xref: /freebsd-12.1/sys/dev/ubsec/ubsec.c (revision 81e6c8e7)
1 /*	$OpenBSD: ubsec.c,v 1.115 2002/09/24 18:33:26 jason Exp $	*/
2 
3 /*-
4  * Copyright (c) 2000 Jason L. Wright ([email protected])
5  * Copyright (c) 2000 Theo de Raadt ([email protected])
6  * Copyright (c) 2001 Patrik Lindergren ([email protected])
7  *
8  * All rights reserved.
9  *
10  * Redistribution and use in source and binary forms, with or without
11  * modification, are permitted provided that the following conditions
12  * are met:
13  * 1. Redistributions of source code must retain the above copyright
14  *    notice, this list of conditions and the following disclaimer.
15  * 2. Redistributions in binary form must reproduce the above copyright
16  *    notice, this list of conditions and the following disclaimer in the
17  *    documentation and/or other materials provided with the distribution.
18  * 3. All advertising materials mentioning features or use of this software
19  *    must display the following acknowledgement:
20  *	This product includes software developed by Jason L. Wright
21  * 4. The name of the author may not be used to endorse or promote products
22  *    derived from this software without specific prior written permission.
23  *
24  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
25  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
26  * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
27  * DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT,
28  * INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
29  * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
30  * SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
31  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
32  * STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN
33  * ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
34  * POSSIBILITY OF SUCH DAMAGE.
35  *
36  * Effort sponsored in part by the Defense Advanced Research Projects
37  * Agency (DARPA) and Air Force Research Laboratory, Air Force
38  * Materiel Command, USAF, under agreement number F30602-01-2-0537.
39  */
40 
41 #include <sys/cdefs.h>
42 __FBSDID("$FreeBSD$");
43 
44 /*
45  * uBsec 5[56]01, 58xx hardware crypto accelerator
46  */
47 
48 #include "opt_ubsec.h"
49 
50 #include <sys/param.h>
51 #include <sys/systm.h>
52 #include <sys/proc.h>
53 #include <sys/errno.h>
54 #include <sys/malloc.h>
55 #include <sys/kernel.h>
56 #include <sys/module.h>
57 #include <sys/mbuf.h>
58 #include <sys/lock.h>
59 #include <sys/mutex.h>
60 #include <sys/sysctl.h>
61 #include <sys/endian.h>
62 
63 #include <vm/vm.h>
64 #include <vm/pmap.h>
65 
66 #include <machine/bus.h>
67 #include <machine/resource.h>
68 #include <sys/bus.h>
69 #include <sys/rman.h>
70 
71 #include <crypto/sha1.h>
72 #include <opencrypto/cryptodev.h>
73 #include <opencrypto/cryptosoft.h>
74 #include <sys/md5.h>
75 #include <sys/random.h>
76 #include <sys/kobj.h>
77 
78 #include "cryptodev_if.h"
79 
80 #include <dev/pci/pcivar.h>
81 #include <dev/pci/pcireg.h>
82 
83 /* grr, #defines for gratuitous incompatibility in queue.h */
84 #define	SIMPLEQ_HEAD		STAILQ_HEAD
85 #define	SIMPLEQ_ENTRY		STAILQ_ENTRY
86 #define	SIMPLEQ_INIT		STAILQ_INIT
87 #define	SIMPLEQ_INSERT_TAIL	STAILQ_INSERT_TAIL
88 #define	SIMPLEQ_EMPTY		STAILQ_EMPTY
89 #define	SIMPLEQ_FIRST		STAILQ_FIRST
90 #define	SIMPLEQ_REMOVE_HEAD	STAILQ_REMOVE_HEAD
91 #define	SIMPLEQ_FOREACH		STAILQ_FOREACH
92 /* ditto for endian.h */
93 #define	letoh16(x)		le16toh(x)
94 #define	letoh32(x)		le32toh(x)
95 
96 #ifdef UBSEC_RNDTEST
97 #include <dev/rndtest/rndtest.h>
98 #endif
99 #include <dev/ubsec/ubsecreg.h>
100 #include <dev/ubsec/ubsecvar.h>
101 
102 /*
103  * Prototypes and count for the pci_device structure
104  */
105 static	int ubsec_probe(device_t);
106 static	int ubsec_attach(device_t);
107 static	int ubsec_detach(device_t);
108 static	int ubsec_suspend(device_t);
109 static	int ubsec_resume(device_t);
110 static	int ubsec_shutdown(device_t);
111 
112 static	int ubsec_newsession(device_t, u_int32_t *, struct cryptoini *);
113 static	int ubsec_freesession(device_t, u_int64_t);
114 static	int ubsec_process(device_t, struct cryptop *, int);
115 static	int ubsec_kprocess(device_t, struct cryptkop *, int);
116 
117 static device_method_t ubsec_methods[] = {
118 	/* Device interface */
119 	DEVMETHOD(device_probe,		ubsec_probe),
120 	DEVMETHOD(device_attach,	ubsec_attach),
121 	DEVMETHOD(device_detach,	ubsec_detach),
122 	DEVMETHOD(device_suspend,	ubsec_suspend),
123 	DEVMETHOD(device_resume,	ubsec_resume),
124 	DEVMETHOD(device_shutdown,	ubsec_shutdown),
125 
126 	/* crypto device methods */
127 	DEVMETHOD(cryptodev_newsession,	ubsec_newsession),
128 	DEVMETHOD(cryptodev_freesession,ubsec_freesession),
129 	DEVMETHOD(cryptodev_process,	ubsec_process),
130 	DEVMETHOD(cryptodev_kprocess,	ubsec_kprocess),
131 
132 	DEVMETHOD_END
133 };
134 static driver_t ubsec_driver = {
135 	"ubsec",
136 	ubsec_methods,
137 	sizeof (struct ubsec_softc)
138 };
139 static devclass_t ubsec_devclass;
140 
141 DRIVER_MODULE(ubsec, pci, ubsec_driver, ubsec_devclass, 0, 0);
142 MODULE_DEPEND(ubsec, crypto, 1, 1, 1);
143 #ifdef UBSEC_RNDTEST
144 MODULE_DEPEND(ubsec, rndtest, 1, 1, 1);
145 #endif
146 
147 static	void ubsec_intr(void *);
148 static	void ubsec_callback(struct ubsec_softc *, struct ubsec_q *);
149 static	void ubsec_feed(struct ubsec_softc *);
150 static	void ubsec_mcopy(struct mbuf *, struct mbuf *, int, int);
151 static	void ubsec_callback2(struct ubsec_softc *, struct ubsec_q2 *);
152 static	int ubsec_feed2(struct ubsec_softc *);
153 static	void ubsec_rng(void *);
154 static	int ubsec_dma_malloc(struct ubsec_softc *, bus_size_t,
155 			     struct ubsec_dma_alloc *, int);
156 #define	ubsec_dma_sync(_dma, _flags) \
157 	bus_dmamap_sync((_dma)->dma_tag, (_dma)->dma_map, (_flags))
158 static	void ubsec_dma_free(struct ubsec_softc *, struct ubsec_dma_alloc *);
159 static	int ubsec_dmamap_aligned(struct ubsec_operand *op);
160 
161 static	void ubsec_reset_board(struct ubsec_softc *sc);
162 static	void ubsec_init_board(struct ubsec_softc *sc);
163 static	void ubsec_init_pciregs(device_t dev);
164 static	void ubsec_totalreset(struct ubsec_softc *sc);
165 
166 static	int ubsec_free_q(struct ubsec_softc *sc, struct ubsec_q *q);
167 
168 static	int ubsec_kprocess_modexp_hw(struct ubsec_softc *, struct cryptkop *, int);
169 static	int ubsec_kprocess_modexp_sw(struct ubsec_softc *, struct cryptkop *, int);
170 static	int ubsec_kprocess_rsapriv(struct ubsec_softc *, struct cryptkop *, int);
171 static	void ubsec_kfree(struct ubsec_softc *, struct ubsec_q2 *);
172 static	int ubsec_ksigbits(struct crparam *);
173 static	void ubsec_kshift_r(u_int, u_int8_t *, u_int, u_int8_t *, u_int);
174 static	void ubsec_kshift_l(u_int, u_int8_t *, u_int, u_int8_t *, u_int);
175 
176 static SYSCTL_NODE(_hw, OID_AUTO, ubsec, CTLFLAG_RD, 0,
177     "Broadcom driver parameters");
178 
179 #ifdef UBSEC_DEBUG
180 static	void ubsec_dump_pb(volatile struct ubsec_pktbuf *);
181 static	void ubsec_dump_mcr(struct ubsec_mcr *);
182 static	void ubsec_dump_ctx2(struct ubsec_ctx_keyop *);
183 
184 static	int ubsec_debug = 0;
185 SYSCTL_INT(_hw_ubsec, OID_AUTO, debug, CTLFLAG_RW, &ubsec_debug,
186 	    0, "control debugging msgs");
187 #endif
188 
189 #define	READ_REG(sc,r) \
190 	bus_space_read_4((sc)->sc_st, (sc)->sc_sh, (r))
191 
192 #define WRITE_REG(sc,reg,val) \
193 	bus_space_write_4((sc)->sc_st, (sc)->sc_sh, reg, val)
194 
195 #define	SWAP32(x) (x) = htole32(ntohl((x)))
196 #define	HTOLE32(x) (x) = htole32(x)
197 
198 struct ubsec_stats ubsecstats;
199 SYSCTL_STRUCT(_hw_ubsec, OID_AUTO, stats, CTLFLAG_RD, &ubsecstats,
200 	    ubsec_stats, "driver statistics");
201 
202 static int
203 ubsec_probe(device_t dev)
204 {
205 	if (pci_get_vendor(dev) == PCI_VENDOR_SUN &&
206 	    (pci_get_device(dev) == PCI_PRODUCT_SUN_5821 ||
207 	     pci_get_device(dev) == PCI_PRODUCT_SUN_SCA1K))
208 		return (BUS_PROBE_DEFAULT);
209 	if (pci_get_vendor(dev) == PCI_VENDOR_BLUESTEEL &&
210 	    (pci_get_device(dev) == PCI_PRODUCT_BLUESTEEL_5501 ||
211 	     pci_get_device(dev) == PCI_PRODUCT_BLUESTEEL_5601))
212 		return (BUS_PROBE_DEFAULT);
213 	if (pci_get_vendor(dev) == PCI_VENDOR_BROADCOM &&
214 	    (pci_get_device(dev) == PCI_PRODUCT_BROADCOM_5801 ||
215 	     pci_get_device(dev) == PCI_PRODUCT_BROADCOM_5802 ||
216 	     pci_get_device(dev) == PCI_PRODUCT_BROADCOM_5805 ||
217 	     pci_get_device(dev) == PCI_PRODUCT_BROADCOM_5820 ||
218 	     pci_get_device(dev) == PCI_PRODUCT_BROADCOM_5821 ||
219 	     pci_get_device(dev) == PCI_PRODUCT_BROADCOM_5822 ||
220 	     pci_get_device(dev) == PCI_PRODUCT_BROADCOM_5823 ||
221 	     pci_get_device(dev) == PCI_PRODUCT_BROADCOM_5825
222 	     ))
223 		return (BUS_PROBE_DEFAULT);
224 	return (ENXIO);
225 }
226 
227 static const char*
228 ubsec_partname(struct ubsec_softc *sc)
229 {
230 	/* XXX sprintf numbers when not decoded */
231 	switch (pci_get_vendor(sc->sc_dev)) {
232 	case PCI_VENDOR_BROADCOM:
233 		switch (pci_get_device(sc->sc_dev)) {
234 		case PCI_PRODUCT_BROADCOM_5801:	return "Broadcom 5801";
235 		case PCI_PRODUCT_BROADCOM_5802:	return "Broadcom 5802";
236 		case PCI_PRODUCT_BROADCOM_5805:	return "Broadcom 5805";
237 		case PCI_PRODUCT_BROADCOM_5820:	return "Broadcom 5820";
238 		case PCI_PRODUCT_BROADCOM_5821:	return "Broadcom 5821";
239 		case PCI_PRODUCT_BROADCOM_5822:	return "Broadcom 5822";
240 		case PCI_PRODUCT_BROADCOM_5823:	return "Broadcom 5823";
241 		case PCI_PRODUCT_BROADCOM_5825:	return "Broadcom 5825";
242 		}
243 		return "Broadcom unknown-part";
244 	case PCI_VENDOR_BLUESTEEL:
245 		switch (pci_get_device(sc->sc_dev)) {
246 		case PCI_PRODUCT_BLUESTEEL_5601: return "Bluesteel 5601";
247 		}
248 		return "Bluesteel unknown-part";
249 	case PCI_VENDOR_SUN:
250 		switch (pci_get_device(sc->sc_dev)) {
251 		case PCI_PRODUCT_SUN_5821: return "Sun Crypto 5821";
252 		case PCI_PRODUCT_SUN_SCA1K: return "Sun Crypto 1K";
253 		}
254 		return "Sun unknown-part";
255 	}
256 	return "Unknown-vendor unknown-part";
257 }
258 
259 static void
260 default_harvest(struct rndtest_state *rsp, void *buf, u_int count)
261 {
262 	random_harvest(buf, count, count*NBBY/2, RANDOM_PURE_UBSEC);
263 }
264 
265 static int
266 ubsec_attach(device_t dev)
267 {
268 	struct ubsec_softc *sc = device_get_softc(dev);
269 	struct ubsec_dma *dmap;
270 	u_int32_t i;
271 	int rid;
272 
273 	bzero(sc, sizeof (*sc));
274 	sc->sc_dev = dev;
275 
276 	SIMPLEQ_INIT(&sc->sc_queue);
277 	SIMPLEQ_INIT(&sc->sc_qchip);
278 	SIMPLEQ_INIT(&sc->sc_queue2);
279 	SIMPLEQ_INIT(&sc->sc_qchip2);
280 	SIMPLEQ_INIT(&sc->sc_q2free);
281 
282 	/* XXX handle power management */
283 
284 	sc->sc_statmask = BS_STAT_MCR1_DONE | BS_STAT_DMAERR;
285 
286 	if (pci_get_vendor(dev) == PCI_VENDOR_BLUESTEEL &&
287 	    pci_get_device(dev) == PCI_PRODUCT_BLUESTEEL_5601)
288 		sc->sc_flags |= UBS_FLAGS_KEY | UBS_FLAGS_RNG;
289 
290 	if (pci_get_vendor(dev) == PCI_VENDOR_BROADCOM &&
291 	    (pci_get_device(dev) == PCI_PRODUCT_BROADCOM_5802 ||
292 	     pci_get_device(dev) == PCI_PRODUCT_BROADCOM_5805))
293 		sc->sc_flags |= UBS_FLAGS_KEY | UBS_FLAGS_RNG;
294 
295 	if (pci_get_vendor(dev) == PCI_VENDOR_BROADCOM &&
296 	    pci_get_device(dev) == PCI_PRODUCT_BROADCOM_5820)
297 		sc->sc_flags |= UBS_FLAGS_KEY | UBS_FLAGS_RNG |
298 		    UBS_FLAGS_LONGCTX | UBS_FLAGS_HWNORM | UBS_FLAGS_BIGKEY;
299 
300 	if ((pci_get_vendor(dev) == PCI_VENDOR_BROADCOM &&
301 	     (pci_get_device(dev) == PCI_PRODUCT_BROADCOM_5821 ||
302 	      pci_get_device(dev) == PCI_PRODUCT_BROADCOM_5822 ||
303 	      pci_get_device(dev) == PCI_PRODUCT_BROADCOM_5823 ||
304 	      pci_get_device(dev) == PCI_PRODUCT_BROADCOM_5825)) ||
305 	    (pci_get_vendor(dev) == PCI_VENDOR_SUN &&
306 	     (pci_get_device(dev) == PCI_PRODUCT_SUN_SCA1K ||
307 	      pci_get_device(dev) == PCI_PRODUCT_SUN_5821))) {
308 		/* NB: the 5821/5822 defines some additional status bits */
309 		sc->sc_statmask |= BS_STAT_MCR1_ALLEMPTY |
310 		    BS_STAT_MCR2_ALLEMPTY;
311 		sc->sc_flags |= UBS_FLAGS_KEY | UBS_FLAGS_RNG |
312 		    UBS_FLAGS_LONGCTX | UBS_FLAGS_HWNORM | UBS_FLAGS_BIGKEY;
313 	}
314 
315 	pci_enable_busmaster(dev);
316 
317 	/*
318 	 * Setup memory-mapping of PCI registers.
