1 /*- 2 * Copyright (c) 2009 Yahoo! Inc. 3 * Copyright (c) 2011-2015 LSI Corp. 4 * Copyright (c) 2013-2016 Avago Technologies 5 * Copyright 2000-2020 Broadcom Inc. 6 * All rights reserved. 7 * 8 * Redistribution and use in source and binary forms, with or without 9 * modification, are permitted provided that the following conditions 10 * are met: 11 * 1. Redistributions of source code must retain the above copyright 12 * notice, this list of conditions and the following disclaimer. 13 * 2. Redistributions in binary form must reproduce the above copyright 14 * notice, this list of conditions and the following disclaimer in the 15 * documentation and/or other materials provided with the distribution. 16 * 17 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND 18 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 19 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 20 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE 21 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 22 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 23 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 24 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 25 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 26 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 27 * SUCH DAMAGE. 28 * 29 * Broadcom Inc. (LSI) MPT-Fusion Host Adapter FreeBSD 30 * 31 */ 32 33 #include <sys/cdefs.h> 34 /* Communications core for Avago Technologies (LSI) MPT3 */ 35 36 /* TODO Move headers to mprvar */ 37 #include <sys/types.h> 38 #include <sys/param.h> 39 #include <sys/systm.h> 40 #include <sys/kernel.h> 41 #include <sys/selinfo.h> 42 #include <sys/lock.h> 43 #include <sys/mutex.h> 44 #include <sys/module.h> 45 #include <sys/bus.h> 46 #include <sys/conf.h> 47 #include <sys/bio.h> 48 #include <sys/malloc.h> 49 #include <sys/uio.h> 50 #include <sys/sysctl.h> 51 #include <sys/smp.h> 52 #include <sys/queue.h> 53 #include <sys/kthread.h> 54 #include <sys/taskqueue.h> 55 #include <sys/endian.h> 56 #include <sys/eventhandler.h> 57 #include <sys/sbuf.h> 58 #include <sys/priv.h> 59 60 #include <machine/bus.h> 61 #include <machine/resource.h> 62 #include <sys/rman.h> 63 #include <sys/proc.h> 64 65 #include <dev/pci/pcivar.h> 66 67 #include <cam/cam.h> 68 #include <cam/cam_ccb.h> 69 #include <cam/scsi/scsi_all.h> 70 71 #include <dev/mpr/mpi/mpi2_type.h> 72 #include <dev/mpr/mpi/mpi2.h> 73 #include <dev/mpr/mpi/mpi2_ioc.h> 74 #include <dev/mpr/mpi/mpi2_sas.h> 75 #include <dev/mpr/mpi/mpi2_pci.h> 76 #include <dev/mpr/mpi/mpi2_cnfg.h> 77 #include <dev/mpr/mpi/mpi2_init.h> 78 #include <dev/mpr/mpi/mpi2_tool.h> 79 #include <dev/mpr/mpr_ioctl.h> 80 #include <dev/mpr/mprvar.h> 81 #include <dev/mpr/mpr_table.h> 82 #include <dev/mpr/mpr_sas.h> 83 84 static int mpr_diag_reset(struct mpr_softc *sc, int sleep_flag); 85 static int mpr_init_queues(struct mpr_softc *sc); 86 static void mpr_resize_queues(struct mpr_softc *sc); 87 static int mpr_message_unit_reset(struct mpr_softc *sc, int sleep_flag); 88 static int mpr_transition_operational(struct mpr_softc *sc); 89 static int mpr_iocfacts_allocate(struct mpr_softc *sc, uint8_t attaching); 90 static void mpr_iocfacts_free(struct mpr_softc *sc); 91 static void mpr_startup(void *arg); 92 static int mpr_send_iocinit(struct mpr_softc *sc); 93 static int mpr_alloc_queues(struct mpr_softc *sc); 94 static int mpr_alloc_hw_queues(struct mpr_softc *sc); 95 static int mpr_alloc_replies(struct mpr_softc *sc); 96 static int mpr_alloc_requests(struct mpr_softc *sc); 97 static int mpr_alloc_nvme_prp_pages(struct mpr_softc *sc); 98 static int mpr_attach_log(struct mpr_softc *sc); 99 static __inline void mpr_complete_command(struct mpr_softc *sc, 100 struct mpr_command *cm); 101 static void mpr_dispatch_event(struct mpr_softc *sc, uintptr_t data, 102 MPI2_EVENT_NOTIFICATION_REPLY *reply); 103 static void mpr_config_complete(struct mpr_softc *sc, struct mpr_command *cm); 104 static void mpr_periodic(void *); 105 static int mpr_reregister_events(struct mpr_softc *sc); 106 static void mpr_enqueue_request(struct mpr_softc *sc, struct mpr_command *cm); 107 static int mpr_get_iocfacts(struct mpr_softc *sc, MPI2_IOC_FACTS_REPLY *facts); 108 static int mpr_wait_db_ack(struct mpr_softc *sc, int timeout, int sleep_flag); 109 static int mpr_debug_sysctl(SYSCTL_HANDLER_ARGS); 110 static int mpr_dump_reqs(SYSCTL_HANDLER_ARGS); 111 static void mpr_parse_debug(struct mpr_softc *sc, char *list); 112 static void adjust_iocfacts_endianness(MPI2_IOC_FACTS_REPLY *facts); 113 114 SYSCTL_NODE(_hw, OID_AUTO, mpr, CTLFLAG_RD | CTLFLAG_MPSAFE, 0, 115 "MPR Driver Parameters"); 116 117 MALLOC_DEFINE(M_MPR, "mpr", "mpr driver memory"); 118 119 /* 120 * Do a "Diagnostic Reset" aka a hard reset. This should get the chip out of 121 * any state and back to its initialization state machine. 122 */ 123 static char mpt2_reset_magic[] = { 0x00, 0x0f, 0x04, 0x0b, 0x02, 0x07, 0x0d }; 124 125 /* 126 * Added this union to smoothly convert le64toh cm->cm_desc.Words. 127 * Compiler only supports uint64_t to be passed as an argument. 128 * Otherwise it will throw this error: 129 * "aggregate value used where an integer was expected" 130 */ 131 typedef union { 132 u64 word; 133 struct { 134 u32 low; 135 u32 high; 136 } u; 137 } request_descriptor_t; 138 139 /* Rate limit chain-fail messages to 1 per minute */ 140 static struct timeval mpr_chainfail_interval = { 60, 0 }; 141 142 /* 143 * sleep_flag can be either CAN_SLEEP or NO_SLEEP. 144 * If this function is called from process context, it can sleep 145 * and there is no harm to sleep, in case if this fuction is called 146 * from Interrupt handler, we can not sleep and need NO_SLEEP flag set. 147 * based on sleep flags driver will call either msleep, pause or DELAY. 148 * msleep and pause are of same variant, but pause is used when mpr_mtx 149 * is not hold by driver. 150 */ 151 static int 152 mpr_diag_reset(struct mpr_softc *sc,int sleep_flag) 153 { 154 uint32_t reg; 155 int i, error, tries = 0; 156 uint8_t first_wait_done = FALSE; 157 158 mpr_dprint(sc, MPR_INIT, "%s entered\n", __func__); 159 160 /* Clear any pending interrupts */ 161 mpr_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0); 162 163 /* 164 * Force NO_SLEEP for threads prohibited to sleep 165 * e.a Thread from interrupt handler are prohibited to sleep. 166 */ 167 if (curthread->td_no_sleeping) 168 sleep_flag = NO_SLEEP; 169 170 mpr_dprint(sc, MPR_INIT, "sequence start, sleep_flag=%d\n", sleep_flag); 171 /* Push the magic sequence */ 172 error = ETIMEDOUT; 173 while (tries++ < 20) { 174 for (i = 0; i < sizeof(mpt2_reset_magic); i++) 175 mpr_regwrite(sc, MPI2_WRITE_SEQUENCE_OFFSET, 176 mpt2_reset_magic[i]); 177 178 /* wait 100 msec */ 179 if (mtx_owned(&sc->mpr_mtx) && sleep_flag == CAN_SLEEP) 180 msleep(&sc->msleep_fake_chan, &sc->mpr_mtx, 0, 181 "mprdiag", hz/10); 182 else if (sleep_flag == CAN_SLEEP) 183 pause("mprdiag", hz/10); 184 else 185 DELAY(100 * 1000); 186 187 reg = mpr_regread(sc, MPI2_HOST_DIAGNOSTIC_OFFSET); 188 if (reg & MPI2_DIAG_DIAG_WRITE_ENABLE) { 189 error = 0; 190 break; 191 } 192 } 193 if (error) { 194 mpr_dprint(sc, MPR_INIT, "sequence failed, error=%d, exit\n", 195 error); 196 return (error); 197 } 198 199 /* Send the actual reset. XXX need to refresh the reg? */ 200 reg |= MPI2_DIAG_RESET_ADAPTER; 201 mpr_dprint(sc, MPR_INIT, "sequence success, sending reset, reg= 0x%x\n", 202 reg); 203 mpr_regwrite(sc, MPI2_HOST_DIAGNOSTIC_OFFSET, reg); 204 205 /* Wait up to 300 seconds in 50ms intervals */ 206 error = ETIMEDOUT; 207 for (i = 0; i < 6000; i++) { 208 /* 209 * Wait 50 msec. If this is the first time through, wait 256 210 * msec to satisfy Diag Reset timing requirements. 211 */ 212 if (first_wait_done) { 213 if (mtx_owned(&sc->mpr_mtx) && sleep_flag == CAN_SLEEP) 214 msleep(&sc->msleep_fake_chan, &sc->mpr_mtx, 0, 215 "mprdiag", hz/20); 216 else if (sleep_flag == CAN_SLEEP) 217 pause("mprdiag", hz/20); 218 else 219 DELAY(50 * 1000); 220 } else { 221 DELAY(256 * 1000); 222 first_wait_done = TRUE; 223 } 224 /* 225 * Check for the RESET_ADAPTER bit to be cleared first, then 226 * wait for the RESET state to be cleared, which takes a little 227 * longer. 228 */ 229 reg = mpr_regread(sc, MPI2_HOST_DIAGNOSTIC_OFFSET); 230 if (reg & MPI2_DIAG_RESET_ADAPTER) { 231 continue; 232 } 233 reg = mpr_regread(sc, MPI2_DOORBELL_OFFSET); 234 if ((reg & MPI2_IOC_STATE_MASK) != MPI2_IOC_STATE_RESET) { 235 error = 0; 236 break; 237 } 238 } 239 if (error) { 240 mpr_dprint(sc, MPR_INIT, "reset failed, error= %d, exit\n", 241 error); 242 return (error); 243 } 244 245 mpr_regwrite(sc, MPI2_WRITE_SEQUENCE_OFFSET, 0x0); 246 mpr_dprint(sc, MPR_INIT, "diag reset success, exit\n"); 247 248 return (0); 249 } 250 251 static int 252 mpr_message_unit_reset(struct mpr_softc *sc, int sleep_flag) 253 { 254 int error; 255 256 MPR_FUNCTRACE(sc); 257 258 mpr_dprint(sc, MPR_INIT, "%s entered\n", __func__); 259 260 error = 0; 261 mpr_regwrite(sc, MPI2_DOORBELL_OFFSET, 262 MPI2_FUNCTION_IOC_MESSAGE_UNIT_RESET << 263 MPI2_DOORBELL_FUNCTION_SHIFT); 264 265 if (mpr_wait_db_ack(sc, 5, sleep_flag) != 0) { 266 mpr_dprint(sc, MPR_INIT|MPR_FAULT, 267 "Doorbell handshake failed\n"); 268 error = ETIMEDOUT; 269 } 270 271 mpr_dprint(sc, MPR_INIT, "%s exit\n", __func__); 272 return (error); 273 } 274 275 static int 276 mpr_transition_ready(struct mpr_softc *sc) 277 { 278 uint32_t reg, state; 279 int error, tries = 0; 280 int sleep_flags; 281 282 MPR_FUNCTRACE(sc); 283 /* If we are in attach call, do not sleep */ 284 sleep_flags = (sc->mpr_flags & MPR_FLAGS_ATTACH_DONE) 285 ? CAN_SLEEP : NO_SLEEP; 286 287 error = 0; 288 289 mpr_dprint(sc, MPR_INIT, "%s entered, sleep_flags= %d\n", 290 __func__, sleep_flags); 291 292 while (tries++ < 1200) { 293 reg = mpr_regread(sc, MPI2_DOORBELL_OFFSET); 294 mpr_dprint(sc, MPR_INIT, " Doorbell= 0x%x\n", reg); 295 296 /* 297 * Ensure the IOC is ready to talk. If it's not, try 298 * resetting it. 299 */ 300 if (reg & MPI2_DOORBELL_USED) { 301 mpr_dprint(sc, MPR_INIT, " Not ready, sending diag " 302 "reset\n"); 303 mpr_diag_reset(sc, sleep_flags); 304 DELAY(50000); 305 continue; 306 } 307 308 /* Is the adapter owned by another peer? */ 309 if ((reg & MPI2_DOORBELL_WHO_INIT_MASK) == 310 (MPI2_WHOINIT_PCI_PEER << MPI2_DOORBELL_WHO_INIT_SHIFT)) { 311 mpr_dprint(sc, MPR_INIT|MPR_FAULT, "IOC is under the " 312 "control of another peer host, aborting " 313 "initialization.\n"); 314 error = ENXIO; 315 break; 316 } 317 318 state = reg & MPI2_IOC_STATE_MASK; 319 if (state == MPI2_IOC_STATE_READY) { 320 /* Ready to go! */ 321 error = 0; 322 break; 323 } else if (state == MPI2_IOC_STATE_FAULT) { 324 mpr_dprint(sc, MPR_INIT|MPR_FAULT, "IOC in fault " 325 "state 0x%x, resetting\n", 326 state & MPI2_DOORBELL_FAULT_CODE_MASK); 327 mpr_diag_reset(sc, sleep_flags); 328 } else if (state == MPI2_IOC_STATE_OPERATIONAL) { 329 /* Need to take ownership */ 330 mpr_message_unit_reset(sc, sleep_flags); 331 } else if (state == MPI2_IOC_STATE_RESET) { 332 /* Wait a bit, IOC might be in transition */ 333 mpr_dprint(sc, MPR_INIT|MPR_FAULT, 334 "IOC in unexpected reset state\n"); 335 } else { 336 mpr_dprint(sc, MPR_INIT|MPR_FAULT, 337 "IOC in unknown state 0x%x\n", state); 338 error = EINVAL; 339 break; 340 } 341 342 /* Wait 50ms for things to settle down. */ 343 DELAY(50000); 344 } 345 346 if (error) 347 mpr_dprint(sc, MPR_INIT|MPR_FAULT, 348 "Cannot transition IOC to ready\n"); 349 mpr_dprint(sc, MPR_INIT, "%s exit\n", __func__); 350 return (error); 351 } 352 353 static int 354 mpr_transition_operational(struct mpr_softc *sc) 355 { 356 uint32_t reg, state; 357 int error; 358 359 MPR_FUNCTRACE(sc); 360 361 error = 0; 362 reg = mpr_regread(sc, MPI2_DOORBELL_OFFSET); 363 mpr_dprint(sc, MPR_INIT, "%s entered, Doorbell= 0x%x\n", __func__, reg); 364 365 state = reg & MPI2_IOC_STATE_MASK; 366 if (state != MPI2_IOC_STATE_READY) { 367 mpr_dprint(sc, MPR_INIT, "IOC not ready\n"); 368 if ((error = mpr_transition_ready(sc)) != 0) { 369 mpr_dprint(sc, MPR_INIT|MPR_FAULT, 370 "failed to transition ready, exit\n"); 371 return (error); 372 } 373 } 374 375 error = mpr_send_iocinit(sc); 376 mpr_dprint(sc, MPR_INIT, "%s exit\n", __func__); 377 378 return (error); 379 } 380 381 static void 382 mpr_resize_queues(struct mpr_softc *sc) 383 { 384 u_int reqcr, prireqcr, maxio, sges_per_frame, chain_seg_size; 385 386 /* 387 * Size the queues. Since the reply queues always need one free 388 * entry, we'll deduct one reply message here. The LSI documents 389 * suggest instead to add a count to the request queue, but I think 390 * that it's better to deduct from reply queue. 391 */ 392 prireqcr = MAX(1, sc->max_prireqframes); 393 prireqcr = MIN(prireqcr, sc->facts->HighPriorityCredit); 394 395 reqcr = MAX(2, sc->max_reqframes); 396 reqcr = MIN(reqcr, sc->facts->RequestCredit); 397 398 sc->num_reqs = prireqcr + reqcr; 399 sc->num_prireqs = prireqcr; 400 sc->num_replies = MIN(sc->max_replyframes + sc->max_evtframes, 401 sc->facts->MaxReplyDescriptorPostQueueDepth) - 1; 402 403 /* Store the request frame size in bytes rather than as 32bit words */ 404 sc->reqframesz = sc->facts->IOCRequestFrameSize * 4; 405 406 /* 407 * Gen3 and beyond uses the IOCMaxChainSegmentSize from IOC Facts to 408 * get the size of a Chain Frame. Previous versions use the size as a 409 * Request Frame for the Chain Frame size. If IOCMaxChainSegmentSize 410 * is 0, use the default value. The IOCMaxChainSegmentSize is the 411 * number of 16-byte elelements that can fit in a Chain Frame, which is 412 * the size of an IEEE Simple SGE. 413 */ 414 if (sc->facts->MsgVersion >= MPI2_VERSION_02_05) { 415 chain_seg_size = sc->facts->IOCMaxChainSegmentSize; 416 if (chain_seg_size == 0) 417 chain_seg_size = MPR_DEFAULT_CHAIN_SEG_SIZE; 418 sc->chain_frame_size = chain_seg_size * 419 MPR_MAX_CHAIN_ELEMENT_SIZE; 420 } else { 421 sc->chain_frame_size = sc->reqframesz; 422 } 423 424 /* 425 * Max IO Size is Page Size * the following: 426 * ((SGEs per frame - 1 for chain element) * Max Chain Depth) 427 * + 1 for no chain needed in last frame 428 * 429 * If user suggests a Max IO size to use, use the smaller of the 430 * user's value and the calculated value as long as the user's 431 * value is larger than 0. The user's value is in pages. 432 */ 433 sges_per_frame = sc->chain_frame_size/sizeof(MPI2_IEEE_SGE_SIMPLE64)-1; 434 maxio = (sges_per_frame * sc->facts->MaxChainDepth + 1) * PAGE_SIZE; 435 436 /* 437 * If I/O size limitation requested then use it and pass up to CAM. 438 * If not, use maxphys as an optimization hint, but report HW limit. 439 */ 440 if (sc->max_io_pages > 0) { 441 maxio = min(maxio, sc->max_io_pages * PAGE_SIZE); 442 sc->maxio = maxio; 443 } else { 444 sc->maxio = maxio; 445 maxio = min(maxio, maxphys); 446 } 447 448 sc->num_chains = (maxio / PAGE_SIZE + sges_per_frame - 2) / 449 sges_per_frame * reqcr; 450 if (sc->max_chains > 0 && sc->max_chains < sc->num_chains) 451 sc->num_chains = sc->max_chains; 452 453 /* 454 * Figure out the number of MSIx-based queues. If the firmware or 455 * user has done something crazy and not allowed enough credit for 456 * the queues to be useful then don't enable multi-queue. 457 */ 458 if (sc->facts->MaxMSIxVectors < 2) 459 sc->msi_msgs = 1; 460 461 if (sc->msi_msgs > 1) { 462 sc->msi_msgs = MIN(sc->msi_msgs, mp_ncpus); 463 sc->msi_msgs = MIN(sc->msi_msgs, sc->facts->MaxMSIxVectors); 464 if (sc->num_reqs / sc->msi_msgs < 2) 465 sc->msi_msgs = 1; 466 } 467 468 mpr_dprint(sc, MPR_INIT, "Sized queues to q=%d reqs=%d replies=%d\n", 469 sc->msi_msgs, sc->num_reqs, sc->num_replies); 470 } 471 472 /* 473 * This is called during attach and when re-initializing due to a Diag Reset. 474 * IOC Facts is used to allocate many of the structures needed by the driver. 475 * If called from attach, de-allocation is not required because the driver has 476 * not allocated any structures yet, but if called from a Diag Reset, previously 477 * allocated structures based on IOC Facts will need to be freed and re- 478 * allocated bases on the latest IOC Facts. 479 */ 480 static int 481 mpr_iocfacts_allocate(struct mpr_softc *sc, uint8_t attaching) 482 { 483 int error; 484 Mpi2IOCFactsReply_t saved_facts; 485 uint8_t saved_mode, reallocating; 486 487 mpr_dprint(sc, MPR_INIT|MPR_TRACE, "%s entered\n", __func__); 488 489 /* Save old IOC Facts and then only reallocate if Facts have changed */ 490 if (!attaching) { 491 bcopy(sc->facts, &saved_facts, sizeof(MPI2_IOC_FACTS_REPLY)); 492 } 493 494 /* 495 * Get IOC Facts. In all cases throughout this function, panic if doing 496 * a re-initialization and only return the error if attaching so the OS 497 * can handle it. 498 */ 499 if ((error = mpr_get_iocfacts(sc, sc->facts)) != 0) { 500 if (attaching) { 501 mpr_dprint(sc, MPR_INIT|MPR_FAULT, "Failed to get " 502 "IOC Facts with error %d, exit\n", error); 503 return (error); 504 } else { 505 panic("%s failed to get IOC Facts with error %d\n", 506 __func__, error); 507 } 508 } 509 510 MPR_DPRINT_PAGE(sc, MPR_XINFO, iocfacts, sc->facts); 511 512 snprintf(sc->fw_version, sizeof(sc->fw_version), 513 "%02d.