unsigned long		dest_paddr;
 677	unsigned long		dest_final;
 678	struct {
 679		unsigned long	src_paddr : 44;
 680		unsigned long	src_len : 20;
 681	} arr[N2_CHUNK_ARR_LEN];
 682};
 683
 684struct n2_request_context {
 685	struct skcipher_walk	walk;
 686	struct list_head	chunk_list;
 687	struct n2_crypto_chunk	chunk;
 688	u8			temp_iv[16];
 689};
 690
 691/* The SPU allows some level of flexibility for partial cipher blocks
 692 * being specified in a descriptor.
 693 *
 694 * It merely requires that every descriptor's length field is at least
 695 * as large as the cipher block size.  This means that a cipher block
 696 * can span at most 2 descriptors.  However, this does not allow a
 697 * partial block to span into the final descriptor as that would
 698 * violate the rule (since every descriptor's length must be at lest
 699 * the block size).  So, for example, assuming an 8 byte block size:
 700 *
 701 *	0xe --> 0xa --> 0x8
 702 *
 703 * is a valid length sequence, whereas:
 704 *
 705 *	0xe --> 0xb --> 0x7
 706 *
 707 * is not a valid sequence.
 708 */
 709
 710struct n2_skcipher_alg {
 711	struct list_head	entry;
 712	u8			enc_type;
 713	struct skcipher_alg	skcipher;
 714};
 715
 716static inline struct n2_skcipher_alg *n2_skcipher_alg(struct crypto_skcipher *tfm)
 717{
 718	struct skcipher_alg *alg = crypto_skcipher_alg(tfm);
 719
 720	return container_of(alg, struct n2_skcipher_alg, skcipher);
 721}
 722
 723struct n2_skcipher_request_context {
 724	struct skcipher_walk	walk;
 725};
 726
 727static int n2_aes_setkey(struct crypto_skcipher *skcipher, const u8 *key,
 728			 unsigned int keylen)
 729{
 730	struct crypto_tfm *tfm = crypto_skcipher_tfm(skcipher);
 731	struct n2_skcipher_context *ctx = crypto_tfm_ctx(tfm);
 732	struct n2_skcipher_alg *n2alg = n2_skcipher_alg(skcipher);
 733
 734	ctx->enc_type = (n2alg->enc_type & ENC_TYPE_CHAINING_MASK);
 735
 736	switch (keylen) {
 737	case AES_KEYSIZE_128:
 738		ctx->enc_type |= ENC_TYPE_ALG_AES128;
 739		break;
 740	case AES_KEYSIZE_192:
 741		ctx->enc_type |= ENC_TYPE_ALG_AES192;
 742		break;
 743	case AES_KEYSIZE_256:
 744		ctx->enc_type |= ENC_TYPE_ALG_AES256;
 745		break;
 746	default:
 747		return -EINVAL;
 748	}
 749
 750	ctx->key_len = keylen;
 751	memcpy(ctx->key.aes, key, keylen);
 752	return 0;
 753}
 754
 755static int n2_des_setkey(struct crypto_skcipher *skcipher, const u8 *key,
 756			 unsigned int keylen)
 757{
 758	struct crypto_tfm *tfm = crypto_skcipher_tfm(skcipher);
 759	struct n2_skcipher_context *ctx = crypto_tfm_ctx(tfm);
 760	struct n2_skcipher_alg *n2alg = n2_skcipher_alg(skcipher);
 761	int err;
 762
 763	err = verify_skcipher_des_key(skcipher, key);
 764	if (err)
 765		return err;
 766
 767	ctx->enc_type = n2alg->enc_type;
 768
 769	ctx->key_len = keylen;
 770	memcpy(ctx->key.des, key, keylen);
 771	return 0;
 772}
 773
 774static int n2_3des_setkey(struct crypto_skcipher *skcipher, const u8 *key,
 775			  unsigned int keylen)
 776{
 777	struct crypto_tfm *tfm = crypto_skcipher_tfm(skcipher);
 778	struct n2_skcipher_context *ctx = crypto_tfm_ctx(tfm);
 779	struct n2_skcipher_alg *n2alg = n2_skcipher_alg(skcipher);
 780	int err;
 781
 782	err = verify_skcipher_des3_key(skcipher, key);
 783	if (err)
 784		return err;
 785
 786	ctx->enc_type = n2alg->enc_type;
 787
 788	ctx->key_len = keylen;
 789	memcpy(ctx->key.des3, key, keylen);
 790	return 0;
 791}
 792
 793static inline int skcipher_descriptor_len(int nbytes, unsigned int block_size)
 794{
 795	int this_len = nbytes;
 796
 797	this_len -= (nbytes & (block_size - 1));
 798	return this_len > (1 << 16) ? (1 << 16) : this_len;
 799}
 800
 801static int __n2_crypt_chunk(struct crypto_skcipher *skcipher,
 802			    struct n2_crypto_chunk *cp,
 803			    struct spu_queue *qp, bool encrypt)
 804{
 805	struct n2_skcipher_context *ctx = crypto_skcipher_ctx(skcipher);
 806	struct cwq_initial_entry *ent;
 807	bool in_place;
 808	int i;
 809
 810	ent = spu_queue_alloc(qp, cp->arr_len);
 811	if (!ent) {
 812		pr_info("queue_alloc() of %d fails\n",
 813			cp->arr_len);
 814		return -EBUSY;
 815	}
 816
 817	in_place = (cp->dest_paddr == cp->arr[0].src_paddr);
 818
 819	ent->control = control_word_base(cp->arr[0].src_len,
 820					 0, ctx->enc_type, 0, 0,
 821					 false, true, false, encrypt,
 822					 OPCODE_ENCRYPT |
 823					 (in_place ? OPCODE_INPLACE_BIT : 0));
 824	ent->src_addr = cp->arr[0].src_paddr;
 825	ent->auth_key_addr = 0UL;
 826	ent->auth_iv_addr = 0UL;
 827	ent->final_auth_state_addr = 0UL;
 828	ent->enc_key_addr = __pa(&ctx->key);
 829	ent->enc_iv_addr = cp->iv_paddr;
 830	ent->dest_addr = (in_place ? 0UL : cp->dest_paddr);
 831
 832	for (i = 1; i < cp->arr_len; i++) {
 833		ent = spu_queue_next(qp, ent);
 834
 835		ent->control = cp->arr[i].src_len - 1;
 836		ent->src_addr = cp->arr[i].src_paddr;
 837		ent->auth_key_addr = 0UL;
 838		ent->auth_iv_addr = 0UL;
 839		ent->final_auth_state_addr = 0UL;
 840		ent->enc_key_addr = 0UL;
 841		ent->enc_iv_addr = 0UL;
 842		ent->dest_addr = 0UL;
 843	}
 844	ent->control |= CONTROL_END_OF_BLOCK;
 845
 846	return (spu_queue_submit(qp, ent) != HV_EOK) ? -EINVAL : 0;
 847}
 848
 849static int n2_compute_chunks(struct skcipher_request *req)
 850{
 851	struct n2_request_context *rctx = skcipher_request_ctx(req);
 852	struct skcipher_walk *walk = &rctx->walk;
 853	struct n2_crypto_chunk *chunk;
 854	unsigned long dest_prev;
 855	unsigned int tot_len;
 856	bool prev_in_place;
 857	int err, nbytes;
 858
 859	err = skcipher_walk_async(walk, req);
 860	if (err)
 861		return err;
 862
 863	INIT_LIST_HEAD(&rctx->chunk_list);
 864
 865	chunk = &rctx->chunk;
 866	INIT_LIST_HEAD(&chunk->entry);
 867
 868	chunk->iv_paddr = 0UL;
 869	chunk->arr_len = 0;
 870	chunk->dest_paddr = 0UL;
 871
 872	prev_in_place = false;
 873	dest_prev = ~0UL;
 874	tot_len = 0;
 875
 876	while ((nbytes = walk->nbytes) != 0) {
 877		unsigned long dest_paddr, src_paddr;
 878		bool in_place;
 879		int this_len;
 880
 881		src_paddr = (page_to_phys(walk->src.phys.page) +
 882			     walk->src.phys.offset);
 883		dest_paddr = (page_to_phys(walk->dst.phys.page) +
 884			      walk->dst.phys.offset);
 885		in_place = (src_paddr == dest_paddr);
 886		this_len = skcipher_descriptor_len(nbytes, walk->blocksize);
 887
 888		if (chunk->arr_len != 0) {
 889			if (in_place != prev_in_place ||
 890			    (!prev_in_place &&
 891			     dest_paddr != dest_prev) ||
 892			    chunk->arr_len == N2_CHUNK_ARR_LEN ||
 893			    tot_len + this_len > (1 << 16)) {
 894				chunk->dest_final = dest_prev;
 895				list_add_tail(&chunk->entry,
 896					      &rctx->chunk_list);
 897				chunk = kzalloc(sizeof(*chunk), GFP_ATOMIC);
 898				if (!chunk) {
 899					err = -ENOMEM;
 900					break;
 901				}
 902				INIT_LIST_HEAD(&chunk->entry);
 903			}
 904		}
 905		if (chunk->arr_len == 0) {
 906			chunk->dest_paddr = dest_paddr;
 907			tot_len = 0;
 908		}
 909		chunk->arr[chunk->arr_len].src_paddr = src_paddr;
 910		chunk->arr[chunk->arr_len].src_len = this_len;
 911		chunk->arr_len++;
 912
 913		dest_prev = dest_paddr + this_len;
 914		prev_in_place = in_place;
 915		tot_len += this_len;
 916
 917		err = skcipher_walk_done(walk, nbytes - this_len);
 918		if (err)
 919			break;
 920	}
 921	if (!err && chunk->arr_len != 0) {
 922		chunk->dest_final = dest_prev;
 923		list_add_tail(&chunk->entry, &rctx->chunk_list);
 924	}
 925
 926	return err;
 927}
 928
 929static void n2_chunk_complete(struct skcipher_request *req, void *final_iv)
 930{
 931	struct n2_request_context *rctx = skcipher_request_ctx(req);
 932	struct n2_crypto_chunk *c, *tmp;
 933
 934	if (final_iv)
 935		memcpy(rctx->walk.iv, final_iv, rctx->walk.blocksize);
 936
 937	list_for_each_entry_safe(c, tmp, &rctx->chunk_list, entry) {
 938		list_del(&c->entry);
 939		if (unlikely(c != &rctx->chunk))
 940			kfree(c);
 941	}
 942
 943}
 944
 945static int n2_do_ecb(struct skcipher_request *req, bool encrypt)
 946{
 947	struct n2_request_context *rctx = skcipher_request_ctx(req);
 948	struct crypto_skcipher *tfm = crypto_skcipher_reqtfm(req);
 949	int err = n2_compute_chunks(req);
 950	struct n2_crypto_chunk *c, *tmp;
 951	unsigned long flags, hv_ret;
 952	struct spu_queue *qp;
 953
 954	if (err)
 955		return err;
 956
 957	qp = cpu_to_cwq[get_cpu()];
 958	err = -ENODEV;
 959	if (!qp)
 960		goto out;
 961
 962	spin_lock_irqsave(&qp->lock, flags);
 963
 964	list_for_each_entry_safe(c, tmp, &rctx->chunk_list, entry) {
 965		err = __n2_crypt_chunk(tfm, c, qp, encrypt);
 966		if (err)
 967			break;
 968		list_del(&c->entry);
 969		if (unlikely(c != &rctx->chunk))
 970			kfree(c);
 971	}
 972	if (!err) {
 973		hv_ret = wait_for_tail(qp);
 974		if (hv_ret != HV_EOK)
 975			err = -EINVAL;
 976	}
 977
 978	spin_unlock_irqrestore(&qp->lock, flags);
 979
 980out:
 981	put_cpu();
 982
 983	n2_chunk_complete(req, NULL);
 984	return err;
 985}
 986
 987static int n2_encrypt_ecb(struct skcipher_request *req)
 988{
 989	return n2_do_ecb(req, true);
 990}
 991
 992static int n2_decrypt_ecb(struct skcipher_request *req)
 993{
 994	return n2_do_ecb(req, false);
 995}
 996
 997static int n2_do_chaining(struct skcipher_request *req, bool encrypt)
 998{
 999	struct n2_request_context *rctx = skcipher_request_ctx(req);
1000	struct crypto_skcipher *tfm = crypto_skcipher_reqtfm(req);
1001	unsigned long flags, hv_ret, iv_paddr;
1002	int err = n2_compute_chunks(req);
1003	struct n2_crypto_chunk *c, *tmp;
1004	struct spu_queue *qp;
1005	void *final_iv_addr;
1006
1007	final_iv_addr = NULL;
1008
1009	if (err)
1010		return err;
1011
1012	qp = cpu_to_cwq[get_cpu()];
1013	err = -ENODEV;
1014	if (!qp)
1015		goto out;
1016
1017	spin_lock_irqsave(&qp->lock, flags);
1018
1019	if (encrypt) {
1020		iv_paddr = __pa(rctx->walk.iv);
1021		list_for_each_entry_safe(c, tmp, &rctx->chunk_list,
1022					 entry) {
1023			c->iv_paddr = iv_paddr;
1024			err = __n2_crypt_chunk(tfm, c, qp, true);
1025			if (err)
1026				break;
1027			iv_paddr = c->dest_final - rctx->walk.blocksize;
1028			list_del(&c->entry);
1029			if (unlikely(c != &rctx->chunk))
1030				kfree(c);
1031		}
1032		final_iv_addr = __va(iv_paddr);
1033	} else {
1034		list_for_each_entry_safe_reverse(c, tmp, &rctx->chunk_list,
1035						 entry) {
1036			if (c == &rctx->chunk) {
1037				iv_paddr = __pa(rctx->walk.iv);
1038			} else {
1039				iv_paddr = (tmp->arr[tmp->arr_len-1].src_paddr +
1040					    tmp->arr[tmp->arr_len-1].src_len -
1041					    rctx->walk.blocksize);
1042			}
1043			if (!final_iv_addr) {
1044				unsigned long pa;
1045
1046				pa = (c->arr[c->arr_len-1].src_paddr +
1047				      c->arr[c->arr_len-1].src_len -
1048				      rctx->walk.blocksize);
1049				final_iv_addr = rctx->temp_iv;
1050				memcpy(rctx->temp_iv, __va(pa),
1051				       rctx->walk.blocksize);
1052			}
1053			c->iv_paddr = iv_paddr;
1054			err = __n2_crypt_chunk(tfm, c, qp, false);
1055			if (err)
1056				break;
1057			list_del(&c->entry);
1058			if (unlikely(c != &rctx->chunk))
1059				kfree(c);
1060		}
1061	}
1062	if (!err) {
1063		hv_ret = wait_for_tail(qp);
1064		if (hv_ret != HV_EOK)
1065			err = -EINVAL;
1066	}
1067
1068	spin_unlock_irqrestore(&qp->lock, flags);
1069
1070out:
1071	put_cpu();
1072
1073	n2_chunk_complete(req, err ? NULL : final_iv_addr);
1074	return err;
1075}
1076
1077static int n2_encrypt_chaining(struct skcipher_request *req)
1078{
1079	return n2_do_chaining(req, true);
1080}
1081
1082static int n2_decrypt_chaining(struct skcipher_request *req)
1083{
1084	return n2_do_chaining(req, false);
1085}
1086
1087struct n2_skcipher_tmpl {
1088	const char		*name;
1089	const char		*drv_name;
1090	u8			block_size;
1091	u8			enc_type;
1092	struct skcipher_alg	skcipher;
1093};
1094
1095static const struct n2_skcipher_tmpl skcipher_tmpls[] = {
1096	/* DES: ECB CBC and CFB are supported */
1097	{	.name		= "ecb(des)",
1098		.drv_name	= "ecb-des",
1099		.block_size	= DES_BLOCK_SIZE,
1100		.enc_type	= (ENC_TYPE_ALG_DES |
1101				   ENC_TYPE_CHAINING_ECB),
1102		.skcipher	= {
1103			.min_keysize	= DES_KEY_SIZE,
