sys/dev/liquidio/base/lio_request_manager.c

   1 /*
   2  *   BSD LICENSE
   3  *
   4  *   Copyright(c) 2017 Cavium, Inc.. All rights reserved.
   5  *   All rights reserved.
   6  *
   7  *   Redistribution and use in source and binary forms, with or without
   8  *   modification, are permitted provided that the following conditions
   9  *   are met:
  10  *
  11  *     * Redistributions of source code must retain the above copyright
  12  *       notice, this list of conditions and the following disclaimer.
  13  *     * Redistributions in binary form must reproduce the above copyright
  14  *       notice, this list of conditions and the following disclaimer in
  15  *       the documentation and/or other materials provided with the
  16  *       distribution.
  17  *     * Neither the name of Cavium, Inc. nor the names of its
  18  *       contributors may be used to endorse or promote products derived
  19  *       from this software without specific prior written permission.
  20  *
  21  *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  22  *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  23  *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
  24  *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
  25  *   OWNER(S) OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
  26  *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
  27  *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
  28  *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
  29  *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
  30  *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
  31  *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  32  */
  33 /*$FreeBSD$*/
  34
  35 #include "lio_bsd.h"
  36 #include "lio_common.h"
  37 #include "lio_droq.h"
  38 #include "lio_iq.h"
  39 #include "lio_response_manager.h"
  40 #include "lio_device.h"
  41 #include "lio_main.h"
  42 #include "lio_network.h"
  43 #include "cn23xx_pf_device.h"
  44 #include "lio_rxtx.h"
  45
  46 struct lio_iq_post_status {
  47         int     status;
  48         int     index;
  49 };
  50
  51 static void     lio_check_db_timeout(void *arg, int pending);
  52 static void     __lio_check_db_timeout(struct octeon_device *oct,
  53                                        uint64_t iq_no);
  54
  55 /* Return 0 on success, 1 on failure */
  56 int
  57 lio_init_instr_queue(struct octeon_device *oct, union octeon_txpciq txpciq,
  58                      uint32_t num_descs)
  59 {
  60         struct lio_instr_queue  *iq;
  61         struct lio_iq_config    *conf = NULL;
  62         struct lio_tq           *db_tq;
  63         struct lio_request_list *request_buf;
  64         bus_size_t              max_size;
  65         uint32_t                iq_no = (uint32_t)txpciq.s.q_no;
  66         uint32_t                q_size;
  67         int                     error, i;
  68
  69         if (LIO_CN23XX_PF(oct))
  70                 conf = &(LIO_GET_IQ_CFG(LIO_CHIP_CONF(oct, cn23xx_pf)));
  71         if (conf == NULL) {
  72                 lio_dev_err(oct, "Unsupported Chip %x\n", oct->chip_id);
  73                 return (1);
  74         }
  75
  76         q_size = (uint32_t)conf->instr_type * num_descs;
  77         iq = oct->instr_queue[iq_no];
  78         iq->oct_dev = oct;
  79
  80         max_size = LIO_CN23XX_PKI_MAX_FRAME_SIZE * num_descs;
  81
  82         error = bus_dma_tag_create(bus_get_dma_tag(oct->device),        /* parent */
  83                                    1, 0,                                /* alignment, bounds */
  84                                    BUS_SPACE_MAXADDR,                   /* lowaddr */
  85                                    BUS_SPACE_MAXADDR,                   /* highaddr */
  86                                    NULL, NULL,                          /* filter, filterarg */
  87                                    max_size,                            /* maxsize */
  88                                    LIO_MAX_SG,                          /* nsegments */
  89                                    PAGE_SIZE,                           /* maxsegsize */
  90                                    0,                                   /* flags */
  91                                    NULL,                                /* lockfunc */
  92                                    NULL,                                /* lockfuncarg */
  93                                    &iq->txtag);
  94         if (error) {
  95                 lio_dev_err(oct, "Cannot allocate memory for instr queue %d\n",