%02d.%02d.%02d", 514 sc->facts->FWVersion.Struct.Major, 515 sc->facts->FWVersion.Struct.Minor, 516 sc->facts->FWVersion.Struct.Unit, 517 sc->facts->FWVersion.Struct.Dev); 518 519 snprintf(sc->msg_version, sizeof(sc->msg_version), "%d.%d", 520 (sc->facts->MsgVersion & MPI2_IOCFACTS_MSGVERSION_MAJOR_MASK) >> 521 MPI2_IOCFACTS_MSGVERSION_MAJOR_SHIFT, 522 (sc->facts->MsgVersion & MPI2_IOCFACTS_MSGVERSION_MINOR_MASK) >> 523 MPI2_IOCFACTS_MSGVERSION_MINOR_SHIFT); 524 525 mpr_dprint(sc, MPR_INFO, "Firmware: %s, Driver: %s\n", sc->fw_version, 526 MPR_DRIVER_VERSION); 527 mpr_dprint(sc, MPR_INFO, 528 "IOCCapabilities: %b\n", sc->facts->IOCCapabilities, 529 "\20" "\3ScsiTaskFull" "\4DiagTrace" "\5SnapBuf" "\6ExtBuf" 530 "\7EEDP" "\10BiDirTarg" "\11Multicast" "\14TransRetry" "\15IR" 531 "\16EventReplay" "\17RaidAccel" "\20MSIXIndex" "\21HostDisc" 532 "\22FastPath" "\23RDPQArray" "\24AtomicReqDesc" "\25PCIeSRIOV"); 533 534 /* 535 * If the chip doesn't support event replay then a hard reset will be 536 * required to trigger a full discovery. Do the reset here then 537 * retransition to Ready. A hard reset might have already been done, 538 * but it doesn't hurt to do it again. Only do this if attaching, not 539 * for a Diag Reset. 540 */ 541 if (attaching && ((sc->facts->IOCCapabilities & 542 MPI2_IOCFACTS_CAPABILITY_EVENT_REPLAY) == 0)) { 543 mpr_dprint(sc, MPR_INIT, "No event replay, resetting\n"); 544 mpr_diag_reset(sc, NO_SLEEP); 545 if ((error = mpr_transition_ready(sc)) != 0) { 546 mpr_dprint(sc, MPR_INIT|MPR_FAULT, "Failed to " 547 "transition to ready with error %d, exit\n", 548 error); 549 return (error); 550 } 551 } 552 553 /* 554 * Set flag if IR Firmware is loaded. If the RAID Capability has 555 * changed from the previous IOC Facts, log a warning, but only if 556 * checking this after a Diag Reset and not during attach. 557 */ 558 saved_mode = sc->ir_firmware; 559 if (sc->facts->IOCCapabilities & 560 MPI2_IOCFACTS_CAPABILITY_INTEGRATED_RAID) 561 sc->ir_firmware = 1; 562 if (!attaching) { 563 if (sc->ir_firmware != saved_mode) { 564 mpr_dprint(sc, MPR_INIT|MPR_FAULT, "new IR/IT mode " 565 "in IOC Facts does not match previous mode\n"); 566 } 567 } 568 569 /* Only deallocate and reallocate if relevant IOC Facts have changed */ 570 reallocating = FALSE; 571 sc->mpr_flags &= ~MPR_FLAGS_REALLOCATED; 572 573 if ((!attaching) && 574 ((saved_facts.MsgVersion != sc->facts->MsgVersion) || 575 (saved_facts.HeaderVersion != sc->facts->HeaderVersion) || 576 (saved_facts.MaxChainDepth != sc->facts->MaxChainDepth) || 577 (saved_facts.RequestCredit != sc->facts->RequestCredit) || 578 (saved_facts.ProductID != sc->facts->ProductID) || 579 (saved_facts.IOCCapabilities != sc->facts->IOCCapabilities) || 580 (saved_facts.IOCRequestFrameSize != 581 sc->facts->IOCRequestFrameSize) || 582 (saved_facts.IOCMaxChainSegmentSize != 583 sc->facts->IOCMaxChainSegmentSize) || 584 (saved_facts.MaxTargets != sc->facts->MaxTargets) || 585 (saved_facts.MaxSasExpanders != sc->facts->MaxSasExpanders) || 586 (saved_facts.MaxEnclosures != sc->facts->MaxEnclosures) || 587 (saved_facts.HighPriorityCredit != sc->facts->HighPriorityCredit) || 588 (saved_facts.MaxReplyDescriptorPostQueueDepth != 589 sc->facts->MaxReplyDescriptorPostQueueDepth) || 590 (saved_facts.ReplyFrameSize != sc->facts->ReplyFrameSize) || 591 (saved_facts.MaxVolumes != sc->facts->MaxVolumes) || 592 (saved_facts.MaxPersistentEntries != 593 sc->facts->MaxPersistentEntries))) { 594 reallocating = TRUE; 595 596 /* Record that we reallocated everything */ 597 sc->mpr_flags |= MPR_FLAGS_REALLOCATED; 598 } 599 600 /* 601 * Some things should be done if attaching or re-allocating after a Diag 602 * Reset, but are not needed after a Diag Reset if the FW has not 603 * changed. 604 */ 605 if (attaching || reallocating) { 606 /* 607 * Check if controller supports FW diag buffers and set flag to 608 * enable each type. 609 */ 610 if (sc->facts->IOCCapabilities & 611 MPI2_IOCFACTS_CAPABILITY_DIAG_TRACE_BUFFER) 612 sc->fw_diag_buffer_list[MPI2_DIAG_BUF_TYPE_TRACE]. 613 enabled = TRUE; 614 if (sc->facts->IOCCapabilities & 615 MPI2_IOCFACTS_CAPABILITY_SNAPSHOT_BUFFER) 616 sc->fw_diag_buffer_list[MPI2_DIAG_BUF_TYPE_SNAPSHOT]. 617 enabled = TRUE; 618 if (sc->facts->IOCCapabilities & 619 MPI2_IOCFACTS_CAPABILITY_EXTENDED_BUFFER) 620 sc->fw_diag_buffer_list[MPI2_DIAG_BUF_TYPE_EXTENDED]. 621 enabled = TRUE; 622 623 /* 624 * Set flags for some supported items. 625 */ 626 if (sc->facts->IOCCapabilities & MPI2_IOCFACTS_CAPABILITY_EEDP) 627 sc->eedp_enabled = TRUE; 628 if (sc->facts->IOCCapabilities & MPI2_IOCFACTS_CAPABILITY_TLR) 629 sc->control_TLR = TRUE; 630 if ((sc->facts->IOCCapabilities & 631 MPI26_IOCFACTS_CAPABILITY_ATOMIC_REQ) && 632 (sc->mpr_flags & MPR_FLAGS_SEA_IOC)) 633 sc->atomic_desc_capable = TRUE; 634 635 mpr_resize_queues(sc); 636 637 /* 638 * Initialize all Tail Queues 639 */ 640 TAILQ_INIT(&sc->req_list); 641 TAILQ_INIT(&sc->high_priority_req_list); 642 TAILQ_INIT(&sc->chain_list); 643 TAILQ_INIT(&sc->prp_page_list); 644 TAILQ_INIT(&sc->tm_list); 645 } 646 647 /* 648 * If doing a Diag Reset and the FW is significantly different 649 * (reallocating will be set above in IOC Facts comparison), then all 650 * buffers based on the IOC Facts will need to be freed before they are 651 * reallocated. 652 */ 653 if (reallocating) { 654 mpr_iocfacts_free(sc); 655 mprsas_realloc_targets(sc, saved_facts.MaxTargets + 656 saved_facts.MaxVolumes); 657 } 658 659 /* 660 * Any deallocation has been completed. Now start reallocating 661 * if needed. Will only need to reallocate if attaching or if the new 662 * IOC Facts are different from the previous IOC Facts after a Diag 663 * Reset. Targets have already been allocated above if needed. 664 */ 665 error = 0; 666 while (attaching || reallocating) { 667 if ((error = mpr_alloc_hw_queues(sc)) != 0) 668 break; 669 if ((error = mpr_alloc_replies(sc)) != 0) 670 break; 671 if ((error = mpr_alloc_requests(sc)) != 0) 672 break; 673 if ((error = mpr_alloc_queues(sc)) != 0) 674 break; 675 break; 676 } 677 if (error) { 678 mpr_dprint(sc, MPR_INIT|MPR_ERROR, 679 "Failed to alloc queues with error %d\n", error); 680 mpr_free(sc); 681 return (error); 682 } 683 684 /* Always initialize the queues */ 685 bzero(sc->free_queue, sc->fqdepth * 4); 686 mpr_init_queues(sc); 687 688 /* 689 * Always get the chip out of the reset state, but only panic if not 690 * attaching. If attaching and there is an error, that is handled by 691 * the OS. 692 */ 693 error = mpr_transition_operational(sc); 694 if (error != 0) { 695 mpr_dprint(sc, MPR_INIT|MPR_FAULT, "Failed to " 696 "transition to operational with error %d\n", error); 697 mpr_free(sc); 698 return (error); 699 } 700 701 /* 702 * Finish the queue initialization. 703 * These are set here instead of in mpr_init_queues() because the 704 * IOC resets these values during the state transition in 705 * mpr_transition_operational(). The free index is set to 1 706 * because the corresponding index in the IOC is set to 0, and the 707 * IOC treats the queues as full if both are set to the same value. 708 * Hence the reason that the queue can't hold all of the possible 709 * replies. 710 */ 711 sc->replypostindex = 0; 712 mpr_regwrite(sc, MPI2_REPLY_FREE_HOST_INDEX_OFFSET, sc->replyfreeindex); 713 mpr_regwrite(sc, MPI2_REPLY_POST_HOST_INDEX_OFFSET, 0); 714 715 /* 716 * Attach the subsystems so they can prepare their event masks. 717 * XXX Should be dynamic so that IM/IR and user modules can attach 718 */ 719 error = 0; 720 while (attaching) { 721 mpr_dprint(sc, MPR_INIT, "Attaching subsystems\n"); 722 if ((error = mpr_attach_log(sc)) != 0) 723 break; 724 if ((error = mpr_attach_sas(sc)) != 0) 725 break; 726 if ((error = mpr_attach_user(sc)) != 0) 727 break; 728 break; 729 } 730 if (error) { 731 mpr_dprint(sc, MPR_INIT|MPR_ERROR, 732 "Failed to attach all subsystems: error %d\n", error); 733 mpr_free(sc); 734 return (error); 735 } 736 737 /* 738 * XXX If the number of MSI-X vectors changes during re-init, this 739 * won't see it and adjust. 740 */ 741 if ((attaching || reallocating) && (error = mpr_pci_setup_interrupts(sc)) != 0) { 742 mpr_dprint(sc, MPR_INIT|MPR_ERROR, 743 "Failed to setup interrupts\n"); 744 mpr_free(sc); 745 return (error); 746 } 747 748 return (error); 749 } 750 751 /* 752 * This is called if memory is being free (during detach for example) and when 753 * buffers need to be reallocated due to a Diag Reset. 754 */ 755 static void 756 mpr_iocfacts_free(struct mpr_softc *sc) 757 { 758 struct mpr_command *cm; 759 int i; 760 761 mpr_dprint(sc, MPR_TRACE, "%s\n", __func__); 762 763 if (sc->free_busaddr != 0) 764 bus_dmamap_unload(sc->queues_dmat, sc->queues_map); 765 if (sc->free_queue != NULL) 766 bus_dmamem_free(sc->queues_dmat, sc->free_queue, 767 sc->queues_map); 768 if (sc->queues_dmat != NULL) 769 bus_dma_tag_destroy(sc->queues_dmat); 770 771 if (sc->chain_frames != NULL) { 772 bus_dmamap_unload(sc->chain_dmat, sc->chain_map); 773 bus_dmamem_free(sc->chain_dmat, sc->chain_frames, 774 sc->chain_map); 775 } 776 if (sc->chain_dmat != NULL) 777 bus_dma_tag_destroy(sc->chain_dmat); 778 779 if (sc->sense_busaddr != 0) 780 bus_dmamap_unload(sc->sense_dmat, sc->sense_map); 781 if (sc->sense_frames != NULL) 782 bus_dmamem_free(sc->sense_dmat, sc->sense_frames, 783 sc->sense_map); 784 if (sc->sense_dmat != NULL) 785 bus_dma_tag_destroy(sc->sense_dmat); 786 787 if (sc->prp_page_busaddr != 0) 788 bus_dmamap_unload(sc->prp_page_dmat, sc->prp_page_map); 789 if (sc->prp_pages != NULL) 790 bus_dmamem_free(sc->prp_page_dmat, sc->prp_pages, 791 sc->prp_page_map); 792 if (sc->prp_page_dmat != NULL) 793 bus_dma_tag_destroy(sc->prp_page_dmat); 794 795 if (sc->reply_busaddr != 0) 796 bus_dmamap_unload(sc->reply_dmat, sc->reply_map); 797 if (sc->reply_frames != NULL) 798 bus_dmamem_free(sc->reply_dmat, sc->reply_frames, 799 sc->reply_map); 800 if (sc->reply_dmat != NULL) 801 bus_dma_tag_destroy(sc->reply_dmat); 802 803 if (sc->req_busaddr != 0) 804 bus_dmamap_unload(sc->req_dmat, sc->req_map); 805 if (sc->req_frames != NULL) 806 bus_dmamem_free(sc->req_dmat, sc->req_frames, sc->req_map); 807 if (sc->req_dmat != NULL) 808 bus_dma_tag_destroy(sc->req_dmat); 809 810 if (sc->chains != NULL) 811 free(sc->chains, M_MPR); 812 if (sc->prps != NULL) 813 free(sc->prps, M_MPR); 814 if (sc->commands != NULL) { 815 for (i = 1; i < sc->num_reqs; i++) { 816 cm = &sc->commands[i]; 817 bus_dmamap_destroy(sc->buffer_dmat, cm->cm_dmamap); 818 } 819 free(sc->commands, M_MPR); 820 } 821 if (sc->buffer_dmat != NULL) 822 bus_dma_tag_destroy(sc->buffer_dmat); 823 824 mpr_pci_free_interrupts(sc); 825 free(sc->queues, M_MPR); 826 sc->queues = NULL; 827 } 828 829 /* 830 * The terms diag reset and hard reset are used interchangeably in the MPI 831 * docs to mean resetting the controller chip. In this code diag reset 832 * cleans everything up, and the hard reset function just sends the reset 833 * sequence to the chip. This should probably be refactored so that every 834 * subsystem gets a reset notification of some sort, and can clean up 835 * appropriately. 836 */ 837 int 838 mpr_reinit(struct mpr_softc *sc) 839 { 840 int error; 841 struct mprsas_softc *sassc; 842 843 sassc = sc->sassc; 844 845 MPR_FUNCTRACE(sc); 846 847 mtx_assert(&sc->mpr_mtx, MA_OWNED); 848 849 mpr_dprint(sc, MPR_INIT|MPR_INFO, "Reinitializing controller\n"); 850 if (sc->mpr_flags & MPR_FLAGS_DIAGRESET) { 851 mpr_dprint(sc, MPR_INIT, "Reset already in progress\n"); 852 return 0; 853 } 854 855 /* 856 * Make sure the completion callbacks can recognize they're getting 857 * a NULL cm_reply due to a reset. 858 */ 859 sc->mpr_flags |= MPR_FLAGS_DIAGRESET; 860 861 /* 862 * Mask interrupts here. 863 */ 864 mpr_dprint(sc, MPR_INIT, "Masking interrupts and resetting\n"); 865 mpr_mask_intr(sc); 866 867 error = mpr_diag_reset(sc, CAN_SLEEP); 868 if (error != 0) { 869 panic("%s hard reset failed with error %d\n", __func__, error); 870 } 871 872 /* Restore the PCI state, including the MSI-X registers */ 873 mpr_pci_restore(sc); 874 875 /* Give the I/O subsystem special priority to get itself prepared */ 876 mprsas_handle_reinit(sc); 877 878 /* 879 * Get IOC Facts and allocate all structures based on this information. 880 * The attach function will also call mpr_iocfacts_allocate at startup. 881 * If relevant values have changed in IOC Facts, this function will free 882 * all of the memory based on IOC Facts and reallocate that memory. 883 */ 884 if ((error = mpr_iocfacts_allocate(sc, FALSE)) != 0) { 885 panic("%s IOC Facts based allocation failed with error %d\n", 886 __func__, error); 887 } 888 889 /* 890 * Mapping structures will be re-allocated after getting IOC Page8, so 891 * free these structures here. 892 */ 893 mpr_mapping_exit(sc); 894 895 /* 896 * The static page function currently read is IOC Page8. Others can be 897 * added in future. It's possible that the values in IOC Page8 have 898 * changed after a Diag Reset due to user modification, so always read 899 * these. Interrupts are masked, so unmask them before getting config 900 * pages. 901 */ 902 mpr_unmask_intr(sc); 903 sc->mpr_flags &= ~MPR_FLAGS_DIAGRESET; 904 mpr_base_static_config_pages(sc); 905 906 /* 907 * Some mapping info is based in IOC Page8 data, so re-initialize the 908 * mapping tables. 909 */ 910 mpr_mapping_initialize(sc); 911 912 /* 913 * Restart will reload the event masks clobbered by the reset, and 914 * then enable the port. 915 */ 916 mpr_reregister_events(sc); 917 918 /* the end of discovery will release the simq, so we're done. */ 919 mpr_dprint(sc, MPR_INIT|MPR_XINFO, "Finished sc %p post %u free %u\n", 920 sc, sc->replypostindex, sc->replyfreeindex); 921 mprsas_release_simq_reinit(sassc); 922 mpr_dprint(sc, MPR_INIT, "%s exit error= %d\n", __func__, error); 923 924 return 0; 925 } 926 927 /* Wait for the chip to ACK a word that we've put into its FIFO 928 * Wait for <timeout> seconds. In single loop wait for busy loop 929 * for 500 microseconds. 930 * Total is [ 0.5 * (2000 * <timeout>) ] in miliseconds. 931 * */ 932 static int 933 mpr_wait_db_ack(struct mpr_softc *sc, int timeout, int sleep_flag) 934 { 935 u32 cntdn, count; 936 u32 int_status; 937 u32 doorbell; 938 939 count = 0; 940 cntdn = (sleep_flag == CAN_SLEEP) ? 1000*timeout : 2000*timeout; 941 do { 942 int_status = mpr_regread(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET); 943 if (!(int_status & MPI2_HIS_SYS2IOC_DB_STATUS)) { 944 mpr_dprint(sc, MPR_TRACE, "%s: successful count(%d), " 945 "timeout(%d)\n", __func__, count, timeout); 946 return 0; 947 } else if (int_status & MPI2_HIS_IOC2SYS_DB_STATUS) { 948 doorbell = mpr_regread(sc, MPI2_DOORBELL_OFFSET); 949 if ((doorbell & MPI2_IOC_STATE_MASK) == 950 MPI2_IOC_STATE_FAULT) { 951 mpr_dprint(sc, MPR_FAULT, 952 "fault_state(0x%04x)!\n", doorbell); 953 return (EFAULT); 954 } 955 } else if (int_status == 0xFFFFFFFF) 956 goto out; 957 958 /* 959 * If it can sleep, sleep for 1 milisecond, else busy loop for 960 * 0.5 milisecond 961 */ 962 if (mtx_owned(&sc->mpr_mtx) && sleep_flag == CAN_SLEEP) 963 msleep(&sc->msleep_fake_chan, &sc->mpr_mtx, 0, "mprdba", 964 hz/1000); 965 else if (sleep_flag == CAN_SLEEP) 966 pause("mprdba", hz/1000); 967 else 968 DELAY(500); 969 count++; 970 } while (--cntdn); 971 972 out: 973 mpr_dprint(sc, MPR_FAULT, "%s: failed due to timeout count(%d), " 974 "int_status(%x)!\n", __func__, count, int_status); 975 return (ETIMEDOUT); 976 } 977 978 /* Wait for the chip to signal that the next word in its FIFO can be fetched */ 979 static int 980 mpr_wait_db_int(struct mpr_softc *sc) 981 { 982 int retry; 983 984 for (retry = 0; retry < MPR_DB_MAX_WAIT; retry++) { 985 if ((mpr_regread(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET) & 986 MPI2_HIS_IOC2SYS_DB_STATUS) != 0) 987 return (0); 988 DELAY(2000); 989 } 990 return (ETIMEDOUT); 991 } 992 993 /* Step through the synchronous command state machine, i.e. "Doorbell mode" */ 994 static int 995 mpr_request_sync(struct mpr_softc *sc, void *req, MPI2_DEFAULT_REPLY *reply, 996 int req_sz, int reply_sz, int timeout) 997 { 998 uint32_t *data32; 999 uint16_t *data16; 1000 int i, count, ioc_sz, residual; 1001 int sleep_flags = CAN_SLEEP; 1002 1003 if (curthread->td_no_sleeping) 1004 sleep_flags = NO_SLEEP; 1005 1006 /* Step 1 */ 1007 mpr_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0); 1008 1009 /* Step 2 */ 1010 if (mpr_regread(sc, MPI2_DOORBELL_OFFSET) & MPI2_DOORBELL_USED) 1011 return (EBUSY); 1012 1013 /* Step 3 1014 * Announce that a message is coming through the doorbell. Messages 1015 * are pushed at 32bit words, so round up if needed. 