1104			.max_keysize	= DES_KEY_SIZE,
1105			.setkey		= n2_des_setkey,
1106			.encrypt	= n2_encrypt_ecb,
1107			.decrypt	= n2_decrypt_ecb,
1108		},
1109	},
1110	{	.name		= "cbc(des)",
1111		.drv_name	= "cbc-des",
1112		.block_size	= DES_BLOCK_SIZE,
1113		.enc_type	= (ENC_TYPE_ALG_DES |
1114				   ENC_TYPE_CHAINING_CBC),
1115		.skcipher	= {
1116			.ivsize		= DES_BLOCK_SIZE,
1117			.min_keysize	= DES_KEY_SIZE,
1118			.max_keysize	= DES_KEY_SIZE,
1119			.setkey		= n2_des_setkey,
1120			.encrypt	= n2_encrypt_chaining,
1121			.decrypt	= n2_decrypt_chaining,
1122		},
1123	},
1124
1125	/* 3DES: ECB CBC and CFB are supported */
1126	{	.name		= "ecb(des3_ede)",
1127		.drv_name	= "ecb-3des",
1128		.block_size	= DES_BLOCK_SIZE,
1129		.enc_type	= (ENC_TYPE_ALG_3DES |
1130				   ENC_TYPE_CHAINING_ECB),
1131		.skcipher	= {
1132			.min_keysize	= 3 * DES_KEY_SIZE,
1133			.max_keysize	= 3 * DES_KEY_SIZE,
1134			.setkey		= n2_3des_setkey,
1135			.encrypt	= n2_encrypt_ecb,
1136			.decrypt	= n2_decrypt_ecb,
1137		},
1138	},
1139	{	.name		= "cbc(des3_ede)",
1140		.drv_name	= "cbc-3des",
1141		.block_size	= DES_BLOCK_SIZE,
1142		.enc_type	= (ENC_TYPE_ALG_3DES |
1143				   ENC_TYPE_CHAINING_CBC),
1144		.skcipher	= {
1145			.ivsize		= DES_BLOCK_SIZE,
1146			.min_keysize	= 3 * DES_KEY_SIZE,
1147			.max_keysize	= 3 * DES_KEY_SIZE,
1148			.setkey		= n2_3des_setkey,
1149			.encrypt	= n2_encrypt_chaining,
1150			.decrypt	= n2_decrypt_chaining,
1151		},
1152	},
1153
1154	/* AES: ECB CBC and CTR are supported */
1155	{	.name		= "ecb(aes)",
1156		.drv_name	= "ecb-aes",
1157		.block_size	= AES_BLOCK_SIZE,
1158		.enc_type	= (ENC_TYPE_ALG_AES128 |
1159				   ENC_TYPE_CHAINING_ECB),
1160		.skcipher	= {
1161			.min_keysize	= AES_MIN_KEY_SIZE,
1162			.max_keysize	= AES_MAX_KEY_SIZE,
1163			.setkey		= n2_aes_setkey,
1164			.encrypt	= n2_encrypt_ecb,
1165			.decrypt	= n2_decrypt_ecb,
1166		},
1167	},
1168	{	.name		= "cbc(aes)",
1169		.drv_name	= "cbc-aes",
1170		.block_size	= AES_BLOCK_SIZE,
1171		.enc_type	= (ENC_TYPE_ALG_AES128 |
1172				   ENC_TYPE_CHAINING_CBC),
1173		.skcipher	= {
1174			.ivsize		= AES_BLOCK_SIZE,
1175			.min_keysize	= AES_MIN_KEY_SIZE,
1176			.max_keysize	= AES_MAX_KEY_SIZE,
1177			.setkey		= n2_aes_setkey,
1178			.encrypt	= n2_encrypt_chaining,
1179			.decrypt	= n2_decrypt_chaining,
1180		},
1181	},
1182	{	.name		= "ctr(aes)",
1183		.drv_name	= "ctr-aes",
1184		.block_size	= AES_BLOCK_SIZE,
1185		.enc_type	= (ENC_TYPE_ALG_AES128 |
1186				   ENC_TYPE_CHAINING_COUNTER),
1187		.skcipher	= {
1188			.ivsize		= AES_BLOCK_SIZE,
1189			.min_keysize	= AES_MIN_KEY_SIZE,
1190			.max_keysize	= AES_MAX_KEY_SIZE,
1191			.setkey		= n2_aes_setkey,
1192			.encrypt	= n2_encrypt_chaining,
1193			.decrypt	= n2_encrypt_chaining,
1194		},
1195	},
1196
1197};
1198#define NUM_CIPHER_TMPLS ARRAY_SIZE(skcipher_tmpls)
1199
1200static LIST_HEAD(skcipher_algs);
1201
1202struct n2_hash_tmpl {
1203	const char	*name;
1204	const u8	*hash_zero;
1205	const u8	*hash_init;
1206	u8		hw_op_hashsz;
1207	u8		digest_size;
1208	u8		statesize;
1209	u8		block_size;
1210	u8		auth_type;
1211	u8		hmac_type;
1212};
1213
1214static const __le32 n2_md5_init[MD5_HASH_WORDS] = {
1215	cpu_to_le32(MD5_H0),
1216	cpu_to_le32(MD5_H1),
1217	cpu_to_le32(MD5_H2),
1218	cpu_to_le32(MD5_H3),
1219};
1220static const u32 n2_sha1_init[SHA1_DIGEST_SIZE / 4] = {
1221	SHA1_H0, SHA1_H1, SHA1_H2, SHA1_H3, SHA1_H4,
1222};
1223static const u32 n2_sha256_init[SHA256_DIGEST_SIZE / 4] = {
1224	SHA256_H0, SHA256_H1, SHA256_H2, SHA256_H3,
1225	SHA256_H4, SHA256_H5, SHA256_H6, SHA256_H7,
1226};
1227static const u32 n2_sha224_init[SHA256_DIGEST_SIZE / 4] = {
1228	SHA224_H0, SHA224_H1, SHA224_H2, SHA224_H3,
1229	SHA224_H4, SHA224_H5, SHA224_H6, SHA224_H7,
1230};
1231
1232static const struct n2_hash_tmpl hash_tmpls[] = {
1233	{ .name		= "md5",
1234	  .hash_zero	= md5_zero_message_hash,
1235	  .hash_init	= (u8 *)n2_md5_init,
1236	  .auth_type	= AUTH_TYPE_MD5,
1237	  .hmac_type	= AUTH_TYPE_HMAC_MD5,
1238	  .hw_op_hashsz	= MD5_DIGEST_SIZE,
1239	  .digest_size	= MD5_DIGEST_SIZE,
1240	  .statesize	= sizeof(struct md5_state),
1241	  .block_size	= MD5_HMAC_BLOCK_SIZE },
1242	{ .name		= "sha1",
1243	  .hash_zero	= sha1_zero_message_hash,
1244	  .hash_init	= (u8 *)n2_sha1_init,
1245	  .auth_type	= AUTH_TYPE_SHA1,
1246	  .hmac_type	= AUTH_TYPE_HMAC_SHA1,
1247	  .hw_op_hashsz	= SHA1_DIGEST_SIZE,
1248	  .digest_size	= SHA1_DIGEST_SIZE,
1249	  .statesize	= sizeof(struct sha1_state),
1250	  .block_size	= SHA1_BLOCK_SIZE },
1251	{ .name		= "sha256",
1252	  .hash_zero	= sha256_zero_message_hash,
1253	  .hash_init	= (u8 *)n2_sha256_init,
1254	  .auth_type	= AUTH_TYPE_SHA256,
1255	  .hmac_type	= AUTH_TYPE_HMAC_SHA256,
1256	  .hw_op_hashsz	= SHA256_DIGEST_SIZE,
1257	  .digest_size	= SHA256_DIGEST_SIZE,
1258	  .statesize	= sizeof(struct sha256_state),
1259	  .block_size	= SHA256_BLOCK_SIZE },
1260	{ .name		= "sha224",
1261	  .hash_zero	= sha224_zero_message_hash,
1262	  .hash_init	= (u8 *)n2_sha224_init,
1263	  .auth_type	= AUTH_TYPE_SHA256,
1264	  .hmac_type	= AUTH_TYPE_RESERVED,
1265	  .hw_op_hashsz	= SHA256_DIGEST_SIZE,
1266	  .digest_size	= SHA224_DIGEST_SIZE,
1267	  .statesize	= sizeof(struct sha256_state),
1268	  .block_size	= SHA224_BLOCK_SIZE },
1269};
1270#define NUM_HASH_TMPLS ARRAY_SIZE(hash_tmpls)
1271
1272static LIST_HEAD(ahash_algs);
1273static LIST_HEAD(hmac_algs);
1274
1275static int algs_registered;
1276
1277static void __n2_unregister_algs(void)
1278{
1279	struct n2_skcipher_alg *skcipher, *skcipher_tmp;
1280	struct n2_ahash_alg *alg, *alg_tmp;
1281	struct n2_hmac_alg *hmac, *hmac_tmp;
1282
1283	list_for_each_entry_safe(skcipher, skcipher_tmp, &skcipher_algs, entry) {
1284		crypto_unregister_skcipher(&skcipher->skcipher);
1285		list_del(&skcipher->entry);
1286		kfree(skcipher);
1287	}
1288	list_for_each_entry_safe(hmac, hmac_tmp, &hmac_algs, derived.entry) {
1289		crypto_unregister_ahash(&hmac->derived.alg);
1290		list_del(&hmac->derived.entry);
1291		kfree(hmac);
1292	}
1293	list_for_each_entry_safe(alg, alg_tmp, &ahash_algs, entry) {
1294		crypto_unregister_ahash(&alg->alg);
1295		list_del(&alg->entry);
1296		kfree(alg);
1297	}
1298}
1299
1300static int n2_skcipher_init_tfm(struct crypto_skcipher *tfm)
1301{
1302	crypto_skcipher_set_reqsize(tfm, sizeof(struct n2_request_context));
1303	return 0;
1304}
1305
1306static int __n2_register_one_skcipher(const struct n2_skcipher_tmpl *tmpl)
1307{
1308	struct n2_skcipher_alg *p = kzalloc(sizeof(*p), GFP_KERNEL);
1309	struct skcipher_alg *alg;
1310	int err;
1311
1312	if (!p)
1313		return -ENOMEM;
1314
1315	alg = &p->skcipher;
1316	*alg = tmpl->skcipher;
1317
1318	snprintf(alg->base.cra_name, CRYPTO_MAX_ALG_NAME, "%s", tmpl->name);
1319	snprintf(alg->base.cra_driver_name, CRYPTO_MAX_ALG_NAME, "%s-n2", tmpl->drv_name);
1320	alg->base.cra_priority = N2_CRA_PRIORITY;
1321	alg->base.cra_flags = CRYPTO_ALG_KERN_DRIVER_ONLY | CRYPTO_ALG_ASYNC |
1322			      CRYPTO_ALG_ALLOCATES_MEMORY;
1323	alg->base.cra_blocksize = tmpl->block_size;
1324	p->enc_type = tmpl->enc_type;
1325	alg->base.cra_ctxsize = sizeof(struct n2_skcipher_context);
1326	alg->base.cra_module = THIS_MODULE;
1327	alg->init = n2_skcipher_init_tfm;
1328
1329	list_add(&p->entry, &skcipher_algs);
1330	err = crypto_register_skcipher(alg);
1331	if (err) {
1332		pr_err("%s alg registration failed\n", alg->base.cra_name);
1333		list_del(&p->entry);
1334		kfree(p);
1335	} else {
1336		pr_info("%s alg registered\n", alg->base.cra_name);
1337	}
1338	return err;
1339}
1340
1341static int __n2_register_one_hmac(struct n2_ahash_alg *n2ahash)
1342{
1343	struct n2_hmac_alg *p = kzalloc(sizeof(*p), GFP_KERNEL);
1344	struct ahash_alg *ahash;
1345	struct crypto_alg *base;
1346	int err;
1347
1348	if (!p)
1349		return -ENOMEM;
1350
1351	p->child_alg = n2ahash->alg.halg.base.cra_name;
1352	memcpy(&p->derived, n2ahash, sizeof(struct n2_ahash_alg));
1353	INIT_LIST_HEAD(&p->derived.entry);
1354
1355	ahash = &p->derived.alg;
1356	ahash->digest = n2_hmac_async_digest;
1357	ahash->setkey = n2_hmac_async_setkey;
1358
1359	base = &ahash->halg.base;
1360	if (snprintf(base->cra_name, CRYPTO_MAX_ALG_NAME, "hmac(%s)",
1361		     p->child_alg) >= CRYPTO_MAX_ALG_NAME)
1362		goto out_free_p;
1363	if (snprintf(base->cra_driver_name, CRYPTO_MAX_ALG_NAME, "hmac-%s-n2",
1364		     p->child_alg) >= CRYPTO_MAX_ALG_NAME)
1365		goto out_free_p;
1366
1367	base->cra_ctxsize = sizeof(struct n2_hmac_ctx);
1368	base->cra_init = n2_hmac_cra_init;
1369	base->cra_exit = n2_hmac_cra_exit;
1370
1371	list_add(&p->derived.entry, &hmac_algs);
1372	err = crypto_register_ahash(ahash);
1373	if (err) {
1374		pr_err("%s alg registration failed\n", base->cra_name);
1375		list_del(&p->derived.entry);
1376out_free_p:
1377		kfree(p);
1378	} else {
1379		pr_info("%s alg registered\n", base->cra_name);
1380	}
1381	return err;
1382}
1383
1384static int __n2_register_one_ahash(const struct n2_hash_tmpl *tmpl)
1385{
1386	struct n2_ahash_alg *p = kzalloc(sizeof(*p), GFP_KERNEL);
1387	struct hash_alg_common *halg;
1388	struct crypto_alg *base;
1389	struct ahash_alg *ahash;
1390	int err;
1391
1392	if (!p)
1393		return -ENOMEM;
1394
1395	p->hash_zero = tmpl->hash_zero;
1396	p->hash_init = tmpl->hash_init;
1397	p->auth_type = tmpl->auth_type;
1398	p->hmac_type = tmpl->hmac_type;
1399	p->hw_op_hashsz = tmpl->hw_op_hashsz;
1400	p->digest_size = tmpl->digest_size;
1401
1402	ahash = &p->alg;
1403	ahash->init = n2_hash_async_init;
1404	ahash->update = n2_hash_async_update;
1405	ahash->final = n2_hash_async_final;
1406	ahash->finup = n2_hash_async_finup;
1407	ahash->digest = n2_hash_async_digest;
1408	ahash->export = n2_hash_async_noexport;
1409	ahash->import = n2_hash_async_noimport;
1410
1411	halg = &ahash->halg;
1412	halg->digestsize = tmpl->digest_size;
1413	halg->statesize = tmpl->statesize;
1414
1415	base = &halg->base;
1416	snprintf(base->cra_name, CRYPTO_MAX_ALG_NAME, "%s", tmpl->name);
1417	snprintf(base->cra_driver_name, CRYPTO_MAX_ALG_NAME, "%s-n2", tmpl->name);
1418	base->cra_priority = N2_CRA_PRIORITY;
1419	base->cra_flags = CRYPTO_ALG_KERN_DRIVER_ONLY |
1420			  CRYPTO_ALG_NEED_FALLBACK;
1421	base->cra_blocksize = tmpl->block_size;
1422	base->cra_ctxsize = sizeof(struct n2_hash_ctx);
1423	base->cra_module = THIS_MODULE;
1424	base->cra_init = n2_hash_cra_init;
1425	base->cra_exit = n2_hash_cra_exit;
1426
1427	list_add(&p->entry, &ahash_algs);
1428	err = crypto_register_ahash(ahash);
1429	if (err) {
1430		pr_err("%s alg registration failed\n", base->cra_name);
1431		list_del(&p->entry);
1432		kfree(p);
1433	} else {
1434		pr_info("%s alg registered\n", base->cra_name);
1435	}
1436	if (!err && p->hmac_type != AUTH_TYPE_RESERVED)
1437		err = __n2_register_one_hmac(p);
1438	return err;
1439}
1440
1441static int n2_register_algs(void)
1442{
1443	int i, err = 0;
1444
1445	mutex_lock(&spu_lock);
1446	if (algs_registered++)
1447		goto out;
1448
1449	for (i = 0; i < NUM_HASH_TMPLS; i++) {
1450		err = __n2_register_one_ahash(&hash_tmpls[i]);
1451		if (err) {
1452			__n2_unregister_algs();
1453			goto out;
1454		}
1455	}
1456	for (i = 0; i < NUM_CIPHER_TMPLS; i++) {
1457		err = __n2_register_one_skcipher(&skcipher_tmpls[i]);
1458		if (err) {
1459			__n2_unregister_algs();
1460			goto out;
1461		}
1462	}
1463
1464out:
1465	mutex_unlock(&spu_lock);
1466	return err;
1467}
1468
1469static void n2_unregister_algs(void)
1470{
1471	mutex_lock(&spu_lock);
1472	if (!--algs_registered)
1473		__n2_unregister_algs();
1474	mutex_unlock(&spu_lock);
1475}
1476
1477/* To map CWQ queues to interrupt sources, the hypervisor API provides
1478 * a devino.  This isn't very useful to us because all of the
1479 * interrupts listed in the device_node have been translated to
1480 * Linux virtual IRQ cookie numbers.
1481 *
1482 * So we have to back-translate, going through the 'intr' and 'ino'
1483 * property tables of the n2cp MDESC node, matching it with the OF
1484 * 'interrupts' property entries, in order to figure out which
1485 * devino goes to which already-translated IRQ.
