drivers/net/ethernet/mellanox/mlx5/core/eq.c

   1 /*
   2  * Copyright (c) 2013-2015, Mellanox Technologies. All rights reserved.
   3  *
   4  * This software is available to you under a choice of one of two
   5  * licenses.  You may choose to be licensed under the terms of the GNU
   6  * General Public License (GPL) Version 2, available from the file
   7  * COPYING in the main directory of this source tree, or the
   8  * OpenIB.org BSD license below:
   9  *
  10  *     Redistribution and use in source and binary forms, with or
  11  *     without modification, are permitted provided that the following
  12  *     conditions are met:
  13  *
  14  *      - Redistributions of source code must retain the above
  15  *        copyright notice, this list of conditions and the following
  16  *        disclaimer.
  17  *
  18  *      - Redistributions in binary form must reproduce the above
  19  *        copyright notice, this list of conditions and the following
  20  *        disclaimer in the documentation and/or other materials
  21  *        provided with the distribution.
  22  *
  23  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  24  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  25  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  26  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  27  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  28  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  29  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  30  * SOFTWARE.
  31  */
  32
  33 #include <linux/interrupt.h>
  34 #include <linux/module.h>
  35 #include <linux/mlx5/driver.h>
  36 #include <linux/mlx5/cmd.h>
  37 #include "mlx5_core.h"
  38 #include "fpga/core.h"
  39 #ifdef CONFIG_MLX5_CORE_EN
  40 #include "eswitch.h"
  41 #endif
  42
  43 enum {
  44         MLX5_EQE_SIZE           = sizeof(struct mlx5_eqe),
  45         MLX5_EQE_OWNER_INIT_VAL = 0x1,
  46 };
  47
  48 enum {
  49         MLX5_EQ_STATE_ARMED             = 0x9,
  50         MLX5_EQ_STATE_FIRED             = 0xa,
  51         MLX5_EQ_STATE_ALWAYS_ARMED      = 0xb,
  52 };
  53
  54 enum {
  55         MLX5_NUM_SPARE_EQE      = 0x80,
  56         MLX5_NUM_ASYNC_EQE      = 0x100,
  57         MLX5_NUM_CMD_EQE        = 32,
  58         MLX5_NUM_PF_DRAIN       = 64,
  59 };
  60
  61 enum {
  62         MLX5_EQ_DOORBEL_OFFSET  = 0x40,
  63 };
  64
  65 #define MLX5_ASYNC_EVENT_MASK ((1ull << MLX5_EVENT_TYPE_PATH_MIG)           | \
  66                                (1ull << MLX5_EVENT_TYPE_COMM_EST)           | \
  67                                (1ull << MLX5_EVENT_TYPE_SQ_DRAINED)         | \
  68                                (1ull << MLX5_EVENT_TYPE_CQ_ERROR)           | \
  69                                (1ull << MLX5_EVENT_TYPE_WQ_CATAS_ERROR)     | \
  70                                (1ull << MLX5_EVENT_TYPE_PATH_MIG_FAILED)    | \
  71                                (1ull << MLX5_EVENT_TYPE_WQ_INVAL_REQ_ERROR) | \
  72                                (1ull << MLX5_EVENT_TYPE_WQ_ACCESS_ERROR)    | \
  73                                (1ull << MLX5_EVENT_TYPE_PORT_CHANGE)        | \
  74                                (1ull << MLX5_EVENT_TYPE_SRQ_CATAS_ERROR)    | \
  75                                (1ull << MLX5_EVENT_TYPE_SRQ_LAST_WQE)       | \
  76                                (1ull << MLX5_EVENT_TYPE_SRQ_RQ_LIMIT))
  77