319 	 */
320 	rid = BS_BAR;
321 	sc->sc_sr = bus_alloc_resource_any(dev, SYS_RES_MEMORY, &rid,
322 					   RF_ACTIVE);
323 	if (sc->sc_sr == NULL) {
324 		device_printf(dev, "cannot map register space\n");
325 		goto bad;
326 	}
327 	sc->sc_st = rman_get_bustag(sc->sc_sr);
328 	sc->sc_sh = rman_get_bushandle(sc->sc_sr);
329 
330 	/*
331 	 * Arrange interrupt line.
332 	 */
333 	rid = 0;
334 	sc->sc_irq = bus_alloc_resource_any(dev, SYS_RES_IRQ, &rid,
335 					    RF_SHAREABLE|RF_ACTIVE);
336 	if (sc->sc_irq == NULL) {
337 		device_printf(dev, "could not map interrupt\n");
338 		goto bad1;
339 	}
340 	/*
341 	 * NB: Network code assumes we are blocked with splimp()
342 	 *     so make sure the IRQ is mapped appropriately.
343 	 */
344 	if (bus_setup_intr(dev, sc->sc_irq, INTR_TYPE_NET | INTR_MPSAFE,
345 			   NULL, ubsec_intr, sc, &sc->sc_ih)) {
346 		device_printf(dev, "could not establish interrupt\n");
347 		goto bad2;
348 	}
349 
350 	sc->sc_cid = crypto_get_driverid(dev, CRYPTOCAP_F_HARDWARE);
351 	if (sc->sc_cid < 0) {
352 		device_printf(dev, "could not get crypto driver id\n");
353 		goto bad3;
354 	}
355 
356 	/*
357 	 * Setup DMA descriptor area.
358 	 */
359 	if (bus_dma_tag_create(bus_get_dma_tag(dev),	/* parent */
360 			       1, 0,			/* alignment, bounds */
361 			       BUS_SPACE_MAXADDR_32BIT,	/* lowaddr */
362 			       BUS_SPACE_MAXADDR,	/* highaddr */
363 			       NULL, NULL,		/* filter, filterarg */
364 			       0x3ffff,			/* maxsize */
365 			       UBS_MAX_SCATTER,		/* nsegments */
366 			       0xffff,			/* maxsegsize */
367 			       BUS_DMA_ALLOCNOW,	/* flags */
368 			       NULL, NULL,		/* lockfunc, lockarg */
369 			       &sc->sc_dmat)) {
370 		device_printf(dev, "cannot allocate DMA tag\n");
371 		goto bad4;
372 	}
373 	SIMPLEQ_INIT(&sc->sc_freequeue);
374 	dmap = sc->sc_dmaa;
375 	for (i = 0; i < UBS_MAX_NQUEUE; i++, dmap++) {
376 		struct ubsec_q *q;
377 
378 		q = (struct ubsec_q *)malloc(sizeof(struct ubsec_q),
379 		    M_DEVBUF, M_NOWAIT);
380 		if (q == NULL) {
381 			device_printf(dev, "cannot allocate queue buffers\n");
382 			break;
383 		}
384 
385 		if (ubsec_dma_malloc(sc, sizeof(struct ubsec_dmachunk),
386 		    &dmap->d_alloc, 0)) {
387 			device_printf(dev, "cannot allocate dma buffers\n");
388 			free(q, M_DEVBUF);
389 			break;
390 		}
391 		dmap->d_dma = (struct ubsec_dmachunk *)dmap->d_alloc.dma_vaddr;
392 
393 		q->q_dma = dmap;
394 		sc->sc_queuea[i] = q;
395 
396 		SIMPLEQ_INSERT_TAIL(&sc->sc_freequeue, q, q_next);
397 	}
398 	mtx_init(&sc->sc_mcr1lock, device_get_nameunit(dev),
399 		"mcr1 operations", MTX_DEF);
400 	mtx_init(&sc->sc_freeqlock, device_get_nameunit(dev),
401 		"mcr1 free q", MTX_DEF);
402 
403 	device_printf(sc->sc_dev, "%s\n", ubsec_partname(sc));
404 
405 	crypto_register(sc->sc_cid, CRYPTO_3DES_CBC, 0, 0);
406 	crypto_register(sc->sc_cid, CRYPTO_DES_CBC, 0, 0);
407 	crypto_register(sc->sc_cid, CRYPTO_MD5_HMAC, 0, 0);
408 	crypto_register(sc->sc_cid, CRYPTO_SHA1_HMAC, 0, 0);
409 
410 	/*
411 	 * Reset Broadcom chip
412 	 */
413 	ubsec_reset_board(sc);
414 
415 	/*
416 	 * Init Broadcom specific PCI settings
417 	 */
418 	ubsec_init_pciregs(dev);
419 
420 	/*
421 	 * Init Broadcom chip
422 	 */
423 	ubsec_init_board(sc);
424 
425 #ifndef UBSEC_NO_RNG
426 	if (sc->sc_flags & UBS_FLAGS_RNG) {
427 		sc->sc_statmask |= BS_STAT_MCR2_DONE;
428 #ifdef UBSEC_RNDTEST
429 		sc->sc_rndtest = rndtest_attach(dev);
430 		if (sc->sc_rndtest)
431 			sc->sc_harvest = rndtest_harvest;
432 		else
433 			sc->sc_harvest = default_harvest;
434 #else
435 		sc->sc_harvest = default_harvest;
436 #endif
437 
438 		if (ubsec_dma_malloc(sc, sizeof(struct ubsec_mcr),
439 		    &sc->sc_rng.rng_q.q_mcr, 0))
440 			goto skip_rng;
441 
442 		if (ubsec_dma_malloc(sc, sizeof(struct ubsec_ctx_rngbypass),
443 		    &sc->sc_rng.rng_q.q_ctx, 0)) {
444 			ubsec_dma_free(sc, &sc->sc_rng.rng_q.q_mcr);
445 			goto skip_rng;
446 		}
447 
448 		if (ubsec_dma_malloc(sc, sizeof(u_int32_t) *
449 		    UBSEC_RNG_BUFSIZ, &sc->sc_rng.rng_buf, 0)) {
450 			ubsec_dma_free(sc, &sc->sc_rng.rng_q.q_ctx);
451 			ubsec_dma_free(sc, &sc->sc_rng.rng_q.q_mcr);
452 			goto skip_rng;
453 		}
454 
455 		if (hz >= 100)
456 			sc->sc_rnghz = hz / 100;
457 		else
458 			sc->sc_rnghz = 1;
459 		callout_init(&sc->sc_rngto, CALLOUT_MPSAFE);
460 		callout_reset(&sc->sc_rngto, sc->sc_rnghz, ubsec_rng, sc);
461 skip_rng:
462 	;
463 	}
464 #endif /* UBSEC_NO_RNG */
465 	mtx_init(&sc->sc_mcr2lock, device_get_nameunit(dev),
466 		"mcr2 operations", MTX_DEF);
467 
468 	if (sc->sc_flags & UBS_FLAGS_KEY) {
469 		sc->sc_statmask |= BS_STAT_MCR2_DONE;
470 
471 		crypto_kregister(sc->sc_cid, CRK_MOD_EXP, 0);
472 #if 0
473 		crypto_kregister(sc->sc_cid, CRK_MOD_EXP_CRT, 0);
474 #endif
475 	}
476 	return (0);
477 bad4:
478 	crypto_unregister_all(sc->sc_cid);
479 bad3:
480 	bus_teardown_intr(dev, sc->sc_irq, sc->sc_ih);
481 bad2:
482 	bus_release_resource(dev, SYS_RES_IRQ, 0, sc->sc_irq);
483 bad1:
484 	bus_release_resource(dev, SYS_RES_MEMORY, BS_BAR, sc->sc_sr);
485 bad:
486 	return (ENXIO);
487 }
488 
489 /*
490  * Detach a device that successfully probed.
491  */
492 static int
493 ubsec_detach(device_t dev)
494 {
495 	struct ubsec_softc *sc = device_get_softc(dev);
496 
497 	/* XXX wait/abort active ops */
498 
499 	/* disable interrupts */
500 	WRITE_REG(sc, BS_CTRL, READ_REG(sc, BS_CTRL) &~
501 		(BS_CTRL_MCR2INT | BS_CTRL_MCR1INT | BS_CTRL_DMAERR));
502 
503 	callout_stop(&sc->sc_rngto);
504 
505 	crypto_unregister_all(sc->sc_cid);
506 
507 #ifdef UBSEC_RNDTEST
508 	if (sc->sc_rndtest)
509 		rndtest_detach(sc->sc_rndtest);
510 #endif
511 
512 	while (!SIMPLEQ_EMPTY(&sc->sc_freequeue)) {
513 		struct ubsec_q *q;
514 
515 		q = SIMPLEQ_FIRST(&sc->sc_freequeue);
516 		SIMPLEQ_REMOVE_HEAD(&sc->sc_freequeue, q_next);
517 		ubsec_dma_free(sc, &q->q_dma->d_alloc);
518 		free(q, M_DEVBUF);
519 	}
520 	mtx_destroy(&sc->sc_mcr1lock);
521 	mtx_destroy(&sc->sc_freeqlock);
522 #ifndef UBSEC_NO_RNG
523 	if (sc->sc_flags & UBS_FLAGS_RNG) {
524 		ubsec_dma_free(sc, &sc->sc_rng.rng_q.q_mcr);
525 		ubsec_dma_free(sc, &sc->sc_rng.rng_q.q_ctx);
526 		ubsec_dma_free(sc, &sc->sc_rng.rng_buf);
527 	}
528 #endif /* UBSEC_NO_RNG */
529 	mtx_destroy(&sc->sc_mcr2lock);
530 
531 	bus_generic_detach(dev);
532 	bus_teardown_intr(dev, sc->sc_irq, sc->sc_ih);
533 	bus_release_resource(dev, SYS_RES_IRQ, 0, sc->sc_irq);
534 
535 	bus_dma_tag_destroy(sc->sc_dmat);
536 	bus_release_resource(dev, SYS_RES_MEMORY, BS_BAR, sc->sc_sr);
537 
538 	return (0);
539 }
540 
541 /*
542  * Stop all chip i/o so that the kernel's probe routines don't
543  * get confused by errant DMAs when rebooting.
544  */
545 static int
546 ubsec_shutdown(device_t dev)
547 {
548 #ifdef notyet
549 	ubsec_stop(device_get_softc(dev));
550 #endif
551 	return (0);
552 }
553 
554 /*
555  * Device suspend routine.
556  */
557 static int
558 ubsec_suspend(device_t dev)
559 {
560 	struct ubsec_softc *sc = device_get_softc(dev);
561 
562 #ifdef notyet
563 	/* XXX stop the device and save PCI settings */
564 #endif
565 	sc->sc_suspended = 1;
566 
567 	return (0);
568 }
569 
570 static int
571 ubsec_resume(device_t dev)
572 {
573 	struct ubsec_softc *sc = device_get_softc(dev);
574 
575 #ifdef notyet
576 	/* XXX retore PCI settings and start the device */
577 #endif
578 	sc->sc_suspended = 0;
579 	return (0);
580 }
581 
582 /*
583  * UBSEC Interrupt routine
584  */
585 static void
586 ubsec_intr(void *arg)
587 {
588 	struct ubsec_softc *sc = arg;
589 	volatile u_int32_t stat;
590 	struct ubsec_q *q;
591 	struct ubsec_dma *dmap;
592 	int npkts = 0, i;
593 
594 	stat = READ_REG(sc, BS_STAT);
595 	stat &= sc->sc_statmask;
596 	if (stat == 0)
597 		return;
598 
599 	WRITE_REG(sc, BS_STAT, stat);		/* IACK */
600 
601 	/*
602 	 * Check to see if we have any packets waiting for us
603 	 */
604 	if ((stat & BS_STAT_MCR1_DONE)) {
605 		mtx_lock(&sc->sc_mcr1lock);
606 		while (!SIMPLEQ_EMPTY(&sc->sc_qchip)) {
607 			q = SIMPLEQ_FIRST(&sc->sc_qchip);
608 			dmap = q->q_dma;
609 
610 			if ((dmap->d_dma->d_mcr.mcr_flags & htole16(UBS_MCR_DONE)) == 0)
611 				break;
612 
613 			SIMPLEQ_REMOVE_HEAD(&sc->sc_qchip, q_next);
614 
615 			npkts = q->q_nstacked_mcrs;
616 			sc->sc_nqchip -= 1+npkts;
617 			/*
618 			 * search for further sc_qchip ubsec_q's that share
619 			 * the same MCR, and complete them too, they must be
620 			 * at the top.
621 			 */
622 			for (i = 0; i < npkts; i++) {
623 				if(q->q_stacked_mcr[i]) {
624 					ubsec_callback(sc, q->q_stacked_mcr[i]);
625 				} else {
626 					break;
627 				}
628 			}
629 			ubsec_callback(sc, q);
630 		}
631 		/*
632 		 * Don't send any more packet to chip if there has been
633 		 * a DMAERR.
634 		 */
635 		if (!(stat & BS_STAT_DMAERR))
636 			ubsec_feed(sc);
637 		mtx_unlock(&sc->sc_mcr1lock);
638 	}
639 
640 	/*
641 	 * Check to see if we have any key setups/rng's waiting for us
642 	 */
643 	if ((sc->sc_flags & (UBS_FLAGS_KEY|UBS_FLAGS_RNG)) &&
644 	    (stat & BS_STAT_MCR2_DONE)) {
645 		struct ubsec_q2 *q2;
646 		struct ubsec_mcr *mcr;
647 
648 		mtx_lock(&sc->sc_mcr2lock);
649 		while (!SIMPLEQ_EMPTY(&sc->sc_qchip2)) {
650 			q2 = SIMPLEQ_FIRST(&sc->sc_qchip2);
651 
652 			ubsec_dma_sync(&q2->q_mcr,
653 			    BUS_DMASYNC_POSTREAD|BUS_DMASYNC_POSTWRITE);
654 
655 			mcr = (struct ubsec_mcr *)q2->q_mcr.dma_vaddr;
656 			if ((mcr->mcr_flags & htole16(UBS_MCR_DONE)) == 0) {
657 				ubsec_dma_sync(&q2->q_mcr,
658 				    BUS_DMASYNC_PREREAD|BUS_DMASYNC_PREWRITE);
659 				break;
660 			}
661 			SIMPLEQ_REMOVE_HEAD(&sc->sc_qchip2, q_next);
662 			ubsec_callback2(sc, q2);
663 			/*
664 			 * Don't send any more packet to chip if there has been
665 			 * a DMAERR.
666 			 */
667 			if (!(stat & BS_STAT_DMAERR))
668 				ubsec_feed2(sc);
669 		}
670 		mtx_unlock(&sc->sc_mcr2lock);
671 	}
672 
673 	/*
674 	 * Check to see if we got any DMA Error
675 	 */
676 	if (stat & BS_STAT_DMAERR) {
677 #ifdef UBSEC_DEBUG
678 		if (ubsec_debug) {
679 			volatile u_int32_t a = READ_REG(sc, BS_ERR);
680 
681 			printf("dmaerr %s@%08x\n",
682 			    (a & BS_ERR_READ) ? "read" : "write",
683 			    a & BS_ERR_ADDR);
684 		}
685 #endif /* UBSEC_DEBUG */
686 		ubsecstats.hst_dmaerr++;
687 		mtx_lock(&sc->sc_mcr1lock);
688 		ubsec_totalreset(sc);
689 		ubsec_feed(sc);
690 		mtx_unlock(&sc->sc_mcr1lock);
691 	}
692 
693 	if (sc->sc_needwakeup) {		/* XXX check high watermark */
694 		int wakeup;
695 
696 		mtx_lock(&sc->sc_freeqlock);
697 		wakeup = sc->sc_needwakeup & (CRYPTO_SYMQ|CRYPTO_ASYMQ);
698 #ifdef UBSEC_DEBUG
699 		if (ubsec_debug)
700 			device_printf(sc->sc_dev, "wakeup crypto (%x)\n",
701 				sc->sc_needwakeup);
702 #endif /* UBSEC_DEBUG */
703 		sc->sc_needwakeup &= ~wakeup;
704 		mtx_unlock(&sc->sc_freeqlock);
705 		crypto_unblock(sc->sc_cid, wakeup);
706 	}
707 }
708 
709 /*
710  * ubsec_feed() - aggregate and post requests to chip
711  */
712 static void
713 ubsec_feed(struct ubsec_softc *sc)
714 {
715 	struct ubsec_q *q, *q2;
716 	int npkts, i;
717 	void *v;
718 	u_int32_t stat;
719 
720 	/*
721 	 * Decide how many ops to combine in a single MCR.  We cannot
722 	 * aggregate more than UBS_MAX_AGGR because this is the number
723 	 * of slots defined in the data structure.  Note that
724 	 * aggregation only happens if ops are marked batch'able.