  96                             iq_no);
  97                 return (1);
  98         }
  99
 100         iq->base_addr = lio_dma_alloc(q_size, &iq->base_addr_dma);
 101         if (!iq->base_addr) {
 102                 lio_dev_err(oct, "Cannot allocate memory for instr queue %d\n",
 103                             iq_no);
 104                 return (1);
 105         }
 106
 107         iq->max_count = num_descs;
 108
 109         /*
 110          * Initialize a list to holds requests that have been posted to
 111          * Octeon but has yet to be fetched by octeon
 112          */
 113         iq->request_list = malloc(sizeof(*iq->request_list) * num_descs,
 114                                   M_DEVBUF, M_NOWAIT | M_ZERO);
 115         if (iq->request_list == NULL) {
 116                 lio_dev_err(oct, "Alloc failed for IQ[%d] nr free list\n",
 117                             iq_no);
 118                 return (1);
 119         }
 120
 121         lio_dev_dbg(oct, "IQ[%d]: base: %p basedma: %lx count: %d\n",
 122                     iq_no, iq->base_addr, iq->base_addr_dma, iq->max_count);
 123
 124         /* Create the descriptor buffer dma maps */
 125         request_buf = iq->request_list;
 126         for (i = 0; i < num_descs; i++, request_buf++) {
 127                 error = bus_dmamap_create(iq->txtag, 0, &request_buf->map);
 128                 if (error) {
 129                         lio_dev_err(oct, "Unable to create TX DMA map\n");
 130                         return (1);
 131                 }
 132         }
 133
 134         iq->txpciq.txpciq64 = txpciq.txpciq64;
 135         iq->fill_cnt = 0;
 136         iq->host_write_index = 0;
 137         iq->octeon_read_index = 0;
 138         iq->flush_index = 0;
 139         iq->last_db_time = 0;
 140         iq->db_timeout = (uint32_t)conf->db_timeout;
 141         atomic_store_rel_int(&iq->instr_pending, 0);
 142
 143         /* Initialize the lock for this instruction queue */
 144         mtx_init(&iq->lock, "Tx_lock", NULL, MTX_DEF);
 145         mtx_init(&iq->post_lock, "iq_post_lock", NULL, MTX_DEF);
 146         mtx_init(&iq->enq_lock, "enq_lock", NULL, MTX_DEF);
 147
 148         mtx_init(&iq->iq_flush_running_lock, "iq_flush_running_lock", NULL,
 149                  MTX_DEF);
 150
 151         oct->io_qmask.iq |= BIT_ULL(iq_no);
 152
 153         /* Set the 32B/64B mode for each input queue */
 154         oct->io_qmask.iq64B |= ((conf->instr_type == 64) << iq_no);
 155         iq->iqcmd_64B = (conf->instr_type == 64);
 156
 157         oct->fn_list.setup_iq_regs(oct, iq_no);
 158
 159         db_tq = &oct->check_db_tq[iq_no];
 160         db_tq->tq = taskqueue_create("lio_check_db_timeout", M_WAITOK,
 161                                      taskqueue_thread_enqueue, &db_tq->tq);
 162         if (db_tq->tq == NULL) {
 163                 lio_dev_err(oct, "check db wq create failed for iq %d\n",
 164                             iq_no);
 165                 return (1);
 166         }
 167
 168         TIMEOUT_TASK_INIT(db_tq->tq, &db_tq->work, 0, lio_check_db_timeout,
 169                           (void *)db_tq);
 170         db_tq->ctxul = iq_no;
 171         db_tq->ctxptr = oct;
 172
 173         taskqueue_start_threads(&db_tq->tq, 1, PI_NET,
 174                                 "lio%d_check_db_timeout:%d",
 175                                 oct->octeon_id, iq_no);
 176         taskqueue_enqueue_timeout(db_tq->tq, &db_tq->work, 1);
 177
 178         /* Allocate a buf ring */
 179         oct->instr_queue[iq_no]->br =
 180                 buf_ring_alloc(LIO_BR_SIZE, M_DEVBUF, M_WAITOK,
 181                                &oct->instr_queue[iq_no]->enq_lock);
 182         if (oct->instr_queue[iq_no]->br == NULL) {
 183                 lio_dev_err(oct, "Critical Failure setting up buf ring\n");
 184                 return (1);
 185         }
 186
 187         return (0);
 188 }
 189
 190 int
 191 lio_delete_instr_queue(struct octeon_device *oct, uint32_t iq_no)
 192 {
 193         struct lio_instr_queue          *iq = oct->instr_queue[iq_no];
 194         struct lio_request_list         *request_buf;
 195         struct lio_mbuf_free_info       *finfo;
 196         uint64_t                        desc_size = 0, q_size;
 197         int                             i;
 198
 199         lio_dev_dbg(oct, "%s[%d]\n", __func__, iq_no);
 200
 201         if (oct->check_db_tq[iq_no].tq != NULL) {
 202                 while (taskqueue_cancel_timeout(oct->check_db_tq[iq_no].tq,
 203                                                 &oct->check_db_tq[iq_no].work,
