1/*- 2 * Copyright (c) 2003 Sam Leffler, Errno Consulting 3 * Copyright (c) 2003 Global Technology Associates, Inc. 4 * All rights reserved. 5 * 6 * Redistribution and use in source and binary forms, with or without 7 * modification, are permitted provided that the following conditions 8 * are met: 9 * 1. Redistributions of source code must retain the above copyright 10 * notice, this list of conditions and the following disclaimer. 11 * 2. Redistributions in binary form must reproduce the above copyright 12 * notice, this list of conditions and the following disclaimer in the 13 * documentation and/or other materials provided with the distribution. 14 * 15 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND 16 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 17 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 18 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE 19 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 20 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 21 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 22 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 23 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 24 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 25 * SUCH DAMAGE. 26 */ 27 28#include <sys/cdefs.h> 29__FBSDID("$FreeBSD$"); 30 31/* 32 * SafeNet SafeXcel-1141 hardware crypto accelerator 33 */ 34#include "opt_safe.h" 35 36#include <sys/param.h> 37#include <sys/systm.h> 38#include <sys/proc.h> 39#include <sys/errno.h> 40#include <sys/malloc.h> 41#include <sys/kernel.h> 42#include <sys/mbuf.h> 43#include <sys/module.h> 44#include <sys/lock.h> 45#include <sys/mutex.h> 46#include <sys/sysctl.h> 47#include <sys/endian.h> 48 49#include <vm/vm.h> 50#include <vm/pmap.h> 51 52#include <machine/bus.h> 53#include <machine/resource.h> 54#include <sys/bus.h> 55#include <sys/rman.h> 56 57#include <crypto/sha1.h> 58#include <opencrypto/cryptodev.h> 59#include <opencrypto/cryptosoft.h> 60#include <sys/md5.h> 61#include <sys/random.h> 62#include <sys/kobj.h> 63 64#include "cryptodev_if.h" 65 66#include <dev/pci/pcivar.h> 67#include <dev/pci/pcireg.h> 68 69#ifdef SAFE_RNDTEST 70#include <dev/rndtest/rndtest.h> 71#endif 72#include <dev/safe/safereg.h> 73#include <dev/safe/safevar.h> 74 75#ifndef bswap32 76#define bswap32 NTOHL 77#endif 78 79/* 80 * Prototypes and count for the pci_device structure 81 */ 82static int safe_probe(device_t); 83static int safe_attach(device_t); 84static int safe_detach(device_t); 85static int safe_suspend(device_t); 86static int safe_resume(device_t); 87static int safe_shutdown(device_t); 88 89static int safe_newsession(device_t, u_int32_t *, struct cryptoini *); 90static int safe_freesession(device_t, u_int64_t); 91static int safe_process(device_t, struct cryptop *, int); 92 93static device_method_t safe_methods[] = { 94 /* Device interface */ 95 DEVMETHOD(device_probe, safe_probe), 96 DEVMETHOD(device_attach, safe_attach), 97 DEVMETHOD(device_detach, safe_detach), 98 DEVMETHOD(device_suspend, safe_suspend), 99 DEVMETHOD(device_resume, safe_resume), 100 DEVMETHOD(device_shutdown, safe_shutdown), 101 102 /* crypto device methods */ 103 DEVMETHOD(cryptodev_newsession, safe_newsession), 104 DEVMETHOD(cryptodev_freesession,safe_freesession), 105 DEVMETHOD(cryptodev_process, safe_process), 106 107 DEVMETHOD_END 108}; 109static driver_t safe_driver = { 110 "safe", 111 safe_methods, 112 sizeof (struct safe_softc) 113}; 114static devclass_t safe_devclass; 115 116DRIVER_MODULE(safe, pci, safe_driver, safe_devclass, 0, 0); 117MODULE_DEPEND(safe, crypto, 1, 1, 1); 118#ifdef SAFE_RNDTEST 119MODULE_DEPEND(safe, rndtest, 1, 1, 1); 120#endif 121 122static void safe_intr(void *); 123static void safe_callback(struct safe_softc *, struct safe_ringentry *); 124static void safe_feed(struct safe_softc *, struct safe_ringentry *); 125static void safe_mcopy(struct mbuf *, struct mbuf *, u_int); 126#ifndef SAFE_NO_RNG 127static void safe_rng_init(struct safe_softc *); 128static void safe_rng(void *); 129#endif /* SAFE_NO_RNG */ 130static int safe_dma_malloc(struct safe_softc *, bus_size_t, 131 struct safe_dma_alloc *, int); 132#define safe_dma_sync(_dma, _flags) \ 133 bus_dmamap_sync((_dma)->dma_tag, (_dma)->dma_map, (_flags)) 134static void safe_dma_free(struct safe_softc *, struct safe_dma_alloc *); 135static int safe_dmamap_aligned(const struct safe_operand *); 136static int safe_dmamap_uniform(const struct safe_operand *); 137 138static void safe_reset_board(struct safe_softc *); 139static void safe_init_board(struct safe_softc *); 140static void safe_init_pciregs(device_t dev); 141static void safe_cleanchip(struct safe_softc *); 142static void safe_totalreset(struct safe_softc *); 143 144static int safe_free_entry(struct safe_softc *, struct safe_ringentry *); 145 146static SYSCTL_NODE(_hw, OID_AUTO, safe, CTLFLAG_RD, 0, 147 "SafeNet driver parameters"); 148 149#ifdef SAFE_DEBUG 150static void safe_dump_dmastatus(struct safe_softc *, const char *); 151static void safe_dump_ringstate(struct safe_softc *, const char *); 152static void safe_dump_intrstate(struct safe_softc *, const char *); 153static void safe_dump_request(struct safe_softc *, const char *, 154 struct safe_ringentry *); 155 156static struct safe_softc *safec; /* for use by hw.safe.dump */ 157 158static int safe_debug = 0; 159SYSCTL_INT(_hw_safe, OID_AUTO, debug, CTLFLAG_RW, &safe_debug, 160 0, "control debugging msgs"); 161#define DPRINTF(_x) if (safe_debug) printf _x 162#else 163#define DPRINTF(_x) 164#endif 165 166#define READ_REG(sc,r) \ 167 bus_space_read_4((sc)->sc_st, (sc)->sc_sh, (r)) 168 169#define WRITE_REG(sc,reg,val) \ 170 bus_space_write_4((sc)->sc_st, (sc)->sc_sh, reg, val) 171 172struct safe_stats safestats; 173SYSCTL_STRUCT(_hw_safe, OID_AUTO, stats, CTLFLAG_RD, &safestats, 174 safe_stats, "driver statistics"); 175#ifndef SAFE_NO_RNG 176static int safe_rnginterval = 1; /* poll once a second */ 177SYSCTL_INT(_hw_safe, OID_AUTO, rnginterval, CTLFLAG_RW, &safe_rnginterval, 178 0, "RNG polling interval (secs)"); 179static int safe_rngbufsize = 16; /* 64 bytes each poll */ 180SYSCTL_INT(_hw_safe, OID_AUTO, rngbufsize, CTLFLAG_RW, &safe_rngbufsize, 181 0, "RNG polling buffer size (32-bit words)"); 182static int safe_rngmaxalarm = 8; /* max alarms before reset */ 183SYSCTL_INT(_hw_safe, OID_AUTO, rngmaxalarm, CTLFLAG_RW, &safe_rngmaxalarm, 184 0, "RNG max alarms before reset"); 185#endif /* SAFE_NO_RNG */ 186 187static int 188safe_probe(device_t dev) 189{ 190 if (pci_get_vendor(dev) == PCI_VENDOR_SAFENET && 191 pci_get_device(dev) == PCI_PRODUCT_SAFEXCEL) 192 return (BUS_PROBE_DEFAULT); 193 return (ENXIO); 194} 195 196static const char* 197safe_partname(struct safe_softc *sc) 198{ 199 /* XXX sprintf numbers when not decoded */ 200 switch (pci_get_vendor(sc->sc_dev)) { 201 case PCI_VENDOR_SAFENET: 202 switch (pci_get_device(sc->sc_dev)) { 203 case PCI_PRODUCT_SAFEXCEL: return "SafeNet SafeXcel-1141"; 204 } 205 return "SafeNet unknown-part"; 206 } 207 return "Unknown-vendor unknown-part"; 208} 209 210#ifndef SAFE_NO_RNG 211static void 212default_harvest(struct rndtest_state *rsp, void *buf, u_int count) 213{ 214 /* MarkM: FIX!! Check that this does not swamp the harvester! */ 215 random_harvest_queue(buf, count, count*NBBY/2, RANDOM_PURE_SAFE); 216} 217#endif /* SAFE_NO_RNG */ 218 219static int 220safe_attach(device_t dev) 221{ 222 struct safe_softc *sc = device_get_softc(dev); 223 u_int32_t raddr; 224 u_int32_t i, devinfo; 225 int rid; 226 227 bzero(sc, sizeof (*sc)); 228 sc->sc_dev = dev; 229 230 /* XXX handle power management */ 231 232 pci_enable_busmaster(dev); 233 234 /* 235 * Setup memory-mapping of PCI registers. 236 */ 237 rid = BS_BAR; 238 sc->sc_sr = bus_alloc_resource_any(dev, SYS_RES_MEMORY, &rid, 239 RF_ACTIVE); 240 if (sc->sc_sr == NULL) { 241 device_printf(dev, "cannot map register space\n"); 242 goto bad; 243 } 244 sc->sc_st = rman_get_bustag(sc->sc_sr); 245 sc->sc_sh = rman_get_bushandle(sc->sc_sr); 246 247 /* 248 * Arrange interrupt line. 249 */ 250 rid = 0; 251 sc->sc_irq = bus_alloc_resource_any(dev, SYS_RES_IRQ, &rid, 252 RF_SHAREABLE|RF_ACTIVE); 253 if (sc->sc_irq == NULL) { 254 device_printf(dev, "could not map interrupt\n"); 255 goto bad1; 256 } 257 /* 258 * NB: Network code assumes we are blocked with splimp() 259 * so make sure the IRQ is mapped appropriately. 