1016 */ 1017 count = (req_sz + 3) / 4; 1018 mpr_regwrite(sc, MPI2_DOORBELL_OFFSET, 1019 (MPI2_FUNCTION_HANDSHAKE << MPI2_DOORBELL_FUNCTION_SHIFT) | 1020 (count << MPI2_DOORBELL_ADD_DWORDS_SHIFT)); 1021 1022 /* Step 4 */ 1023 if (mpr_wait_db_int(sc) || 1024 (mpr_regread(sc, MPI2_DOORBELL_OFFSET) & MPI2_DOORBELL_USED) == 0) { 1025 mpr_dprint(sc, MPR_FAULT, "Doorbell failed to activate\n"); 1026 return (ENXIO); 1027 } 1028 mpr_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0); 1029 if (mpr_wait_db_ack(sc, 5, sleep_flags) != 0) { 1030 mpr_dprint(sc, MPR_FAULT, "Doorbell handshake failed\n"); 1031 return (ENXIO); 1032 } 1033 1034 /* Step 5 */ 1035 /* Clock out the message data synchronously in 32-bit dwords*/ 1036 data32 = (uint32_t *)req; 1037 for (i = 0; i < count; i++) { 1038 mpr_regwrite(sc, MPI2_DOORBELL_OFFSET, htole32(data32[i])); 1039 if (mpr_wait_db_ack(sc, 5, sleep_flags) != 0) { 1040 mpr_dprint(sc, MPR_FAULT, 1041 "Timeout while writing doorbell\n"); 1042 return (ENXIO); 1043 } 1044 } 1045 1046 /* Step 6 */ 1047 /* Clock in the reply in 16-bit words. The total length of the 1048 * message is always in the 4th byte, so clock out the first 2 words 1049 * manually, then loop the rest. 1050 */ 1051 data16 = (uint16_t *)reply; 1052 if (mpr_wait_db_int(sc) != 0) { 1053 mpr_dprint(sc, MPR_FAULT, "Timeout reading doorbell 0\n"); 1054 return (ENXIO); 1055 } 1056 1057 /* 1058 * If in a BE platform, swap bytes using le16toh to not 1059 * disturb 8 bit field neighbors in destination structure 1060 * pointed by data16. 1061 */ 1062 data16[0] = 1063 le16toh(mpr_regread(sc, MPI2_DOORBELL_OFFSET)) & MPI2_DOORBELL_DATA_MASK; 1064 mpr_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0); 1065 if (mpr_wait_db_int(sc) != 0) { 1066 mpr_dprint(sc, MPR_FAULT, "Timeout reading doorbell 1\n"); 1067 return (ENXIO); 1068 } 1069 data16[1] = 1070 le16toh(mpr_regread(sc, MPI2_DOORBELL_OFFSET)) & MPI2_DOORBELL_DATA_MASK; 1071 mpr_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0); 1072 1073 /* Number of 32bit words in the message */ 1074 ioc_sz = reply->MsgLength; 1075 1076 /* 1077 * Figure out how many 16bit words to clock in without overrunning. 1078 * The precision loss with dividing reply_sz can safely be 1079 * ignored because the messages can only be multiples of 32bits. 1080 */ 1081 residual = 0; 1082 count = MIN((reply_sz / 4), ioc_sz) * 2; 1083 if (count < ioc_sz * 2) { 1084 residual = ioc_sz * 2 - count; 1085 mpr_dprint(sc, MPR_ERROR, "Driver error, throwing away %d " 1086 "residual message words\n", residual); 1087 } 1088 1089 for (i = 2; i < count; i++) { 1090 if (mpr_wait_db_int(sc) != 0) { 1091 mpr_dprint(sc, MPR_FAULT, 1092 "Timeout reading doorbell %d\n", i); 1093 return (ENXIO); 1094 } 1095 data16[i] = le16toh(mpr_regread(sc, MPI2_DOORBELL_OFFSET)) & 1096 MPI2_DOORBELL_DATA_MASK; 1097 mpr_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0); 1098 } 1099 1100 /* 1101 * Pull out residual words that won't fit into the provided buffer. 1102 * This keeps the chip from hanging due to a driver programming 1103 * error. 1104 */ 1105 while (residual--) { 1106 if (mpr_wait_db_int(sc) != 0) { 1107 mpr_dprint(sc, MPR_FAULT, "Timeout reading doorbell\n"); 1108 return (ENXIO); 1109 } 1110 (void)mpr_regread(sc, MPI2_DOORBELL_OFFSET); 1111 mpr_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0); 1112 } 1113 1114 /* Step 7 */ 1115 if (mpr_wait_db_int(sc) != 0) { 1116 mpr_dprint(sc, MPR_FAULT, "Timeout waiting to exit doorbell\n"); 1117 return (ENXIO); 1118 } 1119 if (mpr_regread(sc, MPI2_DOORBELL_OFFSET) & MPI2_DOORBELL_USED) 1120 mpr_dprint(sc, MPR_FAULT, "Warning, doorbell still active\n"); 1121 mpr_regwrite(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET, 0x0); 1122 1123 return (0); 1124 } 1125 1126 static void 1127 mpr_enqueue_request(struct mpr_softc *sc, struct mpr_command *cm) 1128 { 1129 request_descriptor_t rd; 1130 1131 MPR_FUNCTRACE(sc); 1132 mpr_dprint(sc, MPR_TRACE, "SMID %u cm %p ccb %p\n", 1133 cm->cm_desc.Default.SMID, cm, cm->cm_ccb); 1134 1135 if (sc->mpr_flags & MPR_FLAGS_ATTACH_DONE && !(sc->mpr_flags & 1136 MPR_FLAGS_SHUTDOWN)) 1137 mtx_assert(&sc->mpr_mtx, MA_OWNED); 1138 1139 if (++sc->io_cmds_active > sc->io_cmds_highwater) 1140 sc->io_cmds_highwater++; 1141 1142 KASSERT(cm->cm_state == MPR_CM_STATE_BUSY, 1143 ("command not busy, state = %u\n", cm->cm_state)); 1144 cm->cm_state = MPR_CM_STATE_INQUEUE; 1145 1146 if (sc->atomic_desc_capable) { 1147 rd.u.low = cm->cm_desc.Words.Low; 1148 mpr_regwrite(sc, MPI26_ATOMIC_REQUEST_DESCRIPTOR_POST_OFFSET, 1149 rd.u.low); 1150 } else { 1151 rd.u.low = htole32(cm->cm_desc.Words.Low); 1152 rd.u.high = htole32(cm->cm_desc.Words.High); 1153 mpr_regwrite(sc, MPI2_REQUEST_DESCRIPTOR_POST_LOW_OFFSET, 1154 rd.u.low); 1155 mpr_regwrite(sc, MPI2_REQUEST_DESCRIPTOR_POST_HIGH_OFFSET, 1156 rd.u.high); 1157 } 1158 } 1159 1160 /* 1161 * Ioc facts are read in 16 bit words and and stored with le16toh, 1162 * this takes care of proper U8 fields endianness in 1163 * MPI2_IOC_FACTS_REPLY, but we still need to swap back U16 fields. 1164 */ 1165 static void 1166 adjust_iocfacts_endianness(MPI2_IOC_FACTS_REPLY *facts) 1167 { 1168 facts->HeaderVersion = le16toh(facts->HeaderVersion); 1169 facts->Reserved1 = le16toh(facts->Reserved1); 1170 facts->IOCExceptions = le16toh(facts->IOCExceptions); 1171 facts->IOCStatus = le16toh(facts->IOCStatus); 1172 facts->IOCLogInfo = le32toh(facts->IOCLogInfo); 1173 facts->RequestCredit = le16toh(facts->RequestCredit); 1174 facts->ProductID = le16toh(facts->ProductID); 1175 facts->IOCCapabilities = le32toh(facts->IOCCapabilities); 1176 facts->IOCRequestFrameSize = le16toh(facts->IOCRequestFrameSize); 1177 facts->IOCMaxChainSegmentSize = le16toh(facts->IOCMaxChainSegmentSize); 1178 facts->MaxInitiators = le16toh(facts->MaxInitiators); 1179 facts->MaxTargets = le16toh(facts->MaxTargets); 1180 facts->MaxSasExpanders = le16toh(facts->MaxSasExpanders); 1181 facts->MaxEnclosures = le16toh(facts->MaxEnclosures); 1182 facts->ProtocolFlags = le16toh(facts->ProtocolFlags); 1183 facts->HighPriorityCredit = le16toh(facts->HighPriorityCredit); 1184 facts->MaxReplyDescriptorPostQueueDepth = le16toh(facts->MaxReplyDescriptorPostQueueDepth); 1185 facts->MaxDevHandle = le16toh(facts->MaxDevHandle); 1186 facts->MaxPersistentEntries = le16toh(facts->MaxPersistentEntries); 1187 facts->MinDevHandle = le16toh(facts->MinDevHandle); 1188 } 1189 1190 /* 1191 * Just the FACTS, ma'am. 1192 */ 1193 static int 1194 mpr_get_iocfacts(struct mpr_softc *sc, MPI2_IOC_FACTS_REPLY *facts) 1195 { 1196 MPI2_DEFAULT_REPLY *reply; 1197 MPI2_IOC_FACTS_REQUEST request; 1198 int error, req_sz, reply_sz; 1199 1200 MPR_FUNCTRACE(sc); 1201 mpr_dprint(sc, MPR_INIT, "%s entered\n", __func__); 1202 1203 req_sz = sizeof(MPI2_IOC_FACTS_REQUEST); 1204 reply_sz = sizeof(MPI2_IOC_FACTS_REPLY); 1205 reply = (MPI2_DEFAULT_REPLY *)facts; 1206 1207 bzero(&request, req_sz); 1208 request.Function = MPI2_FUNCTION_IOC_FACTS; 1209 error = mpr_request_sync(sc, &request, reply, req_sz, reply_sz, 5); 1210 1211 adjust_iocfacts_endianness(facts); 1212 mpr_dprint(sc, MPR_TRACE, "facts->IOCCapabilities 0x%x\n", facts->IOCCapabilities); 1213 1214 mpr_dprint(sc, MPR_INIT, "%s exit, error= %d\n", __func__, error); 1215 return (error); 1216 } 1217 1218 static int 1219 mpr_send_iocinit(struct mpr_softc *sc) 1220 { 1221 MPI2_IOC_INIT_REQUEST init; 1222 MPI2_DEFAULT_REPLY reply; 1223 int req_sz, reply_sz, error; 1224 struct timeval now; 1225 uint64_t time_in_msec; 1226 1227 MPR_FUNCTRACE(sc); 1228 mpr_dprint(sc, MPR_INIT, "%s entered\n", __func__); 1229 1230 /* Do a quick sanity check on proper initialization */ 1231 if ((sc->pqdepth == 0) || (sc->fqdepth == 0) || (sc->reqframesz == 0) 1232 || (sc->replyframesz == 0)) { 1233 mpr_dprint(sc, MPR_INIT|MPR_ERROR, 1234 "Driver not fully initialized for IOCInit\n"); 1235 return (EINVAL); 1236 } 1237 1238 req_sz = sizeof(MPI2_IOC_INIT_REQUEST); 1239 reply_sz = sizeof(MPI2_IOC_INIT_REPLY); 1240 bzero(&init, req_sz); 1241 bzero(&reply, reply_sz); 1242 1243 /* 1244 * Fill in the init block. Note that most addresses are 1245 * deliberately in the lower 32bits of memory. This is a micro- 1246 * optimzation for PCI/PCIX, though it's not clear if it helps PCIe. 1247 */ 1248 init.Function = MPI2_FUNCTION_IOC_INIT; 1249 init.WhoInit = MPI2_WHOINIT_HOST_DRIVER; 1250 init.MsgVersion = htole16(MPI2_VERSION); 1251 init.HeaderVersion = htole16(MPI2_HEADER_VERSION); 1252 init.SystemRequestFrameSize = htole16((uint16_t)(sc->reqframesz / 4)); 1253 init.ReplyDescriptorPostQueueDepth = htole16(sc->pqdepth); 1254 init.ReplyFreeQueueDepth = htole16(sc->fqdepth); 1255 init.SenseBufferAddressHigh = 0; 1256 init.SystemReplyAddressHigh = 0; 1257 init.SystemRequestFrameBaseAddress.High = 0; 1258 init.SystemRequestFrameBaseAddress.Low = 1259 htole32((uint32_t)sc->req_busaddr); 1260 init.ReplyDescriptorPostQueueAddress.High = 0; 1261 init.ReplyDescriptorPostQueueAddress.Low = 1262 htole32((uint32_t)sc->post_busaddr); 1263 init.ReplyFreeQueueAddress.High = 0; 1264 init.ReplyFreeQueueAddress.Low = htole32((uint32_t)sc->free_busaddr); 1265 getmicrotime(&now); 1266 time_in_msec = (now.tv_sec * 1000 + now.tv_usec/1000); 1267 init.TimeStamp.High = htole32((time_in_msec >> 32) & 0xFFFFFFFF); 1268 init.TimeStamp.Low = htole32(time_in_msec & 0xFFFFFFFF); 1269 init.HostPageSize = HOST_PAGE_SIZE_4K; 1270 1271 error = mpr_request_sync(sc, &init, &reply, req_sz, reply_sz, 5); 1272 if ((le16toh(reply.IOCStatus) & MPI2_IOCSTATUS_MASK) != MPI2_IOCSTATUS_SUCCESS) 1273 error = ENXIO; 1274 1275 mpr_dprint(sc, MPR_INIT, "IOCInit status= 0x%x\n", le16toh(reply.IOCStatus)); 1276 mpr_dprint(sc, MPR_INIT, "%s exit\n", __func__); 1277 return (error); 1278 } 1279 1280 void 1281 mpr_memaddr_cb(void *arg, bus_dma_segment_t *segs, int nsegs, int error) 1282 { 1283 bus_addr_t *addr; 1284 1285 addr = arg; 1286 *addr = segs[0].ds_addr; 1287 } 1288 1289 void 1290 mpr_memaddr_wait_cb(void *arg, bus_dma_segment_t *segs, int nsegs, int error) 1291 { 1292 struct mpr_busdma_context *ctx; 1293 int need_unload, need_free; 1294 1295 ctx = (struct mpr_busdma_context *)arg; 1296 need_unload = 0; 1297 need_free = 0; 1298 1299 mpr_lock(ctx->softc); 1300 ctx->error = error; 1301 ctx->completed = 1; 1302 if ((error == 0) && (ctx->abandoned == 0)) { 1303 *ctx->addr = segs[0].ds_addr; 1304 } else { 1305 if (nsegs != 0) 1306 need_unload = 1; 1307 if (ctx->abandoned != 0) 1308 need_free = 1; 1309 } 1310 if (need_free == 0) 1311 wakeup(ctx); 1312 1313 mpr_unlock(ctx->softc); 1314 1315 if (need_unload != 0) { 1316 bus_dmamap_unload(ctx->buffer_dmat, 1317 ctx->buffer_dmamap); 1318 *ctx->addr = 0; 1319 } 1320 1321 if (need_free != 0) 1322 free(ctx, M_MPR); 1323 } 1324 1325 static int 1326 mpr_alloc_queues(struct mpr_softc *sc) 1327 { 1328 struct mpr_queue *q; 1329 int nq, i; 1330 1331 nq = sc->msi_msgs; 1332 mpr_dprint(sc, MPR_INIT|MPR_XINFO, "Allocating %d I/O queues\n", nq); 1333 1334 sc->queues = malloc(sizeof(struct mpr_queue) * nq, M_MPR, 1335 M_NOWAIT|M_ZERO); 1336 if (sc->queues == NULL) 1337 return (ENOMEM); 1338 1339 for (i = 0; i < nq; i++) { 1340 q = &sc->queues[i]; 1341 mpr_dprint(sc, MPR_INIT, "Configuring queue %d %p\n", i, q); 1342 q->sc = sc; 1343 q->qnum = i; 1344 } 1345 return (0); 1346 } 1347 1348 static int 1349 mpr_alloc_hw_queues(struct mpr_softc *sc) 1350 { 1351 bus_dma_template_t t; 1352 bus_addr_t queues_busaddr; 1353 uint8_t *queues; 1354 int qsize, fqsize, pqsize; 1355 1356 /* 1357 * The reply free queue contains 4 byte entries in multiples of 16 and 1358 * aligned on a 16 byte boundary. There must always be an unused entry. 1359 * This queue supplies fresh reply frames for the firmware to use. 1360 * 1361 * The reply descriptor post queue contains 8 byte entries in 1362 * multiples of 16 and aligned on a 16 byte boundary. This queue 1363 * contains filled-in reply frames sent from the firmware to the host. 1364 * 1365 * These two queues are allocated together for simplicity. 1366 */ 1367 sc->fqdepth = roundup2(sc->num_replies + 1, 16); 1368 sc->pqdepth = roundup2(sc->num_replies + 1, 16); 1369 fqsize= sc->fqdepth * 4; 1370 pqsize = sc->pqdepth * 8; 1371 qsize = fqsize + pqsize; 1372 1373 bus_dma_template_init(&t, sc->mpr_parent_dmat); 1374 BUS_DMA_TEMPLATE_FILL(&t, BD_ALIGNMENT(16), BD_MAXSIZE(qsize), 1375 BD_MAXSEGSIZE(qsize), BD_NSEGMENTS(1), 1376 BD_LOWADDR(BUS_SPACE_MAXADDR_32BIT)); 1377 if (bus_dma_template_tag(&t, &sc->queues_dmat)) { 1378 mpr_dprint(sc, MPR_ERROR, "Cannot allocate queues DMA tag\n"); 1379 return (ENOMEM); 1380 } 1381 if (bus_dmamem_alloc(sc->queues_dmat, (void **)&queues, BUS_DMA_NOWAIT, 1382 &sc->queues_map)) { 1383 mpr_dprint(sc, MPR_ERROR, "Cannot allocate queues memory\n"); 1384 return (ENOMEM); 1385 } 1386 bzero(queues, qsize); 1387 bus_dmamap_load(sc->queues_dmat, sc->queues_map, queues, qsize, 1388 mpr_memaddr_cb, &queues_busaddr, 0); 1389 1390 sc->free_queue = (uint32_t *)queues; 1391 sc->free_busaddr = queues_busaddr; 1392 sc->post_queue = (MPI2_REPLY_DESCRIPTORS_UNION *)(queues + fqsize); 1393 sc->post_busaddr = queues_busaddr + fqsize; 1394 mpr_dprint(sc, MPR_INIT, "free queue busaddr= %#016jx size= %d\n", 1395 (uintmax_t)sc->free_busaddr, fqsize); 1396 mpr_dprint(sc, MPR_INIT, "reply queue busaddr= %#016jx size= %d\n", 1397 (uintmax_t)sc->post_busaddr, pqsize); 1398 1399 return (0); 1400 } 1401 1402 static int 1403 mpr_alloc_replies(struct mpr_softc *sc) 1404 { 1405 bus_dma_template_t t; 1406 int rsize, num_replies; 1407 1408 /* Store the reply frame size in bytes rather than as 32bit words */ 1409 sc->replyframesz = sc->facts->ReplyFrameSize * 4; 1410 1411 /* 1412 * sc->num_replies should be one less than sc->fqdepth. We need to 1413 * allocate space for sc->fqdepth replies, but only sc->num_replies 1414 * replies can be used at once. 1415 */ 1416 num_replies = max(sc->fqdepth, sc->num_replies); 1417 1418 rsize = sc->replyframesz * num_replies; 1419 bus_dma_template_init(&t, sc->mpr_parent_dmat); 1420 BUS_DMA_TEMPLATE_FILL(&t, BD_ALIGNMENT(4), BD_MAXSIZE(rsize), 1421 BD_MAXSEGSIZE(rsize), BD_NSEGMENTS(1), 1422 BD_LOWADDR(BUS_SPACE_MAXADDR_32BIT)); 1423 if (bus_dma_template_tag(&t, &sc->reply_dmat)) { 1424 mpr_dprint(sc, MPR_ERROR, "Cannot allocate replies DMA tag\n"); 1425 return (ENOMEM); 1426 } 1427 if (bus_dmamem_alloc(sc->reply_dmat, (void **)&sc->reply_frames, 1428 BUS_DMA_NOWAIT, &sc->reply_map)) { 1429 mpr_dprint(sc, MPR_ERROR, "Cannot allocate replies memory\n"); 1430 return (ENOMEM); 1431 } 1432 bzero(sc->reply_frames, rsize); 1433 bus_dmamap_load(sc->reply_dmat, sc->reply_map, sc->reply_frames, rsize, 1434 mpr_memaddr_cb, &sc->reply_busaddr, 0); 1435 mpr_dprint(sc, MPR_INIT, "reply frames busaddr= %#016jx size= %d\n", 1436 (uintmax_t)sc->reply_busaddr, rsize); 1437 1438 return (0); 1439 } 1440 1441 static void 1442 mpr_load_chains_cb(void *arg, bus_dma_segment_t *segs, int nsegs, int error) 1443 { 1444 struct mpr_softc *sc = arg; 1445 struct mpr_chain *chain; 1446 bus_size_t bo; 1447 int i, o, s; 1448 1449 if (error != 0) 1450 return; 1451 1452 for (i = 0, o = 0, s = 0; s < nsegs; s++) { 1453 for (bo = 0; bo + sc->chain_frame_size <= segs[s].ds_len; 1454 bo += sc->chain_frame_size) { 1455 chain = &sc->chains[i++]; 1456 chain->chain =(MPI2_SGE_IO_UNION *)(sc->chain_frames+o); 1457 chain->chain_busaddr = segs[s].ds_addr + bo; 1458 o += sc->chain_frame_size; 1459 mpr_free_chain(sc, chain); 1460 } 1461 if (bo != segs[s].ds_len) 1462 o += segs[s].ds_len - bo; 1463 } 1464 sc->chain_free_lowwater = i; 1465 } 1466 1467 static int 1468 mpr_alloc_requests(struct mpr_softc *sc) 1469 { 1470 bus_dma_template_t t; 1471 struct mpr_command *cm; 1472 int i, rsize, nsegs; 1473 1474 rsize = sc->reqframesz * sc->num_reqs; 1475 bus_dma_template_init(&t, sc->mpr_parent_dmat); 1476 BUS_DMA_TEMPLATE_FILL(&t, BD_ALIGNMENT(16), BD_MAXSIZE(rsize), 1477 BD_MAXSEGSIZE(rsize), BD_NSEGMENTS(1), 1478 BD_LOWADDR(BUS_SPACE_MAXADDR_32BIT)); 1479 if (bus_dma_template_tag(&t, &sc->req_dmat)) { 1480 mpr_dprint(sc, MPR_ERROR, "Cannot allocate request DMA tag\n"); 1481 return (ENOMEM); 1482 } 1483 if (bus_dmamem_alloc(sc->req_dmat, (void **)&sc->req_frames, 1484 BUS_DMA_NOWAIT, &sc->req_map)) { 1485 mpr_dprint(sc, MPR_ERROR, "Cannot allocate request memory\n"); 1486 return (ENOMEM); 1487 } 1488 bzero(sc->req_frames, rsize); 1489 bus_dmamap_load(sc->req_dmat, sc->req_map, sc->req_frames, rsize, 1490 mpr_memaddr_cb, &sc->req_busaddr, 0); 1491 mpr_dprint(sc, MPR_INIT, "request frames busaddr= %#016jx size= %d\n", 1492 (uintmax_t)sc->req_busaddr, rsize); 1493 1494 sc->chains = malloc(sizeof(struct mpr_chain) * sc->num_chains, M_MPR, 1495 M_NOWAIT | M_ZERO); 1496 if (!sc->chains) { 1497 mpr_dprint(sc, MPR_ERROR, "Cannot allocate chain memory\n"); 1498 return (ENOMEM); 1499 } 1500 rsize = sc->chain_frame_size * sc->num_chains; 1501 bus_dma_template_init(&t, sc->mpr_parent_dmat); 1502 BUS_DMA_TEMPLATE_FILL(&t, BD_ALIGNMENT(16), BD_MAXSIZE(rsize), 1503 BD_MAXSEGSIZE(rsize), BD_NSEGMENTS((howmany(rsize, PAGE_SIZE)))); 1504 if (bus_dma_template_tag(&t, &sc->chain_dmat)) { 1505 mpr_dprint(sc, MPR_ERROR, "Cannot allocate chain DMA tag\n"); 1506 return (ENOMEM); 1507 } 1508 if (bus_dmamem_alloc(sc->chain_dmat, (void **)&sc->chain_frames, 1509 BUS_DMA_NOWAIT | BUS_DMA_ZERO, &sc->chain_map)) { 1510 mpr_dprint(sc, MPR_ERROR, "Cannot allocate chain memory\n"); 1511 return (ENOMEM); 1512 } 1513 if (bus_dmamap_load(sc->chain_dmat, sc->chain_map, sc->chain_frames, 1514 rsize, mpr_load_chains_cb, sc, BUS_DMA_NOWAIT)) { 1515 mpr_dprint(sc, MPR_ERROR, "Cannot load chain memory\n"); 1516 bus_dmamem_free(sc->chain_dmat, sc->chain_frames, 1517 sc->chain_map); 1518 return (ENOMEM); 1519 } 1520 1521 rsize = MPR_SENSE_LEN * sc->num_reqs; 1522 bus_dma_template_clone(&t, sc->req_dmat); 1523 BUS_DMA_TEMPLATE_FILL(&t, BD_ALIGNMENT(1), BD_MAXSIZE(rsize), 1524 BD_MAXSEGSIZE(rsize)); 1525 if (bus_dma_template_tag(&t, &sc->sense_dmat)) { 1526 mpr_dprint(sc, MPR_ERROR, "Cannot allocate sense DMA tag\n"); 1527 return (ENOMEM); 1528 } 1529 if (bus_dmamem_alloc(sc->sense_dmat, (void **)&sc->sense_frames, 1530 BUS_DMA_NOWAIT, &sc->sense_map)) { 1531 mpr_dprint(sc, MPR_ERROR, "Cannot allocate sense memory\n"); 1532 return (ENOMEM); 1533 } 1534 bzero(sc->sense_frames, rsize); 1535 bus_dmamap_load(sc->sense_dmat, sc->sense_map, sc->sense_frames, rsize, 1536 mpr_memaddr_cb, &sc->sense_busaddr, 0); 1537 mpr_dprint(sc, MPR_INIT, "sense frames busaddr= %#016jx size= %d\n", 1538 (uintmax_t)sc->sense_busaddr, rsize); 1539 1540 /* 1541 * Allocate NVMe PRP Pages for NVMe SGL support only if the FW supports 1542 * these devices. 