 204                                                 NULL))
 205                         taskqueue_drain_timeout(oct->check_db_tq[iq_no].tq,
 206                                                 &oct->check_db_tq[iq_no].work);
 207                 taskqueue_free(oct->check_db_tq[iq_no].tq);
 208                 oct->check_db_tq[iq_no].tq = NULL;
 209         }
 210
 211         if (LIO_CN23XX_PF(oct))
 212                 desc_size =
 213                     LIO_GET_IQ_INSTR_TYPE_CFG(LIO_CHIP_CONF(oct, cn23xx_pf));
 214
 215         request_buf = iq->request_list;
 216         for (i = 0; i < iq->max_count; i++, request_buf++) {
 217                 if ((request_buf->reqtype == LIO_REQTYPE_NORESP_NET) ||
 218                     (request_buf->reqtype == LIO_REQTYPE_NORESP_NET_SG)) {
 219                         if (request_buf->buf != NULL) {
 220                                 finfo = request_buf->buf;
 221                                 bus_dmamap_sync(iq->txtag, request_buf->map,
 222                                                 BUS_DMASYNC_POSTWRITE);
 223                                 bus_dmamap_unload(iq->txtag,
 224                                                   request_buf->map);
 225                                 m_freem(finfo->mb);
 226                                 request_buf->buf = NULL;
 227                                 if (request_buf->map != NULL) {
 228                                         bus_dmamap_destroy(iq->txtag,
 229                                                            request_buf->map);
 230                                         request_buf->map = NULL;
 231                                 }
 232                         } else if (request_buf->map != NULL) {
 233                                 bus_dmamap_unload(iq->txtag, request_buf->map);
 234                                 bus_dmamap_destroy(iq->txtag, request_buf->map);
 235                                 request_buf->map = NULL;
 236                         }
 237                 }
 238         }
 239
 240         if (iq->br != NULL) {
 241                 buf_ring_free(iq->br, M_DEVBUF);
 242                 iq->br = NULL;
 243         }
 244
 245         if (iq->request_list != NULL) {
 246                 free(iq->request_list, M_DEVBUF);
 247                 iq->request_list = NULL;
 248         }
 249
 250         if (iq->txtag != NULL) {
 251                 bus_dma_tag_destroy(iq->txtag);
 252                 iq->txtag = NULL;
 253         }
 254
 255         if (iq->base_addr) {
 256                 q_size = iq->max_count * desc_size;
 257                 lio_dma_free((uint32_t)q_size, iq->base_addr);
 258
 259                 oct->io_qmask.iq &= ~(1ULL << iq_no);
 260                 bzero(oct->instr_queue[iq_no], sizeof(struct lio_instr_queue));
 261                 oct->num_iqs--;
 262
 263                 return (0);
 264         }
 265
 266         return (1);
 267 }
 268
 269 /* Return 0 on success, 1 on failure */
 270 int
 271 lio_setup_iq(struct octeon_device *oct, int ifidx, int q_index,
 272              union octeon_txpciq txpciq, uint32_t num_descs)
 273 {
 274         uint32_t        iq_no = (uint32_t)txpciq.s.q_no;
 275
 276         if (oct->instr_queue[iq_no]->oct_dev != NULL) {
 277                 lio_dev_dbg(oct, "IQ is in use. Cannot create the IQ: %d again\n",
 278                             iq_no);
 279                 oct->instr_queue[iq_no]->txpciq.txpciq64 = txpciq.txpciq64;
 280                 return (0);
 281         }
 282
 283         oct->instr_queue[iq_no]->q_index = q_index;
 284         oct->instr_queue[iq_no]->ifidx = ifidx;
 285
 286         if (lio_init_instr_queue(oct, txpciq, num_descs)) {
 287                 lio_delete_instr_queue(oct, iq_no);
 288                 return (1);
 289         }
 290
 291         oct->num_iqs++;
 292         if (oct->fn_list.enable_io_queues(oct))
 293                 return (1);
 294
 295         return (0);
 296 }
 297
 298 int
 299 lio_wait_for_instr_fetch(struct octeon_device *oct)
 300 {
 301         int     i, retry = 1000, pending, instr_cnt = 0;
 302
 303         do {
 304                 instr_cnt = 0;
 305
 306                 for (i = 0; i < LIO_MAX_INSTR_QUEUES(oct); i++) {
 307                         if (!(oct->io_qmask.iq & BIT_ULL(i)))
 308                                 continue;
 309                         pending = atomic_load_acq_int(
 310                                         &oct->instr_queue[i]->instr_pending);
 311                         if (pending)
 312                                 __lio_check_db_timeout(oct, i);