725 	 * Aggregating ops reduces the number of interrupts to the host
726 	 * but also (potentially) increases the latency for processing
727 	 * completed ops as we only get an interrupt when all aggregated
728 	 * ops have completed.
729 	 */
730 	if (sc->sc_nqueue == 0)
731 		return;
732 	if (sc->sc_nqueue > 1) {
733 		npkts = 0;
734 		SIMPLEQ_FOREACH(q, &sc->sc_queue, q_next) {
735 			npkts++;
736 			if ((q->q_crp->crp_flags & CRYPTO_F_BATCH) == 0)
737 				break;
738 		}
739 	} else
740 		npkts = 1;
741 	/*
742 	 * Check device status before going any further.
743 	 */
744 	if ((stat = READ_REG(sc, BS_STAT)) & (BS_STAT_MCR1_FULL | BS_STAT_DMAERR)) {
745 		if (stat & BS_STAT_DMAERR) {
746 			ubsec_totalreset(sc);
747 			ubsecstats.hst_dmaerr++;
748 		} else
749 			ubsecstats.hst_mcr1full++;
750 		return;
751 	}
752 	if (sc->sc_nqueue > ubsecstats.hst_maxqueue)
753 		ubsecstats.hst_maxqueue = sc->sc_nqueue;
754 	if (npkts > UBS_MAX_AGGR)
755 		npkts = UBS_MAX_AGGR;
756 	if (npkts < 2)				/* special case 1 op */
757 		goto feed1;
758 
759 	ubsecstats.hst_totbatch += npkts-1;
760 #ifdef UBSEC_DEBUG
761 	if (ubsec_debug)
762 		printf("merging %d records\n", npkts);
763 #endif /* UBSEC_DEBUG */
764 
765 	q = SIMPLEQ_FIRST(&sc->sc_queue);
766 	SIMPLEQ_REMOVE_HEAD(&sc->sc_queue, q_next);
767 	--sc->sc_nqueue;
768 
769 	bus_dmamap_sync(sc->sc_dmat, q->q_src_map, BUS_DMASYNC_PREWRITE);
770 	if (q->q_dst_map != NULL)
771 		bus_dmamap_sync(sc->sc_dmat, q->q_dst_map, BUS_DMASYNC_PREREAD);
772 
773 	q->q_nstacked_mcrs = npkts - 1;		/* Number of packets stacked */
774 
775 	for (i = 0; i < q->q_nstacked_mcrs; i++) {
776 		q2 = SIMPLEQ_FIRST(&sc->sc_queue);
777 		bus_dmamap_sync(sc->sc_dmat, q2->q_src_map,
778 		    BUS_DMASYNC_PREWRITE);
779 		if (q2->q_dst_map != NULL)
780 			bus_dmamap_sync(sc->sc_dmat, q2->q_dst_map,
781 			    BUS_DMASYNC_PREREAD);
782 		SIMPLEQ_REMOVE_HEAD(&sc->sc_queue, q_next);
783 		--sc->sc_nqueue;
784 
785 		v = (void*)(((char *)&q2->q_dma->d_dma->d_mcr) + sizeof(struct ubsec_mcr) -
786 		    sizeof(struct ubsec_mcr_add));
787 		bcopy(v, &q->q_dma->d_dma->d_mcradd[i], sizeof(struct ubsec_mcr_add));
788 		q->q_stacked_mcr[i] = q2;
789 	}
790 	q->q_dma->d_dma->d_mcr.mcr_pkts = htole16(npkts);
791 	SIMPLEQ_INSERT_TAIL(&sc->sc_qchip, q, q_next);
792 	sc->sc_nqchip += npkts;
793 	if (sc->sc_nqchip > ubsecstats.hst_maxqchip)
794 		ubsecstats.hst_maxqchip = sc->sc_nqchip;
795 	ubsec_dma_sync(&q->q_dma->d_alloc,
796 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
797 	WRITE_REG(sc, BS_MCR1, q->q_dma->d_alloc.dma_paddr +
798 	    offsetof(struct ubsec_dmachunk, d_mcr));
799 	return;
800 feed1:
801 	q = SIMPLEQ_FIRST(&sc->sc_queue);
802 
803 	bus_dmamap_sync(sc->sc_dmat, q->q_src_map, BUS_DMASYNC_PREWRITE);
804 	if (q->q_dst_map != NULL)
805 		bus_dmamap_sync(sc->sc_dmat, q->q_dst_map, BUS_DMASYNC_PREREAD);
806 	ubsec_dma_sync(&q->q_dma->d_alloc,
807 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
808 
809 	WRITE_REG(sc, BS_MCR1, q->q_dma->d_alloc.dma_paddr +
810 	    offsetof(struct ubsec_dmachunk, d_mcr));
811 #ifdef UBSEC_DEBUG
812 	if (ubsec_debug)
813 		printf("feed1: q->chip %p %08x stat %08x\n",
814 		      q, (u_int32_t)vtophys(&q->q_dma->d_dma->d_mcr),
815 		      stat);
816 #endif /* UBSEC_DEBUG */
817 	SIMPLEQ_REMOVE_HEAD(&sc->sc_queue, q_next);
818 	--sc->sc_nqueue;
819 	SIMPLEQ_INSERT_TAIL(&sc->sc_qchip, q, q_next);
820 	sc->sc_nqchip++;
821 	if (sc->sc_nqchip > ubsecstats.hst_maxqchip)
822 		ubsecstats.hst_maxqchip = sc->sc_nqchip;
823 	return;
824 }
825 
826 static void
827 ubsec_setup_enckey(struct ubsec_session *ses, int algo, caddr_t key)
828 {
829 
830 	/* Go ahead and compute key in ubsec's byte order */
831 	if (algo == CRYPTO_DES_CBC) {
832 		bcopy(key, &ses->ses_deskey[0], 8);
833 		bcopy(key, &ses->ses_deskey[2], 8);
834 		bcopy(key, &ses->ses_deskey[4], 8);
835 	} else
836 		bcopy(key, ses->ses_deskey, 24);
837 
838 	SWAP32(ses->ses_deskey[0]);
839 	SWAP32(ses->ses_deskey[1]);
840 	SWAP32(ses->ses_deskey[2]);
841 	SWAP32(ses->ses_deskey[3]);
842 	SWAP32(ses->ses_deskey[4]);
843 	SWAP32(ses->ses_deskey[5]);
844 }
845 
846 static void
847 ubsec_setup_mackey(struct ubsec_session *ses, int algo, caddr_t key, int klen)
848 {
849 	MD5_CTX md5ctx;
850 	SHA1_CTX sha1ctx;
851 	int i;
852 
853 	for (i = 0; i < klen; i++)
854 		key[i] ^= HMAC_IPAD_VAL;
855 
856 	if (algo == CRYPTO_MD5_HMAC) {
857 		MD5Init(&md5ctx);
858 		MD5Update(&md5ctx, key, klen);
859 		MD5Update(&md5ctx, hmac_ipad_buffer, MD5_HMAC_BLOCK_LEN - klen);
860 		bcopy(md5ctx.state, ses->ses_hminner, sizeof(md5ctx.state));
861 	} else {
862 		SHA1Init(&sha1ctx);
863 		SHA1Update(&sha1ctx, key, klen);
864 		SHA1Update(&sha1ctx, hmac_ipad_buffer,
865 		    SHA1_HMAC_BLOCK_LEN - klen);
866 		bcopy(sha1ctx.h.b32, ses->ses_hminner, sizeof(sha1ctx.h.b32));
867 	}
868 
869 	for (i = 0; i < klen; i++)
870 		key[i] ^= (HMAC_IPAD_VAL ^ HMAC_OPAD_VAL);
871 
872 	if (algo == CRYPTO_MD5_HMAC) {
873 		MD5Init(&md5ctx);
874 		MD5Update(&md5ctx, key, klen);
875 		MD5Update(&md5ctx, hmac_opad_buffer, MD5_HMAC_BLOCK_LEN - klen);
876 		bcopy(md5ctx.state, ses->ses_hmouter, sizeof(md5ctx.state));
877 	} else {
878 		SHA1Init(&sha1ctx);
879 		SHA1Update(&sha1ctx, key, klen);
880 		SHA1Update(&sha1ctx, hmac_opad_buffer,
881 		    SHA1_HMAC_BLOCK_LEN - klen);
882 		bcopy(sha1ctx.h.b32, ses->ses_hmouter, sizeof(sha1ctx.h.b32));
883 	}
884 
885 	for (i = 0; i < klen; i++)
886 		key[i] ^= HMAC_OPAD_VAL;
887 }
888 
889 /*
890  * Allocate a new 'session' and return an encoded session id.  'sidp'
891  * contains our registration id, and should contain an encoded session
892  * id on successful allocation.
893  */
894 static int
895 ubsec_newsession(device_t dev, u_int32_t *sidp, struct cryptoini *cri)
896 {
897 	struct ubsec_softc *sc = device_get_softc(dev);
898 	struct cryptoini *c, *encini = NULL, *macini = NULL;
899 	struct ubsec_session *ses = NULL;
900 	int sesn;
901 
902 	if (sidp == NULL || cri == NULL || sc == NULL)
903 		return (EINVAL);
904 
905 	for (c = cri; c != NULL; c = c->cri_next) {
906 		if (c->cri_alg == CRYPTO_MD5_HMAC ||
907 		    c->cri_alg == CRYPTO_SHA1_HMAC) {
908 			if (macini)
909 				return (EINVAL);
910 			macini = c;
911 		} else if (c->cri_alg == CRYPTO_DES_CBC ||
912 		    c->cri_alg == CRYPTO_3DES_CBC) {
913 			if (encini)
914 				return (EINVAL);
915 			encini = c;
916 		} else
917 			return (EINVAL);
918 	}
919 	if (encini == NULL && macini == NULL)
920 		return (EINVAL);
921 
922 	if (sc->sc_sessions == NULL) {
923 		ses = sc->sc_sessions = (struct ubsec_session *)malloc(
924 		    sizeof(struct ubsec_session), M_DEVBUF, M_NOWAIT);
925 		if (ses == NULL)
926 			return (ENOMEM);
927 		sesn = 0;
928 		sc->sc_nsessions = 1;
929 	} else {
930 		for (sesn = 0; sesn < sc->sc_nsessions; sesn++) {
931 			if (sc->sc_sessions[sesn].ses_used == 0) {
932 				ses = &sc->sc_sessions[sesn];
933 				break;
934 			}
935 		}
936 
937 		if (ses == NULL) {
938 			sesn = sc->sc_nsessions;
939 			ses = (struct ubsec_session *)malloc((sesn + 1) *
940 			    sizeof(struct ubsec_session), M_DEVBUF, M_NOWAIT);
941 			if (ses == NULL)
942 				return (ENOMEM);
943 			bcopy(sc->sc_sessions, ses, sesn *
944 			    sizeof(struct ubsec_session));
945 			bzero(sc->sc_sessions, sesn *
946 			    sizeof(struct ubsec_session));
947 			free(sc->sc_sessions, M_DEVBUF);
948 			sc->sc_sessions = ses;
949 			ses = &sc->sc_sessions[sesn];
950 			sc->sc_nsessions++;
951 		}
952 	}
953 	bzero(ses, sizeof(struct ubsec_session));
954 	ses->ses_used = 1;
955 
956 	if (encini) {
957 		/* get an IV, network byte order */
958 		/* XXX may read fewer than requested */
959 		read_random(ses->ses_iv, sizeof(ses->ses_iv));
960 
961 		if (encini->cri_key != NULL) {
962 			ubsec_setup_enckey(ses, encini->cri_alg,
963 			    encini->cri_key);
964 		}
965 	}
966 
967 	if (macini) {
968 		ses->ses_mlen = macini->cri_mlen;
969 		if (ses->ses_mlen == 0) {
970 			if (macini->cri_alg == CRYPTO_MD5_HMAC)
971 				ses->ses_mlen = MD5_HASH_LEN;
972 			else
973 				ses->ses_mlen = SHA1_HASH_LEN;
974 		}
975 
976 		if (macini->cri_key != NULL) {
977 			ubsec_setup_mackey(ses, macini->cri_alg,
978 			    macini->cri_key, macini->cri_klen / 8);
979 		}
980 	}
981 
982 	*sidp = UBSEC_SID(device_get_unit(sc->sc_dev), sesn);
983 	return (0);
984 }
985 
986 /*
987  * Deallocate a session.