1486 */
1487static int find_devino_index(struct platform_device *dev, struct spu_mdesc_info *ip,
1488			     unsigned long dev_ino)
1489{
1490	const unsigned int *dev_intrs;
1491	unsigned int intr;
1492	int i;
1493
1494	for (i = 0; i < ip->num_intrs; i++) {
1495		if (ip->ino_table[i].ino == dev_ino)
1496			break;
1497	}
1498	if (i == ip->num_intrs)
1499		return -ENODEV;
1500
1501	intr = ip->ino_table[i].intr;
1502
1503	dev_intrs = of_get_property(dev->dev.of_node, "interrupts", NULL);
1504	if (!dev_intrs)
1505		return -ENODEV;
1506
1507	for (i = 0; i < dev->archdata.num_irqs; i++) {
1508		if (dev_intrs[i] == intr)
1509			return i;
1510	}
1511
1512	return -ENODEV;
1513}
1514
1515static int spu_map_ino(struct platform_device *dev, struct spu_mdesc_info *ip,
1516		       const char *irq_name, struct spu_queue *p,
1517		       irq_handler_t handler)
1518{
1519	unsigned long herr;
1520	int index;
1521
1522	herr = sun4v_ncs_qhandle_to_devino(p->qhandle, &p->devino);
1523	if (herr)
1524		return -EINVAL;
1525
1526	index = find_devino_index(dev, ip, p->devino);
1527	if (index < 0)
1528		return index;
1529
1530	p->irq = dev->archdata.irqs[index];
1531
1532	sprintf(p->irq_name, "%s-%d", irq_name, index);
1533
1534	return request_irq(p->irq, handler, 0, p->irq_name, p);
1535}
1536
1537static struct kmem_cache *queue_cache[2];
1538
1539static void *new_queue(unsigned long q_type)
1540{
1541	return kmem_cache_zalloc(queue_cache[q_type - 1], GFP_KERNEL);
1542}
1543
1544static void free_queue(void *p, unsigned long q_type)
1545{
1546	kmem_cache_free(queue_cache[q_type - 1], p);
1547}
1548
1549static int queue_cache_init(void)
1550{
1551	if (!queue_cache[HV_NCS_QTYPE_MAU - 1])
1552		queue_cache[HV_NCS_QTYPE_MAU - 1] =
1553			kmem_cache_create("mau_queue",
1554					  (MAU_NUM_ENTRIES *
1555					   MAU_ENTRY_SIZE),
1556					  MAU_ENTRY_SIZE, 0, NULL);
1557	if (!queue_cache[HV_NCS_QTYPE_MAU - 1])
1558		return -ENOMEM;
1559
1560	if (!queue_cache[HV_NCS_QTYPE_CWQ - 1])
1561		queue_cache[HV_NCS_QTYPE_CWQ - 1] =
1562			kmem_cache_create("cwq_queue",
1563					  (CWQ_NUM_ENTRIES *
1564					   CWQ_ENTRY_SIZE),
1565					  CWQ_ENTRY_SIZE, 0, NULL);
1566	if (!queue_cache[HV_NCS_QTYPE_CWQ - 1]) {
1567		kmem_cache_destroy(queue_cache[HV_NCS_QTYPE_MAU - 1]);
1568		queue_cache[HV_NCS_QTYPE_MAU - 1] = NULL;
1569		return -ENOMEM;
1570	}
1571	return 0;
1572}
1573
1574static void queue_cache_destroy(void)
1575{
1576	kmem_cache_destroy(queue_cache[HV_NCS_QTYPE_MAU - 1]);
1577	kmem_cache_destroy(queue_cache[HV_NCS_QTYPE_CWQ - 1]);
1578	queue_cache[HV_NCS_QTYPE_MAU - 1] = NULL;
1579	queue_cache[HV_NCS_QTYPE_CWQ - 1] = NULL;
1580}
1581
1582static long spu_queue_register_workfn(void *arg)
1583{
1584	struct spu_qreg *qr = arg;
1585	struct spu_queue *p = qr->queue;
1586	unsigned long q_type = qr->type;
1587	unsigned long hv_ret;
1588
1589	hv_ret = sun4v_ncs_qconf(q_type, __pa(p->q),
1590				 CWQ_NUM_ENTRIES, &p->qhandle);
1591	if (!hv_ret)
1592		sun4v_ncs_sethead_marker(p->qhandle, 0);
1593
1594	return hv_ret ? -EINVAL : 0;
1595}
1596
1597static int spu_queue_register(struct spu_queue *p, unsigned long q_type)
1598{
1599	int cpu = cpumask_any_and(&p->sharing, cpu_online_mask);
1600	struct spu_qreg qr = { .queue = p, .type = q_type };
1601
1602	return work_on_cpu_safe(cpu, spu_queue_register_workfn, &qr);
1603}
1604
1605static int spu_queue_setup(struct spu_queue *p)
1606{
1607	int err;
1608
1609	p->q = new_queue(p->q_type);
1610	if (!p->q)
1611		return -ENOMEM;
1612
1613	err = spu_queue_register(p, p->q_type);
1614	if (err) {
1615		free_queue(p->q, p->q_type);
1616		p->q = NULL;
1617	}
1618
1619	return err;
1620}
1621
1622static void spu_queue_destroy(struct spu_queue *p)
1623{
1624	unsigned long hv_ret;
1625
1626	if (!p->q)
1627		return;
1628
1629	hv_ret = sun4v_ncs_qconf(p->q_type, p->qhandle, 0, &p->qhandle);
1630
1631	if (!hv_ret)
1632		free_queue(p->q, p->q_type);
1633}
1634
1635static void spu_list_destroy(struct list_head *list)
1636{
1637	struct spu_queue *p, *n;
1638
1639	list_for_each_entry_safe(p, n, list, list) {
1640		int i;
1641
1642		for (i = 0; i < NR_CPUS; i++) {
1643			if (cpu_to_cwq[i] == p)
1644				cpu_to_cwq[i] = NULL;
1645		}
1646
1647		if (p->irq) {
1648			free_irq(p->irq, p);
1649			p->irq = 0;
1650		}
1651		spu_queue_destroy(p);
1652		list_del(&p->list);
1653		kfree(p);
1654	}
1655}
1656
1657/* Walk the backward arcs of a CWQ 'exec-unit' node,
1658 * gathering cpu membership information.
1659 */
1660static int spu_mdesc_walk_arcs(struct mdesc_handle *mdesc,
1661			       struct platform_device *dev,
1662			       u64 node, struct spu_queue *p,
1663			       struct spu_queue **table)
1664{
1665	u64 arc;
1666
1667	mdesc_for_each_arc(arc, mdesc, node, MDESC_ARC_TYPE_BACK) {
1668		u64 tgt = mdesc_arc_target(mdesc, arc);
1669		const char *name = mdesc_node_name(mdesc, tgt);
1670		const u64 *id;
1671
1672		if (strcmp(name, "cpu"))
1673			continue;
1674		id = mdesc_get_property(mdesc, tgt, "id", NULL);
1675		if (table[*id] != NULL) {
1676			dev_err(&dev->dev, "%pOF: SPU cpu slot already set.\n",
1677				dev->dev.of_node);
1678			return -EINVAL;
1679		}
1680		cpumask_set_cpu(*id, &p->sharing);
1681		table[*id] = p;
1682	}
1683	return 0;
1684}
1685
1686/* Process an 'exec-unit' MDESC node of type 'cwq'.  */
1687static int handle_exec_unit(struct spu_mdesc_info *ip, struct list_head *list,
1688			    struct platform_device *dev, struct mdesc_handle *mdesc,
1689			    u64 node, const char *iname, unsigned long q_type,
1690			    irq_handler_t handler, struct spu_queue **table)
1691{
1692	struct spu_queue *p;
1693	int err;
1694
1695	p = kzalloc(sizeof(struct spu_queue), GFP_KERNEL);
1696	if (!p) {
1697		dev_err(&dev->dev, "%pOF: Could not allocate SPU queue.\n",
1698			dev->dev.of_node);
1699		return -ENOMEM;
1700	}
1701
1702	cpumask_clear(&p->sharing);
1703	spin_lock_init(&p->lock);
1704	p->q_type = q_type;
1705	INIT_LIST_HEAD(&p->jobs);
1706	list_add(&p->list, list);
1707
1708	err = spu_mdesc_walk_arcs(mdesc, dev, node, p, table);
1709	if (err)
1710		return err;
1711
1712	err = spu_queue_setup(p);
1713	if (err)
1714		return err;
1715
1716	return spu_map_ino(dev, ip, iname, p, handler);
1717}
1718
1719static int spu_mdesc_scan(struct mdesc_handle *mdesc, struct platform_device *dev,
1720			  struct spu_mdesc_info *ip, struct list_head *list,
1721			  const char *exec_name, unsigned long q_type,
1722			  irq_handler_t handler, struct spu_queue **table)
1723{
1724	int err = 0;
1725	u64 node;
1726
1727	mdesc_for_each_node_by_name(mdesc, node, "exec-unit") {
1728		const char *type;
1729
1730		type = mdesc_get_property(mdesc, node, "type", NULL);
1731		if (!type || strcmp(type, exec_name))
1732			continue;
1733
1734		err = handle_exec_unit(ip, list, dev, mdesc, node,
1735				       exec_name, q_type, handler, table);
1736		if (err) {
1737			spu_list_destroy(list);
1738			break;
1739		}
1740	}
1741
1742	return err;
1743}
1744
1745static int get_irq_props(struct mdesc_handle *mdesc, u64 node,
1746			 struct spu_mdesc_info *ip)
1747{
1748	const u64 *ino;
1749	int ino_len;
1750	int i;
1751
1752	ino = mdesc_get_property(mdesc, node, "ino", &ino_len);
1753	if (!ino) {
1754		printk("NO 'ino'\n");
1755		return -ENODEV;
1756	}
1757
1758	ip->num_intrs = ino_len / sizeof(u64);
1759	ip->ino_table = kzalloc((sizeof(struct ino_blob) *
1760				 ip->num_intrs),
1761				GFP_KERNEL);
1762	if (!ip->ino_table)
1763		return -ENOMEM;
1764
1765	for (i = 0; i < ip->num_intrs; i++) {
1766		struct ino_blob *b = &ip->ino_table[i];
1767		b->intr = i + 1;
1768		b->ino = ino[i];
1769	}
1770
1771	return 0;
1772}
1773
1774static int grab_mdesc_irq_props(struct mdesc_handle *mdesc,
1775				struct platform_device *dev,
1776				struct spu_mdesc_info *ip,
1777				const char *node_name)
1778{
1779	u64 node, reg;
1780
1781	if (of_property_read_reg(dev->dev.of_node, 0, &reg, NULL) < 0)
1782		return -ENODEV;
1783
1784	mdesc_for_each_node_by_name(mdesc, node, "virtual-device") {
1785		const char *name;
1786		const u64 *chdl;
1787
1788		name = mdesc_get_property(mdesc, node, "name", NULL);
1789		if (!name || strcmp(name, node_name))
1790			continue;
1791		chdl = mdesc_get_property(mdesc, node, "cfg-handle", NULL);
1792		if (!chdl || (*chdl != reg))
1793			continue;
1794		ip->cfg_handle = *chdl;
1795		return get_irq_props(mdesc, node, ip);
1796	}
1797
1798	return -ENODEV;
1799}
1800
1801static unsigned long n2_spu_hvapi_major;
1802static unsigned long n2_spu_hvapi_minor;
1803
1804static int n2_spu_hvapi_register(void)
1805{
1806	int err;
1807
1808	n2_spu_hvapi_major = 2;
1809	n2_spu_hvapi_minor = 0;
1810
1811	err = sun4v_hvapi_register(HV_GRP_NCS,
1812				   n2_spu_hvapi_major,
1813				   &n2_spu_hvapi_minor);
1814
1815	if (!err)
1816		pr_info("Registered NCS HVAPI version %lu.%lu\n",
1817			n2_spu_hvapi_major,
1818			n2_spu_hvapi_minor);
1819
1820	return err;
1821}
1822
1823static void n2_spu_hvapi_unregister(void)
1824{
1825	sun4v_hvapi_unregister(HV_GRP_NCS);
1826}
1827
1828static int global_ref;
1829
1830static int grab_global_resources(void)
1831{
1832	int err = 0;
1833
1834	mutex_lock(&spu_lock);
1835
1836	if (global_ref++)
1837		goto out;
1838
1839	err = n2_spu_hvapi_register();
1840	if (err)
1841		goto out;
1842
1843	err = queue_cache_init();
1844	if (err)
1845		goto out_hvapi_release;
1846
1847	err = -ENOMEM;
1848	cpu_to_cwq = kcalloc(NR_CPUS, sizeof(struct spu_queue *),
1849			     GFP_KERNEL);
1850	if (!cpu_to_cwq)
1851		goto out_queue_cache_destroy;
1852
1853	cpu_to_mau = kcalloc(NR_CPUS, sizeof(struct spu_queue *),
1854			     GFP_KERNEL);
1855	if (!cpu_to_mau)
1856		goto out_free_cwq_table;
1857
1858	err = 0;
1859
1860out:
1861	if (err)
1862		global_ref--;
1863	mutex_unlock(&spu_lock);
1864	return err;
1865
1866out_free_cwq_table:
1867	kfree(cpu_to_cwq);
1868	cpu_to_cwq = NULL;
1869
1870out_queue_cache_destroy:
1871	queue_cache_destroy();
1872
1873out_hvapi_release:
1874	n2_spu_hvapi_unregister();
1875	goto out;
1876}
1877
1878static void release_global_resources(void)
1879{
1880	mutex_lock(&spu_lock);
1881	if (!--global_ref) {
1882		kfree(cpu_to_cwq);
1883		cpu_to_cwq = NULL;
1884
1885		kfree(cpu_to_mau);
1886		cpu_to_mau = NULL;
1887
1888		queue_cache_destroy();
1889		n2_spu_hvapi_unregister();
1890	}
1891	mutex_unlock(&spu_lock);
1892}
1893
1894static struct n2_crypto *alloc_n2cp(void)
1895{
1896	struct n2_crypto *np = kzalloc(sizeof(struct n2_crypto), GFP_KERNEL);
1897
1898	if (np)
1899		INIT_LIST_HEAD(&np->cwq_list);
1900
1901	return np;
1902}
1903
1904static void free_n2cp(struct n2_crypto *np)
1905{
1906	kfree(np->cwq_info.ino_table);
1907	np->cwq_info.ino_table = NULL;
1908
1909	kfree(np);
1910}
1911
1912static void n2_spu_driver_version(void)
1913{
1914	static int n2_spu_version_printed;
1915
1916	if (n2_spu_version_printed++ == 0)
1917		pr_info("%s", version);
1918}
1919
1920static int n2_crypto_probe(struct platform_device *dev)
1921{
1922	struct mdesc_handle *mdesc;
1923	struct n2_crypto *np;
1924	int err;
1925
1926	n2_spu_driver_version();
1927
1928	pr_info("Found N2CP at %pOF\n", dev->dev.of_node);
1929
1930	np = alloc_n2cp();
1931	if (!np) {
1932		dev_err(&dev->dev, "%pOF: Unable to allocate n2cp.\n",
1933			dev->dev.of_node);
1934		return -ENOMEM;
1935	}
1936
1937	err = grab_global_resources();
1938	if (err) {
1939		dev_err(&dev->dev, "%pOF: Unable to grab global resources.\n",
1940			dev->dev.of_node);
1941		goto out_free_n2cp;
1942	}
1943
1944	mdesc = mdesc_grab();
1945
1946	if (!mdesc) {
1947		dev_err(&dev->dev, "%pOF: Unable to grab MDESC.\n",
1948			dev->dev.of_node);
1949		err = -ENODEV;
1950		goto out_free_global;
1951	}
1952	err = grab_mdesc_irq_props(mdesc, dev, &np->cwq_info, "n2cp");
1953	if (err) {
1954		dev_err(&dev->dev, "%pOF: Unable to grab IRQ props.\n",
1955			dev->dev.of_node);
1956		mdesc_release(mdesc);
1957		goto out_free_global;
1958	}
1959
1960	err = spu_mdesc_scan(mdesc, dev, &np->cwq_info, &np->cwq_list,
1961			     "cwq", HV_NCS_QTYPE_CWQ, cwq_intr,
1962			     cpu_to_cwq);
1963	mdesc_release(mdesc);
1964
1965	if (err) {
1966		dev_err(&dev->dev, "%pOF: CWQ MDESC scan failed.\n",
1967			dev->dev.of_node);
1968		goto out_free_global;
1969	}
1970
1971	err = n2_register_algs();
1972	if (err) {
1973		dev_err(&dev->dev, "%pOF: Unable to register algorithms.\n",
1974			dev->dev.of_node);
1975		goto out_free_spu_list;
1976	}
1977
1978	dev_set_drvdata(&dev->dev, np);
1979
1980	return 0;
1981
1982out_free_spu_list:
1983	spu_list_destroy(&np->cwq_list);
1984
1985out_free_global:
1986	release_global_resources();
1987
1988out_free_n2cp:
1989	free_n2cp(np);
1990
1991	return err;
1992}
1993
1994static void n2_crypto_remove(struct platform_device *dev)
1995{
1996	struct n2_crypto *np = dev_get_drvdata(&dev->dev);
1997
1998	n2_unregister_algs();
1999
2000	spu_list_destroy(&np->cwq_list);
2001
2002	release_global_resources();
2003
2004	free_n2cp(np);
2005}
2006
2007static struct n2_mau *alloc_ncp(void)
2008{
2009	struct n2_mau *mp = kzalloc(sizeof(struct n2_mau), GFP_KERNEL);
2010
2011	if (mp)
2012		INIT_LIST_HEAD(&mp->mau_list);
2013
2014	return mp;
2015}
2016
2017static void free_ncp(struct n2_mau *mp)
2018{
2019	kfree(mp->mau_info.ino_table);
2020	mp->mau_info.ino_table = NULL;
2021
2022	kfree(mp);
2023}
2024
2025static int n2_mau_probe(struct platform_device *dev)
2026{
2027	struct mdesc_handle *mdesc;
2028	struct n2_mau *mp;
2029	int err;
2030
2031	n2_spu_driver_version();
2032
2033	pr_info("Found NCP at %pOF\n", dev->dev.of_node);
2034
2035	mp = alloc_ncp();
2036	if (!mp) {
2037		dev_err(&dev->dev, "%pOF: Unable to allocate ncp.\n",
2038			dev->dev.of_node);
2039		return -ENOMEM;
2040	}
2041
2042	err = grab_global_resources();
2043	if (err) {
2044		dev_err(&dev->dev, "%pOF: Unable to grab global resources.\n",
2045			dev->dev.of_node);
2046		goto out_free_ncp;
2047	}
2048
2049	mdesc = mdesc_grab();
2050
2051	if (!mdesc) {
2052		dev_err(&dev->dev, "%pOF: Unable to grab MDESC.\n",
2053			dev->dev.of_node);
2054		err = -ENODEV;
2055		goto out_free_global;
2056	}
2057
2058	err = grab_mdesc_irq_props(mdesc, dev, &mp->mau_info, "ncp");
2059	if (err) {
2060		dev_err(&dev->dev, "%pOF: Unable to grab IRQ props.\n",
2061			dev->dev.of_node);
2062		mdesc_release(mdesc);
2063		goto out_free_global;
2064	}
2065
2066	err = spu_mdesc_scan(mdesc, dev, &mp->mau_info, &mp->mau_list,
2067			     "mau", HV_NCS_QTYPE_MAU, mau_intr,
2068			     cpu_to_mau);
2069	mdesc_release(mdesc);
2070
2071	if (err) {
2072		dev_err(&dev->dev, "%pOF: MAU MDESC scan failed.\n",
2073			dev->dev.of_node);
2074		goto out_free_global;
2075	}
2076
2077	dev_set_drvdata(&dev->dev, mp);
2078
2079	return 0;
2080
2081out_free_global:
2082	release_global_resources();
2083
2084out_free_ncp:
2085	free_ncp(mp);
2086
2087	return err;
2088}
2089
2090static void n2_mau_remove(struct platform_device *dev)
2091{
2092	struct n2_mau *mp = dev_get_drvdata(&dev->dev);
2093
2094	spu_list_destroy(&mp->mau_list);
2095
2096	release_global_resources();
2097
2098	free_ncp(mp);
2099}
2100
2101static const struct of_device_id n2_crypto_match[] = {
2102	{
2103		.name = "n2cp",
2104		.compatible = "SUNW,n2-cwq",
2105	},
2106	{
2107		.name = "n2cp",
2108		.compatible = "SUNW,vf-cwq",
2109	},
2110	{
2111		.name = "n2cp",
2112		.compatible = "SUNW,kt-cwq",
2113	},
2114	{},
2115};
2116
2117MODULE_DEVICE_TABLE(of, n2_crypto_match);
2118
2119static struct platform_driver n2_crypto_driver = {
2120	.driver = {
2121		.name		=	"n2cp",
2122		.of_match_table	=	n2_crypto_match,
2123	},
2124	.probe		=	n2_crypto_probe,
2125	.remove_new	=	n2_crypto_remove,
2126};
2127
2128static const struct of_device_id n2_mau_match[] = {
2129	{
2130		.name = "ncp",
2131		.compatible = "SUNW,n2-mau",
2132	},
2133	{
2134		.name = "ncp",
2135		.compatible = "SUNW,vf-mau",
2136	},
2137	{
2138		.name = "ncp",
2139		.compatible = "SUNW,kt-mau",
2140	},
2141	{},
2142};
2143
2144MODULE_DEVICE_TABLE(of, n2_mau_match);
2145
2146static struct platform_driver n2_mau_driver = {
2147	.driver = {
2148		.name		=	"ncp",
2149		.of_match_table	=	n2_mau_match,
2150	},
2151	.probe		=	n2_mau_probe,
2152	.remove_new	=	n2_mau_remove,
2153};
2154
2155static struct platform_driver * const drivers[] = {
2156	&n2_crypto_driver,
2157	&n2_mau_driver,
2158};
2159
2160static int __init n2_init(void)
2161{
2162	return platform_register_drivers(drivers, ARRAY_SIZE(drivers));
2163}
2164
2165static void __exit n2_exit(void)
2166{
2167	platform_unregister_drivers(drivers, ARRAY_SIZE(drivers));
2168}
2169
2170module_init(n2_init);
2171module_exit(n2_exit);
v6.8
   1// SPDX-License-Identifier: GPL-2.0-only
   2/* n2_core.c: Niagara2 Stream Processing Unit (SPU) crypto support.