260 */ 261 if (bus_setup_intr(dev, sc->sc_irq, INTR_TYPE_NET | INTR_MPSAFE, 262 NULL, safe_intr, sc, &sc->sc_ih)) { 263 device_printf(dev, "could not establish interrupt\n"); 264 goto bad2; 265 } 266 267 sc->sc_cid = crypto_get_driverid(dev, CRYPTOCAP_F_HARDWARE); 268 if (sc->sc_cid < 0) { 269 device_printf(dev, "could not get crypto driver id\n"); 270 goto bad3; 271 } 272 273 sc->sc_chiprev = READ_REG(sc, SAFE_DEVINFO) & 274 (SAFE_DEVINFO_REV_MAJ | SAFE_DEVINFO_REV_MIN); 275 276 /* 277 * Setup DMA descriptor area. 278 */ 279 if (bus_dma_tag_create(bus_get_dma_tag(dev), /* parent */ 280 1, /* alignment */ 281 SAFE_DMA_BOUNDARY, /* boundary */ 282 BUS_SPACE_MAXADDR_32BIT, /* lowaddr */ 283 BUS_SPACE_MAXADDR, /* highaddr */ 284 NULL, NULL, /* filter, filterarg */ 285 SAFE_MAX_DMA, /* maxsize */ 286 SAFE_MAX_PART, /* nsegments */ 287 SAFE_MAX_SSIZE, /* maxsegsize */ 288 BUS_DMA_ALLOCNOW, /* flags */ 289 NULL, NULL, /* locking */ 290 &sc->sc_srcdmat)) { 291 device_printf(dev, "cannot allocate DMA tag\n"); 292 goto bad4; 293 } 294 if (bus_dma_tag_create(bus_get_dma_tag(dev), /* parent */ 295 1, /* alignment */ 296 SAFE_MAX_DSIZE, /* boundary */ 297 BUS_SPACE_MAXADDR_32BIT, /* lowaddr */ 298 BUS_SPACE_MAXADDR, /* highaddr */ 299 NULL, NULL, /* filter, filterarg */ 300 SAFE_MAX_DMA, /* maxsize */ 301 SAFE_MAX_PART, /* nsegments */ 302 SAFE_MAX_DSIZE, /* maxsegsize */ 303 BUS_DMA_ALLOCNOW, /* flags */ 304 NULL, NULL, /* locking */ 305 &sc->sc_dstdmat)) { 306 device_printf(dev, "cannot allocate DMA tag\n"); 307 goto bad4; 308 } 309 310 /* 311 * Allocate packet engine descriptors. 312 */ 313 if (safe_dma_malloc(sc, 314 SAFE_MAX_NQUEUE * sizeof (struct safe_ringentry), 315 &sc->sc_ringalloc, 0)) { 316 device_printf(dev, "cannot allocate PE descriptor ring\n"); 317 bus_dma_tag_destroy(sc->sc_srcdmat); 318 goto bad4; 319 } 320 /* 321 * Hookup the static portion of all our data structures. 322 */ 323 sc->sc_ring = (struct safe_ringentry *) sc->sc_ringalloc.dma_vaddr; 324 sc->sc_ringtop = sc->sc_ring + SAFE_MAX_NQUEUE; 325 sc->sc_front = sc->sc_ring; 326 sc->sc_back = sc->sc_ring; 327 raddr = sc->sc_ringalloc.dma_paddr; 328 bzero(sc->sc_ring, SAFE_MAX_NQUEUE * sizeof(struct safe_ringentry)); 329 for (i = 0; i < SAFE_MAX_NQUEUE; i++) { 330 struct safe_ringentry *re = &sc->sc_ring[i]; 331 332 re->re_desc.d_sa = raddr + 333 offsetof(struct safe_ringentry, re_sa); 334 re->re_sa.sa_staterec = raddr + 335 offsetof(struct safe_ringentry, re_sastate); 336 337 raddr += sizeof (struct safe_ringentry); 338 } 339 mtx_init(&sc->sc_ringmtx, device_get_nameunit(dev), 340 "packet engine ring", MTX_DEF); 341 342 /* 343 * Allocate scatter and gather particle descriptors. 344 */ 345 if (safe_dma_malloc(sc, SAFE_TOTAL_SPART * sizeof (struct safe_pdesc), 346 &sc->sc_spalloc, 0)) { 347 device_printf(dev, "cannot allocate source particle " 348 "descriptor ring\n"); 349 mtx_destroy(&sc->sc_ringmtx); 350 safe_dma_free(sc, &sc->sc_ringalloc); 351 bus_dma_tag_destroy(sc->sc_srcdmat); 352 goto bad4; 353 } 354 sc->sc_spring = (struct safe_pdesc *) sc->sc_spalloc.dma_vaddr; 355 sc->sc_springtop = sc->sc_spring + SAFE_TOTAL_SPART; 356 sc->sc_spfree = sc->sc_spring; 357 bzero(sc->sc_spring, SAFE_TOTAL_SPART * sizeof(struct safe_pdesc)); 358 359 if (safe_dma_malloc(sc, SAFE_TOTAL_DPART * sizeof (struct safe_pdesc), 360 &sc->sc_dpalloc, 0)) { 361 device_printf(dev, "cannot allocate destination particle " 362 "descriptor ring\n"); 363 mtx_destroy(&sc->sc_ringmtx); 364 safe_dma_free(sc, &sc->sc_spalloc); 365 safe_dma_free(sc, &sc->sc_ringalloc); 366 bus_dma_tag_destroy(sc->sc_dstdmat); 367 goto bad4; 368 } 369 sc->sc_dpring = (struct safe_pdesc *) sc->sc_dpalloc.dma_vaddr; 370 sc->sc_dpringtop = sc->sc_dpring + SAFE_TOTAL_DPART; 371 sc->sc_dpfree = sc->sc_dpring; 372 bzero(sc->sc_dpring, SAFE_TOTAL_DPART * sizeof(struct safe_pdesc)); 373 374 device_printf(sc->sc_dev, "%s", safe_partname(sc)); 375 376 devinfo = READ_REG(sc, SAFE_DEVINFO); 377 if (devinfo & SAFE_DEVINFO_RNG) { 378 sc->sc_flags |= SAFE_FLAGS_RNG; 379 printf(" rng"); 380 } 381 if (devinfo & SAFE_DEVINFO_PKEY) { 382#if 0 383 printf(" key"); 384 sc->sc_flags |= SAFE_FLAGS_KEY; 385 crypto_kregister(sc->sc_cid, CRK_MOD_EXP, 0); 386 crypto_kregister(sc->sc_cid, CRK_MOD_EXP_CRT, 0); 387#endif 388 } 389 if (devinfo & SAFE_DEVINFO_DES) { 390 printf(" des/3des"); 391 crypto_register(sc->sc_cid, CRYPTO_3DES_CBC, 0, 0); 392 crypto_register(sc->sc_cid, CRYPTO_DES_CBC, 0, 0); 393 } 394 if (devinfo & SAFE_DEVINFO_AES) { 395 printf(" aes"); 396 crypto_register(sc->sc_cid, CRYPTO_AES_CBC, 0, 0); 397 } 398 if (devinfo & SAFE_DEVINFO_MD5) { 399 printf(" md5"); 400 crypto_register(sc->sc_cid, CRYPTO_MD5_HMAC, 0, 0); 401 } 402 if (devinfo & SAFE_DEVINFO_SHA1) { 403 printf(" sha1"); 404 crypto_register(sc->sc_cid, CRYPTO_SHA1_HMAC, 0, 0); 405 } 406 printf(" null"); 407 crypto_register(sc->sc_cid, CRYPTO_NULL_CBC, 0, 0); 408 crypto_register(sc->sc_cid, CRYPTO_NULL_HMAC, 0, 0); 409 /* XXX other supported algorithms */ 410 printf("\n"); 411 412 safe_reset_board(sc); /* reset h/w */ 413 safe_init_pciregs(dev); /* init pci settings */ 414 safe_init_board(sc); /* init h/w */ 415 416#ifndef SAFE_NO_RNG 417 if (sc->sc_flags & SAFE_FLAGS_RNG) { 418#ifdef SAFE_RNDTEST 419 sc->sc_rndtest = rndtest_attach(dev); 420 if (sc->sc_rndtest) 421 sc->sc_harvest = rndtest_harvest; 422 else 423 sc->sc_harvest = default_harvest; 424#else 425 sc->sc_harvest = default_harvest; 426#endif 427 safe_rng_init(sc); 428 429 callout_init(&sc->sc_rngto, 1); 430 callout_reset(&sc->sc_rngto, hz*safe_rnginterval, safe_rng, sc); 431 } 432#endif /* SAFE_NO_RNG */ 433#ifdef SAFE_DEBUG 434 safec = sc; /* for use by hw.safe.dump */ 435#endif 436 return (0); 437bad4: 438 crypto_unregister_all(sc->sc_cid); 439bad3: 440 bus_teardown_intr(dev, sc->sc_irq, sc->sc_ih); 441bad2: 442 bus_release_resource(dev, SYS_RES_IRQ, 0, sc->sc_irq); 443bad1: 444 bus_release_resource(dev, SYS_RES_MEMORY, BS_BAR, sc->sc_sr); 445bad: 446 return (ENXIO); 447} 448 449/* 450 * Detach a device that successfully probed. 451 */ 452static int 453safe_detach(device_t dev) 454{ 455 struct safe_softc *sc = device_get_softc(dev); 456 457 /* XXX wait/abort active ops */ 458 459 WRITE_REG(sc, SAFE_HI_MASK, 0); /* disable interrupts */ 460 461 callout_stop(&sc->sc_rngto); 462 463 crypto_unregister_all(sc->sc_cid); 464 465#ifdef SAFE_RNDTEST 466 if (sc->sc_rndtest) 467 rndtest_detach(sc->sc_rndtest); 468#endif 469 470 safe_cleanchip(sc); 471 safe_dma_free(sc, &sc->sc_dpalloc); 472 safe_dma_free(sc, &sc->sc_spalloc); 473 mtx_destroy(&sc->sc_ringmtx); 474 safe_dma_free(sc, &sc->sc_ringalloc); 475 476 bus_generic_detach(dev); 477 bus_teardown_intr(dev, sc->sc_irq, sc->sc_ih); 478 bus_release_resource(dev, SYS_RES_IRQ, 0, sc->sc_irq); 479 480 bus_dma_tag_destroy(sc->sc_srcdmat); 481 bus_dma_tag_destroy(sc->sc_dstdmat); 482 bus_release_resource(dev, SYS_RES_MEMORY, BS_BAR, sc->sc_sr); 483 484 return (0); 485} 486 487/* 488 * Stop all chip i/o so that the kernel's probe routines don't 489 * get confused by errant DMAs when rebooting. 490 */ 491static int 492safe_shutdown(device_t dev) 493{ 494#ifdef notyet 495 safe_stop(device_get_softc(dev)); 496#endif 497 return (0); 498} 499 500/* 501 * Device suspend routine. 502 */ 503static int 504safe_suspend(device_t dev) 505{ 506 struct safe_softc *sc = device_get_softc(dev); 507 508#ifdef notyet 509 /* XXX stop the device and save PCI settings */ 510#endif 511 sc->sc_suspended = 1; 512 513 return (0); 514} 515 516static int 517safe_resume(device_t dev) 518{ 519 struct safe_softc *sc = device_get_softc(dev); 520 521#ifdef notyet 522 /* XXX retore PCI settings and start the device */ 523#endif 524 sc->sc_suspended = 0; 525 return (0); 526} 527 528/* 529 * SafeXcel Interrupt routine 530 */ 531static void 532safe_intr(void *arg) 533{ 534 struct safe_softc *sc = arg; 535 volatile u_int32_t stat; 536 537 stat = READ_REG(sc, SAFE_HM_STAT); 538 if (stat == 0) /* shared irq, not for us */ 539 return; 540 541 WRITE_REG(sc, SAFE_HI_CLR, stat); /* IACK */ 542 543 if ((stat & SAFE_INT_PE_DDONE)) { 544 /* 545 * Descriptor(s) done; scan the ring and 546 * process completed operations. 