988  */
989 static int
990 ubsec_freesession(device_t dev, u_int64_t tid)
991 {
992 	struct ubsec_softc *sc = device_get_softc(dev);
993 	int session, ret;
994 	u_int32_t sid = CRYPTO_SESID2LID(tid);
995 
996 	if (sc == NULL)
997 		return (EINVAL);
998 
999 	session = UBSEC_SESSION(sid);
1000 	if (session < sc->sc_nsessions) {
1001 		bzero(&sc->sc_sessions[session],
1002 			sizeof(sc->sc_sessions[session]));
1003 		ret = 0;
1004 	} else
1005 		ret = EINVAL;
1006 
1007 	return (ret);
1008 }
1009 
1010 static void
1011 ubsec_op_cb(void *arg, bus_dma_segment_t *seg, int nsegs, bus_size_t mapsize, int error)
1012 {
1013 	struct ubsec_operand *op = arg;
1014 
1015 	KASSERT(nsegs <= UBS_MAX_SCATTER,
1016 		("Too many DMA segments returned when mapping operand"));
1017 #ifdef UBSEC_DEBUG
1018 	if (ubsec_debug)
1019 		printf("ubsec_op_cb: mapsize %u nsegs %d error %d\n",
1020 			(u_int) mapsize, nsegs, error);
1021 #endif
1022 	if (error != 0)
1023 		return;
1024 	op->mapsize = mapsize;
1025 	op->nsegs = nsegs;
1026 	bcopy(seg, op->segs, nsegs * sizeof (seg[0]));
1027 }
1028 
1029 static int
1030 ubsec_process(device_t dev, struct cryptop *crp, int hint)
1031 {
1032 	struct ubsec_softc *sc = device_get_softc(dev);
1033 	struct ubsec_q *q = NULL;
1034 	int err = 0, i, j, nicealign;
1035 	struct cryptodesc *crd1, *crd2, *maccrd, *enccrd;
1036 	int encoffset = 0, macoffset = 0, cpskip, cpoffset;
1037 	int sskip, dskip, stheend, dtheend;
1038 	int16_t coffset;
1039 	struct ubsec_session *ses;
1040 	struct ubsec_pktctx ctx;
1041 	struct ubsec_dma *dmap = NULL;
1042 
1043 	if (crp == NULL || crp->crp_callback == NULL || sc == NULL) {
1044 		ubsecstats.hst_invalid++;
1045 		return (EINVAL);
1046 	}
1047 	if (UBSEC_SESSION(crp->crp_sid) >= sc->sc_nsessions) {
1048 		ubsecstats.hst_badsession++;
1049 		return (EINVAL);
1050 	}
1051 
1052 	mtx_lock(&sc->sc_freeqlock);
1053 	if (SIMPLEQ_EMPTY(&sc->sc_freequeue)) {
1054 		ubsecstats.hst_queuefull++;
1055 		sc->sc_needwakeup |= CRYPTO_SYMQ;
1056 		mtx_unlock(&sc->sc_freeqlock);
1057 		return (ERESTART);
1058 	}
1059 	q = SIMPLEQ_FIRST(&sc->sc_freequeue);
1060 	SIMPLEQ_REMOVE_HEAD(&sc->sc_freequeue, q_next);
1061 	mtx_unlock(&sc->sc_freeqlock);
1062 
1063 	dmap = q->q_dma; /* Save dma pointer */
1064 	bzero(q, sizeof(struct ubsec_q));
1065 	bzero(&ctx, sizeof(ctx));
1066 
1067 	q->q_sesn = UBSEC_SESSION(crp->crp_sid);
1068 	q->q_dma = dmap;
1069 	ses = &sc->sc_sessions[q->q_sesn];
1070 
1071 	if (crp->crp_flags & CRYPTO_F_IMBUF) {
1072 		q->q_src_m = (struct mbuf *)crp->crp_buf;
1073 		q->q_dst_m = (struct mbuf *)crp->crp_buf;
1074 	} else if (crp->crp_flags & CRYPTO_F_IOV) {
1075 		q->q_src_io = (struct uio *)crp->crp_buf;
1076 		q->q_dst_io = (struct uio *)crp->crp_buf;
1077 	} else {
1078 		ubsecstats.hst_badflags++;
1079 		err = EINVAL;
1080 		goto errout;	/* XXX we don't handle contiguous blocks! */
1081 	}
1082 
1083 	bzero(&dmap->d_dma->d_mcr, sizeof(struct ubsec_mcr));
1084 
1085 	dmap->d_dma->d_mcr.mcr_pkts = htole16(1);
1086 	dmap->d_dma->d_mcr.mcr_flags = 0;
1087 	q->q_crp = crp;
1088 
1089 	crd1 = crp->crp_desc;
1090 	if (crd1 == NULL) {
1091 		ubsecstats.hst_nodesc++;
1092 		err = EINVAL;
1093 		goto errout;
1094 	}
1095 	crd2 = crd1->crd_next;
1096 
1097 	if (crd2 == NULL) {
1098 		if (crd1->crd_alg == CRYPTO_MD5_HMAC ||
1099 		    crd1->crd_alg == CRYPTO_SHA1_HMAC) {
1100 			maccrd = crd1;
1101 			enccrd = NULL;
1102 		} else if (crd1->crd_alg == CRYPTO_DES_CBC ||
1103 		    crd1->crd_alg == CRYPTO_3DES_CBC) {
1104 			maccrd = NULL;
1105 			enccrd = crd1;
1106 		} else {
1107 			ubsecstats.hst_badalg++;
1108 			err = EINVAL;
1109 			goto errout;
1110 		}
1111 	} else {
1112 		if ((crd1->crd_alg == CRYPTO_MD5_HMAC ||
1113 		    crd1->crd_alg == CRYPTO_SHA1_HMAC) &&
1114 		    (crd2->crd_alg == CRYPTO_DES_CBC ||
1115 			crd2->crd_alg == CRYPTO_3DES_CBC) &&
1116 		    ((crd2->crd_flags & CRD_F_ENCRYPT) == 0)) {
1117 			maccrd = crd1;
1118 			enccrd = crd2;
1119 		} else if ((crd1->crd_alg == CRYPTO_DES_CBC ||
1120 		    crd1->crd_alg == CRYPTO_3DES_CBC) &&
1121 		    (crd2->crd_alg == CRYPTO_MD5_HMAC ||
1122 			crd2->crd_alg == CRYPTO_SHA1_HMAC) &&
1123 		    (crd1->crd_flags & CRD_F_ENCRYPT)) {
1124 			enccrd = crd1;
1125 			maccrd = crd2;
1126 		} else {
1127 			/*
1128 			 * We cannot order the ubsec as requested
1129 			 */
1130 			ubsecstats.hst_badalg++;
1131 			err = EINVAL;
1132 			goto errout;
1133 		}
1134 	}
1135 
1136 	if (enccrd) {
1137 		if (enccrd->crd_flags & CRD_F_KEY_EXPLICIT) {
1138 			ubsec_setup_enckey(ses, enccrd->crd_alg,
1139 			    enccrd->crd_key);
1140 		}
1141 
1142 		encoffset = enccrd->crd_skip;
1143 		ctx.pc_flags |= htole16(UBS_PKTCTX_ENC_3DES);
1144 
1145 		if (enccrd->crd_flags & CRD_F_ENCRYPT) {
1146 			q->q_flags |= UBSEC_QFLAGS_COPYOUTIV;
1147 
1148 			if (enccrd->crd_flags & CRD_F_IV_EXPLICIT)
1149 				bcopy(enccrd->crd_iv, ctx.pc_iv, 8);
1150 			else {
1151 				ctx.pc_iv[0] = ses->ses_iv[0];
1152 				ctx.pc_iv[1] = ses->ses_iv[1];
1153 			}
1154 
1155 			if ((enccrd->crd_flags & CRD_F_IV_PRESENT) == 0) {
1156 				crypto_copyback(crp->crp_flags, crp->crp_buf,
1157 				    enccrd->crd_inject, 8, (caddr_t)ctx.pc_iv);
1158 			}
1159 		} else {
1160 			ctx.pc_flags |= htole16(UBS_PKTCTX_INBOUND);
1161 
1162 			if (enccrd->crd_flags & CRD_F_IV_EXPLICIT)
1163 				bcopy(enccrd->crd_iv, ctx.pc_iv, 8);
1164 			else {
1165 				crypto_copydata(crp->crp_flags, crp->crp_buf,
1166 				    enccrd->crd_inject, 8, (caddr_t)ctx.pc_iv);
1167 			}
1168 		}
1169 
1170 		ctx.pc_deskey[0] = ses->ses_deskey[0];
1171 		ctx.pc_deskey[1] = ses->ses_deskey[1];
1172 		ctx.pc_deskey[2] = ses->ses_deskey[2];
1173 		ctx.pc_deskey[3] = ses->ses_deskey[3];
1174 		ctx.pc_deskey[4] = ses->ses_deskey[4];
1175 		ctx.pc_deskey[5] = ses->ses_deskey[5];
1176 		SWAP32(ctx.pc_iv[0]);
1177 		SWAP32(ctx.pc_iv[1]);
1178 	}
1179 
1180 	if (maccrd) {
1181 		if (maccrd->crd_flags & CRD_F_KEY_EXPLICIT) {
1182 			ubsec_setup_mackey(ses, maccrd->crd_alg,
1183 			    maccrd->crd_key, maccrd->crd_klen / 8);
1184 		}
1185 
1186 		macoffset = maccrd->crd_skip;
1187 
1188 		if (maccrd->crd_alg == CRYPTO_MD5_HMAC)
1189 			ctx.pc_flags |= htole16(UBS_PKTCTX_AUTH_MD5);
1190 		else
1191 			ctx.pc_flags |= htole16(UBS_PKTCTX_AUTH_SHA1);
1192 
1193 		for (i = 0; i < 5; i++) {
1194 			ctx.pc_hminner[i] = ses->ses_hminner[i];
1195 			ctx.pc_hmouter[i] = ses->ses_hmouter[i];
1196 
1197 			HTOLE32(ctx.pc_hminner[i]);
1198 			HTOLE32(ctx.pc_hmouter[i]);
1199 		}
1200 	}
1201 
1202 	if (enccrd && maccrd) {
1203 		/*
1204 		 * ubsec cannot handle packets where the end of encryption
1205 		 * and authentication are not the same, or where the
1206 		 * encrypted part begins before the authenticated part.
1207 		 */
1208 		if ((encoffset + enccrd->crd_len) !=
1209 		    (macoffset + maccrd->crd_len)) {
1210 			ubsecstats.hst_lenmismatch++;
1211 			err = EINVAL;
1212 			goto errout;
1213 		}
1214 		if (enccrd->crd_skip < maccrd->crd_skip) {
1215 			ubsecstats.hst_skipmismatch++;
1216 			err = EINVAL;
1217 			goto errout;
1218 		}
1219 		sskip = maccrd->crd_skip;
1220 		cpskip = dskip = enccrd->crd_skip;
1221 		stheend = maccrd->crd_len;
1222 		dtheend = enccrd->crd_len;
1223 		coffset = enccrd->crd_skip - maccrd->crd_skip;
1224 		cpoffset = cpskip + dtheend;
1225 #ifdef UBSEC_DEBUG
1226 		if (ubsec_debug) {
1227 			printf("mac: skip %d, len %d, inject %d\n",
1228 			    maccrd->crd_skip, maccrd->crd_len, maccrd->crd_inject);
1229 			printf("enc: skip %d, len %d, inject %d\n",
1230 			    enccrd->crd_skip, enccrd->crd_len, enccrd->crd_inject);
1231 			printf("src: skip %d, len %d\n", sskip, stheend);
1232 			printf("dst: skip %d, len %d\n", dskip, dtheend);
1233 			printf("ubs: coffset %d, pktlen %d, cpskip %d, cpoffset %d\n",
1234 			    coffset, stheend, cpskip, cpoffset);
1235 		}
1236 #endif
1237 	} else {
1238 		cpskip = dskip = sskip = macoffset + encoffset;
1239 		dtheend = stheend = (enccrd)?enccrd->crd_len:maccrd->crd_len;
1240 		cpoffset = cpskip + dtheend;
1241 		coffset = 0;
1242 	}
1243 	ctx.pc_offset = htole16(coffset >> 2);
1244 
1245 	if (bus_dmamap_create(sc->sc_dmat, BUS_DMA_NOWAIT, &q->q_src_map)) {
1246 		ubsecstats.hst_nomap++;
1247 		err = ENOMEM;
1248 		goto errout;
1249 	}
1250 	if (crp->crp_flags & CRYPTO_F_IMBUF) {
1251 		if (bus_dmamap_load_mbuf(sc->sc_dmat, q->q_src_map,
1252 		    q->q_src_m, ubsec_op_cb, &q->q_src, BUS_DMA_NOWAIT) != 0) {
1253 			bus_dmamap_destroy(sc->sc_dmat, q->q_src_map);
1254 			q->q_src_map = NULL;
1255 			ubsecstats.hst_noload++;
1256 			err = ENOMEM;
1257 			goto errout;
1258 		}
1259 	} else if (crp->crp_flags & CRYPTO_F_IOV) {
1260 		if (bus_dmamap_load_uio(sc->sc_dmat, q->q_src_map,
1261 		    q->q_src_io, ubsec_op_cb, &q->q_src, BUS_DMA_NOWAIT) != 0) {
1262 			bus_dmamap_destroy(sc->sc_dmat, q->q_src_map);
1263 			q->q_src_map = NULL;
1264 			ubsecstats.hst_noload++;
1265 			err = ENOMEM;
1266 			goto errout;
1267 		}
1268 	}
1269 	nicealign = ubsec_dmamap_aligned(&q->q_src);
1270 
1271 	dmap->d_dma->d_mcr.mcr_pktlen = htole16(stheend);
1272 
1273 #ifdef UBSEC_DEBUG
1274 	if (ubsec_debug)
1275 		printf("src skip: %d nicealign: %u\n", sskip, nicealign);
1276 #endif
1277 	for (i = j = 0; i < q->q_src_nsegs; i++) {
1278 		struct ubsec_pktbuf *pb;
1279 		bus_size_t packl = q->q_src_segs[i].ds_len;
1280 		bus_addr_t packp = q->q_src_segs[i].ds_addr;
1281 
1282 		if (sskip >= packl) {
1283 			sskip -= packl;
1284 			continue;
1285 		}
1286 
1287 		packl -= sskip;
1288 		packp += sskip;
1289 		sskip = 0;
1290 
1291 		if (packl > 0xfffc) {
1292 			err = EIO;
1293 			goto errout;
1294 		}
1295 
1296 		if (j == 0)
1297 			pb = &dmap->d_dma->d_mcr.mcr_ipktbuf;
1298 		else
1299 			pb = &dmap->d_dma->d_sbuf[j - 1];
1300 
1301 		pb->pb_addr = htole32(packp);
1302 
1303 		if (stheend) {
1304 			if (packl > stheend) {
1305 				pb->pb_len = htole32(stheend);
1306 				stheend = 0;
1307 			} else {
1308 				pb->pb_len = htole32(packl);
1309 				stheend -= packl;
1310 			}
1311 		} else
1312 			pb->pb_len = htole32(packl);
1313 
1314 		if ((i + 1) == q->q_src_nsegs)
1315 			pb->pb_next = 0;
1316 		else
1317 			pb->pb_next = htole32(dmap->d_alloc.dma_paddr +
1318 			    offsetof(struct ubsec_dmachunk, d_sbuf[j]));
1319 		j++;
1320 	}
1321 
1322 	if (enccrd == NULL && maccrd != NULL) {
1323 		dmap->d_dma->d_mcr.mcr_opktbuf.pb_addr = 0;
1324 		dmap->d_dma->d_mcr.mcr_opktbuf.pb_len = 0;
1325 		dmap->d_dma->d_mcr.mcr_opktbuf.pb_next = htole32(dmap->d_alloc.dma_paddr +
1326 		    offsetof(struct ubsec_dmachunk, d_macbuf[0]));
1327 #ifdef UBSEC_DEBUG
1328 		if (ubsec_debug)
1329 			printf("opkt: %x %x %x\n",
1330 			    dmap->d_dma->d_mcr.mcr_opktbuf.pb_addr,
1331 			    dmap->d_dma->d_mcr.mcr_opktbuf.pb_len,
1332 			    dmap->d_dma->d_mcr.mcr_opktbuf.pb_next);
1333 #endif
1334 	} else {
1335 		if (crp->crp_flags & CRYPTO_F_IOV) {
1336 			if (!nicealign) {
1337 				ubsecstats.hst_iovmisaligned++;
1338 				err = EINVAL;
1339 				goto errout;
1340 			}
1341 			if (bus_dmamap_create(sc->sc_dmat, BUS_DMA_NOWAIT,
1342 			     &q->q_dst_map)) {
1343 				ubsecstats.hst_nomap++;
1344 				err = ENOMEM;
1345 				goto errout;
1346 			}
1347 			if (bus_dmamap_load_uio(sc->sc_dmat, q->q_dst_map,
1348 			    q->q_dst_io, ubsec_op_cb, &q->q_dst, BUS_DMA_NOWAIT) != 0) {
1349 				bus_dmamap_destroy(sc->sc_dmat, q->q_dst_map);
1350 				q->q_dst_map = NULL;
1351 				ubsecstats.hst_noload++;
1352 				err = ENOMEM;
1353 				goto errout;
1354 			}
1355 		} else if (crp->crp_flags & CRYPTO_F_IMBUF) {
1356 			if (nicealign) {
1357 				q->q_dst = q->q_src;
1358 			} else {
1359 				int totlen, len;
1360 				struct mbuf *m, *top, **mp;
1361 
1362 				ubsecstats.hst_unaligned++;
1363 				totlen = q->q_src_mapsize;
1364 				if (totlen >= MINCLSIZE) {
1365 					m = m_getcl(M_NOWAIT, MT_DATA,
1366 					    q->q_src_m->m_flags & M_PKTHDR);
1367 					len = MCLBYTES;
1368 				} else if (q->q_src_m->m_flags & M_PKTHDR) {
1369 					m = m_gethdr(M_NOWAIT, MT_DATA);
1370 					len = MHLEN;
1371 				} else {
1372 					m = m_get(M_NOWAIT, MT_DATA);
1373 					len = MLEN;
1374 				}
1375 				if (m && q->q_src_m->m_flags & M_PKTHDR &&
1376 				    !m_dup_pkthdr(m, q->q_src_m, M_NOWAIT)) {
1377 					m_free(m);
1378 					m = NULL;
1379 				}
1380 				if (m == NULL) {
1381 					ubsecstats.hst_nombuf++;
1382 					err = sc->sc_nqueue ? ERESTART : ENOMEM;
1383 					goto errout;
1384 				}
1385 				m->m_len = len = min(totlen, len);
1386 				totlen -= len;
1387 				top = m;
1388 				mp = &top;
1389 
1390 				while (totlen > 0) {
1391 					if (totlen >= MINCLSIZE) {
1392 						m = m_getcl(M_NOWAIT,
1393 						    MT_DATA, 0);
1394 						len = MCLBYTES;
1395 					} else {
1396 						m = m_get(M_NOWAIT, MT_DATA);
1397 						len = MLEN;
1398 					}
1399 					if (m == NULL) {
1400 						m_freem(top);
1401 						ubsecstats.hst_nombuf++;
1402 						err = sc->sc_nqueue ? ERESTART : ENOMEM;
1403 						goto errout;
1404 					}
1405 					m->m_len = len = min(totlen, len);