   3 *
   4 * Copyright (C) 2010, 2011 David S. Miller <davem@davemloft.net>
   5 */
   6
   7#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
   8
   9#include <linux/kernel.h>
  10#include <linux/module.h>
  11#include <linux/of.h>
  12#include <linux/of_address.h>
  13#include <linux/platform_device.h>
  14#include <linux/cpumask.h>
  15#include <linux/slab.h>
  16#include <linux/interrupt.h>
  17#include <linux/crypto.h>
  18#include <crypto/md5.h>
  19#include <crypto/sha1.h>
  20#include <crypto/sha2.h>
  21#include <crypto/aes.h>
  22#include <crypto/internal/des.h>
  23#include <linux/mutex.h>
  24#include <linux/delay.h>
  25#include <linux/sched.h>
  26
  27#include <crypto/internal/hash.h>
  28#include <crypto/internal/skcipher.h>
  29#include <crypto/scatterwalk.h>
  30#include <crypto/algapi.h>
  31
  32#include <asm/hypervisor.h>
  33#include <asm/mdesc.h>
  34
  35#include "n2_core.h"
  36
  37#define DRV_MODULE_NAME		"n2_crypto"
  38#define DRV_MODULE_VERSION	"0.2"
  39#define DRV_MODULE_RELDATE	"July 28, 2011"
  40
  41static const char version[] =
  42	DRV_MODULE_NAME ".c:v" DRV_MODULE_VERSION " (" DRV_MODULE_RELDATE ")\n";
  43
  44MODULE_AUTHOR("David S. Miller (davem@davemloft.net)");
  45MODULE_DESCRIPTION("Niagara2 Crypto driver");
  46MODULE_LICENSE("GPL");
  47MODULE_VERSION(DRV_MODULE_VERSION);
  48
  49#define N2_CRA_PRIORITY		200
  50
  51static DEFINE_MUTEX(spu_lock);
  52
  53struct spu_queue {
  54	cpumask_t		sharing;
  55	unsigned long		qhandle;
  56
  57	spinlock_t		lock;
  58	u8			q_type;
  59	void			*q;
  60	unsigned long		head;
  61	unsigned long		tail;
  62	struct list_head	jobs;
  63
  64	unsigned long		devino;
  65
  66	char			irq_name[32];
  67	unsigned int		irq;
  68
  69	struct list_head	list;
  70};
  71
  72struct spu_qreg {
  73	struct spu_queue	*queue;
  74	unsigned long		type;
  75};
  76
  77static struct spu_queue **cpu_to_cwq;
  78static struct spu_queue **cpu_to_mau;
  79
  80static unsigned long spu_next_offset(struct spu_queue *q, unsigned long off)
  81{
  82	if (q->q_type == HV_NCS_QTYPE_MAU) {
  83		off += MAU_ENTRY_SIZE;
  84		if (off == (MAU_ENTRY_SIZE * MAU_NUM_ENTRIES))
  85			off = 0;
  86	} else {
  87		off += CWQ_ENTRY_SIZE;
  88		if (off == (CWQ_ENTRY_SIZE * CWQ_NUM_ENTRIES))
  89			off = 0;
  90	}
  91	return off;
  92}
  93
  94struct n2_request_common {
  95	struct list_head	entry;
  96	unsigned int		offset;
  97};
  98#define OFFSET_NOT_RUNNING	(~(unsigned int)0)
  99
 100/* An async job request records the final tail value it used in
 101 * n2_request_common->offset, test to see if that offset is in
 102 * the range old_head, new_head, inclusive.
 103 */
 104static inline bool job_finished(struct spu_queue *q, unsigned int offset,
 105				unsigned long old_head, unsigned long new_head)
 106{
 107	if (old_head <= new_head) {
 108		if (offset > old_head && offset <= new_head)
 109			return true;
 110	} else {
 111		if (offset > old_head || offset <= new_head)
 112			return true;
 113	}
 114	return false;
 115}
 116
 117/* When the HEAD marker is unequal to the actual HEAD, we get
 118 * a virtual device INO interrupt.  We should process the
 119 * completed CWQ entries and adjust the HEAD marker to clear
 120 * the IRQ.
 121 */
 122static irqreturn_t cwq_intr(int irq, void *dev_id)
 123{
 124	unsigned long off, new_head, hv_ret;
 125	struct spu_queue *q = dev_id;
 126
 127	pr_err("CPU[%d]: Got CWQ interrupt for qhdl[%lx]\n",
 128	       smp_processor_id(), q->qhandle);
 129
 130	spin_lock(&q->lock);
 131
 132	hv_ret = sun4v_ncs_gethead(q->qhandle, &new_head);
 133
 134	pr_err("CPU[%d]: CWQ gethead[%lx] hv_ret[%lu]\n",
 135	       smp_processor_id(), new_head, hv_ret);
 136
 137	for (off = q->head; off != new_head; off = spu_next_offset(q, off)) {
 138		/* XXX ... XXX */
 139	}
 140
 141	hv_ret = sun4v_ncs_sethead_marker(q->qhandle, new_head);
 142	if (hv_ret == HV_EOK)
 143		q->head = new_head;
 144
 145	spin_unlock(&q->lock);
 146
 147	return IRQ_HANDLED;
 148}
 149
 150static irqreturn_t mau_intr(int irq, void *dev_id)
 151{
 152	struct spu_queue *q = dev_id;
 153	unsigned long head, hv_ret;
 154
 155	spin_lock(&q->lock);
 156
 157	pr_err("CPU[%d]: Got MAU interrupt for qhdl[%lx]\n",
 158	       smp_processor_id(), q->qhandle);
 159
 160	hv_ret = sun4v_ncs_gethead(q->qhandle, &head);
 161
 162	pr_err("CPU[%d]: MAU gethead[%lx] hv_ret[%lu]\n",
 163	       smp_processor_id(), head, hv_ret);
 164
 165	sun4v_ncs_sethead_marker(q->qhandle, head);
 166
 167	spin_unlock(&q->lock);
 168
 169	return IRQ_HANDLED;
 170}
 171
 172static void *spu_queue_next(struct spu_queue *q, void *cur)
 173{
 174	return q->q + spu_next_offset(q, cur - q->q);
 175}
 176
 177static int spu_queue_num_free(struct spu_queue *q)
 178{
 179	unsigned long head = q->head;
 180	unsigned long tail = q->tail;
 181	unsigned long end = (CWQ_ENTRY_SIZE * CWQ_NUM_ENTRIES);
 182	unsigned long diff;
 183
 184	if (head > tail)
 185		diff = head - tail;
 186	else
 187		diff = (end - tail) + head;
 188
 189	return (diff / CWQ_ENTRY_SIZE) - 1;
 190}
 191
 192static void *spu_queue_alloc(struct spu_queue *q, int num_entries)
 193{
 194	int avail = spu_queue_num_free(q);
 195
 196	if (avail >= num_entries)
 197		return q->q + q->tail;
 198
 199	return NULL;
 200}
 201
 202static unsigned long spu_queue_submit(struct spu_queue *q, void *last)
 203{
 204	unsigned long hv_ret, new_tail;
 205
 206	new_tail = spu_next_offset(q, last - q->q);
 207
 208	hv_ret = sun4v_ncs_settail(q->qhandle, new_tail);
 209	if (hv_ret == HV_EOK)
 210		q->tail = new_tail;
 211	return hv_ret;
 212}
 213
 214static u64 control_word_base(unsigned int len, unsigned int hmac_key_len,
 215			     int enc_type, int auth_type,
 216			     unsigned int hash_len,
 217			     bool sfas, bool sob, bool eob, bool encrypt,
 218			     int opcode)
 219{
 220	u64 word = (len - 1) & CONTROL_LEN;
 221
 222	word |= ((u64) opcode << CONTROL_OPCODE_SHIFT);
 223	word |= ((u64) enc_type << CONTROL_ENC_TYPE_SHIFT);
 224	word |= ((u64) auth_type << CONTROL_AUTH_TYPE_SHIFT);
 225	if (sfas)
 226		word |= CONTROL_STORE_FINAL_AUTH_STATE;
 227	if (sob)
 228		word |= CONTROL_START_OF_BLOCK;
 229	if (eob)
 230		word |= CONTROL_END_OF_BLOCK;
 231	if (encrypt)
 232		word |= CONTROL_ENCRYPT;
 233	if (hmac_key_len)
 234		word |= ((u64) (hmac_key_len - 1)) << CONTROL_HMAC_KEY_LEN_SHIFT;
 235	if (hash_len)
 236		word |= ((u64) (hash_len - 1)) << CONTROL_HASH_LEN_SHIFT;
 237
 238	return word;
 239}
 240
 241#if 0
 242static inline bool n2_should_run_async(struct spu_queue *qp, int this_len)
 243{
 244	if (this_len >= 64 ||
 245	    qp->head != qp->tail)
 246		return true;
 247	return false;
 248}
 249#endif
 250
 251struct n2_ahash_alg {
 252	struct list_head	entry;
 253	const u8		*hash_zero;
 254	const u8		*hash_init;
 255	u8			hw_op_hashsz;
 256	u8			digest_size;
 257	u8			auth_type;
 258	u8			hmac_type;
 259	struct ahash_alg	alg;
 260};
 261
 262static inline struct n2_ahash_alg *n2_ahash_alg(struct crypto_tfm *tfm)
 263{
 264	struct crypto_alg *alg = tfm->__crt_alg;
 265	struct ahash_alg *ahash_alg;
 266
 267	ahash_alg = container_of(alg, struct ahash_alg, halg.base);
 268
 269	return container_of(ahash_alg, struct n2_ahash_alg, alg);
 270}
 271
 272struct n2_hmac_alg {
 273	const char		*child_alg;
 274	struct n2_ahash_alg	derived;
 275};
 276
 277static inline struct n2_hmac_alg *n2_hmac_alg(struct crypto_tfm *tfm)
 278{
 279	struct crypto_alg *alg = tfm->__crt_alg;
 280	struct ahash_alg *ahash_alg;
 281
 282	ahash_alg = container_of(alg, struct ahash_alg, halg.base);
 283
 284	return container_of(ahash_alg, struct n2_hmac_alg, derived.alg);
 285}
 286
 287struct n2_hash_ctx {
 288	struct crypto_ahash		*fallback_tfm;
 289};
 290
 291#define N2_HASH_KEY_MAX			32 /* HW limit for all HMAC requests */
 292
 293struct n2_hmac_ctx {
 294	struct n2_hash_ctx		base;
 295
 296	struct crypto_shash		*child_shash;
 297
 298	int				hash_key_len;
 299	unsigned char			hash_key[N2_HASH_KEY_MAX];
 300};
 301
 302struct n2_hash_req_ctx {
 303	union {
 304		struct md5_state	md5;
 305		struct sha1_state	sha1;
 306		struct sha256_state	sha256;
 307	} u;
 308
 309	struct ahash_request		fallback_req;
 310};
 311
 312static int n2_hash_async_init(struct ahash_request *req)
 313{
 314	struct n2_hash_req_ctx *rctx = ahash_request_ctx(req);
 315	struct crypto_ahash *tfm = crypto_ahash_reqtfm(req);
 316	struct n2_hash_ctx *ctx = crypto_ahash_ctx(tfm);
 317
 318	ahash_request_set_tfm(&rctx->fallback_req, ctx->fallback_tfm);
 319	rctx->fallback_req.base.flags = req->base.flags & CRYPTO_TFM_REQ_MAY_SLEEP;
 320
 321	return crypto_ahash_init(&rctx->fallback_req);
 322}
 323
 324static int n2_hash_async_update(struct ahash_request *req)
 325{
 326	struct n2_hash_req_ctx *rctx = ahash_request_ctx(req);
 327	struct crypto_ahash *tfm = crypto_ahash_reqtfm(req);
 328	struct n2_hash_ctx *ctx = crypto_ahash_ctx(tfm);
 329
 330	ahash_request_set_tfm(&rctx->fallback_req, ctx->fallback_tfm);
 331	rctx->fallback_req.base.flags = req->base.flags & CRYPTO_TFM_REQ_MAY_SLEEP;
 332	rctx->fallback_req.nbytes = req->nbytes;
 333	rctx->fallback_req.src = req->src;
 334
 335	return crypto_ahash_update(&rctx->fallback_req);
 336}
 337
 338static int n2_hash_async_final(struct ahash_request *req)
 339{
 340	struct n2_hash_req_ctx *rctx = ahash_request_ctx(req);
 341	struct crypto_ahash *tfm = crypto_ahash_reqtfm(req);
 342	struct n2_hash_ctx *ctx = crypto_ahash_ctx(tfm);
 343
 344	ahash_request_set_tfm(&rctx->fallback_req, ctx->fallback_tfm);
 345	rctx->fallback_req.base.flags = req->base.flags & CRYPTO_TFM_REQ_MAY_SLEEP;
 346	rctx->fallback_req.result = req->result;
 347
 348	return crypto_ahash_final(&rctx->fallback_req);
 349}
 350
 351static int n2_hash_async_finup(struct ahash_request *req)
 352{
 353	struct n2_hash_req_ctx *rctx = ahash_request_ctx(req);
 354	struct crypto_ahash *tfm = crypto_ahash_reqtfm(req);
 355	struct n2_hash_ctx *ctx = crypto_ahash_ctx(tfm);
 356
 357	ahash_request_set_tfm(&rctx->fallback_req, ctx->fallback_tfm);
 358	rctx->fallback_req.base.flags = req->base.flags & CRYPTO_TFM_REQ_MAY_SLEEP;
 359	rctx->fallback_req.nbytes = req->nbytes;
 360	rctx->fallback_req.src = req->src;
 361	rctx->fallback_req.result = req->result;
 362
 363	return crypto_ahash_finup(&rctx->fallback_req);