 313                         instr_cnt += pending;
 314                 }
 315
 316                 if (instr_cnt == 0)
 317                         break;
 318
 319                 lio_sleep_timeout(1);
 320
 321         } while (retry-- && instr_cnt);
 322
 323         return (instr_cnt);
 324 }
 325
 326 static inline void
 327 lio_ring_doorbell(struct octeon_device *oct, struct lio_instr_queue *iq)
 328 {
 329
 330         if (atomic_load_acq_int(&oct->status) == LIO_DEV_RUNNING) {
 331                 lio_write_csr32(oct, iq->doorbell_reg, iq->fill_cnt);
 332                 /* make sure doorbell write goes through */
 333                 __compiler_membar();
 334                 iq->fill_cnt = 0;
 335                 iq->last_db_time = ticks;
 336                 return;
 337         }
 338 }
 339
 340 static inline void
 341 __lio_copy_cmd_into_iq(struct lio_instr_queue *iq, uint8_t *cmd)
 342 {
 343         uint8_t *iqptr, cmdsize;
 344
 345         cmdsize = ((iq->iqcmd_64B) ? 64 : 32);
 346         iqptr = iq->base_addr + (cmdsize * iq->host_write_index);
 347
 348         memcpy(iqptr, cmd, cmdsize);
 349 }
 350
 351 static inline struct lio_iq_post_status
 352 __lio_post_command2(struct lio_instr_queue *iq, uint8_t *cmd)
 353 {
 354         struct lio_iq_post_status       st;
 355
 356         st.status = LIO_IQ_SEND_OK;
 357
 358         /*
 359          * This ensures that the read index does not wrap around to the same
 360          * position if queue gets full before Octeon could fetch any instr.
 361          */
 362         if (atomic_load_acq_int(&iq->instr_pending) >=
 363             (int32_t)(iq->max_count - 1)) {
 364                 st.status = LIO_IQ_SEND_FAILED;
 365                 st.index = -1;
 366                 return (st);
 367         }
 368
 369         if (atomic_load_acq_int(&iq->instr_pending) >=
 370             (int32_t)(iq->max_count - 2))
 371                 st.status = LIO_IQ_SEND_STOP;
 372
 373         __lio_copy_cmd_into_iq(iq, cmd);
 374
 375         /* "index" is returned, host_write_index is modified. */
 376         st.index = iq->host_write_index;
 377         iq->host_write_index = lio_incr_index(iq->host_write_index, 1,
 378                                               iq->max_count);
 379         iq->fill_cnt++;
 380
 381         /*
 382          * Flush the command into memory. We need to be sure the data is in
 383          * memory before indicating that the instruction is pending.
 384          */
 385         wmb();
 386
 387         atomic_add_int(&iq->instr_pending, 1);
 388
 389         return (st);
 390 }
 391
 392 static inline void
 393 __lio_add_to_request_list(struct lio_instr_queue *iq, int idx, void *buf,
 394                           int reqtype)
 395 {
 396
 397         iq->request_list[idx].buf = buf;
 398         iq->request_list[idx].reqtype = reqtype;
 399 }
 400
 401 /* Can only run in process context */
 402 int
 403 lio_process_iq_request_list(struct octeon_device *oct,
 404                             struct lio_instr_queue *iq, uint32_t budget)
 405 {
 406         struct lio_soft_command         *sc;
 407         struct octeon_instr_irh         *irh = NULL;
 408         struct lio_mbuf_free_info       *finfo;
 409         void                            *buf;
 410         uint32_t                        inst_count = 0;
 411         uint32_t                        old = iq->flush_index;
 412         int                             reqtype;
 413
 414         while (old != iq->octeon_read_index) {
 415                 reqtype = iq->request_list[old].reqtype;
 416                 buf = iq->request_list[old].buf;
 417                 finfo = buf;
 418
 419                 if (reqtype == LIO_REQTYPE_NONE)
 420                         goto skip_this;
 421
 422                 switch (reqtype) {
 423                 case LIO_REQTYPE_NORESP_NET:
 424                         lio_free_mbuf(iq, buf);
 425                         break;
 426                 case LIO_REQTYPE_NORESP_NET_SG:
 427                         lio_free_sgmbuf(iq, buf);
 428                         break;
 429                 case LIO_REQTYPE_RESP_NET:
 430                 case LIO_REQTYPE_SOFT_COMMAND:
 431                         sc = buf;
 432                         if (LIO_CN23XX_PF(oct))
 433                                 irh = (struct octeon_instr_irh *)
 434                                         &sc->cmd.cmd3.irh;
 435                         if (irh->rflag) {
 436                                 /*
 437                                  * We're expecting a response from Octeon.