547 */ 548 mtx_lock(&sc->sc_ringmtx); 549 while (sc->sc_back != sc->sc_front) { 550 struct safe_ringentry *re = sc->sc_back; 551#ifdef SAFE_DEBUG 552 if (safe_debug) { 553 safe_dump_ringstate(sc, __func__); 554 safe_dump_request(sc, __func__, re); 555 } 556#endif 557 /* 558 * safe_process marks ring entries that were allocated 559 * but not used with a csr of zero. This insures the 560 * ring front pointer never needs to be set backwards 561 * in the event that an entry is allocated but not used 562 * because of a setup error. 563 */ 564 if (re->re_desc.d_csr != 0) { 565 if (!SAFE_PE_CSR_IS_DONE(re->re_desc.d_csr)) 566 break; 567 if (!SAFE_PE_LEN_IS_DONE(re->re_desc.d_len)) 568 break; 569 sc->sc_nqchip--; 570 safe_callback(sc, re); 571 } 572 if (++(sc->sc_back) == sc->sc_ringtop) 573 sc->sc_back = sc->sc_ring; 574 } 575 mtx_unlock(&sc->sc_ringmtx); 576 } 577 578 /* 579 * Check to see if we got any DMA Error 580 */ 581 if (stat & SAFE_INT_PE_ERROR) { 582 DPRINTF(("dmaerr dmastat %08x\n", 583 READ_REG(sc, SAFE_PE_DMASTAT))); 584 safestats.st_dmaerr++; 585 safe_totalreset(sc); 586#if 0 587 safe_feed(sc); 588#endif 589 } 590 591 if (sc->sc_needwakeup) { /* XXX check high watermark */ 592 int wakeup = sc->sc_needwakeup & (CRYPTO_SYMQ|CRYPTO_ASYMQ); 593 DPRINTF(("%s: wakeup crypto %x\n", __func__, 594 sc->sc_needwakeup)); 595 sc->sc_needwakeup &= ~wakeup; 596 crypto_unblock(sc->sc_cid, wakeup); 597 } 598} 599 600/* 601 * safe_feed() - post a request to chip 602 */ 603static void 604safe_feed(struct safe_softc *sc, struct safe_ringentry *re) 605{ 606 bus_dmamap_sync(sc->sc_srcdmat, re->re_src_map, BUS_DMASYNC_PREWRITE); 607 if (re->re_dst_map != NULL) 608 bus_dmamap_sync(sc->sc_dstdmat, re->re_dst_map, 609 BUS_DMASYNC_PREREAD); 610 /* XXX have no smaller granularity */ 611 safe_dma_sync(&sc->sc_ringalloc, 612 BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE); 613 safe_dma_sync(&sc->sc_spalloc, BUS_DMASYNC_PREWRITE); 614 safe_dma_sync(&sc->sc_dpalloc, BUS_DMASYNC_PREWRITE); 615 616#ifdef SAFE_DEBUG 617 if (safe_debug) { 618 safe_dump_ringstate(sc, __func__); 619 safe_dump_request(sc, __func__, re); 620 } 621#endif 622 sc->sc_nqchip++; 623 if (sc->sc_nqchip > safestats.st_maxqchip) 624 safestats.st_maxqchip = sc->sc_nqchip; 625 /* poke h/w to check descriptor ring, any value can be written */ 626 WRITE_REG(sc, SAFE_HI_RD_DESCR, 0); 627} 628 629#define N(a) (sizeof(a) / sizeof (a[0])) 630static void 631safe_setup_enckey(struct safe_session *ses, caddr_t key) 632{ 633 int i; 634 635 bcopy(key, ses->ses_key, ses->ses_klen / 8); 636 637 /* PE is little-endian, insure proper byte order */ 638 for (i = 0; i < N(ses->ses_key); i++) 639 ses->ses_key[i] = htole32(ses->ses_key[i]); 640} 641 642static void 643safe_setup_mackey(struct safe_session *ses, int algo, caddr_t key, int klen) 644{ 645 MD5_CTX md5ctx; 646 SHA1_CTX sha1ctx; 647 int i; 648 649 650 for (i = 0; i < klen; i++) 651 key[i] ^= HMAC_IPAD_VAL; 652 653 if (algo == CRYPTO_MD5_HMAC) { 654 MD5Init(&md5ctx); 655 MD5Update(&md5ctx, key, klen); 656 MD5Update(&md5ctx, hmac_ipad_buffer, MD5_HMAC_BLOCK_LEN - klen); 657 bcopy(md5ctx.state, ses->ses_hminner, sizeof(md5ctx.state)); 658 } else { 659 SHA1Init(&sha1ctx); 660 SHA1Update(&sha1ctx, key, klen); 661 SHA1Update(&sha1ctx, hmac_ipad_buffer, 662 SHA1_HMAC_BLOCK_LEN - klen); 663 bcopy(sha1ctx.h.b32, ses->ses_hminner, sizeof(sha1ctx.h.b32)); 664 } 665 666 for (i = 0; i < klen; i++) 667 key[i] ^= (HMAC_IPAD_VAL ^ HMAC_OPAD_VAL); 668 669 if (algo == CRYPTO_MD5_HMAC) { 670 MD5Init(&md5ctx); 671 MD5Update(&md5ctx, key, klen); 672 MD5Update(&md5ctx, hmac_opad_buffer, MD5_HMAC_BLOCK_LEN - klen); 673 bcopy(md5ctx.state, ses->ses_hmouter, sizeof(md5ctx.state)); 674 } else { 675 SHA1Init(&sha1ctx); 676 SHA1Update(&sha1ctx, key, klen); 677 SHA1Update(&sha1ctx, hmac_opad_buffer, 678 SHA1_HMAC_BLOCK_LEN - klen); 679 bcopy(sha1ctx.h.b32, ses->ses_hmouter, sizeof(sha1ctx.h.b32)); 680 } 681 682 for (i = 0; i < klen; i++) 683 key[i] ^= HMAC_OPAD_VAL; 684 685 /* PE is little-endian, insure proper byte order */ 686 for (i = 0; i < N(ses->ses_hminner); i++) { 687 ses->ses_hminner[i] = htole32(ses->ses_hminner[i]); 688 ses->ses_hmouter[i] = htole32(ses->ses_hmouter[i]); 689 } 690} 691#undef N 692 693/* 694 * Allocate a new 'session' and return an encoded session id. 'sidp' 695 * contains our registration id, and should contain an encoded session 696 * id on successful allocation. 697 */ 698static int 699safe_newsession(device_t dev, u_int32_t *sidp, struct cryptoini *cri) 700{ 701 struct safe_softc *sc = device_get_softc(dev); 702 struct cryptoini *c, *encini = NULL, *macini = NULL; 703 struct safe_session *ses = NULL; 704 int sesn; 705 706 if (sidp == NULL || cri == NULL || sc == NULL) 707 return (EINVAL); 708 709 for (c = cri; c != NULL; c = c->cri_next) { 710 if (c->cri_alg == CRYPTO_MD5_HMAC || 711 c->cri_alg == CRYPTO_SHA1_HMAC || 712 c->cri_alg == CRYPTO_NULL_HMAC) { 713 if (macini) 714 return (EINVAL); 715 macini = c; 716 } else if (c->cri_alg == CRYPTO_DES_CBC || 717 c->cri_alg == CRYPTO_3DES_CBC || 718 c->cri_alg == CRYPTO_AES_CBC || 719 c->cri_alg == CRYPTO_NULL_CBC) { 720 if (encini) 721 return (EINVAL); 722 encini = c; 723 } else 724 return (EINVAL); 725 } 726 if (encini == NULL && macini == NULL) 727 return (EINVAL); 728 if (encini) { /* validate key length */ 729 switch (encini->cri_alg) { 730 case CRYPTO_DES_CBC: 731 if (encini->cri_klen != 64) 732 return (EINVAL); 733 break; 734 case CRYPTO_3DES_CBC: 735 if (encini->cri_klen != 192) 736 return (EINVAL); 737 break; 738 case CRYPTO_AES_CBC: 739 if (encini->cri_klen != 128 && 740 encini->cri_klen != 192 && 741 encini->cri_klen != 256) 742 return (EINVAL); 743 break; 744 } 745 } 746 747 if (sc->sc_sessions == NULL) { 748 ses = sc->sc_sessions = (struct safe_session *)malloc( 749 sizeof(struct safe_session), M_DEVBUF, M_NOWAIT); 750 if (ses == NULL) 751 return (ENOMEM); 752 sesn = 0; 753 sc->sc_nsessions = 1; 754 } else { 755 for (sesn = 0; sesn < sc->sc_nsessions; sesn++) { 756 if (sc->sc_sessions[sesn].ses_used == 0) { 757 ses = &sc->sc_sessions[sesn]; 758 break; 759 } 760 } 761 762 if (ses == NULL) { 763 sesn = sc->sc_nsessions; 764 ses = (struct safe_session *)malloc((sesn + 1) * 765 sizeof(struct safe_session), M_DEVBUF, M_NOWAIT); 766 if (ses == NULL) 767 return (ENOMEM); 768 bcopy(sc->sc_sessions, ses, sesn * 769 sizeof(struct safe_session)); 770 bzero(sc->sc_sessions, sesn * 771 sizeof(struct safe_session)); 772 free(sc->sc_sessions, M_DEVBUF); 773 sc->sc_sessions = ses; 774 ses = &sc->sc_sessions[sesn]; 775 sc->sc_nsessions++; 776 } 777 } 778 779 bzero(ses, sizeof(struct safe_session)); 780 ses->ses_used = 1; 781 782 if (encini) { 783 /* get an IV */ 784 /* XXX may read fewer than requested */ 785 read_random(ses->ses_iv, sizeof(ses->ses_iv)); 786 787 ses->ses_klen = encini->cri_klen; 788 if (encini->cri_key != NULL) 789 safe_setup_enckey(ses, encini->cri_key); 790 } 791 792 if (macini) { 793 ses->ses_mlen = macini->cri_mlen; 794 if (ses->ses_mlen == 0) { 795 if (macini->cri_alg == CRYPTO_MD5_HMAC) 796 ses->ses_mlen = MD5_HASH_LEN; 797 else 798 ses->ses_mlen = SHA1_HASH_LEN; 799 } 800 801 if (macini->cri_key != NULL) { 802 safe_setup_mackey(ses, macini->cri_alg, macini->cri_key, 803 macini->cri_klen / 8); 804 } 805 } 806 807 *sidp = SAFE_SID(device_get_unit(sc->sc_dev), sesn); 808 return (0); 809} 810 811/* 812 * Deallocate a session. 