1543 */ 1544 if ((sc->facts->MsgVersion >= MPI2_VERSION_02_06) && 1545 (sc->facts->ProtocolFlags & MPI2_IOCFACTS_PROTOCOL_NVME_DEVICES)) { 1546 if (mpr_alloc_nvme_prp_pages(sc) == ENOMEM) 1547 return (ENOMEM); 1548 } 1549 1550 nsegs = (sc->maxio / PAGE_SIZE) + 1; 1551 bus_dma_template_init(&t, sc->mpr_parent_dmat); 1552 BUS_DMA_TEMPLATE_FILL(&t, BD_MAXSIZE(BUS_SPACE_MAXSIZE_32BIT), 1553 BD_NSEGMENTS(nsegs), BD_MAXSEGSIZE(BUS_SPACE_MAXSIZE_32BIT), 1554 BD_FLAGS(BUS_DMA_ALLOCNOW), BD_LOCKFUNC(busdma_lock_mutex), 1555 BD_LOCKFUNCARG(&sc->mpr_mtx)); 1556 if (bus_dma_template_tag(&t, &sc->buffer_dmat)) { 1557 mpr_dprint(sc, MPR_ERROR, "Cannot allocate buffer DMA tag\n"); 1558 return (ENOMEM); 1559 } 1560 1561 /* 1562 * SMID 0 cannot be used as a free command per the firmware spec. 1563 * Just drop that command instead of risking accounting bugs. 1564 */ 1565 sc->commands = malloc(sizeof(struct mpr_command) * sc->num_reqs, 1566 M_MPR, M_WAITOK | M_ZERO); 1567 for (i = 1; i < sc->num_reqs; i++) { 1568 cm = &sc->commands[i]; 1569 cm->cm_req = sc->req_frames + i * sc->reqframesz; 1570 cm->cm_req_busaddr = sc->req_busaddr + i * sc->reqframesz; 1571 cm->cm_sense = &sc->sense_frames[i]; 1572 cm->cm_sense_busaddr = sc->sense_busaddr + i * MPR_SENSE_LEN; 1573 cm->cm_desc.Default.SMID = htole16(i); 1574 cm->cm_sc = sc; 1575 cm->cm_state = MPR_CM_STATE_BUSY; 1576 TAILQ_INIT(&cm->cm_chain_list); 1577 TAILQ_INIT(&cm->cm_prp_page_list); 1578 callout_init_mtx(&cm->cm_callout, &sc->mpr_mtx, 0); 1579 1580 /* XXX Is a failure here a critical problem? */ 1581 if (bus_dmamap_create(sc->buffer_dmat, 0, &cm->cm_dmamap) 1582 == 0) { 1583 if (i <= sc->num_prireqs) 1584 mpr_free_high_priority_command(sc, cm); 1585 else 1586 mpr_free_command(sc, cm); 1587 } else { 1588 panic("failed to allocate command %d\n", i); 1589 sc->num_reqs = i; 1590 break; 1591 } 1592 } 1593 1594 return (0); 1595 } 1596 1597 /* 1598 * Allocate contiguous buffers for PCIe NVMe devices for building native PRPs, 1599 * which are scatter/gather lists for NVMe devices. 1600 * 1601 * This buffer must be contiguous due to the nature of how NVMe PRPs are built 1602 * and translated by FW. 1603 * 1604 * returns ENOMEM if memory could not be allocated, otherwise returns 0. 1605 */ 1606 static int 1607 mpr_alloc_nvme_prp_pages(struct mpr_softc *sc) 1608 { 1609 bus_dma_template_t t; 1610 struct mpr_prp_page *prp_page; 1611 int PRPs_per_page, PRPs_required, pages_required; 1612 int rsize, i; 1613 1614 /* 1615 * Assuming a MAX_IO_SIZE of 1MB and a PAGE_SIZE of 4k, the max number 1616 * of PRPs (NVMe's Scatter/Gather Element) needed per I/O is: 1617 * MAX_IO_SIZE / PAGE_SIZE = 256 1618 * 1619 * 1 PRP entry in main frame for PRP list pointer still leaves 255 PRPs 1620 * required for the remainder of the 1MB I/O. 512 PRPs can fit into one 1621 * page (4096 / 8 = 512), so only one page is required for each I/O. 1622 * 1623 * Each of these buffers will need to be contiguous. For simplicity, 1624 * only one buffer is allocated here, which has all of the space 1625 * required for the NVMe Queue Depth. If there are problems allocating 1626 * this one buffer, this function will need to change to allocate 1627 * individual, contiguous NVME_QDEPTH buffers. 1628 * 1629 * The real calculation will use the real max io size. Above is just an 1630 * example. 1631 * 1632 */ 1633 PRPs_required = sc->maxio / PAGE_SIZE; 1634 PRPs_per_page = (PAGE_SIZE / PRP_ENTRY_SIZE) - 1; 1635 pages_required = (PRPs_required / PRPs_per_page) + 1; 1636 1637 sc->prp_buffer_size = PAGE_SIZE * pages_required; 1638 rsize = sc->prp_buffer_size * NVME_QDEPTH; 1639 bus_dma_template_init(&t, sc->mpr_parent_dmat); 1640 BUS_DMA_TEMPLATE_FILL(&t, BD_ALIGNMENT(4), BD_MAXSIZE(rsize), 1641 BD_MAXSEGSIZE(rsize), BD_NSEGMENTS(1), 1642 BD_LOWADDR(BUS_SPACE_MAXADDR_32BIT)); 1643 if (bus_dma_template_tag(&t, &sc->prp_page_dmat)) { 1644 mpr_dprint(sc, MPR_ERROR, "Cannot allocate NVMe PRP DMA " 1645 "tag\n"); 1646 return (ENOMEM); 1647 } 1648 if (bus_dmamem_alloc(sc->prp_page_dmat, (void **)&sc->prp_pages, 1649 BUS_DMA_NOWAIT, &sc->prp_page_map)) { 1650 mpr_dprint(sc, MPR_ERROR, "Cannot allocate NVMe PRP memory\n"); 1651 return (ENOMEM); 1652 } 1653 bzero(sc->prp_pages, rsize); 1654 bus_dmamap_load(sc->prp_page_dmat, sc->prp_page_map, sc->prp_pages, 1655 rsize, mpr_memaddr_cb, &sc->prp_page_busaddr, 0); 1656 1657 sc->prps = malloc(sizeof(struct mpr_prp_page) * NVME_QDEPTH, M_MPR, 1658 M_WAITOK | M_ZERO); 1659 for (i = 0; i < NVME_QDEPTH; i++) { 1660 prp_page = &sc->prps[i]; 1661 prp_page->prp_page = (uint64_t *)(sc->prp_pages + 1662 i * sc->prp_buffer_size); 1663 prp_page->prp_page_busaddr = (uint64_t)(sc->prp_page_busaddr + 1664 i * sc->prp_buffer_size); 1665 mpr_free_prp_page(sc, prp_page); 1666 sc->prp_pages_free_lowwater++; 1667 } 1668 1669 return (0); 1670 } 1671 1672 static int 1673 mpr_init_queues(struct mpr_softc *sc) 1674 { 1675 int i; 1676 1677 memset((uint8_t *)sc->post_queue, 0xff, sc->pqdepth * 8); 1678 1679 /* 1680 * According to the spec, we need to use one less reply than we 1681 * have space for on the queue. So sc->num_replies (the number we 1682 * use) should be less than sc->fqdepth (allocated size). 1683 */ 1684 if (sc->num_replies >= sc->fqdepth) 1685 return (EINVAL); 1686 1687 /* 1688 * Initialize all of the free queue entries. 1689 */ 1690 for (i = 0; i < sc->fqdepth; i++) { 1691 sc->free_queue[i] = htole32(sc->reply_busaddr + (i * sc->replyframesz)); 1692 } 1693 sc->replyfreeindex = sc->num_replies; 1694 1695 return (0); 1696 } 1697 1698 /* Get the driver parameter tunables. Lowest priority are the driver defaults. 1699 * Next are the global settings, if they exist. Highest are the per-unit 1700 * settings, if they exist. 1701 */ 1702 void 1703 mpr_get_tunables(struct mpr_softc *sc) 1704 { 1705 char tmpstr[80], mpr_debug[80]; 1706 1707 /* XXX default to some debugging for now */ 1708 sc->mpr_debug = MPR_INFO | MPR_FAULT; 1709 sc->disable_msix = 0; 1710 sc->disable_msi = 0; 1711 sc->max_msix = MPR_MSIX_MAX; 1712 sc->max_chains = MPR_CHAIN_FRAMES; 1713 sc->max_io_pages = MPR_MAXIO_PAGES; 1714 sc->enable_ssu = MPR_SSU_ENABLE_SSD_DISABLE_HDD; 1715 sc->spinup_wait_time = DEFAULT_SPINUP_WAIT; 1716 sc->use_phynum = 1; 1717 sc->max_reqframes = MPR_REQ_FRAMES; 1718 sc->max_prireqframes = MPR_PRI_REQ_FRAMES; 1719 sc->max_replyframes = MPR_REPLY_FRAMES; 1720 sc->max_evtframes = MPR_EVT_REPLY_FRAMES; 1721 1722 /* 1723 * Grab the global variables. 1724 */ 1725 bzero(mpr_debug, 80); 1726 if (TUNABLE_STR_FETCH("hw.mpr.debug_level", mpr_debug, 80) != 0) 1727 mpr_parse_debug(sc, mpr_debug); 1728 TUNABLE_INT_FETCH("hw.mpr.disable_msix", &sc->disable_msix); 1729 TUNABLE_INT_FETCH("hw.mpr.disable_msi", &sc->disable_msi); 1730 TUNABLE_INT_FETCH("hw.mpr.max_msix", &sc->max_msix); 1731 TUNABLE_INT_FETCH("hw.mpr.max_chains", &sc->max_chains); 1732 TUNABLE_INT_FETCH("hw.mpr.max_io_pages", &sc->max_io_pages); 1733 TUNABLE_INT_FETCH("hw.mpr.enable_ssu", &sc->enable_ssu); 1734 TUNABLE_INT_FETCH("hw.mpr.spinup_wait_time", &sc->spinup_wait_time); 1735 TUNABLE_INT_FETCH("hw.mpr.use_phy_num", &sc->use_phynum); 1736 TUNABLE_INT_FETCH("hw.mpr.max_reqframes", &sc->max_reqframes); 1737 TUNABLE_INT_FETCH("hw.mpr.max_prireqframes", &sc->max_prireqframes); 1738 TUNABLE_INT_FETCH("hw.mpr.max_replyframes", &sc->max_replyframes); 1739 TUNABLE_INT_FETCH("hw.mpr.max_evtframes", &sc->max_evtframes); 1740 1741 /* Grab the unit-instance variables */ 1742 snprintf(tmpstr, sizeof(tmpstr), "dev.mpr.%d.debug_level", 1743 device_get_unit(sc->mpr_dev)); 1744 bzero(mpr_debug, 80); 1745 if (TUNABLE_STR_FETCH(tmpstr, mpr_debug, 80) != 0) 1746 mpr_parse_debug(sc, mpr_debug); 1747 1748 snprintf(tmpstr, sizeof(tmpstr), "dev.mpr.%d.disable_msix", 1749 device_get_unit(sc->mpr_dev)); 1750 TUNABLE_INT_FETCH(tmpstr, &sc->disable_msix); 1751 1752 snprintf(tmpstr, sizeof(tmpstr), "dev.mpr.%d.disable_msi", 1753 device_get_unit(sc->mpr_dev)); 1754 TUNABLE_INT_FETCH(tmpstr, &sc->disable_msi); 1755 1756 snprintf(tmpstr, sizeof(tmpstr), "dev.mpr.%d.max_msix", 1757 device_get_unit(sc->mpr_dev)); 1758 TUNABLE_INT_FETCH(tmpstr, &sc->max_msix); 1759 1760 snprintf(tmpstr, sizeof(tmpstr), "dev.mpr.%d.max_chains", 1761 device_get_unit(sc->mpr_dev)); 1762 TUNABLE_INT_FETCH(tmpstr, &sc->max_chains); 1763 1764 snprintf(tmpstr, sizeof(tmpstr), "dev.mpr.%d.max_io_pages", 1765 device_get_unit(sc->mpr_dev)); 1766 TUNABLE_INT_FETCH(tmpstr, &sc->max_io_pages); 1767 1768 bzero(sc->exclude_ids, sizeof(sc->exclude_ids)); 1769 snprintf(tmpstr, sizeof(tmpstr), "dev.mpr.%d.exclude_ids", 1770 device_get_unit(sc->mpr_dev)); 1771 TUNABLE_STR_FETCH(tmpstr, sc->exclude_ids, sizeof(sc->exclude_ids)); 1772 1773 snprintf(tmpstr, sizeof(tmpstr), "dev.mpr.%d.enable_ssu", 1774 device_get_unit(sc->mpr_dev)); 1775 TUNABLE_INT_FETCH(tmpstr, &sc->enable_ssu); 1776 1777 snprintf(tmpstr, sizeof(tmpstr), "dev.mpr.%d.spinup_wait_time", 1778 device_get_unit(sc->mpr_dev)); 1779 TUNABLE_INT_FETCH(tmpstr, &sc->spinup_wait_time); 1780 1781 snprintf(tmpstr, sizeof(tmpstr), "dev.mpr.%d.use_phy_num", 1782 device_get_unit(sc->mpr_dev)); 1783 TUNABLE_INT_FETCH(tmpstr, &sc->use_phynum); 1784 1785 snprintf(tmpstr, sizeof(tmpstr), "dev.mpr.%d.max_reqframes", 1786 device_get_unit(sc->mpr_dev)); 1787 TUNABLE_INT_FETCH(tmpstr, &sc->max_reqframes); 1788 1789 snprintf(tmpstr, sizeof(tmpstr), "dev.mpr.%d.max_prireqframes", 1790 device_get_unit(sc->mpr_dev)); 1791 TUNABLE_INT_FETCH(tmpstr, &sc->max_prireqframes); 1792 1793 snprintf(tmpstr, sizeof(tmpstr), "dev.mpr.%d.max_replyframes", 1794 device_get_unit(sc->mpr_dev)); 1795 TUNABLE_INT_FETCH(tmpstr, &sc->max_replyframes); 1796 1797 snprintf(tmpstr, sizeof(tmpstr), "dev.mpr.%d.max_evtframes", 1798 device_get_unit(sc->mpr_dev)); 1799 TUNABLE_INT_FETCH(tmpstr, &sc->max_evtframes); 1800 } 1801 1802 static void 1803 mpr_setup_sysctl(struct mpr_softc *sc) 1804 { 1805 struct sysctl_ctx_list *sysctl_ctx = NULL; 1806 struct sysctl_oid *sysctl_tree = NULL; 1807 char tmpstr[80], tmpstr2[80]; 1808 1809 /* 1810 * Setup the sysctl variable so the user can change the debug level 1811 * on the fly. 1812 */ 1813 snprintf(tmpstr, sizeof(tmpstr), "MPR controller %d", 1814 device_get_unit(sc->mpr_dev)); 1815 snprintf(tmpstr2, sizeof(tmpstr2), "%d", device_get_unit(sc->mpr_dev)); 1816 1817 sysctl_ctx = device_get_sysctl_ctx(sc->mpr_dev); 1818 if (sysctl_ctx != NULL) 1819 sysctl_tree = device_get_sysctl_tree(sc->mpr_dev); 1820 1821 if (sysctl_tree == NULL) { 1822 sysctl_ctx_init(&sc->sysctl_ctx); 1823 sc->sysctl_tree = SYSCTL_ADD_NODE(&sc->sysctl_ctx, 1824 SYSCTL_STATIC_CHILDREN(_hw_mpr), OID_AUTO, tmpstr2, 1825 CTLFLAG_RD | CTLFLAG_MPSAFE, 0, tmpstr); 1826 if (sc->sysctl_tree == NULL) 1827 return; 1828 sysctl_ctx = &sc->sysctl_ctx; 1829 sysctl_tree = sc->sysctl_tree; 1830 } 1831 1832 SYSCTL_ADD_PROC(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree), 1833 OID_AUTO, "debug_level", CTLTYPE_STRING | CTLFLAG_RW | CTLFLAG_MPSAFE, 1834 sc, 0, mpr_debug_sysctl, "A", "mpr debug level"); 1835 1836 SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree), 1837 OID_AUTO, "disable_msix", CTLFLAG_RD, &sc->disable_msix, 0, 1838 "Disable the use of MSI-X interrupts"); 1839 1840 SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree), 1841 OID_AUTO, "max_msix", CTLFLAG_RD, &sc->max_msix, 0, 1842 "User-defined maximum number of MSIX queues"); 1843 1844 SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree), 1845 OID_AUTO, "msix_msgs", CTLFLAG_RD, &sc->msi_msgs, 0, 1846 "Negotiated number of MSIX queues"); 1847 1848 SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree), 1849 OID_AUTO, "max_reqframes", CTLFLAG_RD, &sc->max_reqframes, 0, 1850 "Total number of allocated request frames"); 1851 1852 SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree), 1853 OID_AUTO, "max_prireqframes", CTLFLAG_RD, &sc->max_prireqframes, 0, 1854 "Total number of allocated high priority request frames"); 1855 1856 SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree), 1857 OID_AUTO, "max_replyframes", CTLFLAG_RD, &sc->max_replyframes, 0, 1858 "Total number of allocated reply frames"); 1859 1860 SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree), 1861 OID_AUTO, "max_evtframes", CTLFLAG_RD, &sc->max_evtframes, 0, 1862 "Total number of event frames allocated"); 1863 1864 SYSCTL_ADD_STRING(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree), 1865 OID_AUTO, "firmware_version", CTLFLAG_RD, sc->fw_version, 1866 strlen(sc->fw_version), "firmware version"); 1867 1868 SYSCTL_ADD_STRING(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree), 1869 OID_AUTO, "driver_version", CTLFLAG_RD, MPR_DRIVER_VERSION, 1870 strlen(MPR_DRIVER_VERSION), "driver version"); 1871 1872 SYSCTL_ADD_STRING(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree), 1873 OID_AUTO, "msg_version", CTLFLAG_RD, sc->msg_version, 1874 strlen(sc->msg_version), "message interface version (deprecated)"); 1875 1876 SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree), 1877 OID_AUTO, "io_cmds_active", CTLFLAG_RD, 1878 &sc->io_cmds_active, 0, "number of currently active commands"); 1879 1880 SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree), 1881 OID_AUTO, "io_cmds_highwater", CTLFLAG_RD, 1882 &sc->io_cmds_highwater, 0, "maximum active commands seen"); 1883 1884 SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree), 1885 OID_AUTO, "chain_free", CTLFLAG_RD, 1886 &sc->chain_free, 0, "number of free chain elements"); 1887 1888 SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree), 1889 OID_AUTO, "chain_free_lowwater", CTLFLAG_RD, 1890 &sc->chain_free_lowwater, 0,"lowest number of free chain elements"); 1891 1892 SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree), 1893 OID_AUTO, "max_chains", CTLFLAG_RD, 1894 &sc->max_chains, 0,"maximum chain frames that will be allocated"); 1895 1896 SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree), 1897 OID_AUTO, "max_io_pages", CTLFLAG_RD, 1898 &sc->max_io_pages, 0,"maximum pages to allow per I/O (if <1 use " 1899 "IOCFacts)"); 1900 1901 SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree), 1902 OID_AUTO, "enable_ssu", CTLFLAG_RW, &sc->enable_ssu, 0, 1903 "enable SSU to SATA SSD/HDD at shutdown"); 1904 1905 SYSCTL_ADD_UQUAD(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree), 1906 OID_AUTO, "chain_alloc_fail", CTLFLAG_RD, 1907 &sc->chain_alloc_fail, "chain allocation failures"); 1908 1909 SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree), 1910 OID_AUTO, "spinup_wait_time", CTLFLAG_RD, 1911 &sc->spinup_wait_time, DEFAULT_SPINUP_WAIT, "seconds to wait for " 1912 "spinup after SATA ID error"); 1913 1914 SYSCTL_ADD_PROC(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree), 1915 OID_AUTO, "dump_reqs", 1916 CTLTYPE_OPAQUE | CTLFLAG_RD | CTLFLAG_SKIP | CTLFLAG_MPSAFE, 1917 sc, 0, mpr_dump_reqs, "I", "Dump Active Requests"); 1918 1919 SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree), 1920 OID_AUTO, "dump_reqs_alltypes", CTLFLAG_RW, 