 438                                  * It's up to lio_process_ordered_list() to
 439                                  * process  sc. Add sc to the ordered soft
 440                                  * command response list because we expect
 441                                  * a response from Octeon.
 442                                  */
 443                                 mtx_lock(&oct->response_list
 444                                          [LIO_ORDERED_SC_LIST].lock);
 445                                 atomic_add_int(&oct->response_list
 446                                                [LIO_ORDERED_SC_LIST].
 447                                                pending_req_count, 1);
 448                                 STAILQ_INSERT_TAIL(&oct->response_list
 449                                                    [LIO_ORDERED_SC_LIST].
 450                                                    head, &sc->node, entries);
 451                                 mtx_unlock(&oct->response_list
 452                                            [LIO_ORDERED_SC_LIST].lock);
 453                         } else {
 454                                 if (sc->callback != NULL) {
 455                                         /* This callback must not sleep */
 456                                         sc->callback(oct, LIO_REQUEST_DONE,
 457                                                      sc->callback_arg);
 458                                 }
 459                         }
 460
 461                         break;
 462                 default:
 463                         lio_dev_err(oct, "%s Unknown reqtype: %d buf: %p at idx %d\n",
 464                                     __func__, reqtype, buf, old);
 465                 }
 466
 467                 iq->request_list[old].buf = NULL;
 468                 iq->request_list[old].reqtype = 0;
 469
 470 skip_this:
 471                 inst_count++;
 472                 old = lio_incr_index(old, 1, iq->max_count);
 473
 474                 if ((budget) && (inst_count >= budget))
 475                         break;
 476         }
 477
 478         iq->flush_index = old;
 479
 480         return (inst_count);
 481 }
 482
 483 /* Can only be called from process context */
 484 int
 485 lio_flush_iq(struct octeon_device *oct, struct lio_instr_queue *iq,
 486              uint32_t budget)
 487 {
 488         uint32_t        inst_processed = 0;
 489         uint32_t        tot_inst_processed = 0;
 490         int             tx_done = 1;
 491
 492         if (!mtx_trylock(&iq->iq_flush_running_lock))
 493                 return (tx_done);
 494
 495         mtx_lock(&iq->lock);
 496
 497         iq->octeon_read_index = oct->fn_list.update_iq_read_idx(iq);
 498
 499         do {
 500                 /* Process any outstanding IQ packets. */
 501                 if (iq->flush_index == iq->octeon_read_index)
 502                         break;
 503
 504                 if (budget)
 505                         inst_processed =
 506                                 lio_process_iq_request_list(oct, iq,
 507                                                             budget -
 508                                                             tot_inst_processed);
 509                 else
 510                         inst_processed =
 511                                 lio_process_iq_request_list(oct, iq, 0);
 512
 513                 if (inst_processed) {
 514                         atomic_subtract_int(&iq->instr_pending, inst_processed);
 515                         iq->stats.instr_processed += inst_processed;
 516                 }
 517                 tot_inst_processed += inst_processed;
 518                 inst_processed = 0;
 519
 520         } while (tot_inst_processed < budget);
 521
 522         if (budget && (tot_inst_processed >= budget))
 523                 tx_done = 0;
 524
 525         iq->last_db_time = ticks;
 526
 527         mtx_unlock(&iq->lock);
 528
 529         mtx_unlock(&iq->iq_flush_running_lock);
 530
 531         return (tx_done);
 532 }
 533
 534 /*
 535  * Process instruction queue after timeout.
 536  * This routine gets called from a taskqueue or when removing the module.
 537  */
 538 static void
 539 __lio_check_db_timeout(struct octeon_device *oct, uint64_t iq_no)
 540 {
 541         struct lio_instr_queue  *iq;
 542         uint64_t                next_time;
 543
 544         if (oct == NULL)
 545                 return;
 546
 547         iq = oct->instr_queue[iq_no];
 548         if (iq == NULL)
 549                 return;
 550
 551         if (atomic_load_acq_int(&iq->instr_pending)) {
 552                 /* If ticks - last_db_time < db_timeout do nothing  */
 553                 next_time = iq->last_db_time + lio_ms_to_ticks(iq->db_timeout);
 554                 if (!lio_check_timeout(ticks, next_time))
 555                         return;
 556
 557                 iq->last_db_time = ticks;
 558
 559                 /* Flush the instruction queue */
 560                 lio_flush_iq(oct, iq, 0);
 561
 562                 lio_enable_irq(NULL, iq);
 563         }
 564
 565         if (oct->props.ifp != NULL && iq->br != NULL) {
 566                 if (mtx_trylock(&iq->enq_lock)) {
 567                         if (!drbr_empty(oct->props.ifp, iq->br))
 568                                 lio_mq_start_locked(oct->props.ifp, iq);
 569
 570                         mtx_unlock(&iq->enq_lock);
 571                 }
 572         }
 573 }
 574
 575 /*
 576  * Called by the Poll thread at regular intervals to check the instruction
 577  * queue for commands to be posted and for commands that were fetched by Octeon.