813 */ 814static int 815safe_freesession(device_t dev, u_int64_t tid) 816{ 817 struct safe_softc *sc = device_get_softc(dev); 818 int session, ret; 819 u_int32_t sid = ((u_int32_t) tid) & 0xffffffff; 820 821 if (sc == NULL) 822 return (EINVAL); 823 824 session = SAFE_SESSION(sid); 825 if (session < sc->sc_nsessions) { 826 bzero(&sc->sc_sessions[session], sizeof(sc->sc_sessions[session])); 827 ret = 0; 828 } else 829 ret = EINVAL; 830 return (ret); 831} 832 833static void 834safe_op_cb(void *arg, bus_dma_segment_t *seg, int nsegs, bus_size_t mapsize, int error) 835{ 836 struct safe_operand *op = arg; 837 838 DPRINTF(("%s: mapsize %u nsegs %d error %d\n", __func__, 839 (u_int) mapsize, nsegs, error)); 840 if (error != 0) 841 return; 842 op->mapsize = mapsize; 843 op->nsegs = nsegs; 844 bcopy(seg, op->segs, nsegs * sizeof (seg[0])); 845} 846 847static int 848safe_process(device_t dev, struct cryptop *crp, int hint) 849{ 850 struct safe_softc *sc = device_get_softc(dev); 851 int err = 0, i, nicealign, uniform; 852 struct cryptodesc *crd1, *crd2, *maccrd, *enccrd; 853 int bypass, oplen, ivsize; 854 caddr_t iv; 855 int16_t coffset; 856 struct safe_session *ses; 857 struct safe_ringentry *re; 858 struct safe_sarec *sa; 859 struct safe_pdesc *pd; 860 u_int32_t cmd0, cmd1, staterec; 861 862 if (crp == NULL || crp->crp_callback == NULL || sc == NULL) { 863 safestats.st_invalid++; 864 return (EINVAL); 865 } 866 if (SAFE_SESSION(crp->crp_sid) >= sc->sc_nsessions) { 867 safestats.st_badsession++; 868 return (EINVAL); 869 } 870 871 mtx_lock(&sc->sc_ringmtx); 872 if (sc->sc_front == sc->sc_back && sc->sc_nqchip != 0) { 873 safestats.st_ringfull++; 874 sc->sc_needwakeup |= CRYPTO_SYMQ; 875 mtx_unlock(&sc->sc_ringmtx); 876 return (ERESTART); 877 } 878 re = sc->sc_front; 879 880 staterec = re->re_sa.sa_staterec; /* save */ 881 /* NB: zero everything but the PE descriptor */ 882 bzero(&re->re_sa, sizeof(struct safe_ringentry) - sizeof(re->re_desc)); 883 re->re_sa.sa_staterec = staterec; /* restore */ 884 885 re->re_crp = crp; 886 re->re_sesn = SAFE_SESSION(crp->crp_sid); 887 888 if (crp->crp_flags & CRYPTO_F_IMBUF) { 889 re->re_src_m = (struct mbuf *)crp->crp_buf; 890 re->re_dst_m = (struct mbuf *)crp->crp_buf; 891 } else if (crp->crp_flags & CRYPTO_F_IOV) { 892 re->re_src_io = (struct uio *)crp->crp_buf; 893 re->re_dst_io = (struct uio *)crp->crp_buf; 894 } else { 895 safestats.st_badflags++; 896 err = EINVAL; 897 goto errout; /* XXX we don't handle contiguous blocks! */ 898 } 899 900 sa = &re->re_sa; 901 ses = &sc->sc_sessions[re->re_sesn]; 902 903 crd1 = crp->crp_desc; 904 if (crd1 == NULL) { 905 safestats.st_nodesc++; 906 err = EINVAL; 907 goto errout; 908 } 909 crd2 = crd1->crd_next; 910 911 cmd0 = SAFE_SA_CMD0_BASIC; /* basic group operation */ 912 cmd1 = 0; 913 if (crd2 == NULL) { 914 if (crd1->crd_alg == CRYPTO_MD5_HMAC || 915 crd1->crd_alg == CRYPTO_SHA1_HMAC || 916 crd1->crd_alg == CRYPTO_NULL_HMAC) { 917 maccrd = crd1; 918 enccrd = NULL; 919 cmd0 |= SAFE_SA_CMD0_OP_HASH; 920 } else if (crd1->crd_alg == CRYPTO_DES_CBC || 921 crd1->crd_alg == CRYPTO_3DES_CBC || 922 crd1->crd_alg == CRYPTO_AES_CBC || 923 crd1->crd_alg == CRYPTO_NULL_CBC) { 924 maccrd = NULL; 925 enccrd = crd1; 926 cmd0 |= SAFE_SA_CMD0_OP_CRYPT; 927 } else { 928 safestats.st_badalg++; 929 err = EINVAL; 930 goto errout; 931 } 932 } else { 933 if ((crd1->crd_alg == CRYPTO_MD5_HMAC || 934 crd1->crd_alg == CRYPTO_SHA1_HMAC || 935 crd1->crd_alg == CRYPTO_NULL_HMAC) && 936 (crd2->crd_alg == CRYPTO_DES_CBC || 937 crd2->crd_alg == CRYPTO_3DES_CBC || 938 crd2->crd_alg == CRYPTO_AES_CBC || 939 crd2->crd_alg == CRYPTO_NULL_CBC) && 940 ((crd2->crd_flags & CRD_F_ENCRYPT) == 0)) { 941 maccrd = crd1; 942 enccrd = crd2; 943 } else if ((crd1->crd_alg == CRYPTO_DES_CBC || 944 crd1->crd_alg == CRYPTO_3DES_CBC || 945 crd1->crd_alg == CRYPTO_AES_CBC || 946 crd1->crd_alg == CRYPTO_NULL_CBC) && 947 (crd2->crd_alg == CRYPTO_MD5_HMAC || 948 crd2->crd_alg == CRYPTO_SHA1_HMAC || 949 crd2->crd_alg == CRYPTO_NULL_HMAC) && 950 (crd1->crd_flags & CRD_F_ENCRYPT)) { 951 enccrd = crd1; 952 maccrd = crd2; 953 } else { 954 safestats.st_badalg++; 955 err = EINVAL; 956 goto errout; 957 } 958 cmd0 |= SAFE_SA_CMD0_OP_BOTH; 959 } 960 961 if (enccrd) { 962 if (enccrd->crd_flags & CRD_F_KEY_EXPLICIT) 963 safe_setup_enckey(ses, enccrd->crd_key); 964 965 if (enccrd->crd_alg == CRYPTO_DES_CBC) { 966 cmd0 |= SAFE_SA_CMD0_DES; 967 cmd1 |= SAFE_SA_CMD1_CBC; 968 ivsize = 2*sizeof(u_int32_t); 969 } else if (enccrd->crd_alg == CRYPTO_3DES_CBC) { 970 cmd0 |= SAFE_SA_CMD0_3DES; 971 cmd1 |= SAFE_SA_CMD1_CBC; 972 ivsize = 2*sizeof(u_int32_t); 973 } else if (enccrd->crd_alg == CRYPTO_AES_CBC) { 974 cmd0 |= SAFE_SA_CMD0_AES; 975 cmd1 |= SAFE_SA_CMD1_CBC; 976 if (ses->ses_klen == 128) 977 cmd1 |= SAFE_SA_CMD1_AES128; 978 else if (ses->ses_klen == 192) 979 cmd1 |= SAFE_SA_CMD1_AES192; 980 else 981 cmd1 |= SAFE_SA_CMD1_AES256; 982 ivsize = 4*sizeof(u_int32_t); 983 } else { 984 cmd0 |= SAFE_SA_CMD0_CRYPT_NULL; 985 ivsize = 0; 986 } 987 988 /* 989 * Setup encrypt/decrypt state. When using basic ops 990 * we can't use an inline IV because hash/crypt offset 991 * must be from the end of the IV to the start of the 992 * crypt data and this leaves out the preceding header 993 * from the hash calculation. Instead we place the IV 994 * in the state record and set the hash/crypt offset to 995 * copy both the header+IV. 996 */ 997 if (enccrd->crd_flags & CRD_F_ENCRYPT) { 998 cmd0 |= SAFE_SA_CMD0_OUTBOUND; 999 1000 if (enccrd->crd_flags & CRD_F_IV_EXPLICIT) 1001 iv = enccrd->crd_iv; 1002 else 1003 iv = (caddr_t) ses->ses_iv; 1004 if ((enccrd->crd_flags & CRD_F_IV_PRESENT) == 0) { 1005 crypto_copyback(crp->crp_flags, crp->crp_buf, 1006 enccrd->crd_inject, ivsize, iv); 1007 } 1008 bcopy(iv, re->re_sastate.sa_saved_iv, ivsize); 1009 cmd0 |= SAFE_SA_CMD0_IVLD_STATE | SAFE_SA_CMD0_SAVEIV; 1010 re->re_flags |= SAFE_QFLAGS_COPYOUTIV; 1011 } else { 1012 cmd0 |= SAFE_SA_CMD0_INBOUND; 1013 1014 if (enccrd->crd_flags & CRD_F_IV_EXPLICIT) { 1015 bcopy(enccrd->crd_iv, 1016 re->re_sastate.sa_saved_iv, ivsize); 1017 } else { 1018 crypto_copydata(crp->crp_flags, crp->crp_buf, 1019 enccrd->crd_inject, ivsize, 1020 (caddr_t)re->re_sastate.sa_saved_iv); 1021 } 1022 cmd0 |= SAFE_SA_CMD0_IVLD_STATE; 1023 } 1024 /* 1025 * For basic encryption use the zero pad algorithm. 1026 * This pads results to an 8-byte boundary and 1027 * suppresses padding verification for inbound (i.e. 1028 * decrypt) operations. 1029 * 1030 * NB: Not sure if the 8-byte pad boundary is a problem. 1031 */ 1032 cmd0 |= SAFE_SA_CMD0_PAD_ZERO; 1033 1034 /* XXX assert key bufs have the same size */ 1035 bcopy(ses->ses_key, sa->sa_key, sizeof(sa->sa_key)); 1036 } 1037 1038 if (maccrd) { 1039 if (maccrd->crd_flags & CRD_F_KEY_EXPLICIT) { 1040 safe_setup_mackey(ses, maccrd->crd_alg, 1041 maccrd->crd_key, maccrd->crd_klen / 8); 1042 } 1043 1044 if (maccrd->crd_alg == CRYPTO_MD5_HMAC) { 1045 cmd0 |= SAFE_SA_CMD0_MD5; 1046 cmd1 |= SAFE_SA_CMD1_HMAC; /* NB: enable HMAC */ 1047 } else if (maccrd->crd_alg == CRYPTO_SHA1_HMAC) { 1048 cmd0 |= SAFE_SA_CMD0_SHA1; 1049 cmd1 |= SAFE_SA_CMD1_HMAC; /* NB: enable HMAC */ 1050 } else { 1051 cmd0 |= SAFE_SA_CMD0_HASH_NULL; 1052 } 1053 /* 1054 * Digest data is loaded from the SA and the hash 1055 * result is saved to the state block where we 1056 * retrieve it for return to the caller. 1057 */ 1058 /* XXX assert digest bufs have the same size */ 1059 bcopy(ses->ses_hminner, sa->sa_indigest, 1060 sizeof(sa->sa_indigest)); 1061 bcopy(ses->ses_hmouter, sa->sa_outdigest, 1062 sizeof(sa->sa_outdigest)); 1063 1064 cmd0 |= SAFE_SA_CMD0_HSLD_SA | SAFE_SA_CMD0_SAVEHASH; 1065 re->re_flags |= SAFE_QFLAGS_COPYOUTICV; 1066 } 1067 1068 if (enccrd && maccrd) { 1069 /* 1070 * The offset from hash data to the start of 1071 * crypt data is the difference in the skips. 1072 */ 1073 bypass = maccrd->crd_skip; 1074 coffset = enccrd->crd_skip - maccrd->crd_skip; 1075 if (coffset < 0) { 1076 DPRINTF(("%s: hash does not precede crypt; " 1077 "mac skip %u enc skip %u\n", 1078 __func__, maccrd->crd_skip, enccrd->crd_skip)); 1079 safestats.st_skipmismatch++; 1080 err = EINVAL; 1081 goto errout; 1082 } 1083 oplen = enccrd->crd_skip + enccrd->crd_len; 1084 if (maccrd->crd_skip + maccrd->crd_len != oplen) { 1085 DPRINTF(("%s: hash amount %u != crypt amount %u\n", 1086 __func__, maccrd->crd_skip + maccrd->crd_len, 1087 oplen)); 1088 safestats.st_lenmismatch++; 1089 err = EINVAL; 1090 goto errout; 1091 } 1092#ifdef SAFE_DEBUG 1093 if (safe_debug) { 1094 printf("mac: skip %d, len %d, inject %d\n", 1095 maccrd->crd_skip, maccrd->crd_len, 1096 maccrd->crd_inject); 1097 printf("enc: skip %d, len %d, inject %d\n", 1098 enccrd->crd_skip, enccrd->crd_len, 1099 enccrd->crd_inject); 1100 printf("bypass %d coffset %d oplen %d\n", 1101 bypass, coffset, oplen); 1102 } 1103#endif 1104 if (coffset & 3) { /* offset must be 32-bit aligned */ 1105 DPRINTF(("%s: coffset %u misaligned\n", 1106 __func__, coffset)); 1107 safestats.st_coffmisaligned++; 1108 err = EINVAL; 1109 goto errout; 1110 } 1111 coffset >>= 2; 1112 if (coffset > 255) { /* offset must be <256 dwords */ 1113 DPRINTF(("%s: coffset %u too big\n", 1114 __func__, coffset)); 1115 safestats.st_cofftoobig++; 1116 err = EINVAL; 1117 goto errout; 1118 } 1119 /* 1120 * Tell the hardware to copy the header to the output. 