 364}
 365
 366static int n2_hash_async_noimport(struct ahash_request *req, const void *in)
 367{
 368	return -ENOSYS;
 369}
 370
 371static int n2_hash_async_noexport(struct ahash_request *req, void *out)
 372{
 373	return -ENOSYS;
 374}
 375
 376static int n2_hash_cra_init(struct crypto_tfm *tfm)
 377{
 378	const char *fallback_driver_name = crypto_tfm_alg_name(tfm);
 379	struct crypto_ahash *ahash = __crypto_ahash_cast(tfm);
 380	struct n2_hash_ctx *ctx = crypto_ahash_ctx(ahash);
 381	struct crypto_ahash *fallback_tfm;
 382	int err;
 383
 384	fallback_tfm = crypto_alloc_ahash(fallback_driver_name, 0,
 385					  CRYPTO_ALG_NEED_FALLBACK);
 386	if (IS_ERR(fallback_tfm)) {
 387		pr_warn("Fallback driver '%s' could not be loaded!\n",
 388			fallback_driver_name);
 389		err = PTR_ERR(fallback_tfm);
 390		goto out;
 391	}
 392
 393	crypto_ahash_set_reqsize(ahash, (sizeof(struct n2_hash_req_ctx) +
 394					 crypto_ahash_reqsize(fallback_tfm)));
 395
 396	ctx->fallback_tfm = fallback_tfm;
 397	return 0;
 398
 399out:
 400	return err;
 401}
 402
 403static void n2_hash_cra_exit(struct crypto_tfm *tfm)
 404{
 405	struct crypto_ahash *ahash = __crypto_ahash_cast(tfm);
 406	struct n2_hash_ctx *ctx = crypto_ahash_ctx(ahash);
 407
 408	crypto_free_ahash(ctx->fallback_tfm);
 409}
 410
 411static int n2_hmac_cra_init(struct crypto_tfm *tfm)
 412{
 413	const char *fallback_driver_name = crypto_tfm_alg_name(tfm);
 414	struct crypto_ahash *ahash = __crypto_ahash_cast(tfm);
 415	struct n2_hmac_ctx *ctx = crypto_ahash_ctx(ahash);
 416	struct n2_hmac_alg *n2alg = n2_hmac_alg(tfm);
 417	struct crypto_ahash *fallback_tfm;
 418	struct crypto_shash *child_shash;
 419	int err;
 420
 421	fallback_tfm = crypto_alloc_ahash(fallback_driver_name, 0,
 422					  CRYPTO_ALG_NEED_FALLBACK);
 423	if (IS_ERR(fallback_tfm)) {
 424		pr_warn("Fallback driver '%s' could not be loaded!\n",
 425			fallback_driver_name);
 426		err = PTR_ERR(fallback_tfm);
 427		goto out;
 428	}
 429
 430	child_shash = crypto_alloc_shash(n2alg->child_alg, 0, 0);
 431	if (IS_ERR(child_shash)) {
 432		pr_warn("Child shash '%s' could not be loaded!\n",
 433			n2alg->child_alg);
 434		err = PTR_ERR(child_shash);
 435		goto out_free_fallback;
 436	}
 437
 438	crypto_ahash_set_reqsize(ahash, (sizeof(struct n2_hash_req_ctx) +
 439					 crypto_ahash_reqsize(fallback_tfm)));
 440
 441	ctx->child_shash = child_shash;
 442	ctx->base.fallback_tfm = fallback_tfm;
 443	return 0;
 444
 445out_free_fallback:
 446	crypto_free_ahash(fallback_tfm);
 447
 448out:
 449	return err;
 450}
 451
 452static void n2_hmac_cra_exit(struct crypto_tfm *tfm)
 453{
 454	struct crypto_ahash *ahash = __crypto_ahash_cast(tfm);
 455	struct n2_hmac_ctx *ctx = crypto_ahash_ctx(ahash);
 456
 457	crypto_free_ahash(ctx->base.fallback_tfm);
 458	crypto_free_shash(ctx->child_shash);
 459}
 460
 461static int n2_hmac_async_setkey(struct crypto_ahash *tfm, const u8 *key,
 462				unsigned int keylen)
 463{
 464	struct n2_hmac_ctx *ctx = crypto_ahash_ctx(tfm);
 465	struct crypto_shash *child_shash = ctx->child_shash;
 466	struct crypto_ahash *fallback_tfm;
 467	int err, bs, ds;
 468
 469	fallback_tfm = ctx->base.fallback_tfm;
 470	err = crypto_ahash_setkey(fallback_tfm, key, keylen);
 471	if (err)
 472		return err;
 473
 474	bs = crypto_shash_blocksize(child_shash);
 475	ds = crypto_shash_digestsize(child_shash);
 476	BUG_ON(ds > N2_HASH_KEY_MAX);
 477	if (keylen > bs) {
 478		err = crypto_shash_tfm_digest(child_shash, key, keylen,
 479					      ctx->hash_key);
 480		if (err)
 481			return err;
 482		keylen = ds;
 483	} else if (keylen <= N2_HASH_KEY_MAX)
 484		memcpy(ctx->hash_key, key, keylen);
 485
 486	ctx->hash_key_len = keylen;
 487
 488	return err;
 489}
 490
 491static unsigned long wait_for_tail(struct spu_queue *qp)
 492{
 493	unsigned long head, hv_ret;
 494
 495	do {
 496		hv_ret = sun4v_ncs_gethead(qp->qhandle, &head);
 497		if (hv_ret != HV_EOK) {
 498			pr_err("Hypervisor error on gethead\n");
 499			break;
 500		}
 501		if (head == qp->tail) {
 502			qp->head = head;
 503			break;
 504		}
 505	} while (1);
 506	return hv_ret;
 507}
 508
 509static unsigned long submit_and_wait_for_tail(struct spu_queue *qp,
 510					      struct cwq_initial_entry *ent)
 511{
 512	unsigned long hv_ret = spu_queue_submit(qp, ent);
 513
 514	if (hv_ret == HV_EOK)
 515		hv_ret = wait_for_tail(qp);
 516
 517	return hv_ret;
 518}
 519
 520static int n2_do_async_digest(struct ahash_request *req,
 521			      unsigned int auth_type, unsigned int digest_size,
 522			      unsigned int result_size, void *hash_loc,
 523			      unsigned long auth_key, unsigned int auth_key_len)
 524{
 525	struct crypto_ahash *tfm = crypto_ahash_reqtfm(req);
 526	struct cwq_initial_entry *ent;
 527	struct crypto_hash_walk walk;
 528	struct spu_queue *qp;
 529	unsigned long flags;
 530	int err = -ENODEV;
 531	int nbytes, cpu;
 532
 533	/* The total effective length of the operation may not
 534	 * exceed 2^16.
 535	 */
 536	if (unlikely(req->nbytes > (1 << 16))) {
 537		struct n2_hash_req_ctx *rctx = ahash_request_ctx(req);
 538		struct n2_hash_ctx *ctx = crypto_ahash_ctx(tfm);
 539
 540		ahash_request_set_tfm(&rctx->fallback_req, ctx->fallback_tfm);
 541		rctx->fallback_req.base.flags =
 542			req->base.flags & CRYPTO_TFM_REQ_MAY_SLEEP;
 543		rctx->fallback_req.nbytes = req->nbytes;
 544		rctx->fallback_req.src = req->src;
 545		rctx->fallback_req.result = req->result;
 546
 547		return crypto_ahash_digest(&rctx->fallback_req);
 548	}
 549
 550	nbytes = crypto_hash_walk_first(req, &walk);
 551
 552	cpu = get_cpu();
 553	qp = cpu_to_cwq[cpu];
 554	if (!qp)
 555		goto out;
 556
 557	spin_lock_irqsave(&qp->lock, flags);
 558
 559	/* XXX can do better, improve this later by doing a by-hand scatterlist
 560	 * XXX walk, etc.
 561	 */
 562	ent = qp->q + qp->tail;
 563
 564	ent->control = control_word_base(nbytes, auth_key_len, 0,
 565					 auth_type, digest_size,
 566					 false, true, false, false,
 567					 OPCODE_INPLACE_BIT |
 568					 OPCODE_AUTH_MAC);
 569	ent->src_addr = __pa(walk.data);
 570	ent->auth_key_addr = auth_key;
 571	ent->auth_iv_addr = __pa(hash_loc);
 572	ent->final_auth_state_addr = 0UL;
 573	ent->enc_key_addr = 0UL;
 574	ent->enc_iv_addr = 0UL;
 575	ent->dest_addr = __pa(hash_loc);
 576
 577	nbytes = crypto_hash_walk_done(&walk, 0);
 578	while (nbytes > 0) {
 579		ent = spu_queue_next(qp, ent);
 580
 581		ent->control = (nbytes - 1);
 582		ent->src_addr = __pa(walk.data);
 583		ent->auth_key_addr = 0UL;
 584		ent->auth_iv_addr = 0UL;
 585		ent->final_auth_state_addr = 0UL;
 586		ent->enc_key_addr = 0UL;
 587		ent->enc_iv_addr = 0UL;
 588		ent->dest_addr = 0UL;
 589
 590		nbytes = crypto_hash_walk_done(&walk, 0);
 591	}
 592	ent->control |= CONTROL_END_OF_BLOCK;
 593
 594	if (submit_and_wait_for_tail(qp, ent) != HV_EOK)
 595		err = -EINVAL;
 596	else
 597		err = 0;
 598
 599	spin_unlock_irqrestore(&qp->lock, flags);
 600
 601	if (!err)
 602		memcpy(req->result, hash_loc, result_size);
 603out:
 604	put_cpu();
 605
 606	return err;
 607}
 608
 609static int n2_hash_async_digest(struct ahash_request *req)
 610{
 611	struct n2_ahash_alg *n2alg = n2_ahash_alg(req->base.tfm);
 612	struct n2_hash_req_ctx *rctx = ahash_request_ctx(req);
 613	int ds;
 614
 615	ds = n2alg->digest_size;
 616	if (unlikely(req->nbytes == 0)) {
 617		memcpy(req->result, n2alg->hash_zero, ds);
 618		return 0;
 619	}
 620	memcpy(&rctx->u, n2alg->hash_init, n2alg->hw_op_hashsz);
 621
 622	return n2_do_async_digest(req, n2alg->auth_type,
 623				  n2alg->hw_op_hashsz, ds,
 624				  &rctx->u, 0UL, 0);
 625}
 626
 627static int n2_hmac_async_digest(struct ahash_request *req)
 628{
 629	struct n2_hmac_alg *n2alg = n2_hmac_alg(req->base.tfm);
 630	struct n2_hash_req_ctx *rctx = ahash_request_ctx(req);
 631	struct crypto_ahash *tfm = crypto_ahash_reqtfm(req);
 632	struct n2_hmac_ctx *ctx = crypto_ahash_ctx(tfm);
 633	int ds;
 634
 635	ds = n2alg->derived.digest_size;
 636	if (unlikely(req->nbytes == 0) ||
 637	    unlikely(ctx->hash_key_len > N2_HASH_KEY_MAX)) {
 638		struct n2_hash_req_ctx *rctx = ahash_request_ctx(req);
 639		struct n2_hash_ctx *ctx = crypto_ahash_ctx(tfm);
 640
 641		ahash_request_set_tfm(&rctx->fallback_req, ctx->fallback_tfm);
 642		rctx->fallback_req.base.flags =
 643			req->base.flags & CRYPTO_TFM_REQ_MAY_SLEEP;
 644		rctx->fallback_req.nbytes = req->nbytes;
 645		rctx->fallback_req.src = req->src;
 646		rctx->fallback_req.result = req->result;
 647
 648		return crypto_ahash_digest(&rctx->fallback_req);
 649	}
 650	memcpy(&rctx->u, n2alg->derived.hash_init,
 651	       n2alg->derived.hw_op_hashsz);
 652
 653	return n2_do_async_digest(req, n2alg->derived.hmac_type,
 654				  n2alg->derived.hw_op_hashsz, ds,
 655				  &rctx->u,
 656				  __pa(&ctx->hash_key),
 657				  ctx->hash_key_len);
 658}
 659
 660struct n2_skcipher_context {
 661	int			key_len;
 662	int			enc_type;
 663	union {
 664		u8		aes[AES_MAX_KEY_SIZE];
 665		u8		des[DES_KEY_SIZE];
 666		u8		des3[3 * DES_KEY_SIZE];
 667	} key;
 668};
 669
 670#define N2_CHUNK_ARR_LEN	16
 671
 672struct n2_crypto_chunk {
 673	struct list_head	entry;
 674	unsigned long		iv_paddr : 44;
 675	unsigned long		arr_len : 20;
 676	unsigned long		dest_paddr;
 677	unsigned long		dest_final;
 678	struct {
 679		unsigned long	src_paddr : 44;
 680		unsigned long	src_len : 20;
 681	} arr[N2_CHUNK_ARR_LEN];
 682};
 683
 684struct n2_request_context {
 685	struct skcipher_walk	walk;
 686	struct list_head	chunk_list;
 687	struct n2_crypto_chunk	chunk;
 688	u8			temp_iv[16];
 689};
 690
 691/* The SPU allows some level of flexibility for partial cipher blocks
 692 * being specified in a descriptor.
 693 *
 694 * It merely requires that every descriptor's length field is at least
 695 * as large as the cipher block size.  This means that a cipher block
 696 * can span at most 2 descriptors.  However, this does not allow a
 697 * partial block to span into the final descriptor as that would
 698 * violate the rule (since every descriptor's length must be at lest
 699 * the block size).  So, for example, assuming an 8 byte block size:
 700 *
 701 *	0xe --> 0xa --> 0x8
 702 *
 703 * is a valid length sequence, whereas:
 704 *
 705 *	0xe --> 0xb --> 0x7
 706 *
 707 * is not a valid sequence.