 578  */
 579 static void
 580 lio_check_db_timeout(void *arg, int pending)
 581 {
 582         struct lio_tq           *db_tq = (struct lio_tq *)arg;
 583         struct octeon_device    *oct = db_tq->ctxptr;
 584         uint64_t                iq_no = db_tq->ctxul;
 585         uint32_t                delay = 10;
 586
 587         __lio_check_db_timeout(oct, iq_no);
 588         taskqueue_enqueue_timeout(db_tq->tq, &db_tq->work,
 589                                   lio_ms_to_ticks(delay));
 590 }
 591
 592 int
 593 lio_send_command(struct octeon_device *oct, uint32_t iq_no,
 594                  uint32_t force_db, void *cmd, void *buf,
 595                  uint32_t datasize, uint32_t reqtype)
 596 {
 597         struct lio_iq_post_status       st;
 598         struct lio_instr_queue          *iq = oct->instr_queue[iq_no];
 599
 600         /*
 601          * Get the lock and prevent other tasks and tx interrupt handler
 602          * from running.
 603          */
 604         mtx_lock(&iq->post_lock);
 605
 606         st = __lio_post_command2(iq, cmd);
 607
 608         if (st.status != LIO_IQ_SEND_FAILED) {
 609                 __lio_add_to_request_list(iq, st.index, buf, reqtype);
 610                 LIO_INCR_INSTRQUEUE_PKT_COUNT(oct, iq_no, bytes_sent, datasize);
 611                 LIO_INCR_INSTRQUEUE_PKT_COUNT(oct, iq_no, instr_posted, 1);
 612
 613                 if (force_db || (st.status == LIO_IQ_SEND_STOP))
 614                         lio_ring_doorbell(oct, iq);
 615         } else {
 616                 LIO_INCR_INSTRQUEUE_PKT_COUNT(oct, iq_no, instr_dropped, 1);
 617         }
 618
 619         mtx_unlock(&iq->post_lock);
 620
 621         /*
 622          * This is only done here to expedite packets being flushed for
 623          * cases where there are no IQ completion interrupts.
 624          */
 625
 626         return (st.status);
 627 }
 628
 629 void
 630 lio_prepare_soft_command(struct octeon_device *oct, struct lio_soft_command *sc,
 631                          uint8_t opcode, uint8_t subcode, uint32_t irh_ossp,
 632                          uint64_t ossp0, uint64_t ossp1)
 633 {
 634         struct lio_config               *lio_cfg;
 635         struct octeon_instr_ih3         *ih3;
 636         struct octeon_instr_pki_ih3     *pki_ih3;
 637         struct octeon_instr_irh         *irh;
 638         struct octeon_instr_rdp         *rdp;
 639
 640         KASSERT(opcode <= 15, ("%s, %d, opcode > 15", __func__, __LINE__));
 641         KASSERT(subcode <= 127, ("%s, %d, opcode > 127", __func__, __LINE__));
 642
 643         lio_cfg = lio_get_conf(oct);
 644
 645         if (LIO_CN23XX_PF(oct)) {
 646                 ih3 = (struct octeon_instr_ih3 *)&sc->cmd.cmd3.ih3;
 647
 648                 ih3->pkind = oct->instr_queue[sc->iq_no]->txpciq.s.pkind;
 649
 650                 pki_ih3 = (struct octeon_instr_pki_ih3 *)&sc->cmd.cmd3.pki_ih3;
 651
 652                 pki_ih3->w = 1;
 653                 pki_ih3->raw = 1;
 654                 pki_ih3->utag = 1;
 655                 pki_ih3->uqpg = oct->instr_queue[sc->iq_no]->txpciq.s.use_qpg;
 656                 pki_ih3->utt = 1;
 657                 pki_ih3->tag = LIO_CONTROL;
 658                 pki_ih3->tagtype = LIO_ATOMIC_TAG;
 659                 pki_ih3->qpg = oct->instr_queue[sc->iq_no]->txpciq.s.qpg;
 660                 pki_ih3->pm = 0x7;
 661                 pki_ih3->sl = 8;
 662
 663                 if (sc->datasize)
 664                         ih3->dlengsz = sc->datasize;
 665
 666                 irh = (struct octeon_instr_irh *)&sc->cmd.cmd3.irh;
 667                 irh->opcode = opcode;
 668                 irh->subcode = subcode;