1921 &sc->dump_reqs_alltypes, 0, 1922 "dump all request types not just inqueue"); 1923 1924 SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree), 1925 OID_AUTO, "use_phy_num", CTLFLAG_RD, &sc->use_phynum, 0, 1926 "Use the phy number for enumeration"); 1927 1928 SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree), 1929 OID_AUTO, "prp_pages_free", CTLFLAG_RD, 1930 &sc->prp_pages_free, 0, "number of free PRP pages"); 1931 1932 SYSCTL_ADD_INT(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree), 1933 OID_AUTO, "prp_pages_free_lowwater", CTLFLAG_RD, 1934 &sc->prp_pages_free_lowwater, 0,"lowest number of free PRP pages"); 1935 1936 SYSCTL_ADD_UQUAD(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree), 1937 OID_AUTO, "prp_page_alloc_fail", CTLFLAG_RD, 1938 &sc->prp_page_alloc_fail, "PRP page allocation failures"); 1939 } 1940 1941 static struct mpr_debug_string { 1942 char *name; 1943 int flag; 1944 } mpr_debug_strings[] = { 1945 {"info", MPR_INFO}, 1946 {"fault", MPR_FAULT}, 1947 {"event", MPR_EVENT}, 1948 {"log", MPR_LOG}, 1949 {"recovery", MPR_RECOVERY}, 1950 {"error", MPR_ERROR}, 1951 {"init", MPR_INIT}, 1952 {"xinfo", MPR_XINFO}, 1953 {"user", MPR_USER}, 1954 {"mapping", MPR_MAPPING}, 1955 {"trace", MPR_TRACE} 1956 }; 1957 1958 enum mpr_debug_level_combiner { 1959 COMB_NONE, 1960 COMB_ADD, 1961 COMB_SUB 1962 }; 1963 1964 static int 1965 mpr_debug_sysctl(SYSCTL_HANDLER_ARGS) 1966 { 1967 struct mpr_softc *sc; 1968 struct mpr_debug_string *string; 1969 struct sbuf *sbuf; 1970 char *buffer; 1971 size_t sz; 1972 int i, len, debug, error; 1973 1974 sc = (struct mpr_softc *)arg1; 1975 1976 error = sysctl_wire_old_buffer(req, 0); 1977 if (error != 0) 1978 return (error); 1979 1980 sbuf = sbuf_new_for_sysctl(NULL, NULL, 128, req); 1981 debug = sc->mpr_debug; 1982 1983 sbuf_printf(sbuf, "%#x", debug); 1984 1985 sz = sizeof(mpr_debug_strings) / sizeof(mpr_debug_strings[0]); 1986 for (i = 0; i < sz; i++) { 1987 string = &mpr_debug_strings[i]; 1988 if (debug & string->flag) 1989 sbuf_printf(sbuf, ",%s", string->name); 1990 } 1991 1992 error = sbuf_finish(sbuf); 1993 sbuf_delete(sbuf); 1994 1995 if (error || req->newptr == NULL) 1996 return (error); 1997 1998 len = req->newlen - req->newidx; 1999 if (len == 0) 2000 return (0); 2001 2002 buffer = malloc(len, M_MPR, M_ZERO|M_WAITOK); 2003 error = SYSCTL_IN(req, buffer, len); 2004 2005 mpr_parse_debug(sc, buffer); 2006 2007 free(buffer, M_MPR); 2008 return (error); 2009 } 2010 2011 static void 2012 mpr_parse_debug(struct mpr_softc *sc, char *list) 2013 { 2014 struct mpr_debug_string *string; 2015 enum mpr_debug_level_combiner op; 2016 char *token, *endtoken; 2017 size_t sz; 2018 int flags, i; 2019 2020 if (list == NULL || *list == '\0') 2021 return; 2022 2023 if (*list == '+') { 2024 op = COMB_ADD; 2025 list++; 2026 } else if (*list == '-') { 2027 op = COMB_SUB; 2028 list++; 2029 } else 2030 op = COMB_NONE; 2031 if (*list == '\0') 2032 return; 2033 2034 flags = 0; 2035 sz = sizeof(mpr_debug_strings) / sizeof(mpr_debug_strings[0]); 2036 while ((token = strsep(&list, ":,")) != NULL) { 2037 /* Handle integer flags */ 2038 flags |= strtol(token, &endtoken, 0); 2039 if (token != endtoken) 2040 continue; 2041 2042 /* Handle text flags */ 2043 for (i = 0; i < sz; i++) { 2044 string = &mpr_debug_strings[i]; 2045 if (strcasecmp(token, string->name) == 0) { 2046 flags |= string->flag; 2047 break; 2048 } 2049 } 2050 } 2051 2052 switch (op) { 2053 case COMB_NONE: 2054 sc->mpr_debug = flags; 2055 break; 2056 case COMB_ADD: 2057 sc->mpr_debug |= flags; 2058 break; 2059 case COMB_SUB: 2060 sc->mpr_debug &= (~flags); 2061 break; 2062 } 2063 return; 2064 } 2065 2066 struct mpr_dumpreq_hdr { 2067 uint32_t smid; 2068 uint32_t state; 2069 uint32_t numframes; 2070 uint32_t deschi; 2071 uint32_t desclo; 2072 }; 2073 2074 static int 2075 mpr_dump_reqs(SYSCTL_HANDLER_ARGS) 2076 { 2077 struct mpr_softc *sc; 2078 struct mpr_chain *chain, *chain1; 2079 struct mpr_command *cm; 2080 struct mpr_dumpreq_hdr hdr; 2081 struct sbuf *sb; 2082 uint32_t smid, state; 2083 int i, numreqs, error = 0; 2084 2085 sc = (struct mpr_softc *)arg1; 2086 2087 if ((error = priv_check(curthread, PRIV_DRIVER)) != 0) { 2088 printf("priv check error %d\n", error); 2089 return (error); 2090 } 2091 2092 state = MPR_CM_STATE_INQUEUE; 2093 smid = 1; 2094 numreqs = sc->num_reqs; 2095 2096 if (req->newptr != NULL) 2097 return (EINVAL); 2098 2099 if (smid == 0 || smid > sc->num_reqs) 2100 return (EINVAL); 2101 if (numreqs <= 0 || (numreqs + smid > sc->num_reqs)) 2102 numreqs = sc->num_reqs; 2103 sb = sbuf_new_for_sysctl(NULL, NULL, 4096, req); 2104 2105 /* Best effort, no locking */ 2106 for (i = smid; i < numreqs; i++) { 2107 cm = &sc->commands[i]; 2108 if ((sc->dump_reqs_alltypes == 0) && (cm->cm_state != state)) 2109 continue; 2110 hdr.smid = i; 2111 hdr.state = cm->cm_state; 2112 hdr.numframes = 1; 2113 hdr.deschi = cm->cm_desc.Words.High; 2114 hdr.desclo = cm->cm_desc.Words.Low; 2115 TAILQ_FOREACH_SAFE(chain, &cm->cm_chain_list, chain_link, 2116 chain1) 2117 hdr.numframes++; 2118 sbuf_bcat(sb, &hdr, sizeof(hdr)); 2119 sbuf_bcat(sb, cm->cm_req, 128); 2120 TAILQ_FOREACH_SAFE(chain, &cm->cm_chain_list, chain_link, 2121 chain1) 2122 sbuf_bcat(sb, chain->chain, 128); 2123 } 2124 2125 error = sbuf_finish(sb); 2126 sbuf_delete(sb); 2127 return (error); 2128 } 2129 2130 int 2131 mpr_attach(struct mpr_softc *sc) 2132 { 2133 int error; 2134 2135 MPR_FUNCTRACE(sc); 2136 mpr_dprint(sc, MPR_INIT, "%s entered\n", __func__); 2137 2138 mtx_init(&sc->mpr_mtx, "MPR lock", NULL, MTX_DEF); 2139 callout_init_mtx(&sc->periodic, &sc->mpr_mtx, 0); 2140 callout_init_mtx(&sc->device_check_callout, &sc->mpr_mtx, 0); 2141 TAILQ_INIT(&sc->event_list); 2142 timevalclear(&sc->lastfail); 2143 2144 if ((error = mpr_transition_ready(sc)) != 0) { 2145 mpr_dprint(sc, MPR_INIT|MPR_FAULT, 2146 "Failed to transition ready\n"); 2147 return (error); 2148 } 2149 2150 sc->facts = malloc(sizeof(MPI2_IOC_FACTS_REPLY), M_MPR, 2151 M_ZERO|M_NOWAIT); 2152 if (!sc->facts) { 2153 mpr_dprint(sc, MPR_INIT|MPR_FAULT, 2154 "Cannot allocate memory, exit\n"); 2155 return (ENOMEM); 2156 } 2157 2158 /* 2159 * Get IOC Facts and allocate all structures based on this information. 2160 * A Diag Reset will also call mpr_iocfacts_allocate and re-read the IOC 2161 * Facts. If relevant values have changed in IOC Facts, this function 2162 * will free all of the memory based on IOC Facts and reallocate that 2163 * memory. If this fails, any allocated memory should already be freed. 2164 */ 2165 if ((error = mpr_iocfacts_allocate(sc, TRUE)) != 0) { 2166 mpr_dprint(sc, MPR_INIT|MPR_FAULT, "IOC Facts allocation " 2167 "failed with error %d\n", error); 2168 return (error); 2169 } 2170 2171 /* Start the periodic watchdog check on the IOC Doorbell */ 2172 mpr_periodic(sc); 2173 2174 /* 2175 * The portenable will kick off discovery events that will drive the 2176 * rest of the initialization process. The CAM/SAS module will 2177 * hold up the boot sequence until discovery is complete. 2178 */ 2179 sc->mpr_ich.ich_func = mpr_startup; 2180 sc->mpr_ich.ich_arg = sc; 2181 if (config_intrhook_establish(&sc->mpr_ich) != 0) { 2182 mpr_dprint(sc, MPR_INIT|MPR_ERROR, 2183 "Cannot establish MPR config hook\n"); 2184 error = EINVAL; 2185 } 2186 2187 /* 2188 * Allow IR to shutdown gracefully when shutdown occurs. 2189 */ 2190 sc->shutdown_eh = EVENTHANDLER_REGISTER(shutdown_final, 2191 mprsas_ir_shutdown, sc, SHUTDOWN_PRI_DEFAULT); 2192 2193 if (sc->shutdown_eh == NULL) 2194 mpr_dprint(sc, MPR_INIT|MPR_ERROR, 2195 "shutdown event registration failed\n"); 2196 2197 mpr_setup_sysctl(sc); 2198 2199 sc->mpr_flags |= MPR_FLAGS_ATTACH_DONE; 2200 mpr_dprint(sc, MPR_INIT, "%s exit error= %d\n", __func__, error); 2201 2202 return (error); 2203 } 2204 2205 /* Run through any late-start handlers. */ 2206 static void 2207 mpr_startup(void *arg) 2208 { 2209 struct mpr_softc *sc; 2210 2211 sc = (struct mpr_softc *)arg; 2212 mpr_dprint(sc, MPR_INIT, "%s entered\n", __func__); 2213 2214 mpr_lock(sc); 2215 mpr_unmask_intr(sc); 2216 2217 /* initialize device mapping tables */ 2218 mpr_base_static_config_pages(sc); 2219 mpr_mapping_initialize(sc); 2220 mprsas_startup(sc); 2221 mpr_unlock(sc); 2222 2223 mpr_dprint(sc, MPR_INIT, "disestablish config intrhook\n"); 2224 config_intrhook_disestablish(&sc->mpr_ich); 2225 sc->mpr_ich.ich_arg = NULL; 2226 2227 mpr_dprint(sc, MPR_INIT, "%s exit\n", __func__); 2228 } 2229 2230 /* Periodic watchdog. Is called with the driver lock already held. */ 2231 static void 2232 mpr_periodic(void *arg) 2233 { 2234 struct mpr_softc *sc; 2235 uint32_t db; 2236 2237 sc = (struct mpr_softc *)arg; 2238 if (sc->mpr_flags & MPR_FLAGS_SHUTDOWN) 2239 return; 2240 2241 db = mpr_regread(sc, MPI2_DOORBELL_OFFSET); 2242 if ((db & MPI2_IOC_STATE_MASK) == MPI2_IOC_STATE_FAULT) { 2243 if ((db & MPI2_DOORBELL_FAULT_CODE_MASK) == 2244 IFAULT_IOP_OVER_TEMP_THRESHOLD_EXCEEDED) { 2245 panic("TEMPERATURE FAULT: STOPPING."); 2246 } 2247 mpr_dprint(sc, MPR_FAULT, "IOC Fault 0x%08x, Resetting\n", db); 2248 mpr_reinit(sc); 2249 } 2250 2251 callout_reset_sbt(&sc->periodic, MPR_PERIODIC_DELAY * SBT_1S, 0, 2252 mpr_periodic, sc, C_PREL(1)); 2253 } 2254 2255 static void 2256 mpr_log_evt_handler(struct mpr_softc *sc, uintptr_t data, 2257 MPI2_EVENT_NOTIFICATION_REPLY *event) 2258 { 2259 MPI2_EVENT_DATA_LOG_ENTRY_ADDED *entry; 2260 2261 MPR_DPRINT_EVENT(sc, generic, event); 2262 2263 switch (event->Event) { 2264 case MPI2_EVENT_LOG_DATA: 2265 mpr_dprint(sc, MPR_EVENT, "MPI2_EVENT_LOG_DATA:\n"); 2266 if (sc->mpr_debug & MPR_EVENT) 2267 hexdump(event->EventData, event->EventDataLength, NULL, 2268 0); 2269 break; 2270 case MPI2_EVENT_LOG_ENTRY_ADDED: 2271 entry = (MPI2_EVENT_DATA_LOG_ENTRY_ADDED *)event->EventData; 2272 mpr_dprint(sc, MPR_EVENT, "MPI2_EVENT_LOG_ENTRY_ADDED event " 2273 "0x%x Sequence %d:\n", entry->LogEntryQualifier, 2274 entry->LogSequence); 2275 break; 2276 default: 2277 break; 2278 } 2279 return; 2280 } 2281 2282 static int 2283 mpr_attach_log(struct mpr_softc *sc) 2284 { 2285 uint8_t events[16]; 2286 2287 bzero(events, 16); 2288 setbit(events, MPI2_EVENT_LOG_DATA); 2289 setbit(events, MPI2_EVENT_LOG_ENTRY_ADDED); 2290 2291 mpr_register_events(sc, events, mpr_log_evt_handler, NULL, 2292 &sc->mpr_log_eh); 2293 2294 return (0); 2295 } 2296 2297 static int 2298 mpr_detach_log(struct mpr_softc *sc) 2299 { 2300 2301 if (sc->mpr_log_eh != NULL) 2302 mpr_deregister_events(sc, sc->mpr_log_eh); 2303 return (0); 2304 } 2305 2306 /* 2307 * Free all of the driver resources and detach submodules. Should be called 2308 * without the lock held. 2309 */ 2310 int 2311 mpr_free(struct mpr_softc *sc) 2312 { 2313 int error; 2314 2315 mpr_dprint(sc, MPR_INIT, "%s entered\n", __func__); 2316 /* Turn off the watchdog */ 2317 mpr_lock(sc); 2318 sc->mpr_flags |= MPR_FLAGS_SHUTDOWN; 2319 mpr_unlock(sc); 2320 /* Lock must not be held for this */ 2321 callout_drain(&sc->periodic); 2322 callout_drain(&sc->device_check_callout); 2323 2324 if (((error = mpr_detach_log(sc)) != 0) || 2325 ((error = mpr_detach_sas(sc)) != 0)) { 2326 mpr_dprint(sc, MPR_INIT|MPR_FAULT, "failed to detach " 2327 "subsystems, error= %d, exit\n", error); 2328 return (error); 2329 } 2330 2331 mpr_detach_user(sc); 2332 2333 /* Put the IOC back in the READY state. */ 2334 mpr_lock(sc); 2335 if ((error = mpr_transition_ready(sc)) != 0) { 2336 mpr_unlock(sc); 2337 return (error); 2338 } 2339 mpr_unlock(sc); 2340 2341 if (sc->facts != NULL) 2342 free(sc->facts, M_MPR); 2343 2344 /* 2345 * Free all buffers that are based on IOC Facts. A Diag Reset may need 2346 * to free these buffers too. 2347 */ 2348 mpr_iocfacts_free(sc); 2349 2350 if (sc->sysctl_tree != NULL) 2351 sysctl_ctx_free(&sc->sysctl_ctx); 2352 2353 /* Deregister the shutdown function */ 2354 if (sc->shutdown_eh != NULL) 2355 EVENTHANDLER_DEREGISTER(shutdown_final, sc->shutdown_eh); 2356 2357 mtx_destroy(&sc->mpr_mtx); 2358 mpr_dprint(sc, MPR_INIT, "%s exit\n", __func__); 2359 2360 return (0); 2361 } 2362 2363 static __inline void 2364 mpr_complete_command(struct mpr_softc *sc, struct mpr_command *cm) 2365 { 2366 MPR_FUNCTRACE(sc); 2367 2368 if (cm == NULL) { 2369 mpr_dprint(sc, MPR_ERROR, "Completing NULL command\n"); 2370 return; 2371 } 2372 2373 KASSERT(cm->cm_state == MPR_CM_STATE_INQUEUE, 2374 ("command not inqueue, state = %u\n", cm->cm_state)); 2375 cm->cm_state = MPR_CM_STATE_BUSY; 2376 if (cm->cm_flags & MPR_CM_FLAGS_POLLED) 2377 cm->cm_flags |= MPR_CM_FLAGS_COMPLETE; 2378 2379 if (cm->cm_complete != NULL) { 2380 mpr_dprint(sc, MPR_TRACE, 2381 "%s cm %p calling cm_complete %p data %p reply %p\n", 2382 __func__, cm, cm->cm_complete, cm->cm_complete_data, 2383 cm->cm_reply); 2384 cm->cm_complete(sc, cm); 2385 } 2386 2387 if (cm->cm_flags & MPR_CM_FLAGS_WAKEUP) { 2388 mpr_dprint(sc, MPR_TRACE, "waking up %p\n", cm); 2389 wakeup(cm); 2390 } 2391 2392 if (sc->io_cmds_active != 0) { 2393 sc->io_cmds_active--; 2394 } else { 2395 mpr_dprint(sc, MPR_ERROR, "Warning: io_cmds_active is " 2396 "out of sync - resynching to 0\n"); 2397 } 2398 } 2399 2400 static void 2401 mpr_sas_log_info(struct mpr_softc *sc , u32 log_info) 2402 { 2403 union loginfo_type { 2404 u32 loginfo; 2405 struct { 2406 u32 subcode:16; 2407 u32 code:8; 2408 u32 originator:4; 2409 u32 bus_type:4; 2410 } dw; 2411 }; 2412 union loginfo_type sas_loginfo; 2413 char *originator_str = NULL; 2414 2415 sas_loginfo.loginfo = log_info; 2416 if (sas_loginfo.dw.bus_type != 3 /*SAS*/) 2417 return; 2418 2419 /* each nexus loss loginfo */ 2420 if (log_info == 0x31170000) 2421 return; 2422 2423 /* eat the loginfos associated with task aborts */ 2424 if ((log_info == 30050000) || (log_info == 0x31140000) || 2425 (log_info == 0x31130000)) 2426 return; 2427 2428 switch (sas_loginfo.dw.originator) { 2429 case 0: 2430 originator_str = "IOP"; 2431 break; 2432 case 1: 2433 originator_str = "PL"; 2434 break; 2435 case 2: 2436 originator_str = "IR"; 2437 break; 2438 } 2439 2440 mpr_dprint(sc, MPR_LOG, "log_info(0x%08x): originator(%s), " 2441 "code(0x%02x), sub_code(0x%04x)\n", log_info, originator_str, 2442 sas_loginfo.dw.code, sas_loginfo.dw.subcode); 2443 } 2444 2445 static void 2446 mpr_display_reply_info(struct mpr_softc *sc, uint8_t *reply) 2447 { 2448 MPI2DefaultReply_t *mpi_reply; 2449 u16 sc_status; 2450 2451 mpi_reply = (MPI2DefaultReply_t*)reply; 2452 sc_status = le16toh(mpi_reply->IOCStatus); 2453 if (sc_status & MPI2_IOCSTATUS_FLAG_LOG_INFO_AVAILABLE) 2454 mpr_sas_log_info(sc, le32toh(mpi_reply->IOCLogInfo)); 2455 } 2456 2457 void 2458 mpr_intr(void *data) 2459 { 2460 struct mpr_softc *sc; 2461 uint32_t status; 2462 2463 sc = (struct mpr_softc *)data; 2464 mpr_dprint(sc, MPR_TRACE, "%s\n", __func__); 2465 2466 /* 2467 * Check interrupt status register to flush the bus. This is 2468 * needed for both INTx interrupts and driver-driven polling 2469 */ 2470 status = mpr_regread(sc, MPI2_HOST_INTERRUPT_STATUS_OFFSET); 2471 if ((status & MPI2_HIS_REPLY_DESCRIPTOR_INTERRUPT) == 0) 2472 return; 2473 2474 mpr_lock(sc); 2475 mpr_intr_locked(data); 2476 mpr_unlock(sc); 2477 return; 2478 } 2479 2480 /* 2481 * In theory, MSI/MSIX interrupts shouldn't need to read any registers on the 2482 * chip. Hopefully this theory is correct. 2483 */ 2484 void 2485 mpr_intr_msi(void *data) 2486 { 2487 struct mpr_softc *sc; 2488 2489 sc = (struct mpr_softc *)data; 2490 mpr_dprint(sc, MPR_TRACE, "%s\n", __func__); 2491 mpr_lock(sc); 2492 mpr_intr_locked(data); 2493 mpr_unlock(sc); 2494 return; 2495 } 2496 2497 /* 2498 * The locking is overly broad and simplistic, but easy to deal with for now. 2499 */ 2500 void 2501 mpr_intr_locked(void *data) 2502 { 2503 MPI2_REPLY_DESCRIPTORS_UNION *desc; 2504 MPI2_DIAG_RELEASE_REPLY *rel_rep; 2505 mpr_fw_diagnostic_buffer_t *pBuffer; 2506 struct mpr_softc *sc; 2507 uint64_t tdesc; 2508 struct mpr_command *cm = NULL; 2509 uint8_t flags; 2510 u_int pq; 2511 2512 sc = (struct mpr_softc *)data; 2513 2514 pq = sc->replypostindex; 2515 mpr_dprint(sc, MPR_TRACE, 2516 "%s sc %p starting with replypostindex %u\n", 2517 __func__, sc, sc->replypostindex); 2518 2519 for ( ;; ) { 2520 cm = NULL; 2521 desc = &sc->post_queue[sc->replypostindex]; 2522 2523 /* 2524 * Copy and clear out the descriptor so that any reentry will 2525 * immediately know that this descriptor has already been 2526 * looked at. There is unfortunate casting magic because the 2527 * MPI API doesn't have a cardinal 64bit type. 2528 */ 2529 tdesc = 0xffffffffffffffff; 2530 tdesc = atomic_swap_64((uint64_t *)desc, tdesc); 2531 desc = (MPI2_REPLY_DESCRIPTORS_UNION *)&tdesc; 2532 2533 flags = desc->Default.ReplyFlags & 2534 MPI2_RPY_DESCRIPT_FLAGS_TYPE_MASK; 2535 if ((flags == MPI2_RPY_DESCRIPT_FLAGS_UNUSED) || 2536 (le32toh(desc->Words.High) == 0xffffffff)) 2537 break; 2538 2539 /* increment the replypostindex now, so that event handlers 2540 * and cm completion handlers which decide to do a diag 2541 * reset can zero it without it getting incremented again 2542 * afterwards, and we break out of this loop on the next 2543 * iteration since the reply post queue has been cleared to 2544 * 0xFF and all descriptors look unused (which they are). 