 708 */
 709
 710struct n2_skcipher_alg {
 711	struct list_head	entry;
 712	u8			enc_type;
 713	struct skcipher_alg	skcipher;
 714};
 715
 716static inline struct n2_skcipher_alg *n2_skcipher_alg(struct crypto_skcipher *tfm)
 717{
 718	struct skcipher_alg *alg = crypto_skcipher_alg(tfm);
 719
 720	return container_of(alg, struct n2_skcipher_alg, skcipher);
 721}
 722
 723struct n2_skcipher_request_context {
 724	struct skcipher_walk	walk;
 725};
 726
 727static int n2_aes_setkey(struct crypto_skcipher *skcipher, const u8 *key,
 728			 unsigned int keylen)
 729{
 730	struct crypto_tfm *tfm = crypto_skcipher_tfm(skcipher);
 731	struct n2_skcipher_context *ctx = crypto_tfm_ctx(tfm);
 732	struct n2_skcipher_alg *n2alg = n2_skcipher_alg(skcipher);
 733
 734	ctx->enc_type = (n2alg->enc_type & ENC_TYPE_CHAINING_MASK);
 735
 736	switch (keylen) {
 737	case AES_KEYSIZE_128:
 738		ctx->enc_type |= ENC_TYPE_ALG_AES128;
 739		break;
 740	case AES_KEYSIZE_192:
 741		ctx->enc_type |= ENC_TYPE_ALG_AES192;
 742		break;
 743	case AES_KEYSIZE_256:
 744		ctx->enc_type |= ENC_TYPE_ALG_AES256;
 745		break;
 746	default:
 747		return -EINVAL;
 748	}
 749
 750	ctx->key_len = keylen;
 751	memcpy(ctx->key.aes, key, keylen);
 752	return 0;
 753}
 754
 755static int n2_des_setkey(struct crypto_skcipher *skcipher, const u8 *key,
 756			 unsigned int keylen)
 757{
 758	struct crypto_tfm *tfm = crypto_skcipher_tfm(skcipher);
 759	struct n2_skcipher_context *ctx = crypto_tfm_ctx(tfm);
 760	struct n2_skcipher_alg *n2alg = n2_skcipher_alg(skcipher);
 761	int err;
 762
 763	err = verify_skcipher_des_key(skcipher, key);
 764	if (err)
 765		return err;
 766
 767	ctx->enc_type = n2alg->enc_type;
 768
 769	ctx->key_len = keylen;
 770	memcpy(ctx->key.des, key, keylen);
 771	return 0;
 772}
 773
 774static int n2_3des_setkey(struct crypto_skcipher *skcipher, const u8 *key,
 775			  unsigned int keylen)
 776{
 777	struct crypto_tfm *tfm = crypto_skcipher_tfm(skcipher);
 778	struct n2_skcipher_context *ctx = crypto_tfm_ctx(tfm);
 779	struct n2_skcipher_alg *n2alg = n2_skcipher_alg(skcipher);
 780	int err;
 781
 782	err = verify_skcipher_des3_key(skcipher, key);
 783	if (err)
 784		return err;
 785
 786	ctx->enc_type = n2alg->enc_type;
 787
 788	ctx->key_len = keylen;
 789	memcpy(ctx->key.des3, key, keylen);
 790	return 0;
 791}
 792
 793static inline int skcipher_descriptor_len(int nbytes, unsigned int block_size)
 794{
 795	int this_len = nbytes;
 796
 797	this_len -= (nbytes & (block_size - 1));
 798	return this_len > (1 << 16) ? (1 << 16) : this_len;
 799}
 800
 801static int __n2_crypt_chunk(struct crypto_skcipher *skcipher,
 802			    struct n2_crypto_chunk *cp,
 803			    struct spu_queue *qp, bool encrypt)
 804{
 805	struct n2_skcipher_context *ctx = crypto_skcipher_ctx(skcipher);
 806	struct cwq_initial_entry *ent;
 807	bool in_place;
 808	int i;
 809
 810	ent = spu_queue_alloc(qp, cp->arr_len);
 811	if (!ent) {
 812		pr_info("queue_alloc() of %d fails\n",
 813			cp->arr_len);
 814		return -EBUSY;
 815	}
 816
 817	in_place = (cp->dest_paddr == cp->arr[0].src_paddr);
 818
 819	ent->control = control_word_base(cp->arr[0].src_len,
 820					 0, ctx->enc_type, 0, 0,
 821					 false, true, false, encrypt,
 822					 OPCODE_ENCRYPT |
 823					 (in_place ? OPCODE_INPLACE_BIT : 0));
 824	ent->src_addr = cp->arr[0].src_paddr;
 825	ent->auth_key_addr = 0UL;
 826	ent->auth_iv_addr = 0UL;
 827	ent->final_auth_state_addr = 0UL;
 828	ent->enc_key_addr = __pa(&ctx->key);
 829	ent->enc_iv_addr = cp->iv_paddr;
 830	ent->dest_addr = (in_place ? 0UL : cp->dest_paddr);
 831
 832	for (i = 1; i < cp->arr_len; i++) {
 833		ent = spu_queue_next(qp, ent);
 834
 835		ent->control = cp->arr[i].src_len - 1;
 836		ent->src_addr = cp->arr[i].src_paddr;
 837		ent->auth_key_addr = 0UL;
 838		ent->auth_iv_addr = 0UL;
 839		ent->final_auth_state_addr = 0UL;
 840		ent->enc_key_addr = 0UL;
 841		ent->enc_iv_addr = 0UL;
 842		ent->dest_addr = 0UL;
 843	}
 844	ent->control |= CONTROL_END_OF_BLOCK;
 845
 846	return (spu_queue_submit(qp, ent) != HV_EOK) ? -EINVAL : 0;
 847}
 848
 849static int n2_compute_chunks(struct skcipher_request *req)
 850{
 851	struct n2_request_context *rctx = skcipher_request_ctx(req);
 852	struct skcipher_walk *walk = &rctx->walk;
 853	struct n2_crypto_chunk *chunk;
 854	unsigned long dest_prev;
 855	unsigned int tot_len;
 856	bool prev_in_place;
 857	int err, nbytes;
 858
 859	err = skcipher_walk_async(walk, req);
 860	if (err)
 861		return err;
 862
 863	INIT_LIST_HEAD(&rctx->chunk_list);
 864
 865	chunk = &rctx->chunk;
 866	INIT_LIST_HEAD(&chunk->entry);
 867
 868	chunk->iv_paddr = 0UL;
 869	chunk->arr_len = 0;
 870	chunk->dest_paddr = 0UL;
 871
 872	prev_in_place = false;
 873	dest_prev = ~0UL;
 874	tot_len = 0;
 875
 876	while ((nbytes = walk->nbytes) != 0) {
 877		unsigned long dest_paddr, src_paddr;
 878		bool in_place;
 879		int this_len;
 880
 881		src_paddr = (page_to_phys(walk->src.phys.page) +
 882			     walk->src.phys.offset);
 883		dest_paddr = (page_to_phys(walk->dst.phys.page) +
 884			      walk->dst.phys.offset);
 885		in_place = (src_paddr == dest_paddr);
 886		this_len = skcipher_descriptor_len(nbytes, walk->blocksize);
 887
 888		if (chunk->arr_len != 0) {
 889			if (in_place != prev_in_place ||
 890			    (!prev_in_place &&
 891			     dest_paddr != dest_prev) ||
 892			    chunk->arr_len == N2_CHUNK_ARR_LEN ||
 893			    tot_len + this_len > (1 << 16)) {
 894				chunk->dest_final = dest_prev;
 895				list_add_tail(&chunk->entry,
 896					      &rctx->chunk_list);
 897				chunk = kzalloc(sizeof(*chunk), GFP_ATOMIC);
 898				if (!chunk) {
 899					err = -ENOMEM;
 900					break;
 901				}
 902				INIT_LIST_HEAD(&chunk->entry);
 903			}
 904		}
 905		if (chunk->arr_len == 0) {
 906			chunk->dest_paddr = dest_paddr;
 907			tot_len = 0;
 908		}
 909		chunk->arr[chunk->arr_len].src_paddr = src_paddr;
 910		chunk->arr[chunk->arr_len].src_len = this_len;
 911		chunk->arr_len++;
 912
 913		dest_prev = dest_paddr + this_len;
 914		prev_in_place = in_place;
 915		tot_len += this_len;
 916
 917		err = skcipher_walk_done(walk, nbytes - this_len);
 918		if (err)
 919			break;
 920	}
 921	if (!err && chunk->arr_len != 0) {
 922		chunk->dest_final = dest_prev;
 923		list_add_tail(&chunk->entry, &rctx->chunk_list);
 924	}
 925
 926	return err;
 927}
 928
 929static void n2_chunk_complete(struct skcipher_request *req, void *final_iv)
 930{
 931	struct n2_request_context *rctx = skcipher_request_ctx(req);
 932	struct n2_crypto_chunk *c, *tmp;
 933
 934	if (final_iv)
 935		memcpy(rctx->walk.iv, final_iv, rctx->walk.blocksize);
 936
 937	list_for_each_entry_safe(c, tmp, &rctx->chunk_list, entry) {
 938		list_del(&c->entry);
 939		if (unlikely(c != &rctx->chunk))
 940			kfree(c);
 941	}
 942
 943}
 944
 945static int n2_do_ecb(struct skcipher_request *req, bool encrypt)
 946{
 947	struct n2_request_context *rctx = skcipher_request_ctx(req);
 948	struct crypto_skcipher *tfm = crypto_skcipher_reqtfm(req);
 949	int err = n2_compute_chunks(req);
 950	struct n2_crypto_chunk *c, *tmp;
 951	unsigned long flags, hv_ret;
 952	struct spu_queue *qp;
 953
 954	if (err)
 955		return err;
 956
 957	qp = cpu_to_cwq[get_cpu()];
 958	err = -ENODEV;
 959	if (!qp)
 960		goto out;
 961
 962	spin_lock_irqsave(&qp->lock, flags);
 963
 964	list_for_each_entry_safe(c, tmp, &rctx->chunk_list, entry) {
 965		err = __n2_crypt_chunk(tfm, c, qp, encrypt);
 966		if (err)
 967			break;
 968		list_del(&c->entry);
 969		if (unlikely(c != &rctx->chunk))
 970			kfree(c);
 971	}
 972	if (!err) {
 973		hv_ret = wait_for_tail(qp);
 974		if (hv_ret != HV_EOK)
 975			err = -EINVAL;
 976	}
 977
 978	spin_unlock_irqrestore(&qp->lock, flags);
 979
 980out:
 981	put_cpu();
 982
 983	n2_chunk_complete(req, NULL);
 984	return err;
 985}
 986
 987static int n2_encrypt_ecb(struct skcipher_request *req)
 988{
 989	return n2_do_ecb(req, true);
 990}
 991
 992static int n2_decrypt_ecb(struct skcipher_request *req)
 993{
 994	return n2_do_ecb(req, false);
 995}
 996
 997static int n2_do_chaining(struct skcipher_request *req, bool encrypt)
 998{
 999	struct n2_request_context *rctx = skcipher_request_ctx(req);
1000	struct crypto_skcipher *tfm = crypto_skcipher_reqtfm(req);
1001	unsigned long flags, hv_ret, iv_paddr;
1002	int err = n2_compute_chunks(req);
1003	struct n2_crypto_chunk *c, *tmp;
1004	struct spu_queue *qp;
1005	void *final_iv_addr;
1006
1007	final_iv_addr = NULL;
1008
1009	if (err)
1010		return err;
1011
1012	qp = cpu_to_cwq[get_cpu()];
1013	err = -ENODEV;
1014	if (!qp)
1015		goto out;
1016
1017	spin_lock_irqsave(&qp->lock, flags);
1018
1019	if (encrypt) {
1020		iv_paddr = __pa(rctx->walk.iv);
1021		list_for_each_entry_safe(c, tmp, &rctx->chunk_list,
1022					 entry) {
1023			c->iv_paddr = iv_paddr;
1024			err = __n2_crypt_chunk(tfm, c, qp, true);
1025			if (err)
1026				break;
1027			iv_paddr = c->dest_final - rctx->walk.blocksize;
1028			list_del(&c->entry);
1029			if (unlikely(c != &rctx->chunk))
1030				kfree(c);
1031		}
1032		final_iv_addr = __va(iv_paddr);
1033	} else {
1034		list_for_each_entry_safe_reverse(c, tmp, &rctx->chunk_list,
1035						 entry) {
1036			if (c == &rctx->chunk) {
1037				iv_paddr = __pa(rctx->walk.iv);
1038			} else {
1039				iv_paddr = (tmp->arr[tmp->arr_len-1].src_paddr +
1040					    tmp->arr[tmp->arr_len-1].src_len -
1041					    rctx->walk.blocksize);
1042			}
1043			if (!final_iv_addr) {
1044				unsigned long pa;
1045
1046				pa = (c->arr[c->arr_len-1].src_paddr +
1047				      c->arr[c->arr_len-1].src_len -
1048				      rctx->walk.blocksize);
1049				final_iv_addr = rctx->temp_iv;
1050				memcpy(rctx->temp_iv, __va(pa),
1051				       rctx->walk.blocksize);
1052			}
1053			c->iv_paddr = iv_paddr;
1054			err = __n2_crypt_chunk(tfm, c, qp, false);
1055			if (err)
1056				break;
1057			list_del(&c->entry);
1058			if (unlikely(c != &rctx->chunk))
1059				kfree(c);
1060		}
1061	}
1062	if (!err) {
1063		hv_ret = wait_for_tail(qp);
1064		if (hv_ret != HV_EOK)
1065			err = -EINVAL;
1066	}
1067
1068	spin_unlock_irqrestore(&qp->lock, flags);
1069
1070out:
1071	put_cpu();
1072
1073	n2_chunk_complete(req, err ? NULL : final_iv_addr);
1074	return err;
1075}
1076
1077static int n2_encrypt_chaining(struct skcipher_request *req)
1078{
1079	return n2_do_chaining(req, true);
1080}
1081
1082static int n2_decrypt_chaining(struct skcipher_request *req)
1083{
1084	return n2_do_chaining(req, false);
1085}
1086
1087struct n2_skcipher_tmpl {
1088	const char		*name;
1089	const char		*drv_name;
1090	u8			block_size;
1091	u8			enc_type;
1092	struct skcipher_alg	skcipher;
1093};
1094
1095static const struct n2_skcipher_tmpl skcipher_tmpls[] = {
1096	/* DES: ECB CBC and CFB are supported */
1097	{	.name		= "ecb(des)",
1098		.drv_name	= "ecb-des",
1099		.block_size	= DES_BLOCK_SIZE,
1100		.enc_type	= (ENC_TYPE_ALG_DES |
1101				   ENC_TYPE_CHAINING_ECB),
1102		.skcipher	= {
1103			.min_keysize	= DES_KEY_SIZE,
1104			.max_keysize	= DES_KEY_SIZE,
1105			.setkey		= n2_des_setkey,
1106			.encrypt	= n2_encrypt_ecb,
1107			.decrypt	= n2_decrypt_ecb,
1108		},
1109	},
1110	{	.name		= "cbc(des)",
1111		.drv_name	= "cbc-des",
1112		.block_size	= DES_BLOCK_SIZE,
1113		.enc_type	= (ENC_TYPE_ALG_DES |
1114				   ENC_TYPE_CHAINING_CBC),
1115		.skcipher	= {
1116			.ivsize		= DES_BLOCK_SIZE,
1117			.min_keysize	= DES_KEY_SIZE,
1118			.max_keysize	= DES_KEY_SIZE,
1119			.setkey		= n2_des_setkey,
1120			.encrypt	= n2_encrypt_chaining,
1121			.decrypt	= n2_decrypt_chaining,
1122		},
1123	},
1124
1125	/* 3DES: ECB CBC and CFB are supported */
1126	{	.name		= "ecb(des3_ede)",
1127		.drv_name	= "ecb-3des",
1128		.block_size	= DES_BLOCK_SIZE,
1129		.enc_type	= (ENC_TYPE_ALG_3DES |
1130				   ENC_TYPE_CHAINING_ECB),
1131		.skcipher	= {
1132			.min_keysize	= 3 * DES_KEY_SIZE,
1133			.max_keysize	= 3 * DES_KEY_SIZE,
1134			.setkey		= n2_3des_setkey,
1135			.encrypt	= n2_encrypt_ecb,
1136			.decrypt	= n2_decrypt_ecb,
1137		},
1138	},
1139	{	.name		= "cbc(des3_ede)",
1140		.drv_name	= "cbc-3des",
1141		.block_size	= DES_BLOCK_SIZE,
1142		.enc_type	= (ENC_TYPE_ALG_3DES |
1143				   ENC_TYPE_CHAINING_CBC),
1144		.skcipher	= {
1145			.ivsize		= DES_BLOCK_SIZE,
1146			.min_keysize	= 3 * DES_KEY_SIZE,
1147			.max_keysize	= 3 * DES_KEY_SIZE,
1148			.setkey		= n2_3des_setkey,
1149			.encrypt	= n2_encrypt_chaining,
1150			.decrypt	= n2_decrypt_chaining,
1151		},
1152	},
1153
1154	/* AES: ECB CBC and CTR are supported */
1155	{	.name		= "ecb(aes)",
1156		.drv_name	= "ecb-aes",
1157		.block_size	= AES_BLOCK_SIZE,
1158		.enc_type	= (ENC_TYPE_ALG_AES128 |
1159				   ENC_TYPE_CHAINING_ECB),
1160		.skcipher	= {
1161			.min_keysize	= AES_MIN_KEY_SIZE,
1162			.max_keysize	= AES_MAX_KEY_SIZE,
1163			.setkey		= n2_aes_setkey,
1164			.encrypt	= n2_encrypt_ecb,
1165			.decrypt	= n2_decrypt_ecb,
1166		},
1167	},
1168	{	.name		= "cbc(aes)",
1169		.drv_name	= "cbc-aes",
1170		.block_size	= AES_BLOCK_SIZE,
1171		.enc_type	= (ENC_TYPE_ALG_AES128 |
1172				   ENC_TYPE_CHAINING_CBC),
1173		.skcipher	= {
1174			.ivsize		= AES_BLOCK_SIZE,
1175			.min_keysize	= AES_MIN_KEY_SIZE,
1176			.max_keysize	= AES_MAX_KEY_SIZE,
1177			.setkey		= n2_aes_setkey,
1178			.encrypt	= n2_encrypt_chaining,
1179			.decrypt	= n2_decrypt_chaining,
1180		},
1181	},
1182	{	.name		= "ctr(aes)",
1183		.drv_name	= "ctr-aes",
1184		.block_size	= AES_BLOCK_SIZE,
1185		.enc_type	= (ENC_TYPE_ALG_AES128 |
1186				   ENC_TYPE_CHAINING_COUNTER),
1187		.skcipher	= {
1188			.ivsize		= AES_BLOCK_SIZE,
1189			.min_keysize	= AES_MIN_KEY_SIZE,
1190			.max_keysize	= AES_MAX_KEY_SIZE,
1191			.setkey		= n2_aes_setkey,
1192			.encrypt	= n2_encrypt_chaining,
1193			.decrypt	= n2_encrypt_chaining,
1194		},
1195	},