1121 * The header is defined as the data from the end of 1122 * the bypass to the start of data to be encrypted. 1123 * Typically this is the inline IV. Note that you need 1124 * to do this even if src+dst are the same; it appears 1125 * that w/o this bit the crypted data is written 1126 * immediately after the bypass data. 1127 */ 1128 cmd1 |= SAFE_SA_CMD1_HDRCOPY; 1129 /* 1130 * Disable IP header mutable bit handling. This is 1131 * needed to get correct HMAC calculations. 1132 */ 1133 cmd1 |= SAFE_SA_CMD1_MUTABLE; 1134 } else { 1135 if (enccrd) { 1136 bypass = enccrd->crd_skip; 1137 oplen = bypass + enccrd->crd_len; 1138 } else { 1139 bypass = maccrd->crd_skip; 1140 oplen = bypass + maccrd->crd_len; 1141 } 1142 coffset = 0; 1143 } 1144 /* XXX verify multiple of 4 when using s/g */ 1145 if (bypass > 96) { /* bypass offset must be <= 96 bytes */ 1146 DPRINTF(("%s: bypass %u too big\n", __func__, bypass)); 1147 safestats.st_bypasstoobig++; 1148 err = EINVAL; 1149 goto errout; 1150 } 1151 1152 if (bus_dmamap_create(sc->sc_srcdmat, BUS_DMA_NOWAIT, &re->re_src_map)) { 1153 safestats.st_nomap++; 1154 err = ENOMEM; 1155 goto errout; 1156 } 1157 if (crp->crp_flags & CRYPTO_F_IMBUF) { 1158 if (bus_dmamap_load_mbuf(sc->sc_srcdmat, re->re_src_map, 1159 re->re_src_m, safe_op_cb, 1160 &re->re_src, BUS_DMA_NOWAIT) != 0) { 1161 bus_dmamap_destroy(sc->sc_srcdmat, re->re_src_map); 1162 re->re_src_map = NULL; 1163 safestats.st_noload++; 1164 err = ENOMEM; 1165 goto errout; 1166 } 1167 } else if (crp->crp_flags & CRYPTO_F_IOV) { 1168 if (bus_dmamap_load_uio(sc->sc_srcdmat, re->re_src_map, 1169 re->re_src_io, safe_op_cb, 1170 &re->re_src, BUS_DMA_NOWAIT) != 0) { 1171 bus_dmamap_destroy(sc->sc_srcdmat, re->re_src_map); 1172 re->re_src_map = NULL; 1173 safestats.st_noload++; 1174 err = ENOMEM; 1175 goto errout; 1176 } 1177 } 1178 nicealign = safe_dmamap_aligned(&re->re_src); 1179 uniform = safe_dmamap_uniform(&re->re_src); 1180 1181 DPRINTF(("src nicealign %u uniform %u nsegs %u\n", 1182 nicealign, uniform, re->re_src.nsegs)); 1183 if (re->re_src.nsegs > 1) { 1184 re->re_desc.d_src = sc->sc_spalloc.dma_paddr + 1185 ((caddr_t) sc->sc_spfree - (caddr_t) sc->sc_spring); 1186 for (i = 0; i < re->re_src_nsegs; i++) { 1187 /* NB: no need to check if there's space */ 1188 pd = sc->sc_spfree; 1189 if (++(sc->sc_spfree) == sc->sc_springtop) 1190 sc->sc_spfree = sc->sc_spring; 1191 1192 KASSERT((pd->pd_flags&3) == 0 || 1193 (pd->pd_flags&3) == SAFE_PD_DONE, 1194 ("bogus source particle descriptor; flags %x", 1195 pd->pd_flags)); 1196 pd->pd_addr = re->re_src_segs[i].ds_addr; 1197 pd->pd_size = re->re_src_segs[i].ds_len; 1198 pd->pd_flags = SAFE_PD_READY; 1199 } 1200 cmd0 |= SAFE_SA_CMD0_IGATHER; 1201 } else { 1202 /* 1203 * No need for gather, reference the operand directly. 1204 */ 1205 re->re_desc.d_src = re->re_src_segs[0].ds_addr; 1206 } 1207 1208 if (enccrd == NULL && maccrd != NULL) { 1209 /* 1210 * Hash op; no destination needed. 1211 */ 1212 } else { 1213 if (crp->crp_flags & CRYPTO_F_IOV) { 1214 if (!nicealign) { 1215 safestats.st_iovmisaligned++; 1216 err = EINVAL; 1217 goto errout; 1218 } 1219 if (uniform != 1) { 1220 /* 1221 * Source is not suitable for direct use as 1222 * the destination. Create a new scatter/gather 1223 * list based on the destination requirements 1224 * and check if that's ok. 1225 */ 1226 if (bus_dmamap_create(sc->sc_dstdmat, 1227 BUS_DMA_NOWAIT, &re->re_dst_map)) { 1228 safestats.st_nomap++; 1229 err = ENOMEM; 1230 goto errout; 1231 } 1232 if (bus_dmamap_load_uio(sc->sc_dstdmat, 1233 re->re_dst_map, re->re_dst_io, 1234 safe_op_cb, &re->re_dst, 1235 BUS_DMA_NOWAIT) != 0) { 1236 bus_dmamap_destroy(sc->sc_dstdmat, 1237 re->re_dst_map); 1238 re->re_dst_map = NULL; 1239 safestats.st_noload++; 1240 err = ENOMEM; 1241 goto errout; 1242 } 1243 uniform = safe_dmamap_uniform(&re->re_dst); 1244 if (!uniform) { 1245 /* 1246 * There's no way to handle the DMA 1247 * requirements with this uio. We 1248 * could create a separate DMA area for 1249 * the result and then copy it back, 1250 * but for now we just bail and return 1251 * an error. Note that uio requests 1252 * > SAFE_MAX_DSIZE are handled because 1253 * the DMA map and segment list for the 1254 * destination wil result in a 1255 * destination particle list that does 1256 * the necessary scatter DMA. 1257 */ 1258 safestats.st_iovnotuniform++; 1259 err = EINVAL; 1260 goto errout; 1261 } 1262 } else 1263 re->re_dst = re->re_src; 1264 } else if (crp->crp_flags & CRYPTO_F_IMBUF) { 1265 if (nicealign && uniform == 1) { 1266 /* 1267 * Source layout is suitable for direct 1268 * sharing of the DMA map and segment list. 1269 */ 1270 re->re_dst = re->re_src; 1271 } else if (nicealign && uniform == 2) { 1272 /* 1273 * The source is properly aligned but requires a 1274 * different particle list to handle DMA of the 1275 * result. Create a new map and do the load to 1276 * create the segment list. The particle 1277 * descriptor setup code below will handle the 1278 * rest. 1279 */ 1280 if (bus_dmamap_create(sc->sc_dstdmat, 1281 BUS_DMA_NOWAIT, &re->re_dst_map)) { 1282 safestats.st_nomap++; 1283 err = ENOMEM; 1284 goto errout; 1285 } 1286 if (bus_dmamap_load_mbuf(sc->sc_dstdmat, 1287 re->re_dst_map, re->re_dst_m, 1288 safe_op_cb, &re->re_dst, 1289 BUS_DMA_NOWAIT) != 0) { 1290 bus_dmamap_destroy(sc->sc_dstdmat, 1291 re->re_dst_map); 1292 re->re_dst_map = NULL; 1293 safestats.st_noload++; 1294 err = ENOMEM; 1295 goto errout; 1296 } 1297 } else { /* !(aligned and/or uniform) */ 1298 int totlen, len; 1299 struct mbuf *m, *top, **mp; 1300 1301 /* 1302 * DMA constraints require that we allocate a 1303 * new mbuf chain for the destination. We 1304 * allocate an entire new set of mbufs of 1305 * optimal/required size and then tell the 1306 * hardware to copy any bits that are not 1307 * created as a byproduct of the operation. 1308 */ 1309 if (!nicealign) 1310 safestats.st_unaligned++; 1311 if (!uniform) 1312 safestats.st_notuniform++; 1313 totlen = re->re_src_mapsize; 1314 if (re->re_src_m->m_flags & M_PKTHDR) { 1315 len = MHLEN; 1316 MGETHDR(m, M_NOWAIT, MT_DATA); 1317 if (m && !m_dup_pkthdr(m, re->re_src_m, 1318 M_NOWAIT)) { 1319 m_free(m); 1320 m = NULL; 1321 } 1322 } else { 1323 len = MLEN; 1324 MGET(m, M_NOWAIT, MT_DATA); 1325 } 1326 if (m == NULL) { 1327 safestats.st_nombuf++; 1328 err = sc->sc_nqchip ? ERESTART : ENOMEM; 1329 goto errout; 1330 } 1331 if (totlen >= MINCLSIZE) { 1332 if (!(MCLGET(m, M_NOWAIT))) { 1333 m_free(m); 1334 safestats.st_nomcl++; 1335 err = sc->sc_nqchip ? 1336 ERESTART : ENOMEM; 1337 goto errout; 1338 } 1339 len = MCLBYTES; 1340 } 1341 m->m_len = len; 1342 top = NULL; 1343 mp = ⊤ 1344 1345 while (totlen > 0) { 1346 if (top) { 1347 MGET(m, M_NOWAIT, MT_DATA); 1348 if (m == NULL) { 1349 m_freem(top); 1350 safestats.st_nombuf++; 1351 err = sc->sc_nqchip ? 1352 ERESTART : ENOMEM; 1353 goto errout; 1354 } 1355 len = MLEN; 1356 } 1357 if (top && totlen >= MINCLSIZE) { 1358 if (!(MCLGET(m, M_NOWAIT))) { 1359 *mp = m; 1360 m_freem(top); 1361 safestats.st_nomcl++; 1362 err = sc->sc_nqchip ? 1363 ERESTART : ENOMEM; 1364 goto errout; 1365 } 1366 len = MCLBYTES; 1367 } 1368 m->m_len = len = min(totlen, len); 1369 totlen -= len; 1370 *mp = m; 1371 mp = &m->m_next; 1372 } 1373 re->re_dst_m = top; 1374 if (bus_dmamap_create(sc->sc_dstdmat, 1375 BUS_DMA_NOWAIT, &re->re_dst_map) != 0) { 1376 safestats.st_nomap++; 1377 err = ENOMEM; 1378 goto errout; 1379 } 1380 if (bus_dmamap_load_mbuf(sc->sc_dstdmat, 1381 re->re_dst_map, re->re_dst_m, 1382 safe_op_cb, &re->re_dst, 1383 BUS_DMA_NOWAIT) != 0) { 1384 bus_dmamap_destroy(sc->sc_dstdmat, 1385 re->re_dst_map); 1386 re->re_dst_map = NULL; 1387 safestats.st_noload++; 1388 err = ENOMEM; 1389 goto errout; 1390 } 1391 if (re->re_src.mapsize > oplen) { 1392 /* 1393 * There's data following what the 1394 * hardware will copy for us. If this 1395 * isn't just the ICV (that's going to 1396 * be written on completion), copy it 1397 * to the new mbufs 1398 */ 1399 if (!