1406 					totlen -= len;
1407 					*mp = m;
1408 					mp = &m->m_next;
1409 				}
1410 				q->q_dst_m = top;
1411 				ubsec_mcopy(q->q_src_m, q->q_dst_m,
1412 				    cpskip, cpoffset);
1413 				if (bus_dmamap_create(sc->sc_dmat,
1414 				    BUS_DMA_NOWAIT, &q->q_dst_map) != 0) {
1415 					ubsecstats.hst_nomap++;
1416 					err = ENOMEM;
1417 					goto errout;
1418 				}
1419 				if (bus_dmamap_load_mbuf(sc->sc_dmat,
1420 				    q->q_dst_map, q->q_dst_m,
1421 				    ubsec_op_cb, &q->q_dst,
1422 				    BUS_DMA_NOWAIT) != 0) {
1423 					bus_dmamap_destroy(sc->sc_dmat,
1424 					q->q_dst_map);
1425 					q->q_dst_map = NULL;
1426 					ubsecstats.hst_noload++;
1427 					err = ENOMEM;
1428 					goto errout;
1429 				}
1430 			}
1431 		} else {
1432 			ubsecstats.hst_badflags++;
1433 			err = EINVAL;
1434 			goto errout;
1435 		}
1436 
1437 #ifdef UBSEC_DEBUG
1438 		if (ubsec_debug)
1439 			printf("dst skip: %d\n", dskip);
1440 #endif
1441 		for (i = j = 0; i < q->q_dst_nsegs; i++) {
1442 			struct ubsec_pktbuf *pb;
1443 			bus_size_t packl = q->q_dst_segs[i].ds_len;
1444 			bus_addr_t packp = q->q_dst_segs[i].ds_addr;
1445 
1446 			if (dskip >= packl) {
1447 				dskip -= packl;
1448 				continue;
1449 			}
1450 
1451 			packl -= dskip;
1452 			packp += dskip;
1453 			dskip = 0;
1454 
1455 			if (packl > 0xfffc) {
1456 				err = EIO;
1457 				goto errout;
1458 			}
1459 
1460 			if (j == 0)
1461 				pb = &dmap->d_dma->d_mcr.mcr_opktbuf;
1462 			else
1463 				pb = &dmap->d_dma->d_dbuf[j - 1];
1464 
1465 			pb->pb_addr = htole32(packp);
1466 
1467 			if (dtheend) {
1468 				if (packl > dtheend) {
1469 					pb->pb_len = htole32(dtheend);
1470 					dtheend = 0;
1471 				} else {
1472 					pb->pb_len = htole32(packl);
1473 					dtheend -= packl;
1474 				}
1475 			} else
1476 				pb->pb_len = htole32(packl);
1477 
1478 			if ((i + 1) == q->q_dst_nsegs) {
1479 				if (maccrd)
1480 					pb->pb_next = htole32(dmap->d_alloc.dma_paddr +
1481 					    offsetof(struct ubsec_dmachunk, d_macbuf[0]));
1482 				else
1483 					pb->pb_next = 0;
1484 			} else
1485 				pb->pb_next = htole32(dmap->d_alloc.dma_paddr +
1486 				    offsetof(struct ubsec_dmachunk, d_dbuf[j]));
1487 			j++;
1488 		}
1489 	}
1490 
1491 	dmap->d_dma->d_mcr.mcr_cmdctxp = htole32(dmap->d_alloc.dma_paddr +
1492 	    offsetof(struct ubsec_dmachunk, d_ctx));
1493 
1494 	if (sc->sc_flags & UBS_FLAGS_LONGCTX) {
1495 		struct ubsec_pktctx_long *ctxl;
1496 
1497 		ctxl = (struct ubsec_pktctx_long *)(dmap->d_alloc.dma_vaddr +
1498 		    offsetof(struct ubsec_dmachunk, d_ctx));
1499 
1500 		/* transform small context into long context */
1501 		ctxl->pc_len = htole16(sizeof(struct ubsec_pktctx_long));
1502 		ctxl->pc_type = htole16(UBS_PKTCTX_TYPE_IPSEC);
1503 		ctxl->pc_flags = ctx.pc_flags;
1504 		ctxl->pc_offset = ctx.pc_offset;
1505 		for (i = 0; i < 6; i++)
1506 			ctxl->pc_deskey[i] = ctx.pc_deskey[i];
1507 		for (i = 0; i < 5; i++)
1508 			ctxl->pc_hminner[i] = ctx.pc_hminner[i];
1509 		for (i = 0; i < 5; i++)
1510 			ctxl->pc_hmouter[i] = ctx.pc_hmouter[i];
1511 		ctxl->pc_iv[0] = ctx.pc_iv[0];
1512 		ctxl->pc_iv[1] = ctx.pc_iv[1];
1513 	} else
1514 		bcopy(&ctx, dmap->d_alloc.dma_vaddr +
1515 		    offsetof(struct ubsec_dmachunk, d_ctx),
1516 		    sizeof(struct ubsec_pktctx));
1517 
1518 	mtx_lock(&sc->sc_mcr1lock);
1519 	SIMPLEQ_INSERT_TAIL(&sc->sc_queue, q, q_next);
1520 	sc->sc_nqueue++;
1521 	ubsecstats.hst_ipackets++;
1522 	ubsecstats.hst_ibytes += dmap->d_alloc.dma_size;
1523 	if ((hint & CRYPTO_HINT_MORE) == 0 || sc->sc_nqueue >= UBS_MAX_AGGR)
1524 		ubsec_feed(sc);
1525 	mtx_unlock(&sc->sc_mcr1lock);
1526 	return (0);
1527 
1528 errout:
1529 	if (q != NULL) {
1530 		if ((q->q_dst_m != NULL) && (q->q_src_m != q->q_dst_m))
1531 			m_freem(q->q_dst_m);
1532 
1533 		if (q->q_dst_map != NULL && q->q_dst_map != q->q_src_map) {
1534 			bus_dmamap_unload(sc->sc_dmat, q->q_dst_map);
1535 			bus_dmamap_destroy(sc->sc_dmat, q->q_dst_map);
1536 		}
1537 		if (q->q_src_map != NULL) {
1538 			bus_dmamap_unload(sc->sc_dmat, q->q_src_map);
1539 			bus_dmamap_destroy(sc->sc_dmat, q->q_src_map);
1540 		}
1541 	}
1542 	if (q != NULL || err == ERESTART) {
1543 		mtx_lock(&sc->sc_freeqlock);
1544 		if (q != NULL)
1545 			SIMPLEQ_INSERT_TAIL(&sc->sc_freequeue, q, q_next);
1546 		if (err == ERESTART)
1547 			sc->sc_needwakeup |= CRYPTO_SYMQ;
1548 		mtx_unlock(&sc->sc_freeqlock);
1549 	}
1550 	if (err != ERESTART) {
1551 		crp->crp_etype = err;
1552 		crypto_done(crp);
1553 	}
1554 	return (err);
1555 }
1556 
1557 static void
1558 ubsec_callback(struct ubsec_softc *sc, struct ubsec_q *q)
1559 {
1560 	struct cryptop *crp = (struct cryptop *)q->q_crp;
1561 	struct cryptodesc *crd;
1562 	struct ubsec_dma *dmap = q->q_dma;
1563 
1564 	ubsecstats.hst_opackets++;
1565 	ubsecstats.hst_obytes += dmap->d_alloc.dma_size;
1566 
1567 	ubsec_dma_sync(&dmap->d_alloc,
1568 	    BUS_DMASYNC_POSTREAD|BUS_DMASYNC_POSTWRITE);
1569 	if (q->q_dst_map != NULL && q->q_dst_map != q->q_src_map) {
1570 		bus_dmamap_sync(sc->sc_dmat, q->q_dst_map,
1571 		    BUS_DMASYNC_POSTREAD);
1572 		bus_dmamap_unload(sc->sc_dmat, q->q_dst_map);
1573 		bus_dmamap_destroy(sc->sc_dmat, q->q_dst_map);
1574 	}
1575 	bus_dmamap_sync(sc->sc_dmat, q->q_src_map, BUS_DMASYNC_POSTWRITE);
1576 	bus_dmamap_unload(sc->sc_dmat, q->q_src_map);
1577 	bus_dmamap_destroy(sc->sc_dmat, q->q_src_map);
1578 
1579 	if ((crp->crp_flags & CRYPTO_F_IMBUF) && (q->q_src_m != q->q_dst_m)) {
1580 		m_freem(q->q_src_m);
1581 		crp->crp_buf = (caddr_t)q->q_dst_m;
1582 	}
1583 
1584 	/* copy out IV for future use */
1585 	if (q->q_flags & UBSEC_QFLAGS_COPYOUTIV) {
1586 		for (crd = crp->crp_desc; crd; crd = crd->crd_next) {
1587 			if (crd->crd_alg != CRYPTO_DES_CBC &&
1588 			    crd->crd_alg != CRYPTO_3DES_CBC)
1589 				continue;
1590 			crypto_copydata(crp->crp_flags, crp->crp_buf,
1591 			    crd->crd_skip + crd->crd_len - 8, 8,
1592 			    (caddr_t)sc->sc_sessions[q->q_sesn].ses_iv);
1593 			break;
1594 		}
1595 	}
1596 
1597 	for (crd = crp->crp_desc; crd; crd = crd->crd_next) {
1598 		if (crd->crd_alg != CRYPTO_MD5_HMAC &&
1599 		    crd->crd_alg != CRYPTO_SHA1_HMAC)
1600 			continue;
1601 		crypto_copyback(crp->crp_flags, crp->crp_buf, crd->crd_inject,
1602 		    sc->sc_sessions[q->q_sesn].ses_mlen,
1603 		    (caddr_t)dmap->d_dma->d_macbuf);
1604 		break;
1605 	}
1606 	mtx_lock(&sc->sc_freeqlock);
1607 	SIMPLEQ_INSERT_TAIL(&sc->sc_freequeue, q, q_next);
1608 	mtx_unlock(&sc->sc_freeqlock);
1609 	crypto_done(crp);
1610 }
1611 
1612 static void
1613 ubsec_mcopy(struct mbuf *srcm, struct mbuf *dstm, int hoffset, int toffset)
1614 {
1615 	int i, j, dlen, slen;
1616 	caddr_t dptr, sptr;
1617 
1618 	j = 0;
1619 	sptr = srcm->m_data;
1620 	slen = srcm->m_len;
1621 	dptr = dstm->m_data;
1622 	dlen = dstm->m_len;
1623 
1624 	while (1) {
1625 		for (i = 0; i < min(slen, dlen); i++) {
1626 			if (j < hoffset || j >= toffset)
1627 				*dptr++ = *sptr++;
1628 			slen--;
1629 			dlen--;
1630 			j++;
1631 		}
1632 		if (slen == 0) {
1633 			srcm = srcm->m_next;
1634 			if (srcm == NULL)
1635 				return;
1636 			sptr = srcm->m_data;
1637 			slen = srcm->m_len;
1638 		}
1639 		if (dlen == 0) {
1640 			dstm = dstm->m_next;
1641 			if (dstm == NULL)
1642 				return;
1643 			dptr = dstm->m_data;
1644 			dlen = dstm->m_len;
1645 		}
1646 	}
1647 }
1648 
1649 /*
1650  * feed the key generator, must be called at splimp() or higher.
1651  */
1652 static int
1653 ubsec_feed2(struct ubsec_softc *sc)
1654 {
1655 	struct ubsec_q2 *q;
1656 
1657 	while (!SIMPLEQ_EMPTY(&sc->sc_queue2)) {
1658 		if (READ_REG(sc, BS_STAT) & BS_STAT_MCR2_FULL)
1659 			break;
1660 		q = SIMPLEQ_FIRST(&sc->sc_queue2);
1661 
1662 		ubsec_dma_sync(&q->q_mcr,
1663 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1664 		ubsec_dma_sync(&q->q_ctx, BUS_DMASYNC_PREWRITE);
1665 
1666 		WRITE_REG(sc, BS_MCR2, q->q_mcr.dma_paddr);
1667 		SIMPLEQ_REMOVE_HEAD(&sc->sc_queue2, q_next);
1668 		--sc->sc_nqueue2;
1669 		SIMPLEQ_INSERT_TAIL(&sc->sc_qchip2, q, q_next);
1670 	}
1671 	return (0);
1672 }
1673 
1674 /*
1675  * Callback for handling random numbers
1676  */
1677 static void
1678 ubsec_callback2(struct ubsec_softc *sc, struct ubsec_q2 *q)
1679 {
1680 	struct cryptkop *krp;
1681 	struct ubsec_ctx_keyop *ctx;
1682 
1683 	ctx = (struct ubsec_ctx_keyop *)q->q_ctx.dma_vaddr;
1684 	ubsec_dma_sync(&q->q_ctx, BUS_DMASYNC_POSTWRITE);
1685 
1686 	switch (q->q_type) {
1687 #ifndef UBSEC_NO_RNG
1688 	case UBS_CTXOP_RNGBYPASS: {
1689 		struct ubsec_q2_rng *rng = (struct ubsec_q2_rng *)q;
1690 
1691 		ubsec_dma_sync(&rng->rng_buf, BUS_DMASYNC_POSTREAD);
1692 		(*sc->sc_harvest)(sc->sc_rndtest,
1693 			rng->rng_buf.dma_vaddr,
1694 			UBSEC_RNG_BUFSIZ*sizeof (u_int32_t));
1695 		rng->rng_used = 0;
1696 		callout_reset(&sc->sc_rngto, sc->sc_rnghz, ubsec_rng, sc);
1697 		break;
1698 	}
1699 #endif
1700 	case UBS_CTXOP_MODEXP: {
1701 		struct ubsec_q2_modexp *me = (struct ubsec_q2_modexp *)q;
1702 		u_int rlen, clen;
1703 
1704 		krp = me->me_krp;
1705 		rlen = (me->me_modbits + 7) / 8;
1706 		clen = (krp->krp_param[krp->krp_iparams].crp_nbits + 7) / 8;
1707 
1708 		ubsec_dma_sync(&me->me_M, BUS_DMASYNC_POSTWRITE);
1709 		ubsec_dma_sync(&me->me_E, BUS_DMASYNC_POSTWRITE);
1710 		ubsec_dma_sync(&me->me_C, BUS_DMASYNC_POSTREAD);
1711 		ubsec_dma_sync(&me->me_epb, BUS_DMASYNC_POSTWRITE);
1712 
1713 		if (clen < rlen)
1714 			krp->krp_status = E2BIG;
1715 		else {
1716 			if (sc->sc_flags & UBS_FLAGS_HWNORM) {
1717 				bzero(krp->krp_param[krp->krp_iparams].crp_p,
1718 				    (krp->krp_param[krp->krp_iparams].crp_nbits
1719 					+ 7) / 8);
1720 				bcopy(me->me_C.dma_vaddr,
1721 				    krp->krp_param[krp->krp_iparams].crp_p,
1722 				    (me->me_modbits + 7) / 8);
1723 			} else
1724 				ubsec_kshift_l(me->me_shiftbits,
1725 				    me->me_C.dma_vaddr, me->me_normbits,
1726 				    krp->krp_param[krp->krp_iparams].crp_p,
1727 				    krp->krp_param[krp->krp_iparams].crp_nbits);
1728 		}
1729 
1730 		crypto_kdone(krp);
1731 
1732 		/* bzero all potentially sensitive data */
1733 		bzero(me->me_E.dma_vaddr, me->me_E.dma_size);
1734 		bzero(me->me_M.dma_vaddr, me->me_M.dma_size);
1735 		bzero(me->me_C.dma_vaddr, me->me_C.dma_size);
1736 		bzero(me->me_q.q_ctx.dma_vaddr, me->me_q.q_ctx.dma_size);
1737 
1738 		/* Can't free here, so put us on the free list. */
1739 		SIMPLEQ_INSERT_TAIL(&sc->sc_q2free, &me->me_q, q_next);
1740 		break;
1741 	}
1742 	case UBS_CTXOP_RSAPRIV: {
1743 		struct ubsec_q2_rsapriv *rp = (struct ubsec_q2_rsapriv *)q;
1744 		u_int len;
1745 
1746 		krp = rp->rpr_krp;
1747 		ubsec_dma_sync(&rp->rpr_msgin, BUS_DMASYNC_POSTWRITE);
1748 		ubsec_dma_sync(&rp->rpr_msgout, BUS_DMASYNC_POSTREAD);
1749 
1750 		len = (krp->krp_param[UBS_RSAPRIV_PAR_MSGOUT].crp_nbits + 7) / 8;
1751 		bcopy(rp->rpr_msgout.dma_vaddr,
1752 		    krp->krp_param[UBS_RSAPRIV_PAR_MSGOUT].crp_p, len);
1753 
1754 		crypto_kdone(krp);
1755 
1756 		bzero(rp->rpr_msgin.dma_vaddr, rp->rpr_msgin.dma_size);
1757 		bzero(rp->rpr_msgout.dma_vaddr, rp->rpr_msgout.dma_size);
1758 		bzero(rp->rpr_q.q_ctx.dma_vaddr, rp->rpr_q.q_ctx.dma_size);
1759 
1760 		/* Can't free here, so put us on the free list. */
1761 		SIMPLEQ_INSERT_TAIL(&sc->sc_q2free, &rp->rpr_q, q_next);
1762 		break;
1763 	}
1764 	default:
1765 		device_printf(sc->sc_dev, "unknown ctx op: %x\n",
1766 		    letoh16(ctx->ctx_op));
1767 		break;
1768 	}
1769 }
1770 
1771 #ifndef UBSEC_NO_RNG
1772 static void
1773 ubsec_rng(void *vsc)
1774 {
1775 	struct ubsec_softc *sc = vsc;
1776 	struct ubsec_q2_rng *rng = &sc->sc_rng;
1777 	struct ubsec_mcr *mcr;
1778 	struct ubsec_ctx_rngbypass *ctx;
1779 
1780 	mtx_lock(&sc->sc_mcr2lock);
1781 	if (rng->rng_used) {
1782 		mtx_unlock(&sc->sc_mcr2lock);
1783 		return;
1784 	}
1785 	sc->sc_nqueue2++;
1786 	if (sc->sc_nqueue2 >= UBS_MAX_NQUEUE)
1787 		goto out;
1788 
1789 	mcr = (struct ubsec_mcr *)rng->rng_q.q_mcr.dma_vaddr;
1790 	ctx = (struct ubsec_ctx_rngbypass *)rng->rng_q.q_ctx.dma_vaddr;
1791 
1792 	mcr->mcr_pkts = htole16(1);
1793 	mcr->mcr_flags = 0;
1794 	mcr->mcr_cmdctxp = htole32(rng->rng_q.q_ctx.dma_paddr);
1795 	mcr->mcr_ipktbuf.pb_addr = mcr->mcr_ipktbuf.pb_next = 0;
1796 	mcr->mcr_ipktbuf.pb_len = 0;
1797 	mcr->mcr_reserved = mcr->mcr_pktlen = 0;
1798 	mcr->mcr_opktbuf.pb_addr = htole32(rng->rng_buf.dma_paddr);
1799 	mcr->mcr_opktbuf.pb_len = htole32(((sizeof(u_int32_t) * UBSEC_RNG_BUFSIZ)) &
1800 	    UBS_PKTBUF_LEN);
1801 	mcr->mcr_opktbuf.pb_next = 0;
1802 
1803 	ctx->rbp_len = htole16(sizeof(struct ubsec_ctx_rngbypass));
1804 	ctx->rbp_op = htole16(UBS_CTXOP_RNGBYPASS);
1805 	rng->rng_q.q_type = UBS_CTXOP_RNGBYPASS;
1806 
1807 	ubsec_dma_sync(&rng->rng_buf, BUS_DMASYNC_PREREAD);
1808 
1809 	SIMPLEQ_INSERT_TAIL(&sc->sc_queue2, &rng->rng_q, q_next);
1810 	rng->rng_used = 1;
1811 	ubsec_feed2(sc);
1812 	ubsecstats.hst_rng++;
1813 	mtx_unlock(&sc->sc_mcr2lock);
1814 
1815 	return;
1816 
1817 out:
1818 	/*
1819 	 * Something weird happened, generate our own call back.