1196
1197};
1198#define NUM_CIPHER_TMPLS ARRAY_SIZE(skcipher_tmpls)
1199
1200static LIST_HEAD(skcipher_algs);
1201
1202struct n2_hash_tmpl {
1203	const char	*name;
1204	const u8	*hash_zero;
1205	const u8	*hash_init;
1206	u8		hw_op_hashsz;
1207	u8		digest_size;
1208	u8		statesize;
1209	u8		block_size;
1210	u8		auth_type;
1211	u8		hmac_type;
1212};
1213
1214static const __le32 n2_md5_init[MD5_HASH_WORDS] = {
1215	cpu_to_le32(MD5_H0),
1216	cpu_to_le32(MD5_H1),
1217	cpu_to_le32(MD5_H2),
1218	cpu_to_le32(MD5_H3),
1219};
1220static const u32 n2_sha1_init[SHA1_DIGEST_SIZE / 4] = {
1221	SHA1_H0, SHA1_H1, SHA1_H2, SHA1_H3, SHA1_H4,
1222};
1223static const u32 n2_sha256_init[SHA256_DIGEST_SIZE / 4] = {
1224	SHA256_H0, SHA256_H1, SHA256_H2, SHA256_H3,
1225	SHA256_H4, SHA256_H5, SHA256_H6, SHA256_H7,
1226};
1227static const u32 n2_sha224_init[SHA256_DIGEST_SIZE / 4] = {
1228	SHA224_H0, SHA224_H1, SHA224_H2, SHA224_H3,
1229	SHA224_H4, SHA224_H5, SHA224_H6, SHA224_H7,
1230};
1231
1232static const struct n2_hash_tmpl hash_tmpls[] = {
1233	{ .name		= "md5",
1234	  .hash_zero	= md5_zero_message_hash,
1235	  .hash_init	= (u8 *)n2_md5_init,
1236	  .auth_type	= AUTH_TYPE_MD5,
1237	  .hmac_type	= AUTH_TYPE_HMAC_MD5,
1238	  .hw_op_hashsz	= MD5_DIGEST_SIZE,
1239	  .digest_size	= MD5_DIGEST_SIZE,
1240	  .statesize	= sizeof(struct md5_state),
1241	  .block_size	= MD5_HMAC_BLOCK_SIZE },
1242	{ .name		= "sha1",
1243	  .hash_zero	= sha1_zero_message_hash,
1244	  .hash_init	= (u8 *)n2_sha1_init,
1245	  .auth_type	= AUTH_TYPE_SHA1,
1246	  .hmac_type	= AUTH_TYPE_HMAC_SHA1,
1247	  .hw_op_hashsz	= SHA1_DIGEST_SIZE,
1248	  .digest_size	= SHA1_DIGEST_SIZE,
1249	  .statesize	= sizeof(struct sha1_state),
1250	  .block_size	= SHA1_BLOCK_SIZE },
1251	{ .name		= "sha256",
1252	  .hash_zero	= sha256_zero_message_hash,
1253	  .hash_init	= (u8 *)n2_sha256_init,
1254	  .auth_type	= AUTH_TYPE_SHA256,
1255	  .hmac_type	= AUTH_TYPE_HMAC_SHA256,
1256	  .hw_op_hashsz	= SHA256_DIGEST_SIZE,
1257	  .digest_size	= SHA256_DIGEST_SIZE,
1258	  .statesize	= sizeof(struct sha256_state),
1259	  .block_size	= SHA256_BLOCK_SIZE },
1260	{ .name		= "sha224",
1261	  .hash_zero	= sha224_zero_message_hash,
1262	  .hash_init	= (u8 *)n2_sha224_init,
1263	  .auth_type	= AUTH_TYPE_SHA256,
1264	  .hmac_type	= AUTH_TYPE_RESERVED,
1265	  .hw_op_hashsz	= SHA256_DIGEST_SIZE,
1266	  .digest_size	= SHA224_DIGEST_SIZE,
1267	  .statesize	= sizeof(struct sha256_state),
1268	  .block_size	= SHA224_BLOCK_SIZE },
1269};
1270#define NUM_HASH_TMPLS ARRAY_SIZE(hash_tmpls)
1271
1272static LIST_HEAD(ahash_algs);
1273static LIST_HEAD(hmac_algs);
1274
1275static int algs_registered;
1276
1277static void __n2_unregister_algs(void)
1278{
1279	struct n2_skcipher_alg *skcipher, *skcipher_tmp;
1280	struct n2_ahash_alg *alg, *alg_tmp;
1281	struct n2_hmac_alg *hmac, *hmac_tmp;
1282
1283	list_for_each_entry_safe(skcipher, skcipher_tmp, &skcipher_algs, entry) {
1284		crypto_unregister_skcipher(&skcipher->skcipher);
1285		list_del(&skcipher->entry);
1286		kfree(skcipher);
1287	}
1288	list_for_each_entry_safe(hmac, hmac_tmp, &hmac_algs, derived.entry) {
1289		crypto_unregister_ahash(&hmac->derived.alg);
1290		list_del(&hmac->derived.entry);
1291		kfree(hmac);
1292	}
1293	list_for_each_entry_safe(alg, alg_tmp, &ahash_algs, entry) {
1294		crypto_unregister_ahash(&alg->alg);
1295		list_del(&alg->entry);
1296		kfree(alg);
1297	}
1298}
1299
1300static int n2_skcipher_init_tfm(struct crypto_skcipher *tfm)
1301{
1302	crypto_skcipher_set_reqsize(tfm, sizeof(struct n2_request_context));
1303	return 0;
1304}
1305
1306static int __n2_register_one_skcipher(const struct n2_skcipher_tmpl *tmpl)
1307{
1308	struct n2_skcipher_alg *p = kzalloc(sizeof(*p), GFP_KERNEL);
1309	struct skcipher_alg *alg;
1310	int err;
1311
1312	if (!p)
1313		return -ENOMEM;
1314
1315	alg = &p->skcipher;
1316	*alg = tmpl->skcipher;
1317
1318	snprintf(alg->base.cra_name, CRYPTO_MAX_ALG_NAME, "%s", tmpl->name);
1319	snprintf(alg->base.cra_driver_name, CRYPTO_MAX_ALG_NAME, "%s-n2", tmpl->drv_name);
1320	alg->base.cra_priority = N2_CRA_PRIORITY;
1321	alg->base.cra_flags = CRYPTO_ALG_KERN_DRIVER_ONLY | CRYPTO_ALG_ASYNC |
1322			      CRYPTO_ALG_ALLOCATES_MEMORY;
1323	alg->base.cra_blocksize = tmpl->block_size;
1324	p->enc_type = tmpl->enc_type;
1325	alg->base.cra_ctxsize = sizeof(struct n2_skcipher_context);
1326	alg->base.cra_module = THIS_MODULE;
1327	alg->init = n2_skcipher_init_tfm;
1328
1329	list_add(&p->entry, &skcipher_algs);
1330	err = crypto_register_skcipher(alg);
1331	if (err) {
1332		pr_err("%s alg registration failed\n", alg->base.cra_name);
1333		list_del(&p->entry);
1334		kfree(p);
1335	} else {
1336		pr_info("%s alg registered\n", alg->base.cra_name);
1337	}
1338	return err;
1339}
1340
1341static int __n2_register_one_hmac(struct n2_ahash_alg *n2ahash)
1342{
1343	struct n2_hmac_alg *p = kzalloc(sizeof(*p), GFP_KERNEL);
1344	struct ahash_alg *ahash;
1345	struct crypto_alg *base;
1346	int err;
1347
1348	if (!p)
1349		return -ENOMEM;
1350
1351	p->child_alg = n2ahash->alg.halg.base.cra_name;
1352	memcpy(&p->derived, n2ahash, sizeof(struct n2_ahash_alg));
1353	INIT_LIST_HEAD(&p->derived.entry);
1354
1355	ahash = &p->derived.alg;
1356	ahash->digest = n2_hmac_async_digest;
1357	ahash->setkey = n2_hmac_async_setkey;
1358
1359	base = &ahash->halg.base;
1360	if (snprintf(base->cra_name, CRYPTO_MAX_ALG_NAME, "hmac(%s)",
1361		     p->child_alg) >= CRYPTO_MAX_ALG_NAME)
1362		goto out_free_p;
1363	if (snprintf(base->cra_driver_name, CRYPTO_MAX_ALG_NAME, "hmac-%s-n2",
1364		     p->child_alg) >= CRYPTO_MAX_ALG_NAME)
1365		goto out_free_p;
1366
1367	base->cra_ctxsize = sizeof(struct n2_hmac_ctx);
1368	base->cra_init = n2_hmac_cra_init;
1369	base->cra_exit = n2_hmac_cra_exit;
1370
1371	list_add(&p->derived.entry, &hmac_algs);
1372	err = crypto_register_ahash(ahash);
1373	if (err) {
1374		pr_err("%s alg registration failed\n", base->cra_name);
1375		list_del(&p->derived.entry);
1376out_free_p:
1377		kfree(p);
1378	} else {
1379		pr_info("%s alg registered\n", base->cra_name);
1380	}
1381	return err;
1382}
1383
1384static int __n2_register_one_ahash(const struct n2_hash_tmpl *tmpl)
1385{
1386	struct n2_ahash_alg *p = kzalloc(sizeof(*p), GFP_KERNEL);
1387	struct hash_alg_common *halg;
1388	struct crypto_alg *base;
1389	struct ahash_alg *ahash;
1390	int err;
1391
1392	if (!p)
1393		return -ENOMEM;
1394
1395	p->hash_zero = tmpl->hash_zero;
1396	p->hash_init = tmpl->hash_init;
1397	p->auth_type = tmpl->auth_type;
1398	p->hmac_type = tmpl->hmac_type;
1399	p->hw_op_hashsz = tmpl->hw_op_hashsz;
1400	p->digest_size = tmpl->digest_size;
1401
1402	ahash = &p->alg;
1403	ahash->init = n2_hash_async_init;
1404	ahash->update = n2_hash_async_update;
1405	ahash->final = n2_hash_async_final;
1406	ahash->finup = n2_hash_async_finup;
1407	ahash->digest = n2_hash_async_digest;
1408	ahash->export = n2_hash_async_noexport;
1409	ahash->import = n2_hash_async_noimport;
1410
1411	halg = &ahash->halg;
1412	halg->digestsize = tmpl->digest_size;
1413	halg->statesize = tmpl->statesize;
1414
1415	base = &halg->base;
1416	snprintf(base->cra_name, CRYPTO_MAX_ALG_NAME, "%s", tmpl->name);
1417	snprintf(base->cra_driver_name, CRYPTO_MAX_ALG_NAME, "%s-n2", tmpl->name);
1418	base->cra_priority = N2_CRA_PRIORITY;
1419	base->cra_flags = CRYPTO_ALG_KERN_DRIVER_ONLY |
1420			  CRYPTO_ALG_NEED_FALLBACK;
1421	base->cra_blocksize = tmpl->block_size;
1422	base->cra_ctxsize = sizeof(struct n2_hash_ctx);
1423	base->cra_module = THIS_MODULE;
1424	base->cra_init = n2_hash_cra_init;
1425	base->cra_exit = n2_hash_cra_exit;
1426
1427	list_add(&p->entry, &ahash_algs);
1428	err = crypto_register_ahash(ahash);
1429	if (err) {
1430		pr_err("%s alg registration failed\n", base->cra_name);
1431		list_del(&p->entry);
1432		kfree(p);
1433	} else {
1434		pr_info("%s alg registered\n", base->cra_name);
1435	}
1436	if (!err && p->hmac_type != AUTH_TYPE_RESERVED)
1437		err = __n2_register_one_hmac(p);
1438	return err;
1439}
1440
1441static int n2_register_algs(void)
1442{
1443	int i, err = 0;
1444
1445	mutex_lock(&spu_lock);
1446	if (algs_registered++)
1447		goto out;
1448
1449	for (i = 0; i < NUM_HASH_TMPLS; i++) {
1450		err = __n2_register_one_ahash(&hash_tmpls[i]);
1451		if (err) {
1452			__n2_unregister_algs();
1453			goto out;
1454		}
1455	}
1456	for (i = 0; i < NUM_CIPHER_TMPLS; i++) {
1457		err = __n2_register_one_skcipher(&skcipher_tmpls[i]);
1458		if (err) {
1459			__n2_unregister_algs();
1460			goto out;
1461		}
1462	}
1463
1464out:
1465	mutex_unlock(&spu_lock);
1466	return err;
1467}
1468
1469static void n2_unregister_algs(void)
1470{
1471	mutex_lock(&spu_lock);
1472	if (!--algs_registered)
1473		__n2_unregister_algs();
1474	mutex_unlock(&spu_lock);
1475}
1476
1477/* To map CWQ queues to interrupt sources, the hypervisor API provides
1478 * a devino.  This isn't very useful to us because all of the
1479 * interrupts listed in the device_node have been translated to
1480 * Linux virtual IRQ cookie numbers.
1481 *
1482 * So we have to back-translate, going through the 'intr' and 'ino'
1483 * property tables of the n2cp MDESC node, matching it with the OF
1484 * 'interrupts' property entries, in order to figure out which
1485 * devino goes to which already-translated IRQ.
1486 */
1487static int find_devino_index(struct platform_device *dev, struct spu_mdesc_info *ip,
1488			     unsigned long dev_ino)
1489{
1490	const unsigned int *dev_intrs;
1491	unsigned int intr;
1492	int i;
1493
1494	for (i = 0; i < ip->num_intrs; i++) {
1495		if (ip->ino_table[i].ino == dev_ino)
1496			break;
1497	}
1498	if (i == ip->num_intrs)
1499		return -ENODEV;
1500
1501	intr = ip->ino_table[i].intr;
1502
1503	dev_intrs = of_get_property(dev->dev.of_node, "interrupts", NULL);
1504	if (!dev_intrs)
1505		return -ENODEV;
1506
1507	for (i = 0; i < dev->archdata.num_irqs; i++) {
1508		if (dev_intrs[i] == intr)
1509			return i;
1510	}
1511
1512	return -ENODEV;
1513}
1514
1515static int spu_map_ino(struct platform_device *dev, struct spu_mdesc_info *ip,
1516		       const char *irq_name, struct spu_queue *p,
1517		       irq_handler_t handler)
1518{
1519	unsigned long herr;
1520	int index;
1521
1522	herr = sun4v_ncs_qhandle_to_devino(p->qhandle, &p->devino);
1523	if (herr)
1524		return -EINVAL;
1525
1526	index = find_devino_index(dev, ip, p->devino);
1527	if (index < 0)
1528		return index;
1529
1530	p->irq = dev->archdata.irqs[index];
1531
1532	sprintf(p->irq_name, "%s-%d", irq_name, index);
1533
1534	return request_irq(p->irq, handler, 0, p->irq_name, p);
1535}
1536
1537static struct kmem_cache *queue_cache[2];
1538
1539static void *new_queue(unsigned long q_type)
1540{
1541	return kmem_cache_zalloc(queue_cache[q_type - 1], GFP_KERNEL);
1542}
1543
1544static void free_queue(void *p, unsigned long q_type)
1545{
1546	kmem_cache_free(queue_cache[q_type - 1], p);
1547}
1548
1549static int queue_cache_init(void)
1550{
1551	if (!queue_cache[HV_NCS_QTYPE_MAU - 1])
1552		queue_cache[HV_NCS_QTYPE_MAU - 1] =
1553			kmem_cache_create("mau_queue",
1554					  (MAU_NUM_ENTRIES *
1555					   MAU_ENTRY_SIZE),
1556					  MAU_ENTRY_SIZE, 0, NULL);
1557	if (!queue_cache[HV_NCS_QTYPE_MAU - 1])
1558		return -ENOMEM;
1559
1560	if (!queue_cache[HV_NCS_QTYPE_CWQ - 1])
1561		queue_cache[HV_NCS_QTYPE_CWQ - 1] =
1562			kmem_cache_create("cwq_queue",
1563					  (CWQ_NUM_ENTRIES *
1564					   CWQ_ENTRY_SIZE),
1565					  CWQ_ENTRY_SIZE, 0, NULL);
1566	if (!queue_cache[HV_NCS_QTYPE_CWQ - 1]) {
1567		kmem_cache_destroy(queue_cache[HV_NCS_QTYPE_MAU - 1]);
1568		queue_cache[HV_NCS_QTYPE_MAU - 1] = NULL;
1569		return -ENOMEM;
1570	}
1571	return 0;
1572}
1573
1574static void queue_cache_destroy(void)
1575{
1576	kmem_cache_destroy(queue_cache[HV_NCS_QTYPE_MAU - 1]);
1577	kmem_cache_destroy(queue_cache[HV_NCS_QTYPE_CWQ - 1]);
1578	queue_cache[HV_NCS_QTYPE_MAU - 1] = NULL;
1579	queue_cache[HV_NCS_QTYPE_CWQ - 1] = NULL;
1580}
1581
1582static long spu_queue_register_workfn(void *arg)
1583{
1584	struct spu_qreg *qr = arg;
1585	struct spu_queue *p = qr->queue;
1586	unsigned long q_type = qr->type;
1587	unsigned long hv_ret;
1588
1589	hv_ret = sun4v_ncs_qconf(q_type, __pa(p->q),
1590				 CWQ_NUM_ENTRIES, &p->qhandle);
1591	if (!hv_ret)
1592		sun4v_ncs_sethead_marker(p->qhandle, 0);
1593
1594	return hv_ret ? -EINVAL : 0;
1595}
1596
1597static int spu_queue_register(struct spu_queue *p, unsigned long q_type)
1598{
1599	int cpu = cpumask_any_and(&p->sharing, cpu_online_mask);
1600	struct spu_qreg qr = { .queue = p, .type = q_type };
1601
1602	return work_on_cpu_safe(cpu, spu_queue_register_workfn, &qr);
1603}
1604
1605static int spu_queue_setup(struct spu_queue *p)
1606{
1607	int err;
1608
1609	p->q = new_queue(p->q_type);
1610	if (!p->q)
1611		return -ENOMEM;
1612
1613	err = spu_queue_register(p, p->q_type);
1614	if (err) {
1615		free_queue(p->q, p->q_type);
1616		p->q = NULL;
1617	}
1618
1619	return err;
1620}
1621
1622static void spu_queue_destroy(struct spu_queue *p)
1623{
1624	unsigned long hv_ret;
1625
1626	if (!p->q)
1627		return;
1628
1629	hv_ret = sun4v_ncs_qconf(p->q_type, p->qhandle, 0, &p->qhandle);
1630
1631	if (!hv_ret)
1632		free_queue(p->q, p->q_type);
1633}
1634
1635static void spu_list_destroy(struct list_head *list)
1636{
1637	struct spu_queue *p, *n;
1638
1639	list_for_each_entry_safe(p, n, list, list) {
1640		int i;
1641
1642		for (i = 0; i < NR_CPUS; i++) {
1643			if (cpu_to_cwq[i] == p)
1644				cpu_to_cwq[i] = NULL;
1645		}
1646
1647		if (p->irq) {
1648			free_irq(p->irq, p);
1649			p->irq = 0;
1650		}
1651		spu_queue_destroy(p);
1652		list_del(&p->list);
1653		kfree(p);
1654	}
1655}
1656
1657/* Walk the backward arcs of a CWQ 'exec-unit' node,
1658 * gathering cpu membership information.