(maccrd && 1400 (re->re_src.mapsize-oplen) == 12 && 1401 maccrd->crd_inject == oplen)) 1402 safe_mcopy(re->re_src_m, 1403 re->re_dst_m, 1404 oplen); 1405 else 1406 safestats.st_noicvcopy++; 1407 } 1408 } 1409 } else { 1410 safestats.st_badflags++; 1411 err = EINVAL; 1412 goto errout; 1413 } 1414 1415 if (re->re_dst.nsegs > 1) { 1416 re->re_desc.d_dst = sc->sc_dpalloc.dma_paddr + 1417 ((caddr_t) sc->sc_dpfree - (caddr_t) sc->sc_dpring); 1418 for (i = 0; i < re->re_dst_nsegs; i++) { 1419 pd = sc->sc_dpfree; 1420 KASSERT((pd->pd_flags&3) == 0 || 1421 (pd->pd_flags&3) == SAFE_PD_DONE, 1422 ("bogus dest particle descriptor; flags %x", 1423 pd->pd_flags)); 1424 if (++(sc->sc_dpfree) == sc->sc_dpringtop) 1425 sc->sc_dpfree = sc->sc_dpring; 1426 pd->pd_addr = re->re_dst_segs[i].ds_addr; 1427 pd->pd_flags = SAFE_PD_READY; 1428 } 1429 cmd0 |= SAFE_SA_CMD0_OSCATTER; 1430 } else { 1431 /* 1432 * No need for scatter, reference the operand directly. 1433 */ 1434 re->re_desc.d_dst = re->re_dst_segs[0].ds_addr; 1435 } 1436 } 1437 1438 /* 1439 * All done with setup; fillin the SA command words 1440 * and the packet engine descriptor. The operation 1441 * is now ready for submission to the hardware. 1442 */ 1443 sa->sa_cmd0 = cmd0 | SAFE_SA_CMD0_IPCI | SAFE_SA_CMD0_OPCI; 1444 sa->sa_cmd1 = cmd1 1445 | (coffset << SAFE_SA_CMD1_OFFSET_S) 1446 | SAFE_SA_CMD1_SAREV1 /* Rev 1 SA data structure */ 1447 | SAFE_SA_CMD1_SRPCI 1448 ; 1449 /* 1450 * NB: the order of writes is important here. In case the 1451 * chip is scanning the ring because of an outstanding request 1452 * it might nab this one too. In that case we need to make 1453 * sure the setup is complete before we write the length 1454 * field of the descriptor as it signals the descriptor is 1455 * ready for processing. 1456 */ 1457 re->re_desc.d_csr = SAFE_PE_CSR_READY | SAFE_PE_CSR_SAPCI; 1458 if (maccrd) 1459 re->re_desc.d_csr |= SAFE_PE_CSR_LOADSA | SAFE_PE_CSR_HASHFINAL; 1460 re->re_desc.d_len = oplen 1461 | SAFE_PE_LEN_READY 1462 | (bypass << SAFE_PE_LEN_BYPASS_S) 1463 ; 1464 1465 safestats.st_ipackets++; 1466 safestats.st_ibytes += oplen; 1467 1468 if (++(sc->sc_front) == sc->sc_ringtop) 1469 sc->sc_front = sc->sc_ring; 1470 1471 /* XXX honor batching */ 1472 safe_feed(sc, re); 1473 mtx_unlock(&sc->sc_ringmtx); 1474 return (0); 1475 1476errout: 1477 if ((re->re_dst_m != NULL) && (re->re_src_m != re->re_dst_m)) 1478 m_freem(re->re_dst_m); 1479 1480 if (re->re_dst_map != NULL && re->re_dst_map != re->re_src_map) { 1481 bus_dmamap_unload(sc->sc_dstdmat, re->re_dst_map); 1482 bus_dmamap_destroy(sc->sc_dstdmat, re->re_dst_map); 1483 } 1484 if (re->re_src_map != NULL) { 1485 bus_dmamap_unload(sc->sc_srcdmat, re->re_src_map); 1486 bus_dmamap_destroy(sc->sc_srcdmat, re->re_src_map); 1487 } 1488 mtx_unlock(&sc->sc_ringmtx); 1489 if (err != ERESTART) { 1490 crp->crp_etype = err; 1491 crypto_done(crp); 1492 } else { 1493 sc->sc_needwakeup |= CRYPTO_SYMQ; 1494 } 1495 return (err); 1496} 1497 1498static void 1499safe_callback(struct safe_softc *sc, struct safe_ringentry *re) 1500{ 1501 struct cryptop *crp = (struct cryptop *)re->re_crp; 1502 struct cryptodesc *crd; 1503 1504 safestats.st_opackets++; 1505 safestats.st_obytes += re->re_dst.mapsize; 1506 1507 safe_dma_sync(&sc->sc_ringalloc, 1508 BUS_DMASYNC_POSTREAD|BUS_DMASYNC_POSTWRITE); 1509 if (re->re_desc.d_csr & SAFE_PE_CSR_STATUS) { 1510 device_printf(sc->sc_dev, "csr 0x%x cmd0 0x%x cmd1 0x%x\n", 1511 re->re_desc.d_csr, 1512 re->re_sa.sa_cmd0, re->re_sa.sa_cmd1); 1513 safestats.st_peoperr++; 1514 crp->crp_etype = EIO; /* something more meaningful? */ 1515 } 1516 if (re->re_dst_map != NULL && re->re_dst_map != re->re_src_map) { 1517 bus_dmamap_sync(sc->sc_dstdmat, re->re_dst_map, 1518 BUS_DMASYNC_POSTREAD); 1519 bus_dmamap_unload(sc->sc_dstdmat, re->re_dst_map); 1520 bus_dmamap_destroy(sc->sc_dstdmat, re->re_dst_map); 1521 } 1522 bus_dmamap_sync(sc->sc_srcdmat, re->re_src_map, BUS_DMASYNC_POSTWRITE); 1523 bus_dmamap_unload(sc->sc_srcdmat, re->re_src_map); 1524 bus_dmamap_destroy(sc->sc_srcdmat, re->re_src_map); 1525 1526 /* 1527 * If result was written to a differet mbuf chain, swap 1528 * it in as the return value and reclaim the original. 1529 */ 1530 if ((crp->crp_flags & CRYPTO_F_IMBUF) && re->re_src_m != re->re_dst_m) { 1531 m_freem(re->re_src_m); 1532 crp->crp_buf = (caddr_t)re->re_dst_m; 1533 } 1534 1535 if (re->re_flags & SAFE_QFLAGS_COPYOUTIV) { 1536 /* copy out IV for future use */ 1537 for (crd = crp->crp_desc; crd; crd = crd->crd_next) { 1538 int ivsize; 1539 1540 if (crd->crd_alg == CRYPTO_DES_CBC || 1541 crd->crd_alg == CRYPTO_3DES_CBC) { 1542 ivsize = 2*sizeof(u_int32_t); 1543 } else if (crd->crd_alg == CRYPTO_AES_CBC) { 1544 ivsize = 4*sizeof(u_int32_t); 1545 } else 1546 continue; 1547 crypto_copydata(crp->crp_flags, crp->crp_buf, 1548 crd->crd_skip + crd->crd_len - ivsize, ivsize, 1549 (caddr_t)sc->sc_sessions[re->re_sesn].ses_iv); 1550 break; 1551 } 1552 } 1553 1554 if (re->re_flags & SAFE_QFLAGS_COPYOUTICV) { 1555 /* copy out ICV result */ 1556 for (crd = crp->crp_desc; crd; crd = crd->crd_next) { 1557 if (!(crd->crd_alg == CRYPTO_MD5_HMAC || 1558 crd->crd_alg == CRYPTO_SHA1_HMAC || 1559 crd->crd_alg == CRYPTO_NULL_HMAC)) 1560 continue; 1561 if (crd->crd_alg == CRYPTO_SHA1_HMAC) { 1562 /* 1563 * SHA-1 ICV's are byte-swapped; fix 'em up 1564 * before copy them to their destination. 1565 */ 1566 re->re_sastate.sa_saved_indigest[0] = 1567 bswap32(re->re_sastate.sa_saved_indigest[0]); 1568 re->re_sastate.sa_saved_indigest[1] = 1569 bswap32(re->re_sastate.sa_saved_indigest[1]); 1570 re->re_sastate.sa_saved_indigest[2] = 1571 bswap32(re->re_sastate.sa_saved_indigest[2]); 1572 } 1573 crypto_copyback(crp->crp_flags, crp->crp_buf, 1574 crd->crd_inject, 1575 sc->sc_sessions[re->re_sesn].ses_mlen, 1576 (caddr_t)re->re_sastate.sa_saved_indigest); 1577 break; 1578 } 1579 } 1580 crypto_done(crp); 1581} 1582 1583/* 1584 * Copy all data past offset from srcm to dstm. 1585 */ 1586static void 1587safe_mcopy(struct mbuf *srcm, struct mbuf *dstm, u_int offset) 1588{ 1589 u_int j, dlen, slen; 1590 caddr_t dptr, sptr; 1591 1592 /* 1593 * Advance src and dst to offset. 1594 */ 1595 j = offset; 1596 while (j >= 0) { 1597 if (srcm->m_len > j) 1598 break; 1599 j -= srcm->m_len; 1600 srcm = srcm->m_next; 1601 if (srcm == NULL) 1602 return; 1603 } 1604 sptr = mtod(srcm, caddr_t) + j; 1605 slen = srcm->m_len - j; 1606 1607 j = offset; 1608 while (j >= 0) { 1609 if (dstm->m_len > j) 1610 break; 1611 j -= dstm->m_len; 1612 dstm = dstm->m_next; 1613 if (dstm == NULL) 1614 return; 1615 } 1616 dptr = mtod(dstm, caddr_t) + j; 1617 dlen = dstm->m_len - j; 1618 1619 /* 1620 * Copy everything that remains. 1621 */ 1622 for (;;) { 1623 j = min(slen, dlen); 1624 bcopy(sptr, dptr, j); 1625 if (slen == j) { 1626 srcm = srcm->m_next; 1627 if (srcm == NULL) 1628 return; 1629 sptr = srcm->m_data; 1630 slen = srcm->m_len; 1631 } else 1632 sptr += j, slen -= j; 1633 if (dlen == j) { 1634 dstm = dstm->m_next; 1635 if (dstm == NULL) 1636 return; 1637 dptr = dstm->m_data; 1638 dlen = dstm->m_len; 1639 } else 1640 dptr += j, dlen -= j; 1641 } 1642} 1643 1644#ifndef SAFE_NO_RNG 1645#define SAFE_RNG_MAXWAIT 1000 1646 1647static void 1648safe_rng_init(struct safe_softc *sc) 1649{ 1650 u_int32_t w, v; 1651 int i; 1652 1653 WRITE_REG(sc, SAFE_RNG_CTRL, 0); 1654 /* use default value according to the manual */ 1655 WRITE_REG(sc, SAFE_RNG_CNFG, 0x834); /* magic from SafeNet */ 1656 WRITE_REG(sc, SAFE_RNG_ALM_CNT, 0); 1657 1658 /* 1659 * There is a bug in rev 1.0 of the 1140 that when the RNG 1660 * is brought out of reset the ready status flag does not 1661 * work until the RNG has finished its internal initialization. 