2545 */ 2546 if (++sc->replypostindex >= sc->pqdepth) 2547 sc->replypostindex = 0; 2548 2549 switch (flags) { 2550 case MPI2_RPY_DESCRIPT_FLAGS_SCSI_IO_SUCCESS: 2551 case MPI25_RPY_DESCRIPT_FLAGS_FAST_PATH_SCSI_IO_SUCCESS: 2552 case MPI26_RPY_DESCRIPT_FLAGS_PCIE_ENCAPSULATED_SUCCESS: 2553 cm = &sc->commands[le16toh(desc->SCSIIOSuccess.SMID)]; 2554 cm->cm_reply = NULL; 2555 break; 2556 case MPI2_RPY_DESCRIPT_FLAGS_ADDRESS_REPLY: 2557 { 2558 uint32_t baddr; 2559 uint8_t *reply; 2560 2561 /* 2562 * Re-compose the reply address from the address 2563 * sent back from the chip. The ReplyFrameAddress 2564 * is the lower 32 bits of the physical address of 2565 * particular reply frame. Convert that address to 2566 * host format, and then use that to provide the 2567 * offset against the virtual address base 2568 * (sc->reply_frames). 2569 */ 2570 baddr = le32toh(desc->AddressReply.ReplyFrameAddress); 2571 reply = sc->reply_frames + 2572 (baddr - ((uint32_t)sc->reply_busaddr)); 2573 /* 2574 * Make sure the reply we got back is in a valid 2575 * range. If not, go ahead and panic here, since 2576 * we'll probably panic as soon as we deference the 2577 * reply pointer anyway. 2578 */ 2579 if ((reply < sc->reply_frames) 2580 || (reply > (sc->reply_frames + 2581 (sc->fqdepth * sc->replyframesz)))) { 2582 printf("%s: WARNING: reply %p out of range!\n", 2583 __func__, reply); 2584 printf("%s: reply_frames %p, fqdepth %d, " 2585 "frame size %d\n", __func__, 2586 sc->reply_frames, sc->fqdepth, 2587 sc->replyframesz); 2588 printf("%s: baddr %#x,\n", __func__, baddr); 2589 /* LSI-TODO. See Linux Code for Graceful exit */ 2590 panic("Reply address out of range"); 2591 } 2592 if (le16toh(desc->AddressReply.SMID) == 0) { 2593 if (((MPI2_DEFAULT_REPLY *)reply)->Function == 2594 MPI2_FUNCTION_DIAG_BUFFER_POST) { 2595 /* 2596 * If SMID is 0 for Diag Buffer Post, 2597 * this implies that the reply is due to 2598 * a release function with a status that 2599 * the buffer has been released. Set 2600 * the buffer flags accordingly. 2601 */ 2602 rel_rep = 2603 (MPI2_DIAG_RELEASE_REPLY *)reply; 2604 if ((le16toh(rel_rep->IOCStatus) & 2605 MPI2_IOCSTATUS_MASK) == 2606 MPI2_IOCSTATUS_DIAGNOSTIC_RELEASED) 2607 { 2608 pBuffer = 2609 &sc->fw_diag_buffer_list[ 2610 rel_rep->BufferType]; 2611 pBuffer->valid_data = TRUE; 2612 pBuffer->owned_by_firmware = 2613 FALSE; 2614 pBuffer->immediate = FALSE; 2615 } 2616 } else 2617 mpr_dispatch_event(sc, baddr, 2618 (MPI2_EVENT_NOTIFICATION_REPLY *) 2619 reply); 2620 } else { 2621 cm = &sc->commands[ 2622 le16toh(desc->AddressReply.SMID)]; 2623 if (cm->cm_state == MPR_CM_STATE_INQUEUE) { 2624 cm->cm_reply = reply; 2625 cm->cm_reply_data = 2626 le32toh(desc->AddressReply. 2627 ReplyFrameAddress); 2628 } else { 2629 mpr_dprint(sc, MPR_RECOVERY, 2630 "Bad state for ADDRESS_REPLY status," 2631 " ignoring state %d cm %p\n", 2632 cm->cm_state, cm); 2633 } 2634 } 2635 break; 2636 } 2637 case MPI2_RPY_DESCRIPT_FLAGS_TARGETASSIST_SUCCESS: 2638 case MPI2_RPY_DESCRIPT_FLAGS_TARGET_COMMAND_BUFFER: 2639 case MPI2_RPY_DESCRIPT_FLAGS_RAID_ACCELERATOR_SUCCESS: 2640 default: 2641 /* Unhandled */ 2642 mpr_dprint(sc, MPR_ERROR, "Unhandled reply 0x%x\n", 2643 desc->Default.ReplyFlags); 2644 cm = NULL; 2645 break; 2646 } 2647 2648 if (cm != NULL) { 2649 // Print Error reply frame 2650 if (cm->cm_reply) 2651 mpr_display_reply_info(sc,cm->cm_reply); 2652 mpr_complete_command(sc, cm); 2653 } 2654 } 2655 2656 if (pq != sc->replypostindex) { 2657 mpr_dprint(sc, MPR_TRACE, "%s sc %p writing postindex %d\n", 2658 __func__, sc, sc->replypostindex); 2659 mpr_regwrite(sc, MPI2_REPLY_POST_HOST_INDEX_OFFSET, 2660 sc->replypostindex); 2661 } 2662 2663 return; 2664 } 2665 2666 static void 2667 mpr_dispatch_event(struct mpr_softc *sc, uintptr_t data, 2668 MPI2_EVENT_NOTIFICATION_REPLY *reply) 2669 { 2670 struct mpr_event_handle *eh; 2671 int event, handled = 0; 2672 2673 event = le16toh(reply->Event); 2674 TAILQ_FOREACH(eh, &sc->event_list, eh_list) { 2675 if (isset(eh->mask, event)) { 2676 eh->callback(sc, data, reply); 2677 handled++; 2678 } 2679 } 2680 2681 if (handled == 0) 2682 mpr_dprint(sc, MPR_EVENT, "Unhandled event 0x%x\n", 2683 le16toh(event)); 2684 2685 /* 2686 * This is the only place that the event/reply should be freed. 2687 * Anything wanting to hold onto the event data should have 2688 * already copied it into their own storage. 2689 */ 2690 mpr_free_reply(sc, data); 2691 } 2692 2693 static void 2694 mpr_reregister_events_complete(struct mpr_softc *sc, struct mpr_command *cm) 2695 { 2696 mpr_dprint(sc, MPR_TRACE, "%s\n", __func__); 2697 2698 if (cm->cm_reply) 2699 MPR_DPRINT_EVENT(sc, generic, 2700 (MPI2_EVENT_NOTIFICATION_REPLY *)cm->cm_reply); 2701 2702 mpr_free_command(sc, cm); 2703 2704 /* next, send a port enable */ 2705 mprsas_startup(sc); 2706 } 2707 2708 /* 2709 * For both register_events and update_events, the caller supplies a bitmap 2710 * of events that it _wants_. These functions then turn that into a bitmask 2711 * suitable for the controller. 2712 */ 2713 int 2714 mpr_register_events(struct mpr_softc *sc, uint8_t *mask, 2715 mpr_evt_callback_t *cb, void *data, struct mpr_event_handle **handle) 2716 { 2717 struct mpr_event_handle *eh; 2718 int error = 0; 2719 2720 eh = malloc(sizeof(struct mpr_event_handle), M_MPR, M_WAITOK|M_ZERO); 2721 eh->callback = cb; 2722 eh->data = data; 2723 TAILQ_INSERT_TAIL(&sc->event_list, eh, eh_list); 2724 if (mask != NULL) 2725 error = mpr_update_events(sc, eh, mask); 2726 *handle = eh; 2727 2728 return (error); 2729 } 2730 2731 int 2732 mpr_update_events(struct mpr_softc *sc, struct mpr_event_handle *handle, 2733 uint8_t *mask) 2734 { 2735 MPI2_EVENT_NOTIFICATION_REQUEST *evtreq; 2736 MPI2_EVENT_NOTIFICATION_REPLY *reply = NULL; 2737 struct mpr_command *cm = NULL; 2738 struct mpr_event_handle *eh; 2739 int error, i; 2740 2741 mpr_dprint(sc, MPR_TRACE, "%s\n", __func__); 2742 2743 if ((mask != NULL) && (handle != NULL)) 2744 bcopy(mask, &handle->mask[0], 16); 2745 memset(sc->event_mask, 0xff, 16); 2746 2747 TAILQ_FOREACH(eh, &sc->event_list, eh_list) { 2748 for (i = 0; i < 16; i++) 2749 sc->event_mask[i] &= ~eh->mask[i]; 2750 } 2751 2752 if ((cm = mpr_alloc_command(sc)) == NULL) 2753 return (EBUSY); 2754 evtreq = (MPI2_EVENT_NOTIFICATION_REQUEST *)cm->cm_req; 2755 evtreq->Function = MPI2_FUNCTION_EVENT_NOTIFICATION; 2756 evtreq->MsgFlags = 0; 2757 evtreq->SASBroadcastPrimitiveMasks = 0; 2758 #ifdef MPR_DEBUG_ALL_EVENTS 2759 { 2760 u_char fullmask[sizeof(evtreq->EventMasks)]; 2761 memset(fullmask, 0x00, sizeof(fullmask)); 2762 bcopy(fullmask, (uint8_t *)&evtreq->EventMasks, sizeof(fullmask)); 2763 } 2764 #else 2765 bcopy(sc->event_mask, (uint8_t *)&evtreq->EventMasks, sizeof(sc->event_mask)); 2766 for (i = 0; i < MPI2_EVENT_NOTIFY_EVENTMASK_WORDS; i++) 2767 evtreq->EventMasks[i] = htole32(evtreq->EventMasks[i]); 2768 #endif 2769 cm->cm_desc.Default.RequestFlags = MPI2_REQ_DESCRIPT_FLAGS_DEFAULT_TYPE; 2770 cm->cm_data = NULL; 2771 2772 error = mpr_request_polled(sc, &cm); 2773 if (cm != NULL) 2774 reply = (MPI2_EVENT_NOTIFICATION_REPLY *)cm->cm_reply; 2775 if ((reply == NULL) || 2776 (reply->IOCStatus & MPI2_IOCSTATUS_MASK) != MPI2_IOCSTATUS_SUCCESS) 2777 error = ENXIO; 2778 2779 if (reply) 2780 MPR_DPRINT_EVENT(sc, generic, reply); 2781 2782 mpr_dprint(sc, MPR_TRACE, "%s finished error %d\n", __func__, error); 2783 2784 if (cm != NULL) 2785 mpr_free_command(sc, cm); 2786 return (error); 2787 } 2788 2789 static int 2790 mpr_reregister_events(struct mpr_softc *sc) 2791 { 2792 MPI2_EVENT_NOTIFICATION_REQUEST *evtreq; 2793 struct mpr_command *cm; 2794 struct mpr_event_handle *eh; 2795 int error, i; 2796 2797 mpr_dprint(sc, MPR_TRACE, "%s\n", __func__); 2798 2799 /* first, reregister events */ 2800 2801 memset(sc->event_mask, 0xff, 16); 2802 2803 TAILQ_FOREACH(eh, &sc->event_list, eh_list) { 2804 for (i = 0; i < 16; i++) 2805 sc->event_mask[i] &= ~eh->mask[i]; 2806 } 2807 2808 if ((cm = mpr_alloc_command(sc)) == NULL) 2809 return (EBUSY); 2810 evtreq = (MPI2_EVENT_NOTIFICATION_REQUEST *)cm->cm_req; 2811 evtreq->Function = MPI2_FUNCTION_EVENT_NOTIFICATION; 2812 evtreq->MsgFlags = 0; 2813 evtreq->SASBroadcastPrimitiveMasks = 0; 2814 #ifdef MPR_DEBUG_ALL_EVENTS 2815 { 2816 u_char fullmask[sizeof(evtreq->EventMasks)]; 2817 memset(fullmask, 0x00, sizeof(fullmask)); 2818 bcopy(fullmask, (uint8_t *)&evtreq->EventMasks, sizeof(fullmask)); 2819 } 2820 #else 2821 bcopy(sc->event_mask, (uint8_t *)&evtreq->EventMasks, sizeof(sc->event_mask)); 2822 for (i = 0; i < MPI2_EVENT_NOTIFY_EVENTMASK_WORDS; i++) 2823 evtreq->EventMasks[i] = htole32(evtreq->EventMasks[i]); 2824 #endif 2825 cm->cm_desc.Default.RequestFlags = MPI2_REQ_DESCRIPT_FLAGS_DEFAULT_TYPE; 2826 cm->cm_data = NULL; 2827 cm->cm_complete = mpr_reregister_events_complete; 2828 2829 error = mpr_map_command(sc, cm); 2830 2831 mpr_dprint(sc, MPR_TRACE, "%s finished with error %d\n", __func__, 2832 error); 2833 return (error); 2834 } 2835 2836 int 2837 mpr_deregister_events(struct mpr_softc *sc, struct mpr_event_handle *handle) 2838 { 2839 2840 TAILQ_REMOVE(&sc->event_list, handle, eh_list); 2841 free(handle, M_MPR); 2842 return (mpr_update_events(sc, NULL, NULL)); 2843 } 2844 2845 /** 2846 * mpr_build_nvme_prp - This function is called for NVMe end devices to build a 2847 * native SGL (NVMe PRP). The native SGL is built starting in the first PRP entry 2848 * of the NVMe message (PRP1). If the data buffer is small enough to be described 2849 * entirely using PRP1, then PRP2 is not used. If needed, PRP2 is used to 2850 * describe a larger data buffer. If the data buffer is too large to describe 2851 * using the two PRP entriess inside the NVMe message, then PRP1 describes the 2852 * first data memory segment, and PRP2 contains a pointer to a PRP list located 2853 * elsewhere in memory to describe the remaining data memory segments. The PRP 2854 * list will be contiguous. 2855 2856 * The native SGL for NVMe devices is a Physical Region Page (PRP). A PRP 2857 * consists of a list of PRP entries to describe a number of noncontigous 2858 * physical memory segments as a single memory buffer, just as a SGL does. Note 2859 * however, that this function is only used by the IOCTL call, so the memory 2860 * given will be guaranteed to be contiguous. There is no need to translate 2861 * non-contiguous SGL into a PRP in this case. All PRPs will describe contiguous 2862 * space that is one page size each. 2863 * 2864 * Each NVMe message contains two PRP entries. The first (PRP1) either contains 2865 * a PRP list pointer or a PRP element, depending upon the command. PRP2 contains 2866 * the second PRP element if the memory being described fits within 2 PRP 2867 * entries, or a PRP list pointer if the PRP spans more than two entries. 2868 * 2869 * A PRP list pointer contains the address of a PRP list, structured as a linear 2870 * array of PRP entries. Each PRP entry in this list describes a segment of 2871 * physical memory. 2872 * 2873 * Each 64-bit PRP entry comprises an address and an offset field. The address 2874 * always points to the beginning of a PAGE_SIZE physical memory page, and the 2875 * offset describes where within that page the memory segment begins. Only the 2876 * first element in a PRP list may contain a non-zero offest, implying that all 2877 * memory segments following the first begin at the start of a PAGE_SIZE page. 2878 * 2879 * Each PRP element normally describes a chunck of PAGE_SIZE physical memory, 2880 * with exceptions for the first and last elements in the list. If the memory 2881 * being described by the list begins at a non-zero offset within the first page, 2882 * then the first PRP element will contain a non-zero offset indicating where the 2883 * region begins within the page. The last memory segment may end before the end 2884 * of the PAGE_SIZE segment, depending upon the overall size of the memory being 2885 * described by the PRP list. 2886 * 2887 * Since PRP entries lack any indication of size, the overall data buffer length 2888 * is used to determine where the end of the data memory buffer is located, and 2889 * how many PRP entries are required to describe it. 2890 * 2891 * Returns nothing. 2892 */ 2893 void 2894 mpr_build_nvme_prp(struct mpr_softc *sc, struct mpr_command *cm, 2895 Mpi26NVMeEncapsulatedRequest_t *nvme_encap_request, void *data, 2896 uint32_t data_in_sz, uint32_t data_out_sz) 2897 { 2898 int prp_size = PRP_ENTRY_SIZE; 2899 uint64_t *prp_entry, *prp1_entry, *prp2_entry; 2900 uint64_t *prp_entry_phys, *prp_page, *prp_page_phys; 2901 uint32_t offset, entry_len, page_mask_result, page_mask; 2902 bus_addr_t paddr; 2903 size_t length; 2904 struct mpr_prp_page *prp_page_info = NULL; 2905 2906 /* 2907 * Not all commands require a data transfer. If no data, just return 2908 * without constructing any PRP. 2909 */ 2910 if (!data_in_sz && !data_out_sz) 2911 return; 2912 2913 /* 2914 * Set pointers to PRP1 and PRP2, which are in the NVMe command. PRP1 is 2915 * located at a 24 byte offset from the start of the NVMe command. Then 2916 * set the current PRP entry pointer to PRP1. 2917 */ 2918 prp1_entry = (uint64_t *)(nvme_encap_request->NVMe_Command + 2919 NVME_CMD_PRP1_OFFSET); 2920 prp2_entry = (uint64_t *)(nvme_encap_request->NVMe_Command + 2921 NVME_CMD_PRP2_OFFSET); 2922 prp_entry = prp1_entry; 2923 2924 /* 2925 * For the PRP entries, use the specially allocated buffer of 2926 * contiguous memory. PRP Page allocation failures should not happen 2927 * because there should be enough PRP page buffers to account for the 2928 * possible NVMe QDepth. 2929 */ 2930 prp_page_info = mpr_alloc_prp_page(sc); 2931 KASSERT(prp_page_info != NULL, ("%s: There are no PRP Pages left to be " 2932 "used for building a native NVMe SGL.\n", __func__)); 2933 prp_page = (uint64_t *)prp_page_info->prp_page; 2934 prp_page_phys = (uint64_t *)(uintptr_t)prp_page_info->prp_page_busaddr; 2935 2936 /* 2937 * Insert the allocated PRP page into the command's PRP page list. This 2938 * will be freed when the command is freed. 2939 */ 2940 TAILQ_INSERT_TAIL(&cm->cm_prp_page_list, prp_page_info, prp_page_link); 2941 2942 /* 2943 * Check if we are within 1 entry of a page boundary we don't want our 2944 * first entry to be a PRP List entry. 2945 */ 2946 page_mask = PAGE_SIZE - 1; 2947 page_mask_result = (uintptr_t)((uint8_t *)prp_page + prp_size) & 2948 page_mask; 2949 if (!page_mask_result) 2950 { 2951 /* Bump up to next page boundary. */ 2952 prp_page = (uint64_t *)((uint8_t *)prp_page + prp_size); 2953 prp_page_phys = (uint64_t *)((uint8_t *)prp_page_phys + 2954 prp_size); 2955 } 2956 2957 /* 2958 * Set PRP physical pointer, which initially points to the current PRP 2959 * DMA memory page. 2960 */ 2961 prp_entry_phys = prp_page_phys; 2962 2963 /* Get physical address and length of the data buffer. */ 2964 paddr = (bus_addr_t)(uintptr_t)data; 2965 if (data_in_sz) 2966 length = data_in_sz; 2967 else 2968 length = data_out_sz; 2969 2970 /* Loop while the length is not zero. */ 2971 while (length) 2972 { 2973 /* 2974 * Check if we need to put a list pointer here if we are at page 2975 * boundary - prp_size (8 bytes). 2976 */ 2977 page_mask_result = (uintptr_t)((uint8_t *)prp_entry_phys + 2978 prp_size) & page_mask; 2979 if (!page_mask_result) 2980 { 2981 /* 2982 * This is the last entry in a PRP List, so we need to 2983 * put a PRP list pointer here. What this does is: 2984 * - bump the current memory pointer to the next 2985 * address, which will be the next full page. 2986 * - set the PRP Entry to point to that page. This is 2987 * now the PRP List pointer. 2988 * - bump the PRP Entry pointer the start of the next 2989 * page. Since all of this PRP memory is contiguous, 2990 * no need to get a new page - it's just the next 2991 * address. 2992 */ 2993 prp_entry_phys++; 2994 *prp_entry = 2995 htole64((uint64_t)(uintptr_t)prp_entry_phys); 2996 prp_entry++; 2997 } 2998 2999 /* Need to handle if entry will be part of a page. */ 3000 offset = (uint32_t)paddr & page_mask; 3001 entry_len = PAGE_SIZE - offset; 3002 3003 if (prp_entry == prp1_entry) 3004 { 3005 /* 3006 * Must fill in the first PRP pointer (PRP1) before 3007 * moving on. 3008 */ 3009 *prp1_entry = htole64((uint64_t)paddr); 3010 3011 /* 3012 * Now point to the second PRP entry within the 3013 * command (PRP2). 