 669
 670                 /* opcode/subcode specific parameters (ossp) */
 671                 irh->ossp = irh_ossp;
 672                 sc->cmd.cmd3.ossp[0] = ossp0;
 673                 sc->cmd.cmd3.ossp[1] = ossp1;
 674
 675                 if (sc->rdatasize) {
 676                         rdp = (struct octeon_instr_rdp *)&sc->cmd.cmd3.rdp;
 677                         rdp->pcie_port = oct->pcie_port;
 678                         rdp->rlen = sc->rdatasize;
 679
 680                         irh->rflag = 1;
 681                         /* PKI IH3 */
 682                         /* pki_ih3 irh+ossp[0]+ossp[1]+rdp+rptr = 48 bytes */
 683                         ih3->fsz = LIO_SOFTCMDRESP_IH3;
 684                 } else {
 685                         irh->rflag = 0;
 686                         /* PKI IH3 */
 687                         /* pki_h3 + irh + ossp[0] + ossp[1] = 32 bytes */
 688                         ih3->fsz = LIO_PCICMD_O3;
 689                 }
 690         }
 691 }
 692
 693 int
 694 lio_send_soft_command(struct octeon_device *oct, struct lio_soft_command *sc)
 695 {
 696         struct octeon_instr_ih3 *ih3;
 697         struct octeon_instr_irh *irh;
 698         uint32_t                len = 0;
 699
 700         if (LIO_CN23XX_PF(oct)) {
 701                 ih3 = (struct octeon_instr_ih3 *)&sc->cmd.cmd3.ih3;
 702                 if (ih3->dlengsz) {
 703                         KASSERT(sc->dmadptr, ("%s, %d, sc->dmadptr is NULL",
 704                                               __func__, __LINE__));
 705                         sc->cmd.cmd3.dptr = sc->dmadptr;
 706                 }
 707
 708                 irh = (struct octeon_instr_irh *)&sc->cmd.cmd3.irh;
 709                 if (irh->rflag) {
 710                         KASSERT(sc->dmarptr, ("%s, %d, sc->dmarptr is NULL",
 711                                               __func__, __LINE__));
 712                         KASSERT(sc->status_word, ("%s, %d, sc->status_word is NULL",
 713                                                   __func__, __LINE__));
 714                         *sc->status_word = COMPLETION_WORD_INIT;
 715                         sc->cmd.cmd3.rptr = sc->dmarptr;
 716                 }
 717                 len = (uint32_t)ih3->dlengsz;
 718         }
 719         if (sc->wait_time)
 720                 sc->timeout = ticks + lio_ms_to_ticks(sc->wait_time);
 721
 722         return (lio_send_command(oct, sc->iq_no, 1, &sc->cmd, sc,
 723                                  len, LIO_REQTYPE_SOFT_COMMAND));
 724 }
 725
 726 int
 727 lio_setup_sc_buffer_pool(struct octeon_device *oct)
 728 {
 729         struct lio_soft_command *sc;
 730         uint64_t                dma_addr;
 731         int                     i;
 732
 733         STAILQ_INIT(&oct->sc_buf_pool.head);
 734         mtx_init(&oct->sc_buf_pool.lock, "sc_pool_lock", NULL, MTX_DEF);
 735         atomic_store_rel_int(&oct->sc_buf_pool.alloc_buf_count, 0);
 736
 737         for (i = 0; i < LIO_MAX_SOFT_COMMAND_BUFFERS; i++) {
 738                 sc = (struct lio_soft_command *)
 739                         lio_dma_alloc(LIO_SOFT_COMMAND_BUFFER_SIZE, &dma_addr);
 740                 if (sc == NULL) {
 741                         lio_free_sc_buffer_pool(oct);
 742                         return (1);
 743                 }
 744
 745                 sc->dma_addr = dma_addr;
 746                 sc->size = LIO_SOFT_COMMAND_BUFFER_SIZE;
 747
 748                 STAILQ_INSERT_TAIL(&oct->sc_buf_pool.head, &sc->node, entries);
 749         }
 750
 751         return (0);
 752 }
 753
 754 int
 755 lio_free_sc_buffer_pool(struct octeon_device *oct)