1820 	 */
1821 	sc->sc_nqueue2--;
1822 	mtx_unlock(&sc->sc_mcr2lock);
1823 	callout_reset(&sc->sc_rngto, sc->sc_rnghz, ubsec_rng, sc);
1824 }
1825 #endif /* UBSEC_NO_RNG */
1826 
1827 static void
1828 ubsec_dmamap_cb(void *arg, bus_dma_segment_t *segs, int nseg, int error)
1829 {
1830 	bus_addr_t *paddr = (bus_addr_t*) arg;
1831 	*paddr = segs->ds_addr;
1832 }
1833 
1834 static int
1835 ubsec_dma_malloc(
1836 	struct ubsec_softc *sc,
1837 	bus_size_t size,
1838 	struct ubsec_dma_alloc *dma,
1839 	int mapflags
1840 )
1841 {
1842 	int r;
1843 
1844 	/* XXX could specify sc_dmat as parent but that just adds overhead */
1845 	r = bus_dma_tag_create(bus_get_dma_tag(sc->sc_dev),	/* parent */
1846 			       1, 0,			/* alignment, bounds */
1847 			       BUS_SPACE_MAXADDR_32BIT,	/* lowaddr */
1848 			       BUS_SPACE_MAXADDR,	/* highaddr */
1849 			       NULL, NULL,		/* filter, filterarg */
1850 			       size,			/* maxsize */
1851 			       1,			/* nsegments */
1852 			       size,			/* maxsegsize */
1853 			       BUS_DMA_ALLOCNOW,	/* flags */
1854 			       NULL, NULL,		/* lockfunc, lockarg */
1855 			       &dma->dma_tag);
1856 	if (r != 0) {
1857 		device_printf(sc->sc_dev, "ubsec_dma_malloc: "
1858 			"bus_dma_tag_create failed; error %u\n", r);
1859 		goto fail_1;
1860 	}
1861 
1862 	r = bus_dmamem_alloc(dma->dma_tag, (void**) &dma->dma_vaddr,
1863 			     BUS_DMA_NOWAIT, &dma->dma_map);
1864 	if (r != 0) {
1865 		device_printf(sc->sc_dev, "ubsec_dma_malloc: "
1866 			"bus_dmammem_alloc failed; size %ju, error %u\n",
1867 			(intmax_t)size, r);
1868 		goto fail_2;
1869 	}
1870 
1871 	r = bus_dmamap_load(dma->dma_tag, dma->dma_map, dma->dma_vaddr,
1872 		            size,
1873 			    ubsec_dmamap_cb,
1874 			    &dma->dma_paddr,
1875 			    mapflags | BUS_DMA_NOWAIT);
1876 	if (r != 0) {
1877 		device_printf(sc->sc_dev, "ubsec_dma_malloc: "
1878 			"bus_dmamap_load failed; error %u\n", r);
1879 		goto fail_3;
1880 	}
1881 
1882 	dma->dma_size = size;
1883 	return (0);
1884 
1885 fail_3:
1886 	bus_dmamap_unload(dma->dma_tag, dma->dma_map);
1887 fail_2:
1888 	bus_dmamem_free(dma->dma_tag, dma->dma_vaddr, dma->dma_map);
1889 fail_1:
1890 	bus_dma_tag_destroy(dma->dma_tag);
1891 	dma->dma_tag = NULL;
1892 	return (r);
1893 }
1894 
1895 static void
1896 ubsec_dma_free(struct ubsec_softc *sc, struct ubsec_dma_alloc *dma)
1897 {
1898 	bus_dmamap_unload(dma->dma_tag, dma->dma_map);
1899 	bus_dmamem_free(dma->dma_tag, dma->dma_vaddr, dma->dma_map);
1900 	bus_dma_tag_destroy(dma->dma_tag);
1901 }
1902 
1903 /*
1904  * Resets the board.  Values in the regesters are left as is
1905  * from the reset (i.e. initial values are assigned elsewhere).
1906  */
1907 static void
1908 ubsec_reset_board(struct ubsec_softc *sc)
1909 {
1910     volatile u_int32_t ctrl;
1911 
1912     ctrl = READ_REG(sc, BS_CTRL);
1913     ctrl |= BS_CTRL_RESET;
1914     WRITE_REG(sc, BS_CTRL, ctrl);
1915 
1916     /*
1917      * Wait aprox. 30 PCI clocks = 900 ns = 0.9 us
1918      */
1919     DELAY(10);
1920 }
1921 
1922 /*
1923  * Init Broadcom registers
1924  */
1925 static void
1926 ubsec_init_board(struct ubsec_softc *sc)
1927 {
1928 	u_int32_t ctrl;
1929 
1930 	ctrl = READ_REG(sc, BS_CTRL);
1931 	ctrl &= ~(BS_CTRL_BE32 | BS_CTRL_BE64);
1932 	ctrl |= BS_CTRL_LITTLE_ENDIAN | BS_CTRL_MCR1INT;
1933 
1934 	if (sc->sc_flags & (UBS_FLAGS_KEY|UBS_FLAGS_RNG))
1935 		ctrl |= BS_CTRL_MCR2INT;
1936 	else
1937 		ctrl &= ~BS_CTRL_MCR2INT;
1938 
1939 	if (sc->sc_flags & UBS_FLAGS_HWNORM)
1940 		ctrl &= ~BS_CTRL_SWNORM;
1941 
1942 	WRITE_REG(sc, BS_CTRL, ctrl);
1943 }
1944 
1945 /*
1946  * Init Broadcom PCI registers
1947  */
1948 static void
1949 ubsec_init_pciregs(device_t dev)
1950 {
1951 #if 0
1952 	u_int32_t misc;
1953 
1954 	misc = pci_conf_read(pc, pa->pa_tag, BS_RTY_TOUT);
1955 	misc = (misc & ~(UBS_PCI_RTY_MASK << UBS_PCI_RTY_SHIFT))
1956 	    | ((UBS_DEF_RTY & 0xff) << UBS_PCI_RTY_SHIFT);
1957 	misc = (misc & ~(UBS_PCI_TOUT_MASK << UBS_PCI_TOUT_SHIFT))
1958 	    | ((UBS_DEF_TOUT & 0xff) << UBS_PCI_TOUT_SHIFT);
1959 	pci_conf_write(pc, pa->pa_tag, BS_RTY_TOUT, misc);
1960 #endif
1961 
1962 	/*
1963 	 * This will set the cache line size to 1, this will
1964 	 * force the BCM58xx chip just to do burst read/writes.
1965 	 * Cache line read/writes are to slow
1966 	 */
1967 	pci_write_config(dev, PCIR_CACHELNSZ, UBS_DEF_CACHELINE, 1);
1968 }
1969 
1970 /*
1971  * Clean up after a chip crash.
1972  * It is assumed that the caller in splimp()
1973  */
1974 static void
1975 ubsec_cleanchip(struct ubsec_softc *sc)
1976 {
1977 	struct ubsec_q *q;
1978 
1979 	while (!SIMPLEQ_EMPTY(&sc->sc_qchip)) {
1980 		q = SIMPLEQ_FIRST(&sc->sc_qchip);
1981 		SIMPLEQ_REMOVE_HEAD(&sc->sc_qchip, q_next);
1982 		ubsec_free_q(sc, q);
1983 	}
1984 	sc->sc_nqchip = 0;
1985 }
1986 
1987 /*
1988  * free a ubsec_q
1989  * It is assumed that the caller is within splimp().
1990  */
1991 static int
1992 ubsec_free_q(struct ubsec_softc *sc, struct ubsec_q *q)
1993 {
1994 	struct ubsec_q *q2;
1995 	struct cryptop *crp;
1996 	int npkts;
1997 	int i;
1998 
1999 	npkts = q->q_nstacked_mcrs;
2000 
2001 	for (i = 0; i < npkts; i++) {
2002 		if(q->q_stacked_mcr[i]) {
2003 			q2 = q->q_stacked_mcr[i];
2004 
2005 			if ((q2->q_dst_m != NULL) && (q2->q_src_m != q2->q_dst_m))
2006 				m_freem(q2->q_dst_m);
2007 
2008 			crp = (struct cryptop *)q2->q_crp;
2009 
2010 			SIMPLEQ_INSERT_TAIL(&sc->sc_freequeue, q2, q_next);
2011 
2012 			crp->crp_etype = EFAULT;
2013 			crypto_done(crp);
2014 		} else {
2015 			break;
2016 		}
2017 	}
2018 
2019 	/*
2020 	 * Free header MCR
2021 	 */
2022 	if ((q->q_dst_m != NULL) && (q->q_src_m != q->q_dst_m))
2023 		m_freem(q->q_dst_m);
2024 
2025 	crp = (struct cryptop *)q->q_crp;
2026 
2027 	SIMPLEQ_INSERT_TAIL(&sc->sc_freequeue, q, q_next);
2028 
2029 	crp->crp_etype = EFAULT;
2030 	crypto_done(crp);
2031 	return(0);
2032 }
2033 
2034 /*
2035  * Routine to reset the chip and clean up.
2036  * It is assumed that the caller is in splimp()
2037  */
2038 static void
2039 ubsec_totalreset(struct ubsec_softc *sc)
2040 {
2041 	ubsec_reset_board(sc);
2042 	ubsec_init_board(sc);
2043 	ubsec_cleanchip(sc);
2044 }
2045 
2046 static int
2047 ubsec_dmamap_aligned(struct ubsec_operand *op)
2048 {
2049 	int i;
2050 
2051 	for (i = 0; i < op->nsegs; i++) {
2052 		if (op->segs[i].ds_addr & 3)
2053 			return (0);
2054 		if ((i != (op->nsegs - 1)) &&
2055 		    (op->segs[i].ds_len & 3))
2056 			return (0);
2057 	}
2058 	return (1);
2059 }
2060 
2061 static void
2062 ubsec_kfree(struct ubsec_softc *sc, struct ubsec_q2 *q)
2063 {
2064 	switch (q->q_type) {
2065 	case UBS_CTXOP_MODEXP: {
2066 		struct ubsec_q2_modexp *me = (struct ubsec_q2_modexp *)q;
2067 
2068 		ubsec_dma_free(sc, &me->me_q.q_mcr);
2069 		ubsec_dma_free(sc, &me->me_q.q_ctx);
2070 		ubsec_dma_free(sc, &me->me_M);
2071 		ubsec_dma_free(sc, &me->me_E);
2072 		ubsec_dma_free(sc, &me->me_C);
2073 		ubsec_dma_free(sc, &me->me_epb);
2074 		free(me, M_DEVBUF);
2075 		break;
2076 	}
2077 	case UBS_CTXOP_RSAPRIV: {
2078 		struct ubsec_q2_rsapriv *rp = (struct ubsec_q2_rsapriv *)q;
2079 
2080 		ubsec_dma_free(sc, &rp->rpr_q.q_mcr);
2081 		ubsec_dma_free(sc, &rp->rpr_q.q_ctx);
2082 		ubsec_dma_free(sc, &rp->rpr_msgin);
2083 		ubsec_dma_free(sc, &rp->rpr_msgout);
2084 		free(rp, M_DEVBUF);
2085 		break;
2086 	}
2087 	default:
2088 		device_printf(sc->sc_dev, "invalid kfree 0x%x\n", q->q_type);
2089 		break;
2090 	}
2091 }
2092 
2093 static int
2094 ubsec_kprocess(device_t dev, struct cryptkop *krp, int hint)
2095 {
2096 	struct ubsec_softc *sc = device_get_softc(dev);
2097 	int r;
2098 
2099 	if (krp == NULL || krp->krp_callback == NULL)
2100 		return (EINVAL);
2101 
2102 	while (!SIMPLEQ_EMPTY(&sc->sc_q2free)) {
2103 		struct ubsec_q2 *q;
2104 
2105 		q = SIMPLEQ_FIRST(&sc->sc_q2free);
2106 		SIMPLEQ_REMOVE_HEAD(&sc->sc_q2free, q_next);
2107 		ubsec_kfree(sc, q);
2108 	}
2109 
2110 	switch (krp->krp_op) {
2111 	case CRK_MOD_EXP:
2112 		if (sc->sc_flags & UBS_FLAGS_HWNORM)
2113 			r = ubsec_kprocess_modexp_hw(sc, krp, hint);
2114 		else
2115 			r = ubsec_kprocess_modexp_sw(sc, krp, hint);
2116 		break;
2117 	case CRK_MOD_EXP_CRT:
2118 		return (ubsec_kprocess_rsapriv(sc, krp, hint));
2119 	default:
2120 		device_printf(sc->sc_dev, "kprocess: invalid op 0x%x\n",
2121 		    krp->krp_op);
2122 		krp->krp_status = EOPNOTSUPP;
2123 		crypto_kdone(krp);
2124 		return (0);
2125 	}
2126 	return (0);			/* silence compiler */
2127 }
2128 
2129 /*
2130  * Start computation of cr[C] = (cr[M] ^ cr[E]) mod cr[N] (sw normalization)
2131  */
2132 static int
2133 ubsec_kprocess_modexp_sw(struct ubsec_softc *sc, struct cryptkop *krp, int hint)
2134 {
2135 	struct ubsec_q2_modexp *me;
2136 	struct ubsec_mcr *mcr;
2137 	struct ubsec_ctx_modexp *ctx;
2138 	struct ubsec_pktbuf *epb;
2139 	int err = 0;
2140 	u_int nbits, normbits, mbits, shiftbits, ebits;
2141 
2142 	me = (struct ubsec_q2_modexp *)malloc(sizeof *me, M_DEVBUF, M_NOWAIT);
2143 	if (me == NULL) {
2144 		err = ENOMEM;
2145 		goto errout;
2146 	}
2147 	bzero(me, sizeof *me);
2148 	me->me_krp = krp;
2149 	me->me_q.q_type = UBS_CTXOP_MODEXP;
2150 
2151 	nbits = ubsec_ksigbits(&krp->krp_param[UBS_MODEXP_PAR_N]);
2152 	if (nbits <= 512)
2153 		normbits = 512;
2154 	else if (nbits <= 768)
2155 		normbits = 768;
2156 	else if (nbits <= 1024)
2157 		normbits = 1024;
2158 	else if (sc->sc_flags & UBS_FLAGS_BIGKEY && nbits <= 1536)
2159 		normbits = 1536;
2160 	else if (sc->sc_flags & UBS_FLAGS_BIGKEY && nbits <= 2048)
2161 		normbits = 2048;
2162 	else {
2163 		err = E2BIG;
2164 		goto errout;
2165 	}
2166 
2167 	shiftbits = normbits - nbits;
2168 
2169 	me->me_modbits = nbits;
2170 	me->me_shiftbits = shiftbits;
2171 	me->me_normbits = normbits;
2172 
2173 	/* Sanity check: result bits must be >= true modulus bits. */