3014 */ 3015 prp_entry = prp2_entry; 3016 } 3017 else if (prp_entry == prp2_entry) 3018 { 3019 /* 3020 * Should the PRP2 entry be a PRP List pointer or just a 3021 * regular PRP pointer? If there is more than one more 3022 * page of data, must use a PRP List pointer. 3023 */ 3024 if (length > PAGE_SIZE) 3025 { 3026 /* 3027 * PRP2 will contain a PRP List pointer because 3028 * more PRP's are needed with this command. The 3029 * list will start at the beginning of the 3030 * contiguous buffer. 3031 */ 3032 *prp2_entry = 3033 htole64( 3034 (uint64_t)(uintptr_t)prp_entry_phys); 3035 3036 /* 3037 * The next PRP Entry will be the start of the 3038 * first PRP List. 3039 */ 3040 prp_entry = prp_page; 3041 } 3042 else 3043 { 3044 /* 3045 * After this, the PRP Entries are complete. 3046 * This command uses 2 PRP's and no PRP list. 3047 */ 3048 *prp2_entry = htole64((uint64_t)paddr); 3049 } 3050 } 3051 else 3052 { 3053 /* 3054 * Put entry in list and bump the addresses. 3055 * 3056 * After PRP1 and PRP2 are filled in, this will fill in 3057 * all remaining PRP entries in a PRP List, one per each 3058 * time through the loop. 3059 */ 3060 *prp_entry = htole64((uint64_t)paddr); 3061 prp_entry++; 3062 prp_entry_phys++; 3063 } 3064 3065 /* 3066 * Bump the phys address of the command's data buffer by the 3067 * entry_len. 3068 */ 3069 paddr += entry_len; 3070 3071 /* Decrement length accounting for last partial page. */ 3072 if (entry_len > length) 3073 length = 0; 3074 else 3075 length -= entry_len; 3076 } 3077 } 3078 3079 /* 3080 * mpr_check_pcie_native_sgl - This function is called for PCIe end devices to 3081 * determine if the driver needs to build a native SGL. If so, that native SGL 3082 * is built in the contiguous buffers allocated especially for PCIe SGL 3083 * creation. If the driver will not build a native SGL, return TRUE and a 3084 * normal IEEE SGL will be built. Currently this routine supports NVMe devices 3085 * only. 3086 * 3087 * Returns FALSE (0) if native SGL was built, TRUE (1) if no SGL was built. 3088 */ 3089 static int 3090 mpr_check_pcie_native_sgl(struct mpr_softc *sc, struct mpr_command *cm, 3091 bus_dma_segment_t *segs, int segs_left) 3092 { 3093 uint32_t i, sge_dwords, length, offset, entry_len; 3094 uint32_t num_entries, buff_len = 0, sges_in_segment; 3095 uint32_t page_mask, page_mask_result, *curr_buff; 3096 uint32_t *ptr_sgl, *ptr_first_sgl, first_page_offset; 3097 uint32_t first_page_data_size, end_residual; 3098 uint64_t *msg_phys; 3099 bus_addr_t paddr; 3100 int build_native_sgl = 0, first_prp_entry; 3101 int prp_size = PRP_ENTRY_SIZE; 3102 Mpi25IeeeSgeChain64_t *main_chain_element = NULL; 3103 struct mpr_prp_page *prp_page_info = NULL; 3104 3105 mpr_dprint(sc, MPR_TRACE, "%s\n", __func__); 3106 3107 /* 3108 * Add up the sizes of each segment length to get the total transfer 3109 * size, which will be checked against the Maximum Data Transfer Size. 3110 * If the data transfer length exceeds the MDTS for this device, just 3111 * return 1 so a normal IEEE SGL will be built. F/W will break the I/O 3112 * up into multiple I/O's. [nvme_mdts = 0 means unlimited] 3113 */ 3114 for (i = 0; i < segs_left; i++) 3115 buff_len += htole32(segs[i].ds_len); 3116 if ((cm->cm_targ->MDTS > 0) && (buff_len > cm->cm_targ->MDTS)) 3117 return 1; 3118 3119 /* Create page_mask (to get offset within page) */ 3120 page_mask = PAGE_SIZE - 1; 3121 3122 /* 3123 * Check if the number of elements exceeds the max number that can be 3124 * put in the main message frame (H/W can only translate an SGL that 3125 * is contained entirely in the main message frame). 3126 */ 3127 sges_in_segment = (sc->reqframesz - 3128 offsetof(Mpi25SCSIIORequest_t, SGL)) / sizeof(MPI25_SGE_IO_UNION); 3129 if (segs_left > sges_in_segment) 3130 build_native_sgl = 1; 3131 else 3132 { 3133 /* 3134 * NVMe uses one PRP for each physical page (or part of physical 3135 * page). 3136 * if 4 pages or less then IEEE is OK 3137 * if > 5 pages then we need to build a native SGL 3138 * if > 4 and <= 5 pages, then check the physical address of 3139 * the first SG entry, then if this first size in the page 3140 * is >= the residual beyond 4 pages then use IEEE, 3141 * otherwise use native SGL 3142 */ 3143 if (buff_len > (PAGE_SIZE * 5)) 3144 build_native_sgl = 1; 3145 else if ((buff_len > (PAGE_SIZE * 4)) && 3146 (buff_len <= (PAGE_SIZE * 5)) ) 3147 { 3148 msg_phys = (uint64_t *)(uintptr_t)segs[0].ds_addr; 3149 first_page_offset = 3150 ((uint32_t)(uint64_t)(uintptr_t)msg_phys & 3151 page_mask); 3152 first_page_data_size = PAGE_SIZE - first_page_offset; 3153 end_residual = buff_len % PAGE_SIZE; 3154 3155 /* 3156 * If offset into first page pushes the end of the data 3157 * beyond end of the 5th page, we need the extra PRP 3158 * list. 3159 */ 3160 if (first_page_data_size < end_residual) 3161 build_native_sgl = 1; 3162 3163 /* 3164 * Check if first SG entry size is < residual beyond 4 3165 * pages. 3166 */ 3167 if (htole32(segs[0].ds_len) < 3168 (buff_len - (PAGE_SIZE * 4))) 3169 build_native_sgl = 1; 3170 } 3171 } 3172 3173 /* check if native SGL is needed */ 3174 if (!build_native_sgl) 3175 return 1; 3176 3177 /* 3178 * Native SGL is needed. 3179 * Put a chain element in main message frame that points to the first 3180 * chain buffer. 3181 * 3182 * NOTE: The ChainOffset field must be 0 when using a chain pointer to 3183 * a native SGL. 3184 */ 3185 3186 /* Set main message chain element pointer */ 3187 main_chain_element = (pMpi25IeeeSgeChain64_t)cm->cm_sge; 3188 3189 /* 3190 * For NVMe the chain element needs to be the 2nd SGL entry in the main 3191 * message. 3192 */ 3193 main_chain_element = (Mpi25IeeeSgeChain64_t *) 3194 ((uint8_t *)main_chain_element + sizeof(MPI25_IEEE_SGE_CHAIN64)); 3195 3196 /* 3197 * For the PRP entries, use the specially allocated buffer of 3198 * contiguous memory. PRP Page allocation failures should not happen 3199 * because there should be enough PRP page buffers to account for the 3200 * possible NVMe QDepth. 3201 */ 3202 prp_page_info = mpr_alloc_prp_page(sc); 3203 KASSERT(prp_page_info != NULL, ("%s: There are no PRP Pages left to be " 3204 "used for building a native NVMe SGL.\n", __func__)); 3205 curr_buff = (uint32_t *)prp_page_info->prp_page; 3206 msg_phys = (uint64_t *)(uintptr_t)prp_page_info->prp_page_busaddr; 3207 3208 /* 3209 * Insert the allocated PRP page into the command's PRP page list. This 3210 * will be freed when the command is freed. 3211 */ 3212 TAILQ_INSERT_TAIL(&cm->cm_prp_page_list, prp_page_info, prp_page_link); 3213 3214 /* 3215 * Check if we are within 1 entry of a page boundary we don't want our 3216 * first entry to be a PRP List entry. 3217 */ 3218 page_mask_result = (uintptr_t)((uint8_t *)curr_buff + prp_size) & 3219 page_mask; 3220 if (!page_mask_result) { 3221 /* Bump up to next page boundary. */ 3222 curr_buff = (uint32_t *)((uint8_t *)curr_buff + prp_size); 3223 msg_phys = (uint64_t *)((uint8_t *)msg_phys + prp_size); 3224 } 3225 3226 /* Fill in the chain element and make it an NVMe segment type. */ 3227 main_chain_element->Address.High = 3228 htole32((uint32_t)((uint64_t)(uintptr_t)msg_phys >> 32)); 3229 main_chain_element->Address.Low = 3230 htole32((uint32_t)(uintptr_t)msg_phys); 3231 main_chain_element->NextChainOffset = 0; 3232 main_chain_element->Flags = MPI2_IEEE_SGE_FLAGS_CHAIN_ELEMENT | 3233 MPI2_IEEE_SGE_FLAGS_SYSTEM_ADDR | 3234 MPI26_IEEE_SGE_FLAGS_NSF_NVME_PRP; 3235 3236 /* Set SGL pointer to start of contiguous PCIe buffer. */ 3237 ptr_sgl = curr_buff; 3238 sge_dwords = 2; 3239 num_entries = 0; 3240 3241 /* 3242 * NVMe has a very convoluted PRP format. One PRP is required for each 3243 * page or partial page. We need to split up OS SG entries if they are 3244 * longer than one page or cross a page boundary. We also have to insert 3245 * a PRP list pointer entry as the last entry in each physical page of 3246 * the PRP list. 3247 * 3248 * NOTE: The first PRP "entry" is actually placed in the first SGL entry 3249 * in the main message in IEEE 64 format. The 2nd entry in the main 3250 * message is the chain element, and the rest of the PRP entries are 3251 * built in the contiguous PCIe buffer. 3252 */ 3253 first_prp_entry = 1; 3254 ptr_first_sgl = (uint32_t *)cm->cm_sge; 3255 3256 for (i = 0; i < segs_left; i++) { 3257 /* Get physical address and length of this SG entry. */ 3258 paddr = segs[i].ds_addr; 3259 length = segs[i].ds_len; 3260 3261 /* 3262 * Check whether a given SGE buffer lies on a non-PAGED 3263 * boundary if this is not the first page. If so, this is not 3264 * expected so have FW build the SGL. 3265 */ 3266 if ((i != 0) && (((uint32_t)paddr & page_mask) != 0)) { 3267 mpr_dprint(sc, MPR_ERROR, "Unaligned SGE while " 3268 "building NVMe PRPs, low address is 0x%x\n", 3269 (uint32_t)paddr); 3270 return 1; 3271 } 3272 3273 /* Apart from last SGE, if any other SGE boundary is not page 3274 * aligned then it means that hole exists. Existence of hole 3275 * leads to data corruption. So fallback to IEEE SGEs. 3276 */ 3277 if (i != (segs_left - 1)) { 3278 if (((uint32_t)paddr + length) & page_mask) { 3279 mpr_dprint(sc, MPR_ERROR, "Unaligned SGE " 3280 "boundary while building NVMe PRPs, low " 3281 "address: 0x%x and length: %u\n", 3282 (uint32_t)paddr, length); 3283 return 1; 3284 } 3285 } 3286 3287 /* Loop while the length is not zero. */ 3288 while (length) { 3289 /* 3290 * Check if we need to put a list pointer here if we are 3291 * at page boundary - prp_size. 3292 */ 3293 page_mask_result = (uintptr_t)((uint8_t *)ptr_sgl + 3294 prp_size) & page_mask; 3295 if (!page_mask_result) { 3296 /* 3297 * Need to put a PRP list pointer here. 3298 */ 3299 msg_phys = (uint64_t *)((uint8_t *)msg_phys + 3300 prp_size); 3301 *ptr_sgl = htole32((uintptr_t)msg_phys); 3302 *(ptr_sgl+1) = htole32((uint64_t)(uintptr_t) 3303 msg_phys >> 32); 3304 ptr_sgl += sge_dwords; 3305 num_entries++; 3306 } 3307 3308 /* Need to handle if entry will be part of a page. */ 3309 offset = (uint32_t)paddr & page_mask; 3310 entry_len = PAGE_SIZE - offset; 3311 if (first_prp_entry) { 3312 /* 3313 * Put IEEE entry in first SGE in main message. 3314 * (Simple element, System addr, not end of 3315 * list.) 3316 */ 3317 *ptr_first_sgl = htole32((uint32_t)paddr); 3318 *(ptr_first_sgl + 1) = 3319 htole32((uint32_t)((uint64_t)paddr >> 32)); 3320 *(ptr_first_sgl + 2) = htole32(entry_len); 3321 *(ptr_first_sgl + 3) = 0; 3322 3323 /* No longer the first PRP entry. */ 3324 first_prp_entry = 0; 3325 } else { 3326 /* Put entry in list. */ 3327 *ptr_sgl = htole32((uint32_t)paddr); 3328 *(ptr_sgl + 1) = 3329 htole32((uint32_t)((uint64_t)paddr >> 32)); 3330 3331 /* Bump ptr_sgl, msg_phys, and num_entries. */ 3332 ptr_sgl += sge_dwords; 3333 msg_phys = (uint64_t *)((uint8_t *)msg_phys + 3334 prp_size); 3335 num_entries++; 3336 } 3337 3338 /* Bump the phys address by the entry_len. */ 3339 paddr += entry_len; 3340 3341 /* Decrement length accounting for last partial page. */ 3342 if (entry_len > length) 3343 length = 0; 3344 else 3345 length -= entry_len; 3346 } 3347 } 3348 3349 /* Set chain element Length. */ 3350 main_chain_element->Length = htole32(num_entries * prp_size); 3351 3352 /* Return 0, indicating we built a native SGL. */ 3353 return 0; 3354 } 3355 3356 /* 3357 * Add a chain element as the next SGE for the specified command. 3358 * Reset cm_sge and cm_sgesize to indicate all the available space. Chains are 3359 * only required for IEEE commands. Therefore there is no code for commands 3360 * that have the MPR_CM_FLAGS_SGE_SIMPLE flag set (and those commands 3361 * shouldn't be requesting chains). 3362 */ 3363 static int 3364 mpr_add_chain(struct mpr_command *cm, int segsleft) 3365 { 3366 struct mpr_softc *sc = cm->cm_sc; 3367 MPI2_REQUEST_HEADER *req; 3368 MPI25_IEEE_SGE_CHAIN64 *ieee_sgc; 3369 struct mpr_chain *chain; 3370 int sgc_size, current_segs, rem_segs, segs_per_frame; 3371 uint8_t next_chain_offset = 0; 3372 3373 /* 3374 * Fail if a command is requesting a chain for SIMPLE SGE's. For SAS3 3375 * only IEEE commands should be requesting chains. Return some error 3376 * code other than 0. 3377 */ 3378 if (cm->cm_flags & MPR_CM_FLAGS_SGE_SIMPLE) { 3379 mpr_dprint(sc, MPR_ERROR, "A chain element cannot be added to " 3380 "an MPI SGL.\n"); 3381 return(ENOBUFS); 3382 } 3383 3384 sgc_size = sizeof(MPI25_IEEE_SGE_CHAIN64); 3385 if (cm->cm_sglsize < sgc_size) 3386 panic("MPR: Need SGE Error Code\n"); 3387 3388 chain = mpr_alloc_chain(cm->cm_sc); 3389 if (chain == NULL) 3390 return (ENOBUFS); 3391 3392 /* 3393 * Note: a double-linked list is used to make it easier to walk for 3394 * debugging. 3395 */ 3396 TAILQ_INSERT_TAIL(&cm->cm_chain_list, chain, chain_link); 3397 3398 /* 3399 * Need to know if the number of frames left is more than 1 or not. If 3400 * more than 1 frame is required, NextChainOffset will need to be set, 3401 * which will just be the last segment of the frame. 3402 */ 3403 rem_segs = 0; 3404 if (cm->cm_sglsize < (sgc_size * segsleft)) { 3405 /* 3406 * rem_segs is the number of segment remaining after the 3407 * segments that will go into the current frame. Since it is 3408 * known that at least one more frame is required, account for 3409 * the chain element. To know if more than one more frame is 3410 * required, just check if there will be a remainder after using 3411 * the current frame (with this chain) and the next frame. If 3412 * so the NextChainOffset must be the last element of the next 3413 * frame. 3414 */ 3415 current_segs = (cm->cm_sglsize / sgc_size) - 1; 3416 rem_segs = segsleft - current_segs; 3417 segs_per_frame = sc->chain_frame_size / sgc_size; 3418 if (rem_segs > segs_per_frame) { 3419 next_chain_offset = segs_per_frame - 1; 3420 } 3421 } 3422 ieee_sgc = &((MPI25_SGE_IO_UNION *)cm->cm_sge)->IeeeChain; 3423 ieee_sgc->Length = next_chain_offset ? 3424 htole32((uint32_t)sc->chain_frame_size) : 3425 htole32((uint32_t)rem_segs * (uint32_t)sgc_size); 3426 ieee_sgc->NextChainOffset = next_chain_offset; 3427 ieee_sgc->Flags = (MPI2_IEEE_SGE_FLAGS_CHAIN_ELEMENT | 3428 MPI2_IEEE_SGE_FLAGS_SYSTEM_ADDR); 3429 ieee_sgc->Address.Low = htole32(chain->chain_busaddr); 3430 ieee_sgc->Address.High = htole32(chain->chain_busaddr >> 32); 3431 cm->cm_sge = &((MPI25_SGE_IO_UNION *)chain->chain)->IeeeSimple; 3432 req = (MPI2_REQUEST_HEADER *)cm->cm_req; 3433 req->ChainOffset = (sc->chain_frame_size - sgc_size) >> 4; 3434 3435 cm->cm_sglsize = sc->chain_frame_size; 3436 return (0); 3437 } 3438 3439 /* 3440 * Add one scatter-gather element to the scatter-gather list for a command. 3441 * Maintain cm_sglsize and cm_sge as the remaining size and pointer to the 3442 * next SGE to fill in, respectively. In Gen3, the MPI SGL does not have a 3443 * chain, so don't consider any chain additions. 3444 */ 3445 int 3446 mpr_push_sge(struct mpr_command *cm, MPI2_SGE_SIMPLE64 *sge, size_t len, 3447 int segsleft) 3448 { 3449 uint32_t saved_buf_len, saved_address_low, saved_address_high; 3450 u32 sge_flags; 3451 3452 /* 3453 * case 1: >=1 more segment, no room for anything (error) 3454 * case 2: 1 more segment and enough room for it 3455 */ 3456 3457 if (cm->cm_sglsize < (segsleft * sizeof(MPI2_SGE_SIMPLE64))) { 3458 mpr_dprint(cm->cm_sc, MPR_ERROR, 3459 "%s: warning: Not enough room for MPI SGL in frame.\n", 3460 __func__); 3461 return(ENOBUFS); 3462 } 3463 3464 KASSERT(segsleft == 1, 3465 ("segsleft cannot be more than 1 for an MPI SGL; segsleft = %d\n", 3466 segsleft)); 3467 3468 /* 3469 * There is one more segment left to add for the MPI SGL and there is 3470 * enough room in the frame to add it. This is the normal case because 3471 * MPI SGL's don't have chains, otherwise something is wrong. 3472 * 3473 * If this is a bi-directional request, need to account for that 3474 * here. Save the pre-filled sge values. These will be used 3475 * either for the 2nd SGL or for a single direction SGL. If 3476 * cm_out_len is non-zero, this is a bi-directional request, so 3477 * fill in the OUT SGL first, then the IN SGL, otherwise just 3478 * fill in the IN SGL. Note that at this time, when filling in 3479 * 2 SGL's for a bi-directional request, they both use the same 3480 * DMA buffer (same cm command). 