1662 * 1663 * So in order to determine the device is through its 1664 * initialization we must read the data register, using the 1665 * status reg in the read in case it is initialized. Then read 1666 * the data register until it changes from the first read. 1667 * Once it changes read the data register until it changes 1668 * again. At this time the RNG is considered initialized. 1669 * This could take between 750ms - 1000ms in time. 1670 */ 1671 i = 0; 1672 w = READ_REG(sc, SAFE_RNG_OUT); 1673 do { 1674 v = READ_REG(sc, SAFE_RNG_OUT); 1675 if (v != w) { 1676 w = v; 1677 break; 1678 } 1679 DELAY(10); 1680 } while (++i < SAFE_RNG_MAXWAIT); 1681 1682 /* Wait Until data changes again */ 1683 i = 0; 1684 do { 1685 v = READ_REG(sc, SAFE_RNG_OUT); 1686 if (v != w) 1687 break; 1688 DELAY(10); 1689 } while (++i < SAFE_RNG_MAXWAIT); 1690} 1691 1692static __inline void 1693safe_rng_disable_short_cycle(struct safe_softc *sc) 1694{ 1695 WRITE_REG(sc, SAFE_RNG_CTRL, 1696 READ_REG(sc, SAFE_RNG_CTRL) &~ SAFE_RNG_CTRL_SHORTEN); 1697} 1698 1699static __inline void 1700safe_rng_enable_short_cycle(struct safe_softc *sc) 1701{ 1702 WRITE_REG(sc, SAFE_RNG_CTRL, 1703 READ_REG(sc, SAFE_RNG_CTRL) | SAFE_RNG_CTRL_SHORTEN); 1704} 1705 1706static __inline u_int32_t 1707safe_rng_read(struct safe_softc *sc) 1708{ 1709 int i; 1710 1711 i = 0; 1712 while (READ_REG(sc, SAFE_RNG_STAT) != 0 && ++i < SAFE_RNG_MAXWAIT) 1713 ; 1714 return READ_REG(sc, SAFE_RNG_OUT); 1715} 1716 1717static void 1718safe_rng(void *arg) 1719{ 1720 struct safe_softc *sc = arg; 1721 u_int32_t buf[SAFE_RNG_MAXBUFSIZ]; /* NB: maybe move to softc */ 1722 u_int maxwords; 1723 int i; 1724 1725 safestats.st_rng++; 1726 /* 1727 * Fetch the next block of data. 1728 */ 1729 maxwords = safe_rngbufsize; 1730 if (maxwords > SAFE_RNG_MAXBUFSIZ) 1731 maxwords = SAFE_RNG_MAXBUFSIZ; 1732retry: 1733 for (i = 0; i < maxwords; i++) 1734 buf[i] = safe_rng_read(sc); 1735 /* 1736 * Check the comparator alarm count and reset the h/w if 1737 * it exceeds our threshold. This guards against the 1738 * hardware oscillators resonating with external signals. 1739 */ 1740 if (READ_REG(sc, SAFE_RNG_ALM_CNT) > safe_rngmaxalarm) { 1741 u_int32_t freq_inc, w; 1742 1743 DPRINTF(("%s: alarm count %u exceeds threshold %u\n", __func__, 1744 READ_REG(sc, SAFE_RNG_ALM_CNT), safe_rngmaxalarm)); 1745 safestats.st_rngalarm++; 1746 safe_rng_enable_short_cycle(sc); 1747 freq_inc = 18; 1748 for (i = 0; i < 64; i++) { 1749 w = READ_REG(sc, SAFE_RNG_CNFG); 1750 freq_inc = ((w + freq_inc) & 0x3fL); 1751 w = ((w & ~0x3fL) | freq_inc); 1752 WRITE_REG(sc, SAFE_RNG_CNFG, w); 1753 1754 WRITE_REG(sc, SAFE_RNG_ALM_CNT, 0); 1755 1756 (void) safe_rng_read(sc); 1757 DELAY(25); 1758 1759 if (READ_REG(sc, SAFE_RNG_ALM_CNT) == 0) { 1760 safe_rng_disable_short_cycle(sc); 1761 goto retry; 1762 } 1763 freq_inc = 1; 1764 } 1765 safe_rng_disable_short_cycle(sc); 1766 } else 1767 WRITE_REG(sc, SAFE_RNG_ALM_CNT, 0); 1768 1769 (*sc->sc_harvest)(sc->sc_rndtest, buf, maxwords*sizeof (u_int32_t)); 1770 callout_reset(&sc->sc_rngto, 1771 hz * (safe_rnginterval ? safe_rnginterval : 1), safe_rng, sc); 1772} 1773#endif /* SAFE_NO_RNG */ 1774 1775static void 1776safe_dmamap_cb(void *arg, bus_dma_segment_t *segs, int nseg, int error) 1777{ 1778 bus_addr_t *paddr = (bus_addr_t*) arg; 1779 *paddr = segs->ds_addr; 1780} 1781 1782static int 1783safe_dma_malloc( 1784 struct safe_softc *sc, 1785 bus_size_t size, 1786 struct safe_dma_alloc *dma, 1787 int mapflags 1788) 1789{ 1790 int r; 1791 1792 r = bus_dma_tag_create(bus_get_dma_tag(sc->sc_dev), /* parent */ 1793 sizeof(u_int32_t), 0, /* alignment, bounds */ 1794 BUS_SPACE_MAXADDR_32BIT, /* lowaddr */ 1795 BUS_SPACE_MAXADDR, /* highaddr */ 1796 NULL, NULL, /* filter, filterarg */ 1797 size, /* maxsize */ 1798 1, /* nsegments */ 1799 size, /* maxsegsize */ 1800 BUS_DMA_ALLOCNOW, /* flags */ 1801 NULL, NULL, /* locking */ 1802 &dma->dma_tag); 1803 if (r != 0) { 1804 device_printf(sc->sc_dev, "safe_dma_malloc: " 1805 "bus_dma_tag_create failed; error %u\n", r); 1806 goto fail_0; 1807 } 1808 1809 r = bus_dmamem_alloc(dma->dma_tag, (void**) &dma->dma_vaddr, 1810 BUS_DMA_NOWAIT, &dma->dma_map); 1811 if (r != 0) { 1812 device_printf(sc->sc_dev, "safe_dma_malloc: " 1813 "bus_dmammem_alloc failed; size %ju, error %u\n", 1814 (uintmax_t)size, r); 1815 goto fail_1; 1816 } 1817 1818 r = bus_dmamap_load(dma->dma_tag, dma->dma_map, dma->dma_vaddr, 1819 size, 1820 safe_dmamap_cb, 1821 &dma->dma_paddr, 1822 mapflags | BUS_DMA_NOWAIT); 1823 if (r != 0) { 1824 device_printf(sc->sc_dev, "safe_dma_malloc: " 1825 "bus_dmamap_load failed; error %u\n", r); 1826 goto fail_2; 1827 } 1828 1829 dma->dma_size = size; 1830 return (0); 1831 1832 bus_dmamap_unload(dma->dma_tag, dma->dma_map); 1833fail_2: 1834 bus_dmamem_free(dma->dma_tag, dma->dma_vaddr, dma->dma_map); 1835fail_1: 1836 bus_dma_tag_destroy(dma->dma_tag); 1837fail_0: 1838 dma->dma_tag = NULL; 1839 return (r); 1840} 1841 1842static void 1843safe_dma_free(struct safe_softc *sc, struct safe_dma_alloc *dma) 1844{ 1845 bus_dmamap_unload(dma->dma_tag, dma->dma_map); 1846 bus_dmamem_free(dma->dma_tag, dma->dma_vaddr, dma->dma_map); 1847 bus_dma_tag_destroy(dma->dma_tag); 1848} 1849 1850/* 1851 * Resets the board. Values in the regesters are left as is 1852 * from the reset (i.e. initial values are assigned elsewhere). 1853 */ 1854static void 1855safe_reset_board(struct safe_softc *sc) 1856{ 1857 u_int32_t v; 1858 /* 1859 * Reset the device. The manual says no delay 1860 * is needed between marking and clearing reset. 1861 */ 1862 v = READ_REG(sc, SAFE_PE_DMACFG) &~ 1863 (SAFE_PE_DMACFG_PERESET | SAFE_PE_DMACFG_PDRRESET | 1864 SAFE_PE_DMACFG_SGRESET); 1865 WRITE_REG(sc, SAFE_PE_DMACFG, v 1866 | SAFE_PE_DMACFG_PERESET 1867 | SAFE_PE_DMACFG_PDRRESET 1868 | SAFE_PE_DMACFG_SGRESET); 1869 WRITE_REG(sc, SAFE_PE_DMACFG, v); 1870} 1871 1872/* 1873 * Initialize registers we need to touch only once. 1874 */ 1875static void 1876safe_init_board(struct safe_softc *sc) 1877{ 1878 u_int32_t v, dwords; 1879 1880 v = READ_REG(sc, SAFE_PE_DMACFG); 1881 v &=~ SAFE_PE_DMACFG_PEMODE; 1882 v |= SAFE_PE_DMACFG_FSENA /* failsafe enable */ 1883 | SAFE_PE_DMACFG_GPRPCI /* gather ring on PCI */ 1884 | SAFE_PE_DMACFG_SPRPCI /* scatter ring on PCI */ 1885 | SAFE_PE_DMACFG_ESDESC /* endian-swap descriptors */ 1886 | SAFE_PE_DMACFG_ESSA /* endian-swap SA's */ 1887 | SAFE_PE_DMACFG_ESPDESC /* endian-swap part. desc's */ 1888 ; 1889 WRITE_REG(sc, SAFE_PE_DMACFG, v); 1890#if 0 1891 /* XXX select byte swap based on host byte order */ 1892 WRITE_REG(sc, SAFE_ENDIAN, 0x1b); 1893#endif 1894 if (sc->sc_chiprev == SAFE_REV(1,0)) { 1895 /* 1896 * Avoid large PCI DMA transfers. Rev 1.0 has a bug where 1897 * "target mode transfers" done while the chip is DMA'ing 1898 * >1020 bytes cause the hardware to lockup. To avoid this 1899 * we reduce the max PCI transfer size and use small source 1900 * particle descriptors (<= 256 bytes). 1901 */ 1902 WRITE_REG(sc, SAFE_DMA_CFG, 256); 1903 device_printf(sc->sc_dev, 1904 "Reduce max DMA size to %u words for rev %u.%u WAR\n", 1905 (READ_REG(sc, SAFE_DMA_CFG)>>2) & 0xff, 1906 SAFE_REV_MAJ(sc->sc_chiprev), 1907 SAFE_REV_MIN(sc->sc_chiprev)); 1908 } 1909 1910 /* NB: operands+results are overlaid */ 1911 WRITE_REG(sc, SAFE_PE_PDRBASE, sc->sc_ringalloc.dma_paddr); 1912 WRITE_REG(sc, SAFE_PE_RDRBASE, sc->sc_ringalloc.dma_paddr); 1913 /* 1914 * Configure ring entry size and number of items in the ring. 1915 */ 1916 KASSERT((sizeof(struct safe_ringentry) % sizeof(u_int32_t)) == 0, 1917 ("PE ring entry not 32-bit aligned!")); 1918 dwords = sizeof(struct safe_ringentry) / sizeof(u_int32_t); 1919 WRITE_REG(sc, SAFE_PE_RINGCFG, 1920 (dwords << SAFE_PE_RINGCFG_OFFSET_S) | SAFE_MAX_NQUEUE); 1921 WRITE_REG(sc, SAFE_PE_RINGPOLL, 0); /* disable polling */ 1922 1923 WRITE_REG(sc, SAFE_PE_GRNGBASE, sc->sc_spalloc.dma_paddr); 1924 WRITE_REG(sc, SAFE_PE_SRNGBASE, sc->sc_dpalloc.dma_paddr); 1925 WRITE_REG(sc, SAFE_PE_PARTSIZE, 1926 (SAFE_TOTAL_DPART<<16) | SAFE_TOTAL_SPART); 1927 /* 1928 * NB: destination particles are fixed size. We use 1929 * an mbuf cluster and require all results go to 1930 * clusters or smaller. 