2174 	if (krp->krp_param[krp->krp_iparams].crp_nbits < nbits) {
2175 		err = ERANGE;
2176 		goto errout;
2177 	}
2178 
2179 	if (ubsec_dma_malloc(sc, sizeof(struct ubsec_mcr),
2180 	    &me->me_q.q_mcr, 0)) {
2181 		err = ENOMEM;
2182 		goto errout;
2183 	}
2184 	mcr = (struct ubsec_mcr *)me->me_q.q_mcr.dma_vaddr;
2185 
2186 	if (ubsec_dma_malloc(sc, sizeof(struct ubsec_ctx_modexp),
2187 	    &me->me_q.q_ctx, 0)) {
2188 		err = ENOMEM;
2189 		goto errout;
2190 	}
2191 
2192 	mbits = ubsec_ksigbits(&krp->krp_param[UBS_MODEXP_PAR_M]);
2193 	if (mbits > nbits) {
2194 		err = E2BIG;
2195 		goto errout;
2196 	}
2197 	if (ubsec_dma_malloc(sc, normbits / 8, &me->me_M, 0)) {
2198 		err = ENOMEM;
2199 		goto errout;
2200 	}
2201 	ubsec_kshift_r(shiftbits,
2202 	    krp->krp_param[UBS_MODEXP_PAR_M].crp_p, mbits,
2203 	    me->me_M.dma_vaddr, normbits);
2204 
2205 	if (ubsec_dma_malloc(sc, normbits / 8, &me->me_C, 0)) {
2206 		err = ENOMEM;
2207 		goto errout;
2208 	}
2209 	bzero(me->me_C.dma_vaddr, me->me_C.dma_size);
2210 
2211 	ebits = ubsec_ksigbits(&krp->krp_param[UBS_MODEXP_PAR_E]);
2212 	if (ebits > nbits) {
2213 		err = E2BIG;
2214 		goto errout;
2215 	}
2216 	if (ubsec_dma_malloc(sc, normbits / 8, &me->me_E, 0)) {
2217 		err = ENOMEM;
2218 		goto errout;
2219 	}
2220 	ubsec_kshift_r(shiftbits,
2221 	    krp->krp_param[UBS_MODEXP_PAR_E].crp_p, ebits,
2222 	    me->me_E.dma_vaddr, normbits);
2223 
2224 	if (ubsec_dma_malloc(sc, sizeof(struct ubsec_pktbuf),
2225 	    &me->me_epb, 0)) {
2226 		err = ENOMEM;
2227 		goto errout;
2228 	}
2229 	epb = (struct ubsec_pktbuf *)me->me_epb.dma_vaddr;
2230 	epb->pb_addr = htole32(me->me_E.dma_paddr);
2231 	epb->pb_next = 0;
2232 	epb->pb_len = htole32(normbits / 8);
2233 
2234 #ifdef UBSEC_DEBUG
2235 	if (ubsec_debug) {
2236 		printf("Epb ");
2237 		ubsec_dump_pb(epb);
2238 	}
2239 #endif
2240 
2241 	mcr->mcr_pkts = htole16(1);
2242 	mcr->mcr_flags = 0;
2243 	mcr->mcr_cmdctxp = htole32(me->me_q.q_ctx.dma_paddr);
2244 	mcr->mcr_reserved = 0;
2245 	mcr->mcr_pktlen = 0;
2246 
2247 	mcr->mcr_ipktbuf.pb_addr = htole32(me->me_M.dma_paddr);
2248 	mcr->mcr_ipktbuf.pb_len = htole32(normbits / 8);
2249 	mcr->mcr_ipktbuf.pb_next = htole32(me->me_epb.dma_paddr);
2250 
2251 	mcr->mcr_opktbuf.pb_addr = htole32(me->me_C.dma_paddr);
2252 	mcr->mcr_opktbuf.pb_next = 0;
2253 	mcr->mcr_opktbuf.pb_len = htole32(normbits / 8);
2254 
2255 #ifdef DIAGNOSTIC
2256 	/* Misaligned output buffer will hang the chip. */
2257 	if ((letoh32(mcr->mcr_opktbuf.pb_addr) & 3) != 0)
2258 		panic("%s: modexp invalid addr 0x%x\n",
2259 		    device_get_nameunit(sc->sc_dev),
2260 		    letoh32(mcr->mcr_opktbuf.pb_addr));
2261 	if ((letoh32(mcr->mcr_opktbuf.pb_len) & 3) != 0)
2262 		panic("%s: modexp invalid len 0x%x\n",
2263 		    device_get_nameunit(sc->sc_dev),
2264 		    letoh32(mcr->mcr_opktbuf.pb_len));
2265 #endif
2266 
2267 	ctx = (struct ubsec_ctx_modexp *)me->me_q.q_ctx.dma_vaddr;
2268 	bzero(ctx, sizeof(*ctx));
2269 	ubsec_kshift_r(shiftbits,
2270 	    krp->krp_param[UBS_MODEXP_PAR_N].crp_p, nbits,
2271 	    ctx->me_N, normbits);
2272 	ctx->me_len = htole16((normbits / 8) + (4 * sizeof(u_int16_t)));
2273 	ctx->me_op = htole16(UBS_CTXOP_MODEXP);
2274 	ctx->me_E_len = htole16(nbits);
2275 	ctx->me_N_len = htole16(nbits);
2276 
2277 #ifdef UBSEC_DEBUG
2278 	if (ubsec_debug) {
2279 		ubsec_dump_mcr(mcr);
2280 		ubsec_dump_ctx2((struct ubsec_ctx_keyop *)ctx);
2281 	}
2282 #endif
2283 
2284 	/*
2285 	 * ubsec_feed2 will sync mcr and ctx, we just need to sync
2286 	 * everything else.
2287 	 */
2288 	ubsec_dma_sync(&me->me_M, BUS_DMASYNC_PREWRITE);
2289 	ubsec_dma_sync(&me->me_E, BUS_DMASYNC_PREWRITE);
2290 	ubsec_dma_sync(&me->me_C, BUS_DMASYNC_PREREAD);
2291 	ubsec_dma_sync(&me->me_epb, BUS_DMASYNC_PREWRITE);
2292 
2293 	/* Enqueue and we're done... */
2294 	mtx_lock(&sc->sc_mcr2lock);
2295 	SIMPLEQ_INSERT_TAIL(&sc->sc_queue2, &me->me_q, q_next);
2296 	ubsec_feed2(sc);
2297 	ubsecstats.hst_modexp++;
2298 	mtx_unlock(&sc->sc_mcr2lock);
2299 
2300 	return (0);
2301 
2302 errout:
2303 	if (me != NULL) {
2304 		if (me->me_q.q_mcr.dma_tag != NULL)
2305 			ubsec_dma_free(sc, &me->me_q.q_mcr);
2306 		if (me->me_q.q_ctx.dma_tag != NULL) {
2307 			bzero(me->me_q.q_ctx.dma_vaddr, me->me_q.q_ctx.dma_size);
2308 			ubsec_dma_free(sc, &me->me_q.q_ctx);
2309 		}
2310 		if (me->me_M.dma_tag != NULL) {
2311 			bzero(me->me_M.dma_vaddr, me->me_M.dma_size);
2312 			ubsec_dma_free(sc, &me->me_M);
2313 		}
2314 		if (me->me_E.dma_tag != NULL) {
2315 			bzero(me->me_E.dma_vaddr, me->me_E.dma_size);
2316 			ubsec_dma_free(sc, &me->me_E);
2317 		}
2318 		if (me->me_C.dma_tag != NULL) {
2319 			bzero(me->me_C.dma_vaddr, me->me_C.dma_size);
2320 			ubsec_dma_free(sc, &me->me_C);
2321 		}
2322 		if (me->me_epb.dma_tag != NULL)
2323 			ubsec_dma_free(sc, &me->me_epb);
2324 		free(me, M_DEVBUF);
2325 	}
2326 	krp->krp_status = err;
2327 	crypto_kdone(krp);
2328 	return (0);
2329 }
2330 
2331 /*
2332  * Start computation of cr[C] = (cr[M] ^ cr[E]) mod cr[N] (hw normalization)
2333  */
2334 static int
2335 ubsec_kprocess_modexp_hw(struct ubsec_softc *sc, struct cryptkop *krp, int hint)
2336 {
2337 	struct ubsec_q2_modexp *me;
2338 	struct ubsec_mcr *mcr;
2339 	struct ubsec_ctx_modexp *ctx;
2340 	struct ubsec_pktbuf *epb;
2341 	int err = 0;
2342 	u_int nbits, normbits, mbits, shiftbits, ebits;
2343 
2344 	me = (struct ubsec_q2_modexp *)malloc(sizeof *me, M_DEVBUF, M_NOWAIT);
2345 	if (me == NULL) {
2346 		err = ENOMEM;
2347 		goto errout;
2348 	}
2349 	bzero(me, sizeof *me);
2350 	me->me_krp = krp;
2351 	me->me_q.q_type = UBS_CTXOP_MODEXP;
2352 
2353 	nbits = ubsec_ksigbits(&krp->krp_param[UBS_MODEXP_PAR_N]);
2354 	if (nbits <= 512)
2355 		normbits = 512;
2356 	else if (nbits <= 768)
2357 		normbits = 768;
2358 	else if (nbits <= 1024)
2359 		normbits = 1024;
2360 	else if (sc->sc_flags & UBS_FLAGS_BIGKEY && nbits <= 1536)
2361 		normbits = 1536;
2362 	else if (sc->sc_flags & UBS_FLAGS_BIGKEY && nbits <= 2048)
2363 		normbits = 2048;
2364 	else {
2365 		err = E2BIG;
2366 		goto errout;
2367 	}
2368 
2369 	shiftbits = normbits - nbits;
2370 
2371 	/* XXX ??? */
2372 	me->me_modbits = nbits;
2373 	me->me_shiftbits = shiftbits;
2374 	me->me_normbits = normbits;
2375 
2376 	/* Sanity check: result bits must be >= true modulus bits. */
2377 	if (krp->krp_param[krp->krp_iparams].crp_nbits < nbits) {
2378 		err = ERANGE;
2379 		goto errout;
2380 	}
2381 
2382 	if (ubsec_dma_malloc(sc, sizeof(struct ubsec_mcr),
2383 	    &me->me_q.q_mcr, 0)) {
2384 		err = ENOMEM;
2385 		goto errout;
2386 	}
2387 	mcr = (struct ubsec_mcr *)me->me_q.q_mcr.dma_vaddr;
2388 
2389 	if (ubsec_dma_malloc(sc, sizeof(struct ubsec_ctx_modexp),
2390 	    &me->me_q.q_ctx, 0)) {
2391 		err = ENOMEM;
2392 		goto errout;
2393 	}
2394 
2395 	mbits = ubsec_ksigbits(&krp->krp_param[UBS_MODEXP_PAR_M]);
2396 	if (mbits > nbits) {
2397 		err = E2BIG;
2398 		goto errout;
2399 	}
2400 	if (ubsec_dma_malloc(sc, normbits / 8, &me->me_M, 0)) {
2401 		err = ENOMEM;
2402 		goto errout;
2403 	}
2404 	bzero(me->me_M.dma_vaddr, normbits / 8);
2405 	bcopy(krp->krp_param[UBS_MODEXP_PAR_M].crp_p,
2406 	    me->me_M.dma_vaddr, (mbits + 7) / 8);
2407 
2408 	if (ubsec_dma_malloc(sc, normbits / 8, &me->me_C, 0)) {
2409 		err = ENOMEM;
2410 		goto errout;
2411 	}
2412 	bzero(me->me_C.dma_vaddr, me->me_C.dma_size);
2413 
2414 	ebits = ubsec_ksigbits(&krp->krp_param[UBS_MODEXP_PAR_E]);
2415 	if (ebits > nbits) {
2416 		err = E2BIG;
2417 		goto errout;
2418 	}
2419 	if (ubsec_dma_malloc(sc, normbits / 8, &me->me_E, 0)) {
2420 		err = ENOMEM;
2421 		goto errout;
2422 	}
2423 	bzero(me->me_E.dma_vaddr, normbits / 8);
2424 	bcopy(krp->krp_param[UBS_MODEXP_PAR_E].crp_p,
2425 	    me->me_E.dma_vaddr, (ebits + 7) / 8);
2426 
2427 	if (ubsec_dma_malloc(sc, sizeof(struct ubsec_pktbuf),
2428 	    &me->me_epb, 0)) {
2429 		err = ENOMEM;
2430 		goto errout;
2431 	}
2432 	epb = (struct ubsec_pktbuf *)me->me_epb.dma_vaddr;
2433 	epb->pb_addr = htole32(me->me_E.dma_paddr);
2434 	epb->pb_next = 0;
2435 	epb->pb_len = htole32((ebits + 7) / 8);
2436 
2437 #ifdef UBSEC_DEBUG
2438 	if (ubsec_debug) {
2439 		printf("Epb ");
2440 		ubsec_dump_pb(epb);
2441 	}
2442 #endif
2443 
2444 	mcr->mcr_pkts = htole16(1);
2445 	mcr->mcr_flags = 0;
2446 	mcr->mcr_cmdctxp = htole32(me->me_q.q_ctx.dma_paddr);
2447 	mcr->mcr_reserved = 0;
2448 	mcr->mcr_pktlen = 0;
2449 
2450 	mcr->mcr_ipktbuf.pb_addr = htole32(me->me_M.dma_paddr);
2451 	mcr->mcr_ipktbuf.pb_len = htole32(normbits / 8);
2452 	mcr->mcr_ipktbuf.pb_next = htole32(me->me_epb.dma_paddr);
2453 
2454 	mcr->mcr_opktbuf.pb_addr = htole32(me->me_C.dma_paddr);
2455 	mcr->mcr_opktbuf.pb_next = 0;
2456 	mcr->mcr_opktbuf.pb_len = htole32(normbits / 8);
2457 
2458 #ifdef DIAGNOSTIC
2459 	/* Misaligned output buffer will hang the chip. */
2460 	if ((letoh32(mcr->mcr_opktbuf.pb_addr) & 3) != 0)
2461 		panic("%s: modexp invalid addr 0x%x\n",
2462 		    device_get_nameunit(sc->sc_dev),
2463 		    letoh32(mcr->mcr_opktbuf.pb_addr));
2464 	if ((letoh32(mcr->mcr_opktbuf.pb_len) & 3) != 0)
2465 		panic("%s: modexp invalid len 0x%x\n",
2466 		    device_get_nameunit(sc->sc_dev),
2467 		    letoh32(mcr->mcr_opktbuf.pb_len));
2468 #endif
2469 
2470 	ctx = (struct ubsec_ctx_modexp *)me->me_q.q_ctx.dma_vaddr;
2471 	bzero(ctx, sizeof(*ctx));
2472 	bcopy(krp->krp_param[UBS_MODEXP_PAR_N].crp_p, ctx->me_N,
2473 	    (nbits + 7) / 8);
2474 	ctx->me_len = htole16((normbits / 8) + (4 * sizeof(u_int16_t)));
2475 	ctx->me_op = htole16(UBS_CTXOP_MODEXP);
2476 	ctx->me_E_len = htole16(ebits);
2477 	ctx->me_N_len = htole16(nbits);
2478 
2479 #ifdef UBSEC_DEBUG
2480 	if (ubsec_debug) {
2481 		ubsec_dump_mcr(mcr);
2482 		ubsec_dump_ctx2((struct ubsec_ctx_keyop *)ctx);
2483 	}
2484 #endif
2485 
2486 	/*
2487 	 * ubsec_feed2 will sync mcr and ctx, we just need to sync
2488 	 * everything else.