3481 */ 3482 saved_buf_len = sge->FlagsLength & 0x00FFFFFF; 3483 saved_address_low = sge->Address.Low; 3484 saved_address_high = sge->Address.High; 3485 if (cm->cm_out_len) { 3486 sge->FlagsLength = cm->cm_out_len | 3487 ((uint32_t)(MPI2_SGE_FLAGS_SIMPLE_ELEMENT | 3488 MPI2_SGE_FLAGS_END_OF_BUFFER | 3489 MPI2_SGE_FLAGS_HOST_TO_IOC | 3490 MPI2_SGE_FLAGS_64_BIT_ADDRESSING) << 3491 MPI2_SGE_FLAGS_SHIFT); 3492 cm->cm_sglsize -= len; 3493 /* Endian Safe code */ 3494 sge_flags = sge->FlagsLength; 3495 sge->FlagsLength = htole32(sge_flags); 3496 bcopy(sge, cm->cm_sge, len); 3497 cm->cm_sge = (MPI2_SGE_IO_UNION *)((uintptr_t)cm->cm_sge + len); 3498 } 3499 sge->FlagsLength = saved_buf_len | 3500 ((uint32_t)(MPI2_SGE_FLAGS_SIMPLE_ELEMENT | 3501 MPI2_SGE_FLAGS_END_OF_BUFFER | 3502 MPI2_SGE_FLAGS_LAST_ELEMENT | 3503 MPI2_SGE_FLAGS_END_OF_LIST | 3504 MPI2_SGE_FLAGS_64_BIT_ADDRESSING) << 3505 MPI2_SGE_FLAGS_SHIFT); 3506 if (cm->cm_flags & MPR_CM_FLAGS_DATAIN) { 3507 sge->FlagsLength |= 3508 ((uint32_t)(MPI2_SGE_FLAGS_IOC_TO_HOST) << 3509 MPI2_SGE_FLAGS_SHIFT); 3510 } else { 3511 sge->FlagsLength |= 3512 ((uint32_t)(MPI2_SGE_FLAGS_HOST_TO_IOC) << 3513 MPI2_SGE_FLAGS_SHIFT); 3514 } 3515 sge->Address.Low = saved_address_low; 3516 sge->Address.High = saved_address_high; 3517 3518 cm->cm_sglsize -= len; 3519 /* Endian Safe code */ 3520 sge_flags = sge->FlagsLength; 3521 sge->FlagsLength = htole32(sge_flags); 3522 bcopy(sge, cm->cm_sge, len); 3523 cm->cm_sge = (MPI2_SGE_IO_UNION *)((uintptr_t)cm->cm_sge + len); 3524 return (0); 3525 } 3526 3527 /* 3528 * Add one IEEE scatter-gather element (chain or simple) to the IEEE scatter- 3529 * gather list for a command. Maintain cm_sglsize and cm_sge as the 3530 * remaining size and pointer to the next SGE to fill in, respectively. 3531 */ 3532 int 3533 mpr_push_ieee_sge(struct mpr_command *cm, void *sgep, int segsleft) 3534 { 3535 MPI2_IEEE_SGE_SIMPLE64 *sge = sgep; 3536 int error, ieee_sge_size = sizeof(MPI25_SGE_IO_UNION); 3537 uint32_t saved_buf_len, saved_address_low, saved_address_high; 3538 uint32_t sge_length; 3539 3540 /* 3541 * case 1: No room for chain or segment (error). 3542 * case 2: Two or more segments left but only room for chain. 3543 * case 3: Last segment and room for it, so set flags. 3544 */ 3545 3546 /* 3547 * There should be room for at least one element, or there is a big 3548 * problem. 3549 */ 3550 if (cm->cm_sglsize < ieee_sge_size) 3551 panic("MPR: Need SGE Error Code\n"); 3552 3553 if ((segsleft >= 2) && (cm->cm_sglsize < (ieee_sge_size * 2))) { 3554 if ((error = mpr_add_chain(cm, segsleft)) != 0) 3555 return (error); 3556 } 3557 3558 if (segsleft == 1) { 3559 /* 3560 * If this is a bi-directional request, need to account for that 3561 * here. Save the pre-filled sge values. These will be used 3562 * either for the 2nd SGL or for a single direction SGL. If 3563 * cm_out_len is non-zero, this is a bi-directional request, so 3564 * fill in the OUT SGL first, then the IN SGL, otherwise just 3565 * fill in the IN SGL. Note that at this time, when filling in 3566 * 2 SGL's for a bi-directional request, they both use the same 3567 * DMA buffer (same cm command). 3568 */ 3569 saved_buf_len = sge->Length; 3570 saved_address_low = sge->Address.Low; 3571 saved_address_high = sge->Address.High; 3572 if (cm->cm_out_len) { 3573 sge->Length = cm->cm_out_len; 3574 sge->Flags = (MPI2_IEEE_SGE_FLAGS_SIMPLE_ELEMENT | 3575 MPI2_IEEE_SGE_FLAGS_SYSTEM_ADDR); 3576 cm->cm_sglsize -= ieee_sge_size; 3577 /* Endian Safe code */ 3578 sge_length = sge->Length; 3579 sge->Length = htole32(sge_length); 3580 bcopy(sgep, cm->cm_sge, ieee_sge_size); 3581 cm->cm_sge = 3582 (MPI25_SGE_IO_UNION *)((uintptr_t)cm->cm_sge + 3583 ieee_sge_size); 3584 } 3585 sge->Length = saved_buf_len; 3586 sge->Flags = (MPI2_IEEE_SGE_FLAGS_SIMPLE_ELEMENT | 3587 MPI2_IEEE_SGE_FLAGS_SYSTEM_ADDR | 3588 MPI25_IEEE_SGE_FLAGS_END_OF_LIST); 3589 sge->Address.Low = saved_address_low; 3590 sge->Address.High = saved_address_high; 3591 } 3592 3593 cm->cm_sglsize -= ieee_sge_size; 3594 /* Endian Safe code */ 3595 sge_length = sge->Length; 3596 sge->Length = htole32(sge_length); 3597 bcopy(sgep, cm->cm_sge, ieee_sge_size); 3598 cm->cm_sge = (MPI25_SGE_IO_UNION *)((uintptr_t)cm->cm_sge + 3599 ieee_sge_size); 3600 return (0); 3601 } 3602 3603 /* 3604 * Add one dma segment to the scatter-gather list for a command. 3605 */ 3606 int 3607 mpr_add_dmaseg(struct mpr_command *cm, vm_paddr_t pa, size_t len, u_int flags, 3608 int segsleft) 3609 { 3610 MPI2_SGE_SIMPLE64 sge; 3611 MPI2_IEEE_SGE_SIMPLE64 ieee_sge; 3612 3613 if (!(cm->cm_flags & MPR_CM_FLAGS_SGE_SIMPLE)) { 3614 ieee_sge.Flags = (MPI2_IEEE_SGE_FLAGS_SIMPLE_ELEMENT | 3615 MPI2_IEEE_SGE_FLAGS_SYSTEM_ADDR); 3616 ieee_sge.Length = len; 3617 mpr_from_u64(pa, &ieee_sge.Address); 3618 3619 return (mpr_push_ieee_sge(cm, &ieee_sge, segsleft)); 3620 } else { 3621 /* 3622 * This driver always uses 64-bit address elements for 3623 * simplicity. 3624 */ 3625 flags |= MPI2_SGE_FLAGS_SIMPLE_ELEMENT | 3626 MPI2_SGE_FLAGS_64_BIT_ADDRESSING; 3627 /* Set Endian safe macro in mpr_push_sge */ 3628 sge.FlagsLength = len | (flags << MPI2_SGE_FLAGS_SHIFT); 3629 mpr_from_u64(pa, &sge.Address); 3630 3631 return (mpr_push_sge(cm, &sge, sizeof sge, segsleft)); 3632 } 3633 } 3634 3635 static void 3636 mpr_data_cb(void *arg, bus_dma_segment_t *segs, int nsegs, int error) 3637 { 3638 struct mpr_softc *sc; 3639 struct mpr_command *cm; 3640 u_int i, dir, sflags; 3641 3642 cm = (struct mpr_command *)arg; 3643 sc = cm->cm_sc; 3644 3645 /* 3646 * In this case, just print out a warning and let the chip tell the 3647 * user they did the wrong thing. 3648 */ 3649 if ((cm->cm_max_segs != 0) && (nsegs > cm->cm_max_segs)) { 3650 mpr_dprint(sc, MPR_ERROR, "%s: warning: busdma returned %d " 3651 "segments, more than the %d allowed\n", __func__, nsegs, 3652 cm->cm_max_segs); 3653 } 3654 3655 /* 3656 * Set up DMA direction flags. Bi-directional requests are also handled 3657 * here. In that case, both direction flags will be set. 3658 */ 3659 sflags = 0; 3660 if (cm->cm_flags & MPR_CM_FLAGS_SMP_PASS) { 3661 /* 3662 * We have to add a special case for SMP passthrough, there 3663 * is no easy way to generically handle it. The first 3664 * S/G element is used for the command (therefore the 3665 * direction bit needs to be set). The second one is used 3666 * for the reply. We'll leave it to the caller to make 3667 * sure we only have two buffers. 3668 */ 3669 /* 3670 * Even though the busdma man page says it doesn't make 3671 * sense to have both direction flags, it does in this case. 3672 * We have one s/g element being accessed in each direction. 3673 */ 3674 dir = BUS_DMASYNC_PREWRITE | BUS_DMASYNC_PREREAD; 3675 3676 /* 3677 * Set the direction flag on the first buffer in the SMP 3678 * passthrough request. We'll clear it for the second one. 3679 */ 3680 sflags |= MPI2_SGE_FLAGS_DIRECTION | 3681 MPI2_SGE_FLAGS_END_OF_BUFFER; 3682 } else if (cm->cm_flags & MPR_CM_FLAGS_DATAOUT) { 3683 sflags |= MPI2_SGE_FLAGS_HOST_TO_IOC; 3684 dir = BUS_DMASYNC_PREWRITE; 3685 } else 3686 dir = BUS_DMASYNC_PREREAD; 3687 3688 /* Check if a native SG list is needed for an NVMe PCIe device. */ 3689 if (cm->cm_targ && cm->cm_targ->is_nvme && 3690 mpr_check_pcie_native_sgl(sc, cm, segs, nsegs) == 0) { 3691 /* A native SG list was built, skip to end. */ 3692 goto out; 3693 } 3694 3695 for (i = 0; i < nsegs; i++) { 3696 if ((cm->cm_flags & MPR_CM_FLAGS_SMP_PASS) && (i != 0)) { 3697 sflags &= ~MPI2_SGE_FLAGS_DIRECTION; 3698 } 3699 error = mpr_add_dmaseg(cm, segs[i].ds_addr, segs[i].ds_len, 3700 sflags, nsegs - i); 3701 if (error != 0) { 3702 /* Resource shortage, roll back! */ 3703 if (ratecheck(&sc->lastfail, &mpr_chainfail_interval)) 3704 mpr_dprint(sc, MPR_INFO, "Out of chain frames, " 3705 "consider increasing hw.mpr.max_chains.\n"); 3706 cm->cm_flags |= MPR_CM_FLAGS_CHAIN_FAILED; 3707 /* 3708 * mpr_complete_command can only be called on commands 3709 * that are in the queue. Since this is an error path 3710 * which gets called before we enqueue, update the state 3711 * to meet this requirement before we complete it. 3712 */ 3713 cm->cm_state = MPR_CM_STATE_INQUEUE; 3714 mpr_complete_command(sc, cm); 3715 return; 3716 } 3717 } 3718 3719 out: 3720 bus_dmamap_sync(sc->buffer_dmat, cm->cm_dmamap, dir); 3721 mpr_enqueue_request(sc, cm); 3722 3723 return; 3724 } 3725 3726 static void 3727 mpr_data_cb2(void *arg, bus_dma_segment_t *segs, int nsegs, bus_size_t mapsize, 3728 int error) 3729 { 3730 mpr_data_cb(arg, segs, nsegs, error); 3731 } 3732 3733 /* 3734 * This is the routine to enqueue commands ansynchronously. 3735 * Note that the only error path here is from bus_dmamap_load(), which can 3736 * return EINPROGRESS if it is waiting for resources. Other than this, it's 3737 * assumed that if you have a command in-hand, then you have enough credits 3738 * to use it. 3739 */ 3740 int 3741 mpr_map_command(struct mpr_softc *sc, struct mpr_command *cm) 3742 { 3743 int error = 0; 3744 3745 if (cm->cm_flags & MPR_CM_FLAGS_USE_UIO) { 3746 error = bus_dmamap_load_uio(sc->buffer_dmat, cm->cm_dmamap, 3747 &cm->cm_uio, mpr_data_cb2, cm, 0); 3748 } else if (cm->cm_flags & MPR_CM_FLAGS_USE_CCB) { 3749 error = bus_dmamap_load_ccb(sc->buffer_dmat, cm->cm_dmamap, 3750 cm->cm_data, mpr_data_cb, cm, 0); 3751 } else if ((cm->cm_data != NULL) && (cm->cm_length != 0)) { 3752 error = bus_dmamap_load(sc->buffer_dmat, cm->cm_dmamap, 3753 cm->cm_data, cm->cm_length, mpr_data_cb, cm, 0); 3754 } else { 3755 /* Add a zero-length element as needed */ 3756 if (cm->cm_sge != NULL) 3757 mpr_add_dmaseg(cm, 0, 0, 0, 1); 3758 mpr_enqueue_request(sc, cm); 3759 } 3760 3761 return (error); 3762 } 3763 3764 /* 3765 * This is the routine to enqueue commands synchronously. An error of 3766 * EINPROGRESS from mpr_map_command() is ignored since the command will 3767 * be executed and enqueued automatically. Other errors come from msleep(). 3768 */ 3769 int 3770 mpr_wait_command(struct mpr_softc *sc, struct mpr_command **cmp, int timeout, 3771 int sleep_flag) 3772 { 3773 int error, rc; 3774 struct timeval cur_time, start_time; 3775 struct mpr_command *cm = *cmp; 3776 3777 if (sc->mpr_flags & MPR_FLAGS_DIAGRESET) 3778 return EBUSY; 3779 3780 cm->cm_complete = NULL; 3781 cm->cm_flags |= (MPR_CM_FLAGS_WAKEUP + MPR_CM_FLAGS_POLLED); 3782 error = mpr_map_command(sc, cm); 3783 if ((error != 0) && (error != EINPROGRESS)) 3784 return (error); 3785 3786 // Check for context and wait for 50 mSec at a time until time has 3787 // expired or the command has finished. If msleep can't be used, need 3788 // to poll. 3789 if (curthread->td_no_sleeping) 3790 sleep_flag = NO_SLEEP; 3791 getmicrouptime(&start_time); 3792 if (mtx_owned(&sc->mpr_mtx) && sleep_flag == CAN_SLEEP) { 3793 error = msleep(cm, &sc->mpr_mtx, 0, "mprwait", timeout*hz); 3794 if (error == EWOULDBLOCK) { 3795 /* 3796 * Record the actual elapsed time in the case of a 3797 * timeout for the message below. 3798 */ 3799 getmicrouptime(&cur_time); 3800 timevalsub(&cur_time, &start_time); 3801 } 3802 } else { 3803 while ((cm->cm_flags & MPR_CM_FLAGS_COMPLETE) == 0) { 3804 mpr_intr_locked(sc); 3805 if (sleep_flag == CAN_SLEEP) 3806 pause("mprwait", hz/20); 3807 else 3808 DELAY(50000); 3809 3810 getmicrouptime(&cur_time); 3811 timevalsub(&cur_time, &start_time); 3812 if (cur_time.tv_sec > timeout) { 3813 error = EWOULDBLOCK; 3814 break; 3815 } 3816 } 3817 } 3818 3819 if (error == EWOULDBLOCK) { 3820 if (cm->cm_timeout_handler == NULL) { 3821 mpr_dprint(sc, MPR_FAULT, "Calling Reinit from %s, timeout=%d," 3822 " elapsed=%jd\n", __func__, timeout, 3823 (intmax_t)cur_time.tv_sec); 3824 rc = mpr_reinit(sc); 3825 mpr_dprint(sc, MPR_FAULT, "Reinit %s\n", (rc == 0) ? "success" : 3826 "failed"); 3827 } else 3828 cm->cm_timeout_handler(sc, cm); 3829 if (sc->mpr_flags & MPR_FLAGS_REALLOCATED) { 3830 /* 3831 * Tell the caller that we freed the command in a 3832 * reinit. 3833 */ 3834 *cmp = NULL; 3835 } 3836 error = ETIMEDOUT; 3837 } 3838 return (error); 3839 } 3840 3841 /* 3842 * This is the routine to enqueue a command synchonously and poll for 3843 * completion. Its use should be rare. 3844 */ 3845 int 3846 mpr_request_polled(struct mpr_softc *sc, struct mpr_command **cmp) 3847 { 3848 int error, rc; 3849 struct timeval cur_time, start_time; 3850 struct mpr_command *cm = *cmp; 3851 3852 error = 0; 3853 3854 cm->cm_flags |= MPR_CM_FLAGS_POLLED; 3855 cm->cm_complete = NULL; 3856 mpr_map_command(sc, cm); 3857 3858 getmicrouptime(&start_time); 3859 while ((cm->cm_flags & MPR_CM_FLAGS_COMPLETE) == 0) { 3860 mpr_intr_locked(sc); 3861 3862 if (mtx_owned(&sc->mpr_mtx)) 3863 msleep(&sc->msleep_fake_chan, &sc->mpr_mtx, 0, 3864 "mprpoll", hz/20); 3865 else 3866 pause("mprpoll", hz/20); 3867 3868 /* 3869 * Check for real-time timeout and fail if more than 60 seconds. 3870 */ 3871 getmicrouptime(&cur_time); 3872 timevalsub(&cur_time, &start_time); 3873 if (cur_time.tv_sec > 60) { 3874 mpr_dprint(sc, MPR_FAULT, "polling failed\n"); 3875 error = ETIMEDOUT; 3876 break; 3877 } 3878 } 3879 cm->cm_state = MPR_CM_STATE_BUSY; 3880 if (error) { 3881 mpr_dprint(sc, MPR_FAULT, "Calling Reinit from %s\n", __func__); 3882 rc = mpr_reinit(sc); 3883 mpr_dprint(sc, MPR_FAULT, "Reinit %s\n", (rc == 0) ? "success" : 3884 "failed"); 3885 3886 if (sc->mpr_flags & MPR_FLAGS_REALLOCATED) { 3887 /* 3888 * Tell the caller that we freed the command in a 3889 * reinit. 3890 */ 3891 *cmp = NULL; 3892 } 3893 } 3894 return (error); 3895 } 3896 3897 /* 3898 * The MPT driver had a verbose interface for config pages. In this driver, 3899 * reduce it to much simpler terms, similar to the Linux driver. 3900 */ 3901 int 3902 mpr_read_config_page(struct mpr_softc *sc, struct mpr_config_params *params) 3903 { 3904 MPI2_CONFIG_REQUEST *req; 3905 struct mpr_command *cm; 3906 int error; 3907 3908 if (sc->mpr_flags & MPR_FLAGS_BUSY) { 3909 return (EBUSY); 3910 } 3911 3912 cm = mpr_alloc_command(sc); 3913 if (cm == NULL) { 3914 return (EBUSY); 3915 } 3916 3917 req = (MPI2_CONFIG_REQUEST *)cm->cm_req; 3918 req->Function = MPI2_FUNCTION_CONFIG; 3919 req->Action = params->action; 3920 req->SGLFlags = 0; 3921 req->ChainOffset = 0; 3922 req->PageAddress = params->page_address; 3923 if (params->hdr.Struct.PageType == MPI2_CONFIG_PAGETYPE_EXTENDED) { 3924 MPI2_CONFIG_EXTENDED_PAGE_HEADER *hdr; 3925 3926 hdr = ¶ms->hdr.Ext; 3927 req->ExtPageType = hdr->ExtPageType; 3928 req->ExtPageLength = hdr->ExtPageLength; 3929 req->Header.PageType = MPI2_CONFIG_PAGETYPE_EXTENDED; 3930 req->Header.PageLength = 0; /* Must be set to zero */ 3931 req->Header.PageNumber = hdr->PageNumber; 3932 req->Header.PageVersion = hdr->PageVersion; 3933 } else { 3934 MPI2_CONFIG_PAGE_HEADER *hdr; 3935 3936 hdr = ¶ms->hdr.Struct; 3937 req->Header.PageType = hdr->PageType; 3938 req->Header.PageNumber = hdr->PageNumber; 3939 req->Header.PageLength = hdr->PageLength; 3940 req->Header.PageVersion = hdr->PageVersion; 3941 } 3942 3943 cm->cm_data = params->buffer; 3944 cm->cm_length = params->length; 3945 if (cm->cm_data != NULL) { 3946 cm->cm_sge = &req->PageBufferSGE; 3947 cm->cm_sglsize = sizeof(MPI2_SGE_IO_UNION); 3948 cm->cm_flags = MPR_CM_FLAGS_SGE_SIMPLE | MPR_CM_FLAGS_DATAIN; 3949 } else 3950 cm->cm_sge = NULL; 3951 cm->cm_desc.Default.RequestFlags = MPI2_REQ_DESCRIPT_FLAGS_DEFAULT_TYPE; 3952 3953 cm->cm_complete_data = params; 3954 if (params->callback != NULL) { 3955 cm->cm_complete = mpr_config_complete; 3956 return (mpr_map_command(sc, cm)); 3957 } else { 3958 error = mpr_wait_command(sc, &cm, 0, CAN_SLEEP); 3959 if (error) { 3960 mpr_dprint(sc, MPR_FAULT, 3961 "Error %d reading config page\n", error); 3962 if (cm != NULL) 3963 mpr_free_command(sc, cm); 3964 return (error); 3965 } 3966 mpr_config_complete(sc, cm); 3967 } 3968 3969 return (0); 3970 } 3971 3972 int 3973 mpr_write_config_page(struct mpr_softc *sc, struct mpr_config_params *params) 3974 { 3975 return (EINVAL); 3976 } 3977 3978 static void 3979 mpr_config_complete(struct mpr_softc *sc, struct mpr_command *cm) 3980 { 3981 MPI2_CONFIG_REPLY *reply; 3982 struct mpr_config_params *params; 3983 3984 MPR_FUNCTRACE(sc); 3985 params = cm->cm_complete_data; 3986 3987 if (cm->cm_data != NULL) { 3988 bus_dmamap_sync(sc->buffer_dmat, cm->cm_dmamap, 3989 BUS_DMASYNC_POSTREAD); 3990 bus_dmamap_unload(sc->buffer_dmat, cm->cm_dmamap); 3991 } 3992 3993 /* 3994 * XXX KDM need to do more error recovery? This results in the 3995 * device in question not getting probed. 3996 */ 3997 if ((cm->cm_flags & MPR_CM_FLAGS_ERROR_MASK) != 0) { 3998 params->status = MPI2_IOCSTATUS_BUSY; 3999 goto done; 4000 } 4001 4002 reply = (MPI2_CONFIG_REPLY *)cm->cm_reply; 4003 if (reply == NULL) { 4004 params->status = MPI2_IOCSTATUS_BUSY; 4005 goto done; 4006 } 4007 params->status = reply->IOCStatus; 4008 if (params->hdr.Struct.PageType == MPI2_CONFIG_PAGETYPE_EXTENDED) { 4009 params->hdr.Ext.ExtPageType = reply->ExtPageType; 4010 params->hdr.Ext.ExtPageLength = reply->ExtPageLength; 4011 params->hdr.Ext.PageType = reply->Header.PageType; 4012 params->hdr.Ext.PageNumber = reply->Header.PageNumber; 4013 params->hdr.Ext.PageVersion = reply->Header.PageVersion; 4014 } else { 4015 params->hdr.Struct.PageType = reply->Header.PageType; 4016 params->hdr.Struct.PageNumber = reply->Header.PageNumber; 4017 params->hdr.Struct.PageLength = reply->Header.PageLength; 4018 params->hdr.Struct.PageVersion = reply->Header.PageVersion; 4019 } 4020 4021 done: 4022 mpr_free_command(sc, cm); 4023 if (params->callback != NULL) 4024 params->callback(sc, params); 4025 4026 return; 4027 } 4028