 756 {
 757         struct lio_stailq_node  *tmp, *tmp2;
 758         struct lio_soft_command *sc;
 759
 760         mtx_lock(&oct->sc_buf_pool.lock);
 761
 762         STAILQ_FOREACH_SAFE(tmp, &oct->sc_buf_pool.head, entries, tmp2) {
 763                 sc = LIO_STAILQ_FIRST_ENTRY(&oct->sc_buf_pool.head,
 764                                             struct lio_soft_command, node);
 765
 766                 STAILQ_REMOVE_HEAD(&oct->sc_buf_pool.head, entries);
 767
 768                 lio_dma_free(sc->size, sc);
 769         }
 770
 771         STAILQ_INIT(&oct->sc_buf_pool.head);
 772
 773         mtx_unlock(&oct->sc_buf_pool.lock);
 774
 775         return (0);
 776 }
 777
 778 struct lio_soft_command *
 779 lio_alloc_soft_command(struct octeon_device *oct, uint32_t datasize,
 780                        uint32_t rdatasize, uint32_t ctxsize)
 781 {
 782         struct lio_soft_command *sc = NULL;
 783         struct lio_stailq_node  *tmp;
 784         uint64_t                dma_addr;
 785         uint32_t                size;
 786         uint32_t                offset = sizeof(struct lio_soft_command);
 787
 788         KASSERT((offset + datasize + rdatasize + ctxsize) <=
 789                 LIO_SOFT_COMMAND_BUFFER_SIZE,
 790                 ("%s, %d, offset + datasize + rdatasize + ctxsize > LIO_SOFT_COMMAND_BUFFER_SIZE",
 791                  __func__, __LINE__));
 792
 793         mtx_lock(&oct->sc_buf_pool.lock);
 794
 795         if (STAILQ_EMPTY(&oct->sc_buf_pool.head)) {
 796                 mtx_unlock(&oct->sc_buf_pool.lock);
 797                 return (NULL);
 798         }
 799         tmp = STAILQ_LAST(&oct->sc_buf_pool.head, lio_stailq_node, entries);
 800
 801         STAILQ_REMOVE(&oct->sc_buf_pool.head, tmp, lio_stailq_node, entries);
 802
 803         atomic_add_int(&oct->sc_buf_pool.alloc_buf_count, 1);
 804
 805         mtx_unlock(&oct->sc_buf_pool.lock);
 806
 807         sc = (struct lio_soft_command *)tmp;
 808
 809         dma_addr = sc->dma_addr;
 810         size = sc->size;
 811
 812         bzero(sc, sc->size);
 813
 814         sc->dma_addr = dma_addr;
 815         sc->size = size;
 816
 817         if (ctxsize) {
 818                 sc->ctxptr = (uint8_t *)sc + offset;
 819                 sc->ctxsize = ctxsize;
 820         }
 821
 822         /* Start data at 128 byte boundary */
 823         offset = (offset + ctxsize + 127) & 0xffffff80;
 824
 825         if (datasize) {
 826                 sc->virtdptr = (uint8_t *)sc + offset;
 827                 sc->dmadptr = dma_addr + offset;
 828                 sc->datasize = datasize;
 829         }
 830         /* Start rdata at 128 byte boundary */
 831         offset = (offset + datasize + 127) & 0xffffff80;
 832
 833         if (rdatasize) {
 834                 KASSERT(rdatasize >= 16, ("%s, %d, rdatasize < 16", __func__,
 835                                           __LINE__));
 836                 sc->virtrptr = (uint8_t *)sc + offset;
 837                 sc->dmarptr = dma_addr + offset;
 838                 sc->rdatasize = rdatasize;
 839                 sc->status_word = (uint64_t *)((uint8_t *)(sc->virtrptr) +
 840                                                rdatasize - 8);
 841         }
 842         return (sc);
 843 }
 844
 845 void
 846 lio_free_soft_command(struct octeon_device *oct,
 847                       struct lio_soft_command *sc)
 848 {
 849
 850         mtx_lock(&oct->sc_buf_pool.lock);
 851
 852         STAILQ_INSERT_TAIL(&oct->sc_buf_pool.head, &sc->node, entries);
 853
 854         atomic_subtract_int(&oct->sc_buf_pool.alloc_buf_count, 1);
 855
 856         mtx_unlock(&oct->sc_buf_pool.lock);
 857 }