1931 */ 1932 WRITE_REG(sc, SAFE_PE_PARTCFG, SAFE_MAX_DSIZE); 1933 1934 /* it's now safe to enable PE mode, do it */ 1935 WRITE_REG(sc, SAFE_PE_DMACFG, v | SAFE_PE_DMACFG_PEMODE); 1936 1937 /* 1938 * Configure hardware to use level-triggered interrupts and 1939 * to interrupt after each descriptor is processed. 1940 */ 1941 WRITE_REG(sc, SAFE_HI_CFG, SAFE_HI_CFG_LEVEL); 1942 WRITE_REG(sc, SAFE_HI_DESC_CNT, 1); 1943 WRITE_REG(sc, SAFE_HI_MASK, SAFE_INT_PE_DDONE | SAFE_INT_PE_ERROR); 1944} 1945 1946/* 1947 * Init PCI registers 1948 */ 1949static void 1950safe_init_pciregs(device_t dev) 1951{ 1952} 1953 1954/* 1955 * Clean up after a chip crash. 1956 * It is assumed that the caller in splimp() 1957 */ 1958static void 1959safe_cleanchip(struct safe_softc *sc) 1960{ 1961 1962 if (sc->sc_nqchip != 0) { 1963 struct safe_ringentry *re = sc->sc_back; 1964 1965 while (re != sc->sc_front) { 1966 if (re->re_desc.d_csr != 0) 1967 safe_free_entry(sc, re); 1968 if (++re == sc->sc_ringtop) 1969 re = sc->sc_ring; 1970 } 1971 sc->sc_back = re; 1972 sc->sc_nqchip = 0; 1973 } 1974} 1975 1976/* 1977 * free a safe_q 1978 * It is assumed that the caller is within splimp(). 1979 */ 1980static int 1981safe_free_entry(struct safe_softc *sc, struct safe_ringentry *re) 1982{ 1983 struct cryptop *crp; 1984 1985 /* 1986 * Free header MCR 1987 */ 1988 if ((re->re_dst_m != NULL) && (re->re_src_m != re->re_dst_m)) 1989 m_freem(re->re_dst_m); 1990 1991 crp = (struct cryptop *)re->re_crp; 1992 1993 re->re_desc.d_csr = 0; 1994 1995 crp->crp_etype = EFAULT; 1996 crypto_done(crp); 1997 return(0); 1998} 1999 2000/* 2001 * Routine to reset the chip and clean up. 2002 * It is assumed that the caller is in splimp() 2003 */ 2004static void 2005safe_totalreset(struct safe_softc *sc) 2006{ 2007 safe_reset_board(sc); 2008 safe_init_board(sc); 2009 safe_cleanchip(sc); 2010} 2011 2012/* 2013 * Is the operand suitable aligned for direct DMA. Each 2014 * segment must be aligned on a 32-bit boundary and all 2015 * but the last segment must be a multiple of 4 bytes. 2016 */ 2017static int 2018safe_dmamap_aligned(const struct safe_operand *op) 2019{ 2020 int i; 2021 2022 for (i = 0; i < op->nsegs; i++) { 2023 if (op->segs[i].ds_addr & 3) 2024 return (0); 2025 if (i != (op->nsegs - 1) && (op->segs[i].ds_len & 3)) 2026 return (0); 2027 } 2028 return (1); 2029} 2030 2031/* 2032 * Is the operand suitable for direct DMA as the destination 2033 * of an operation. The hardware requires that each ``particle'' 2034 * but the last in an operation result have the same size. We 2035 * fix that size at SAFE_MAX_DSIZE bytes. This routine returns 2036 * 0 if some segment is not a multiple of of this size, 1 if all 2037 * segments are exactly this size, or 2 if segments are at worst 2038 * a multple of this size. 2039 */ 2040static int 2041safe_dmamap_uniform(const struct safe_operand *op) 2042{ 2043 int result = 1; 2044 2045 if (op->nsegs > 0) { 2046 int i; 2047 2048 for (i = 0; i < op->nsegs-1; i++) { 2049 if (op->segs[i].ds_len % SAFE_MAX_DSIZE) 2050 return (0); 2051 if (op->segs[i].ds_len != SAFE_MAX_DSIZE) 2052 result = 2; 2053 } 2054 } 2055 return (result); 2056} 2057 2058#ifdef SAFE_DEBUG 2059static void 2060safe_dump_dmastatus(struct safe_softc *sc, const char *tag) 2061{ 2062 printf("%s: ENDIAN 0x%x SRC 0x%x DST 0x%x STAT 0x%x\n" 2063 , tag 2064 , READ_REG(sc, SAFE_DMA_ENDIAN) 2065 , READ_REG(sc, SAFE_DMA_SRCADDR) 2066 , READ_REG(sc, SAFE_DMA_DSTADDR) 2067 , READ_REG(sc, SAFE_DMA_STAT) 2068 ); 2069} 2070 2071static void 2072safe_dump_intrstate(struct safe_softc *sc, const char *tag) 2073{ 2074 printf("%s: HI_CFG 0x%x HI_MASK 0x%x HI_DESC_CNT 0x%x HU_STAT 0x%x HM_STAT 0x%x\n" 2075 , tag 2076 , READ_REG(sc, SAFE_HI_CFG) 2077 , READ_REG(sc, SAFE_HI_MASK) 2078 , READ_REG(sc, SAFE_HI_DESC_CNT) 2079 , READ_REG(sc, SAFE_HU_STAT) 2080 , READ_REG(sc, SAFE_HM_STAT) 2081 ); 2082} 2083 2084static void 2085safe_dump_ringstate(struct safe_softc *sc, const char *tag) 2086{ 2087 u_int32_t estat = READ_REG(sc, SAFE_PE_ERNGSTAT); 2088 2089 /* NB: assume caller has lock on ring */ 2090 printf("%s: ERNGSTAT %x (next %u) back %lu front %lu\n", 2091 tag, 2092 estat, (estat >> SAFE_PE_ERNGSTAT_NEXT_S), 2093 (unsigned long)(sc->sc_back - sc->sc_ring), 2094 (unsigned long)(sc->sc_front - sc->sc_ring)); 2095} 2096 2097static void 2098safe_dump_request(struct safe_softc *sc, const char* tag, struct safe_ringentry *re) 2099{ 2100 int ix, nsegs; 2101 2102 ix = re - sc->sc_ring; 2103 printf("%s: %p (%u): csr %x src %x dst %x sa %x len %x\n" 2104 , tag 2105 , re, ix 2106 , re->re_desc.d_csr 2107 , re->re_desc.d_src 2108 , re->re_desc.d_dst 2109 , re->re_desc.d_sa 2110 , re->re_desc.d_len 2111 ); 2112 if (re->re_src.nsegs > 1) { 2113 ix = (re->re_desc.d_src - sc->sc_spalloc.dma_paddr) / 2114 sizeof(struct safe_pdesc); 2115 for (nsegs = re->re_src.nsegs; nsegs; nsegs--) { 2116 printf(" spd[%u] %p: %p size %u flags %x" 2117 , ix, &sc->sc_spring[ix] 2118 , (caddr_t)(uintptr_t) sc->sc_spring[ix].pd_addr 2119 , sc->sc_spring[ix].pd_size 2120 , sc->sc_spring[ix].pd_flags 2121 ); 2122 if (sc->sc_spring[ix].pd_size == 0) 2123 printf(" (zero!)"); 2124 printf("\n"); 2125 if (++ix == SAFE_TOTAL_SPART) 2126 ix = 0; 2127 } 2128 } 2129 if (re->re_dst.nsegs > 1) { 2130 ix = (re->re_desc.d_dst - sc->sc_dpalloc.dma_paddr) / 2131 sizeof(struct safe_pdesc); 2132 for (nsegs = re->re_dst.nsegs; nsegs; nsegs--) { 2133 printf(" dpd[%u] %p: %p flags %x\n" 2134 , ix, &sc->sc_dpring[ix] 2135 , (caddr_t)(uintptr_t) sc->sc_dpring[ix].pd_addr 2136 , sc->sc_dpring[ix].pd_flags 2137 ); 2138 if (++ix == SAFE_TOTAL_DPART) 2139 ix = 0; 2140 } 2141 } 2142 printf("sa: cmd0 %08x cmd1 %08x staterec %x\n", 2143 re->re_sa.sa_cmd0, re->re_sa.sa_cmd1, re->re_sa.sa_staterec); 2144 printf("sa: key %x %x %x %x %x %x %x %x\n" 2145 , re->re_sa.sa_key[0] 2146 , re->re_sa.sa_key[1] 2147 , re->re_sa.sa_key[2] 2148 , re->re_sa.sa_key[3] 2149 , re->re_sa.sa_key[4] 2150 , re->re_sa.sa_key[5] 2151 , re->re_sa.sa_key[6] 2152 , re->re_sa.sa_key[7] 2153 ); 2154 printf("sa: indigest %x %x %x %x %x\n" 2155 , re->re_sa.sa_indigest[0] 2156 , re->re_sa.sa_indigest[1] 2157 , re->re_sa.sa_indigest[2] 2158 , re->re_sa.sa_indigest[3] 2159 , re->re_sa.sa_indigest[4] 2160 ); 2161 printf("sa: outdigest %x %x %x %x %x\n" 2162 , re->re_sa.sa_outdigest[0] 2163 , re->re_sa.sa_outdigest[1] 2164 , re->re_sa.sa_outdigest[2] 2165 , re->re_sa.sa_outdigest[3] 2166 , re->re_sa.sa_outdigest[4] 2167 ); 2168 printf("sr: iv %x %x %x %x\n" 2169 , re->re_sastate.sa_saved_iv[0] 2170 , re->re_sastate.sa_saved_iv[1] 2171 , re->re_sastate.sa_saved_iv[2] 2172 , re->re_sastate.sa_saved_iv[3] 2173 ); 2174 printf("sr: hashbc %u indigest %x %x %x %x %x\n" 2175 , re->re_sastate.sa_saved_hashbc 2176 , re->re_sastate.sa_saved_indigest[0] 2177 , re->re_sastate.sa_saved_indigest[1] 2178 , re->re_sastate.sa_saved_indigest[2] 2179 , re->re_sastate.sa_saved_indigest[3] 2180 , re->re_sastate.sa_saved_indigest[4] 2181 ); 2182} 2183 2184static void 2185safe_dump_ring(struct safe_softc *sc, const char *tag) 2186{ 2187 mtx_lock(&sc->sc_ringmtx); 2188 printf("\nSafeNet Ring State:\n"); 2189 safe_dump_intrstate(sc, tag); 2190 safe_dump_dmastatus(sc, tag); 2191 safe_dump_ringstate(sc, tag); 2192 if (sc->sc_nqchip) { 2193 struct safe_ringentry *re = sc->sc_back; 2194 do { 2195 safe_dump_request(sc, tag, re); 2196 if (++re == sc->sc_ringtop) 2197 re = sc->sc_ring; 2198 } while (re != sc->sc_front); 2199 } 2200 mtx_unlock(&sc->sc_ringmtx); 2201} 2202 2203static int 2204sysctl_hw_safe_dump(SYSCTL_HANDLER_ARGS) 2205{ 2206 char dmode[64]; 2207 int error; 2208 2209 strncpy(dmode, "", sizeof(dmode) - 1); 2210 dmode[sizeof(dmode) - 1] = '\0'; 2211 error = sysctl_handle_string(oidp, &dmode[0], sizeof(dmode), req); 2212 2213 if (error == 0 && req->newptr != NULL) { 2214 struct safe_softc *sc = safec; 2215 2216 if (!sc) 2217 return EINVAL; 2218 if (strncmp(dmode, "dma", 3) == 0) 2219 safe_dump_dmastatus(sc, "safe0"); 2220 else if (strncmp(dmode, "int", 3) == 0) 2221 safe_dump_intrstate(sc, "safe0"); 2222 else if (strncmp(dmode, "ring", 4) == 0) 2223 safe_dump_ring(sc, "safe0"); 2224 else 2225 return EINVAL; 2226 } 2227 return error; 2228} 2229SYSCTL_PROC(_hw_safe, OID_AUTO, dump, CTLTYPE_STRING | CTLFLAG_RW, 2230 0, 0, sysctl_hw_safe_dump, "A", "Dump driver state"); 2231#endif /* SAFE_DEBUG */ 2232