2489 	 */
2490 	ubsec_dma_sync(&me->me_M, BUS_DMASYNC_PREWRITE);
2491 	ubsec_dma_sync(&me->me_E, BUS_DMASYNC_PREWRITE);
2492 	ubsec_dma_sync(&me->me_C, BUS_DMASYNC_PREREAD);
2493 	ubsec_dma_sync(&me->me_epb, BUS_DMASYNC_PREWRITE);
2494 
2495 	/* Enqueue and we're done... */
2496 	mtx_lock(&sc->sc_mcr2lock);
2497 	SIMPLEQ_INSERT_TAIL(&sc->sc_queue2, &me->me_q, q_next);
2498 	ubsec_feed2(sc);
2499 	mtx_unlock(&sc->sc_mcr2lock);
2500 
2501 	return (0);
2502 
2503 errout:
2504 	if (me != NULL) {
2505 		if (me->me_q.q_mcr.dma_tag != NULL)
2506 			ubsec_dma_free(sc, &me->me_q.q_mcr);
2507 		if (me->me_q.q_ctx.dma_tag != NULL) {
2508 			bzero(me->me_q.q_ctx.dma_vaddr, me->me_q.q_ctx.dma_size);
2509 			ubsec_dma_free(sc, &me->me_q.q_ctx);
2510 		}
2511 		if (me->me_M.dma_tag != NULL) {
2512 			bzero(me->me_M.dma_vaddr, me->me_M.dma_size);
2513 			ubsec_dma_free(sc, &me->me_M);
2514 		}
2515 		if (me->me_E.dma_tag != NULL) {
2516 			bzero(me->me_E.dma_vaddr, me->me_E.dma_size);
2517 			ubsec_dma_free(sc, &me->me_E);
2518 		}
2519 		if (me->me_C.dma_tag != NULL) {
2520 			bzero(me->me_C.dma_vaddr, me->me_C.dma_size);
2521 			ubsec_dma_free(sc, &me->me_C);
2522 		}
2523 		if (me->me_epb.dma_tag != NULL)
2524 			ubsec_dma_free(sc, &me->me_epb);
2525 		free(me, M_DEVBUF);
2526 	}
2527 	krp->krp_status = err;
2528 	crypto_kdone(krp);
2529 	return (0);
2530 }
2531 
2532 static int
2533 ubsec_kprocess_rsapriv(struct ubsec_softc *sc, struct cryptkop *krp, int hint)
2534 {
2535 	struct ubsec_q2_rsapriv *rp = NULL;
2536 	struct ubsec_mcr *mcr;
2537 	struct ubsec_ctx_rsapriv *ctx;
2538 	int err = 0;
2539 	u_int padlen, msglen;
2540 
2541 	msglen = ubsec_ksigbits(&krp->krp_param[UBS_RSAPRIV_PAR_P]);
2542 	padlen = ubsec_ksigbits(&krp->krp_param[UBS_RSAPRIV_PAR_Q]);
2543 	if (msglen > padlen)
2544 		padlen = msglen;
2545 
2546 	if (padlen <= 256)
2547 		padlen = 256;
2548 	else if (padlen <= 384)
2549 		padlen = 384;
2550 	else if (padlen <= 512)
2551 		padlen = 512;
2552 	else if (sc->sc_flags & UBS_FLAGS_BIGKEY && padlen <= 768)
2553 		padlen = 768;
2554 	else if (sc->sc_flags & UBS_FLAGS_BIGKEY && padlen <= 1024)
2555 		padlen = 1024;
2556 	else {
2557 		err = E2BIG;
2558 		goto errout;
2559 	}
2560 
2561 	if (ubsec_ksigbits(&krp->krp_param[UBS_RSAPRIV_PAR_DP]) > padlen) {
2562 		err = E2BIG;
2563 		goto errout;
2564 	}
2565 
2566 	if (ubsec_ksigbits(&krp->krp_param[UBS_RSAPRIV_PAR_DQ]) > padlen) {
2567 		err = E2BIG;
2568 		goto errout;
2569 	}
2570 
2571 	if (ubsec_ksigbits(&krp->krp_param[UBS_RSAPRIV_PAR_PINV]) > padlen) {
2572 		err = E2BIG;
2573 		goto errout;
2574 	}
2575 
2576 	rp = (struct ubsec_q2_rsapriv *)malloc(sizeof *rp, M_DEVBUF, M_NOWAIT);
2577 	if (rp == NULL)
2578 		return (ENOMEM);
2579 	bzero(rp, sizeof *rp);
2580 	rp->rpr_krp = krp;
2581 	rp->rpr_q.q_type = UBS_CTXOP_RSAPRIV;
2582 
2583 	if (ubsec_dma_malloc(sc, sizeof(struct ubsec_mcr),
2584 	    &rp->rpr_q.q_mcr, 0)) {
2585 		err = ENOMEM;
2586 		goto errout;
2587 	}
2588 	mcr = (struct ubsec_mcr *)rp->rpr_q.q_mcr.dma_vaddr;
2589 
2590 	if (ubsec_dma_malloc(sc, sizeof(struct ubsec_ctx_rsapriv),
2591 	    &rp->rpr_q.q_ctx, 0)) {
2592 		err = ENOMEM;
2593 		goto errout;
2594 	}
2595 	ctx = (struct ubsec_ctx_rsapriv *)rp->rpr_q.q_ctx.dma_vaddr;
2596 	bzero(ctx, sizeof *ctx);
2597 
2598 	/* Copy in p */
2599 	bcopy(krp->krp_param[UBS_RSAPRIV_PAR_P].crp_p,
2600 	    &ctx->rpr_buf[0 * (padlen / 8)],
2601 	    (krp->krp_param[UBS_RSAPRIV_PAR_P].crp_nbits + 7) / 8);
2602 
2603 	/* Copy in q */
2604 	bcopy(krp->krp_param[UBS_RSAPRIV_PAR_Q].crp_p,
2605 	    &ctx->rpr_buf[1 * (padlen / 8)],
2606 	    (krp->krp_param[UBS_RSAPRIV_PAR_Q].crp_nbits + 7) / 8);
2607 
2608 	/* Copy in dp */
2609 	bcopy(krp->krp_param[UBS_RSAPRIV_PAR_DP].crp_p,
2610 	    &ctx->rpr_buf[2 * (padlen / 8)],
2611 	    (krp->krp_param[UBS_RSAPRIV_PAR_DP].crp_nbits + 7) / 8);
2612 
2613 	/* Copy in dq */
2614 	bcopy(krp->krp_param[UBS_RSAPRIV_PAR_DQ].crp_p,
2615 	    &ctx->rpr_buf[3 * (padlen / 8)],
2616 	    (krp->krp_param[UBS_RSAPRIV_PAR_DQ].crp_nbits + 7) / 8);
2617 
2618 	/* Copy in pinv */
2619 	bcopy(krp->krp_param[UBS_RSAPRIV_PAR_PINV].crp_p,
2620 	    &ctx->rpr_buf[4 * (padlen / 8)],
2621 	    (krp->krp_param[UBS_RSAPRIV_PAR_PINV].crp_nbits + 7) / 8);
2622 
2623 	msglen = padlen * 2;
2624 
2625 	/* Copy in input message (aligned buffer/length). */
2626 	if (ubsec_ksigbits(&krp->krp_param[UBS_RSAPRIV_PAR_MSGIN]) > msglen) {
2627 		/* Is this likely? */
2628 		err = E2BIG;
2629 		goto errout;
2630 	}
2631 	if (ubsec_dma_malloc(sc, (msglen + 7) / 8, &rp->rpr_msgin, 0)) {
2632 		err = ENOMEM;
2633 		goto errout;
2634 	}
2635 	bzero(rp->rpr_msgin.dma_vaddr, (msglen + 7) / 8);
2636 	bcopy(krp->krp_param[UBS_RSAPRIV_PAR_MSGIN].crp_p,
2637 	    rp->rpr_msgin.dma_vaddr,
2638 	    (krp->krp_param[UBS_RSAPRIV_PAR_MSGIN].crp_nbits + 7) / 8);
2639 
2640 	/* Prepare space for output message (aligned buffer/length). */
2641 	if (ubsec_ksigbits(&krp->krp_param[UBS_RSAPRIV_PAR_MSGOUT]) < msglen) {
2642 		/* Is this likely? */
2643 		err = E2BIG;
2644 		goto errout;
2645 	}
2646 	if (ubsec_dma_malloc(sc, (msglen + 7) / 8, &rp->rpr_msgout, 0)) {
2647 		err = ENOMEM;
2648 		goto errout;
2649 	}
2650 	bzero(rp->rpr_msgout.dma_vaddr, (msglen + 7) / 8);
2651 
2652 	mcr->mcr_pkts = htole16(1);
2653 	mcr->mcr_flags = 0;
2654 	mcr->mcr_cmdctxp = htole32(rp->rpr_q.q_ctx.dma_paddr);
2655 	mcr->mcr_ipktbuf.pb_addr = htole32(rp->rpr_msgin.dma_paddr);
2656 	mcr->mcr_ipktbuf.pb_next = 0;
2657 	mcr->mcr_ipktbuf.pb_len = htole32(rp->rpr_msgin.dma_size);
2658 	mcr->mcr_reserved = 0;
2659 	mcr->mcr_pktlen = htole16(msglen);
2660 	mcr->mcr_opktbuf.pb_addr = htole32(rp->rpr_msgout.dma_paddr);
2661 	mcr->mcr_opktbuf.pb_next = 0;
2662 	mcr->mcr_opktbuf.pb_len = htole32(rp->rpr_msgout.dma_size);
2663 
2664 #ifdef DIAGNOSTIC
2665 	if (rp->rpr_msgin.dma_paddr & 3 || rp->rpr_msgin.dma_size & 3) {
2666 		panic("%s: rsapriv: invalid msgin %x(0x%jx)",
2667 		    device_get_nameunit(sc->sc_dev),
2668 		    rp->rpr_msgin.dma_paddr, (uintmax_t)rp->rpr_msgin.dma_size);
2669 	}
2670 	if (rp->rpr_msgout.dma_paddr & 3 || rp->rpr_msgout.dma_size & 3) {
2671 		panic("%s: rsapriv: invalid msgout %x(0x%jx)",
2672 		    device_get_nameunit(sc->sc_dev),
2673 		    rp->rpr_msgout.dma_paddr, (uintmax_t)rp->rpr_msgout.dma_size);
2674 	}
2675 #endif
2676 
2677 	ctx->rpr_len = (sizeof(u_int16_t) * 4) + (5 * (padlen / 8));
2678 	ctx->rpr_op = htole16(UBS_CTXOP_RSAPRIV);
2679 	ctx->rpr_q_len = htole16(padlen);
2680 	ctx->rpr_p_len = htole16(padlen);
2681 
2682 	/*
2683 	 * ubsec_feed2 will sync mcr and ctx, we just need to sync
2684 	 * everything else.
2685 	 */
2686 	ubsec_dma_sync(&rp->rpr_msgin, BUS_DMASYNC_PREWRITE);
2687 	ubsec_dma_sync(&rp->rpr_msgout, BUS_DMASYNC_PREREAD);
2688 
2689 	/* Enqueue and we're done... */
2690 	mtx_lock(&sc->sc_mcr2lock);
2691 	SIMPLEQ_INSERT_TAIL(&sc->sc_queue2, &rp->rpr_q, q_next);
2692 	ubsec_feed2(sc);
2693 	ubsecstats.hst_modexpcrt++;
2694 	mtx_unlock(&sc->sc_mcr2lock);
2695 	return (0);
2696 
2697 errout:
2698 	if (rp != NULL) {
2699 		if (rp->rpr_q.q_mcr.dma_tag != NULL)
2700 			ubsec_dma_free(sc, &rp->rpr_q.q_mcr);
2701 		if (rp->rpr_msgin.dma_tag != NULL) {
2702 			bzero(rp->rpr_msgin.dma_vaddr, rp->rpr_msgin.dma_size);
2703 			ubsec_dma_free(sc, &rp->rpr_msgin);
2704 		}
2705 		if (rp->rpr_msgout.dma_tag != NULL) {
2706 			bzero(rp->rpr_msgout.dma_vaddr, rp->rpr_msgout.dma_size);
2707 			ubsec_dma_free(sc, &rp->rpr_msgout);
2708 		}
2709 		free(rp, M_DEVBUF);
2710 	}
2711 	krp->krp_status = err;
2712 	crypto_kdone(krp);
2713 	return (0);
2714 }
2715 
2716 #ifdef UBSEC_DEBUG
2717 static void
2718 ubsec_dump_pb(volatile struct ubsec_pktbuf *pb)
2719 {
2720 	printf("addr 0x%x (0x%x) next 0x%x\n",
2721 	    pb->pb_addr, pb->pb_len, pb->pb_next);
2722 }
2723 
2724 static void
2725 ubsec_dump_ctx2(struct ubsec_ctx_keyop *c)
2726 {
2727 	printf("CTX (0x%x):\n", c->ctx_len);
2728 	switch (letoh16(c->ctx_op)) {
2729 	case UBS_CTXOP_RNGBYPASS:
2730 	case UBS_CTXOP_RNGSHA1:
2731 		break;
2732 	case UBS_CTXOP_MODEXP:
2733 	{
2734 		struct ubsec_ctx_modexp *cx = (void *)c;
2735 		int i, len;
2736 
2737 		printf(" Elen %u, Nlen %u\n",
2738 		    letoh16(cx->me_E_len), letoh16(cx->me_N_len));
2739 		len = (cx->me_N_len + 7)/8;
2740 		for (i = 0; i < len; i++)
2741 			printf("%s%02x", (i == 0) ? " N: " : ":", cx->me_N[i]);
2742 		printf("\n");
2743 		break;
2744 	}
2745 	default:
2746 		printf("unknown context: %x\n", c->ctx_op);
2747 	}
2748 	printf("END CTX\n");
2749 }
2750 
2751 static void
2752 ubsec_dump_mcr(struct ubsec_mcr *mcr)
2753 {
2754 	volatile struct ubsec_mcr_add *ma;
2755 	int i;
2756 
2757 	printf("MCR:\n");
2758 	printf(" pkts: %u, flags 0x%x\n",
2759 	    letoh16(mcr->mcr_pkts), letoh16(mcr->mcr_flags));
2760 	ma = (volatile struct ubsec_mcr_add *)&mcr->mcr_cmdctxp;
2761 	for (i = 0; i < letoh16(mcr->mcr_pkts); i++) {
2762 		printf(" %d: ctx 0x%x len 0x%x rsvd 0x%x\n", i,
2763 		    letoh32(ma->mcr_cmdctxp), letoh16(ma->mcr_pktlen),
2764 		    letoh16(ma->mcr_reserved));
2765 		printf(" %d: ipkt ", i);
2766 		ubsec_dump_pb(&ma->mcr_ipktbuf);
2767 		printf(" %d: opkt ", i);
2768 		ubsec_dump_pb(&ma->mcr_opktbuf);
2769 		ma++;
2770 	}
2771 	printf("END MCR\n");
2772 }
2773 #endif /* UBSEC_DEBUG */
2774 
2775 /*
2776  * Return the number of significant bits of a big number.
2777  */
2778 static int
2779 ubsec_ksigbits(struct crparam *cr)
2780 {
2781 	u_int plen = (cr->crp_nbits + 7) / 8;
2782 	int i, sig = plen * 8;
2783 	u_int8_t c, *p = cr->crp_p;
2784 
2785 	for (i = plen - 1; i >= 0; i--) {
2786 		c = p[i];
2787 		if (c != 0) {
2788 			while ((c & 0x80) == 0) {
2789 				sig--;
2790 				c <<= 1;
2791 			}
2792 			break;
2793 		}
2794 		sig -= 8;
2795 	}
2796 	return (sig);
2797 }
2798 
2799 static void
2800 ubsec_kshift_r(
2801 	u_int shiftbits,
2802 	u_int8_t *src, u_int srcbits,
2803 	u_int8_t *dst, u_int dstbits)
2804 {
2805 	u_int slen, dlen;
2806 	int i, si, di, n;
2807 
2808 	slen = (srcbits + 7) / 8;
2809 	dlen = (dstbits + 7) / 8;
2810 
2811 	for (i = 0; i < slen; i++)
2812 		dst[i] = src[i];
2813 	for (i = 0; i < dlen - slen; i++)
2814 		dst[slen + i] = 0;
2815 
2816 	n = shiftbits / 8;
2817 	if (n != 0) {
2818 		si = dlen - n - 1;
2819 		di = dlen - 1;
2820 		while (si >= 0)
2821 			dst[di--] = dst[si--];
2822 		while (di >= 0)
2823 			dst[di--] = 0;
2824 	}
2825 
2826 	n = shiftbits % 8;
2827 	if (n != 0) {
2828 		for (i = dlen - 1; i > 0; i--)
2829 			dst[i] = (dst[i] << n) |
2830 			    (dst[i - 1] >> (8 - n));
2831 		dst[0] = dst[0] << n;
2832 	}
2833 }
2834 
2835 static void
2836 ubsec_kshift_l(
2837 	u_int shiftbits,
2838 	u_int8_t *src, u_int srcbits,
2839 	u_int8_t *dst, u_int dstbits)
2840 {
2841 	int slen, dlen, i, n;
2842 
2843 	slen = (srcbits + 7) / 8;
2844 	dlen = (dstbits + 7) / 8;
2845 
2846 	n = shiftbits / 8;
2847 	for (i = 0; i < slen; i++)
2848 		dst[i] = src[i + n];
2849 	for (i = 0; i < dlen - slen; i++)
2850 		dst[slen + i] = 0;
2851 
2852 	n = shiftbits % 8;
2853 	if (n != 0) {
2854 		for (i = 0; i < (dlen - 1); i++)
2855 			dst[i] = (dst[i] >> n) | (dst[i + 1] << (8 - n));
2856 		dst[dlen - 1] = dst[dlen - 1] >> n;
2857 	}
2858 }
2859