1659 */
1660static int spu_mdesc_walk_arcs(struct mdesc_handle *mdesc,
1661			       struct platform_device *dev,
1662			       u64 node, struct spu_queue *p,
1663			       struct spu_queue **table)
1664{
1665	u64 arc;
1666
1667	mdesc_for_each_arc(arc, mdesc, node, MDESC_ARC_TYPE_BACK) {
1668		u64 tgt = mdesc_arc_target(mdesc, arc);
1669		const char *name = mdesc_node_name(mdesc, tgt);
1670		const u64 *id;
1671
1672		if (strcmp(name, "cpu"))
1673			continue;
1674		id = mdesc_get_property(mdesc, tgt, "id", NULL);
1675		if (table[*id] != NULL) {
1676			dev_err(&dev->dev, "%pOF: SPU cpu slot already set.\n",
1677				dev->dev.of_node);
1678			return -EINVAL;
1679		}
1680		cpumask_set_cpu(*id, &p->sharing);
1681		table[*id] = p;
1682	}
1683	return 0;
1684}
1685
1686/* Process an 'exec-unit' MDESC node of type 'cwq'.  */
1687static int handle_exec_unit(struct spu_mdesc_info *ip, struct list_head *list,
1688			    struct platform_device *dev, struct mdesc_handle *mdesc,
1689			    u64 node, const char *iname, unsigned long q_type,
1690			    irq_handler_t handler, struct spu_queue **table)
1691{
1692	struct spu_queue *p;
1693	int err;
1694
1695	p = kzalloc(sizeof(struct spu_queue), GFP_KERNEL);
1696	if (!p) {
1697		dev_err(&dev->dev, "%pOF: Could not allocate SPU queue.\n",
1698			dev->dev.of_node);
1699		return -ENOMEM;
1700	}
1701
1702	cpumask_clear(&p->sharing);
1703	spin_lock_init(&p->lock);
1704	p->q_type = q_type;
1705	INIT_LIST_HEAD(&p->jobs);
1706	list_add(&p->list, list);
1707
1708	err = spu_mdesc_walk_arcs(mdesc, dev, node, p, table);
1709	if (err)
1710		return err;
1711
1712	err = spu_queue_setup(p);
1713	if (err)
1714		return err;
1715
1716	return spu_map_ino(dev, ip, iname, p, handler);
1717}
1718
1719static int spu_mdesc_scan(struct mdesc_handle *mdesc, struct platform_device *dev,
1720			  struct spu_mdesc_info *ip, struct list_head *list,
1721			  const char *exec_name, unsigned long q_type,
1722			  irq_handler_t handler, struct spu_queue **table)
1723{
1724	int err = 0;
1725	u64 node;
1726
1727	mdesc_for_each_node_by_name(mdesc, node, "exec-unit") {
1728		const char *type;
1729
1730		type = mdesc_get_property(mdesc, node, "type", NULL);
1731		if (!type || strcmp(type, exec_name))
1732			continue;
1733
1734		err = handle_exec_unit(ip, list, dev, mdesc, node,
1735				       exec_name, q_type, handler, table);
1736		if (err) {
1737			spu_list_destroy(list);
1738			break;
1739		}
1740	}
1741
1742	return err;
1743}
1744
1745static int get_irq_props(struct mdesc_handle *mdesc, u64 node,
1746			 struct spu_mdesc_info *ip)
1747{
1748	const u64 *ino;
1749	int ino_len;
1750	int i;
1751
1752	ino = mdesc_get_property(mdesc, node, "ino", &ino_len);
1753	if (!ino) {
1754		printk("NO 'ino'\n");
1755		return -ENODEV;
1756	}
1757
1758	ip->num_intrs = ino_len / sizeof(u64);
1759	ip->ino_table = kzalloc((sizeof(struct ino_blob) *
1760				 ip->num_intrs),
1761				GFP_KERNEL);
1762	if (!ip->ino_table)
1763		return -ENOMEM;
1764
1765	for (i = 0; i < ip->num_intrs; i++) {
1766		struct ino_blob *b = &ip->ino_table[i];
1767		b->intr = i + 1;
1768		b->ino = ino[i];
1769	}
1770
1771	return 0;
1772}
1773
1774static int grab_mdesc_irq_props(struct mdesc_handle *mdesc,
1775				struct platform_device *dev,
1776				struct spu_mdesc_info *ip,
1777				const char *node_name)
1778{
1779	u64 node, reg;
1780
1781	if (of_property_read_reg(dev->dev.of_node, 0, &reg, NULL) < 0)
1782		return -ENODEV;
1783
1784	mdesc_for_each_node_by_name(mdesc, node, "virtual-device") {
1785		const char *name;
1786		const u64 *chdl;
1787
1788		name = mdesc_get_property(mdesc, node, "name", NULL);
1789		if (!name || strcmp(name, node_name))
1790			continue;
1791		chdl = mdesc_get_property(mdesc, node, "cfg-handle", NULL);
1792		if (!chdl || (*chdl != reg))
1793			continue;
1794		ip->cfg_handle = *chdl;
1795		return get_irq_props(mdesc, node, ip);
1796	}
1797
1798	return -ENODEV;
1799}
1800
1801static unsigned long n2_spu_hvapi_major;
1802static unsigned long n2_spu_hvapi_minor;
1803
1804static int n2_spu_hvapi_register(void)
1805{
1806	int err;
1807
1808	n2_spu_hvapi_major = 2;
1809	n2_spu_hvapi_minor = 0;
1810
1811	err = sun4v_hvapi_register(HV_GRP_NCS,
1812				   n2_spu_hvapi_major,
1813				   &n2_spu_hvapi_minor);
1814
1815	if (!err)
1816		pr_info("Registered NCS HVAPI version %lu.%lu\n",
1817			n2_spu_hvapi_major,
1818			n2_spu_hvapi_minor);
1819
1820	return err;
1821}
1822
1823static void n2_spu_hvapi_unregister(void)
1824{
1825	sun4v_hvapi_unregister(HV_GRP_NCS);
1826}
1827
1828static int global_ref;
1829
1830static int grab_global_resources(void)
1831{
1832	int err = 0;
1833
1834	mutex_lock(&spu_lock);
1835
1836	if (global_ref++)
1837		goto out;
1838
1839	err = n2_spu_hvapi_register();
1840	if (err)
1841		goto out;
1842
1843	err = queue_cache_init();
1844	if (err)
1845		goto out_hvapi_release;
1846
1847	err = -ENOMEM;
1848	cpu_to_cwq = kcalloc(NR_CPUS, sizeof(struct spu_queue *),
1849			     GFP_KERNEL);
1850	if (!cpu_to_cwq)
1851		goto out_queue_cache_destroy;
1852
1853	cpu_to_mau = kcalloc(NR_CPUS, sizeof(struct spu_queue *),
1854			     GFP_KERNEL);
1855	if (!cpu_to_mau)
1856		goto out_free_cwq_table;
1857
1858	err = 0;
1859
1860out:
1861	if (err)
1862		global_ref--;
1863	mutex_unlock(&spu_lock);
1864	return err;
1865
1866out_free_cwq_table:
1867	kfree(cpu_to_cwq);
1868	cpu_to_cwq = NULL;
1869
1870out_queue_cache_destroy:
1871	queue_cache_destroy();
1872
1873out_hvapi_release:
1874	n2_spu_hvapi_unregister();
1875	goto out;
1876}
1877
1878static void release_global_resources(void)
1879{
1880	mutex_lock(&spu_lock);
1881	if (!--global_ref) {
1882		kfree(cpu_to_cwq);
1883		cpu_to_cwq = NULL;
1884
1885		kfree(cpu_to_mau);
1886		cpu_to_mau = NULL;
1887
1888		queue_cache_destroy();
1889		n2_spu_hvapi_unregister();
1890	}
1891	mutex_unlock(&spu_lock);
1892}
1893
1894static struct n2_crypto *alloc_n2cp(void)
1895{
1896	struct n2_crypto *np = kzalloc(sizeof(struct n2_crypto), GFP_KERNEL);
1897
1898	if (np)
1899		INIT_LIST_HEAD(&np->cwq_list);
1900
1901	return np;
1902}
1903
1904static void free_n2cp(struct n2_crypto *np)
1905{
1906	kfree(np->cwq_info.ino_table);
1907	np->cwq_info.ino_table = NULL;
1908
1909	kfree(np);
1910}
1911
1912static void n2_spu_driver_version(void)
1913{
1914	static int n2_spu_version_printed;
1915
1916	if (n2_spu_version_printed++ == 0)
1917		pr_info("%s", version);
1918}
1919
1920static int n2_crypto_probe(struct platform_device *dev)
1921{
1922	struct mdesc_handle *mdesc;
1923	struct n2_crypto *np;
1924	int err;
1925
1926	n2_spu_driver_version();
1927
1928	pr_info("Found N2CP at %pOF\n", dev->dev.of_node);
1929
1930	np = alloc_n2cp();
1931	if (!np) {
1932		dev_err(&dev->dev, "%pOF: Unable to allocate n2cp.\n",
1933			dev->dev.of_node);
1934		return -ENOMEM;
1935	}
1936
1937	err = grab_global_resources();
1938	if (err) {
1939		dev_err(&dev->dev, "%pOF: Unable to grab global resources.\n",
1940			dev->dev.of_node);
1941		goto out_free_n2cp;
1942	}
1943
1944	mdesc = mdesc_grab();
1945
1946	if (!mdesc) {
1947		dev_err(&dev->dev, "%pOF: Unable to grab MDESC.\n",
1948			dev->dev.of_node);
1949		err = -ENODEV;
1950		goto out_free_global;
1951	}
1952	err = grab_mdesc_irq_props(mdesc, dev, &np->cwq_info, "n2cp");
1953	if (err) {
1954		dev_err(&dev->dev, "%pOF: Unable to grab IRQ props.\n",
1955			dev->dev.of_node);
1956		mdesc_release(mdesc);
1957		goto out_free_global;
1958	}
1959
1960	err = spu_mdesc_scan(mdesc, dev, &np->cwq_info, &np->cwq_list,
1961			     "cwq", HV_NCS_QTYPE_CWQ, cwq_intr,
1962			     cpu_to_cwq);
1963	mdesc_release(mdesc);
1964
1965	if (err) {
1966		dev_err(&dev->dev, "%pOF: CWQ MDESC scan failed.\n",
1967			dev->dev.of_node);
1968		goto out_free_global;
1969	}
1970
1971	err = n2_register_algs();
1972	if (err) {
1973		dev_err(&dev->dev, "%pOF: Unable to register algorithms.\n",
1974			dev->dev.of_node);
1975		goto out_free_spu_list;
1976	}
1977
1978	dev_set_drvdata(&dev->dev, np);
1979
1980	return 0;
1981
1982out_free_spu_list:
1983	spu_list_destroy(&np->cwq_list);
1984
1985out_free_global:
1986	release_global_resources();
1987
1988out_free_n2cp:
1989	free_n2cp(np);
1990
1991	return err;
1992}
1993
1994static void n2_crypto_remove(struct platform_device *dev)
1995{
1996	struct n2_crypto *np = dev_get_drvdata(&dev->dev);
1997
1998	n2_unregister_algs();
1999
2000	spu_list_destroy(&np->cwq_list);
2001
2002	release_global_resources();
2003
2004	free_n2cp(np);
2005}
2006
2007static struct n2_mau *alloc_ncp(void)
2008{
2009	struct n2_mau *mp = kzalloc(sizeof(struct n2_mau), GFP_KERNEL);
2010
2011	if (mp)
2012		INIT_LIST_HEAD(&mp->mau_list);
2013
2014	return mp;
2015}
2016
2017static void free_ncp(struct n2_mau *mp)
2018{
2019	kfree(mp->mau_info.ino_table);
2020	mp->mau_info.ino_table = NULL;
2021
2022	kfree(mp);
2023}
2024
2025static int n2_mau_probe(struct platform_device *dev)
2026{
2027	struct mdesc_handle *mdesc;
2028	struct n2_mau *mp;
2029	int err;
2030
2031	n2_spu_driver_version();
2032
2033	pr_info("Found NCP at %pOF\n", dev->dev.of_node);
2034
2035	mp = alloc_ncp();
2036	if (!mp) {
2037		dev_err(&dev->dev, "%pOF: Unable to allocate ncp.\n",
2038			dev->dev.of_node);
2039		return -ENOMEM;
2040	}
2041
2042	err = grab_global_resources();
2043	if (err) {
2044		dev_err(&dev->dev, "%pOF: Unable to grab global resources.\n",
2045			dev->dev.of_node);
2046		goto out_free_ncp;
2047	}
2048
2049	mdesc = mdesc_grab();
2050
2051	if (!mdesc) {
2052		dev_err(&dev->dev, "%pOF: Unable to grab MDESC.\n",
2053			dev->dev.of_node);
2054		err = -ENODEV;
2055		goto out_free_global;
2056	}
2057
2058	err = grab_mdesc_irq_props(mdesc, dev, &mp->mau_info, "ncp");
2059	if (err) {
2060		dev_err(&dev->dev, "%pOF: Unable to grab IRQ props.\n",
2061			dev->dev.of_node);
2062		mdesc_release(mdesc);
2063		goto out_free_global;
2064	}
2065
2066	err = spu_mdesc_scan(mdesc, dev, &mp->mau_info, &mp->mau_list,
2067			     "mau", HV_NCS_QTYPE_MAU, mau_intr,
2068			     cpu_to_mau);
2069	mdesc_release(mdesc);
2070
2071	if (err) {
2072		dev_err(&dev->dev, "%pOF: MAU MDESC scan failed.\n",
2073			dev->dev.of_node);
2074		goto out_free_global;
2075	}
2076
2077	dev_set_drvdata(&dev->dev, mp);
2078
2079	return 0;
2080
2081out_free_global:
2082	release_global_resources();
2083
2084out_free_ncp:
2085	free_ncp(mp);
2086
2087	return err;
2088}
2089
2090static void n2_mau_remove(struct platform_device *dev)
2091{
2092	struct n2_mau *mp = dev_get_drvdata(&dev->dev);
2093
2094	spu_list_destroy(&mp->mau_list);
2095
2096	release_global_resources();
2097
2098	free_ncp(mp);
2099}
2100
2101static const struct of_device_id n2_crypto_match[] = {
2102	{
2103		.name = "n2cp",
2104		.compatible = "SUNW,n2-cwq",
2105	},
2106	{
2107		.name = "n2cp",
2108		.compatible = "SUNW,vf-cwq",
2109	},
2110	{
2111		.name = "n2cp",
2112		.compatible = "SUNW,kt-cwq",
2113	},
2114	{},
2115};
2116
2117MODULE_DEVICE_TABLE(of, n2_crypto_match);
2118
2119static struct platform_driver n2_crypto_driver = {
2120	.driver = {
2121		.name		=	"n2cp",
2122		.of_match_table	=	n2_crypto_match,
2123	},
2124	.probe		=	n2_crypto_probe,
2125	.remove_new	=	n2_crypto_remove,
2126};
2127
2128static const struct of_device_id n2_mau_match[] = {
2129	{
2130		.name = "ncp",
2131		.compatible = "SUNW,n2-mau",
2132	},
2133	{
2134		.name = "ncp",
2135		.compatible = "SUNW,vf-mau",
2136	},
2137	{
2138		.name = "ncp",
2139		.compatible = "SUNW,kt-mau",
2140	},
2141	{},
2142};
2143
2144MODULE_DEVICE_TABLE(of, n2_mau_match);
2145
2146static struct platform_driver n2_mau_driver = {
2147	.driver = {
2148		.name		=	"ncp",
2149		.of_match_table	=	n2_mau_match,
2150	},
2151	.probe		=	n2_mau_probe,
2152	.remove_new	=	n2_mau_remove,
2153};
2154
2155static struct platform_driver * const drivers[] = {
2156	&n2_crypto_driver,
2157	&n2_mau_driver,
2158};
2159
2160static int __init n2_init(void)
2161{
2162	return platform_register_drivers(drivers, ARRAY_SIZE(drivers));
2163}
2164
2165static void __exit n2_exit(void)
2166{
2167	platform_unregister_drivers(drivers, ARRAY_SIZE(drivers));
2168}
2169
2170module_init(n2_init);
2171module_exit(n2_exit);