  78 struct map_eq_in {
  79         u64     mask;
  80         u32     reserved;
  81         u32     unmap_eqn;
  82 };
  83
  84 struct cre_des_eq {
  85         u8      reserved[15];
  86         u8      eqn;
  87 };
  88
  89 static int mlx5_cmd_destroy_eq(struct mlx5_core_dev *dev, u8 eqn)
  90 {
  91         u32 out[MLX5_ST_SZ_DW(destroy_eq_out)] = {0};
  92         u32 in[MLX5_ST_SZ_DW(destroy_eq_in)]   = {0};
  93
  94         MLX5_SET(destroy_eq_in, in, opcode, MLX5_CMD_OP_DESTROY_EQ);
  95         MLX5_SET(destroy_eq_in, in, eq_number, eqn);
  96         return mlx5_cmd_exec(dev, in, sizeof(in), out, sizeof(out));
  97 }
  98
  99 static struct mlx5_eqe *get_eqe(struct mlx5_eq *eq, u32 entry)
 100 {
 101         return mlx5_buf_offset(&eq->buf, entry * MLX5_EQE_SIZE);
 102 }
 103
 104 static struct mlx5_eqe *next_eqe_sw(struct mlx5_eq *eq)
 105 {
 106         struct mlx5_eqe *eqe = get_eqe(eq, eq->cons_index & (eq->nent - 1));
 107
 108         return ((eqe->owner & 1) ^ !!(eq->cons_index & eq->nent)) ? NULL : eqe;
 109 }
 110
 111 static const char *eqe_type_str(u8 type)
 112 {
 113         switch (type) {
 114         case MLX5_EVENT_TYPE_COMP:
 115                 return "MLX5_EVENT_TYPE_COMP";
 116         case MLX5_EVENT_TYPE_PATH_MIG:
 117                 return "MLX5_EVENT_TYPE_PATH_MIG";
 118         case MLX5_EVENT_TYPE_COMM_EST:
 119                 return "MLX5_EVENT_TYPE_COMM_EST";
 120         case MLX5_EVENT_TYPE_SQ_DRAINED:
 121                 return "MLX5_EVENT_TYPE_SQ_DRAINED";
 122         case MLX5_EVENT_TYPE_SRQ_LAST_WQE:
 123                 return "MLX5_EVENT_TYPE_SRQ_LAST_WQE";
 124         case MLX5_EVENT_TYPE_SRQ_RQ_LIMIT:
 125                 return "MLX5_EVENT_TYPE_SRQ_RQ_LIMIT";
 126         case MLX5_EVENT_TYPE_CQ_ERROR:
 127                 return "MLX5_EVENT_TYPE_CQ_ERROR";
 128         case MLX5_EVENT_TYPE_WQ_CATAS_ERROR:
 129                 return "MLX5_EVENT_TYPE_WQ_CATAS_ERROR";
 130         case MLX5_EVENT_TYPE_PATH_MIG_FAILED:
 131                 return "MLX5_EVENT_TYPE_PATH_MIG_FAILED";
 132         case MLX5_EVENT_TYPE_WQ_INVAL_REQ_ERROR:
 133                 return "MLX5_EVENT_TYPE_WQ_INVAL_REQ_ERROR";
 134         case MLX5_EVENT_TYPE_WQ_ACCESS_ERROR:
 135                 return "MLX5_EVENT_TYPE_WQ_ACCESS_ERROR";
 136         case MLX5_EVENT_TYPE_SRQ_CATAS_ERROR:
 137                 return "MLX5_EVENT_TYPE_SRQ_CATAS_ERROR";
 138         case MLX5_EVENT_TYPE_INTERNAL_ERROR:
 139                 return "MLX5_EVENT_TYPE_INTERNAL_ERROR";
 140         case MLX5_EVENT_TYPE_PORT_CHANGE:
 141                 return "MLX5_EVENT_TYPE_PORT_CHANGE";
 142         case MLX5_EVENT_TYPE_GPIO_EVENT:
 143                 return "MLX5_EVENT_TYPE_GPIO_EVENT";
 144         case MLX5_EVENT_TYPE_PORT_MODULE_EVENT:
 145                 return "MLX5_EVENT_TYPE_PORT_MODULE_EVENT";
 146         case MLX5_EVENT_TYPE_REMOTE_CONFIG:
 147                 return "MLX5_EVENT_TYPE_REMOTE_CONFIG";
 148         case MLX5_EVENT_TYPE_DB_BF_CONGESTION:
 149                 return "MLX5_EVENT_TYPE_DB_BF_CONGESTION";
 150         case MLX5_EVENT_TYPE_STALL_EVENT:
 151                 return "MLX5_EVENT_TYPE_STALL_EVENT";
 152         case MLX5_EVENT_TYPE_CMD:
 153                 return "MLX5_EVENT_TYPE_CMD";
 154         case MLX5_EVENT_TYPE_PAGE_REQUEST:
 155                 return "MLX5_EVENT_TYPE_PAGE_REQUEST";
 156         case MLX5_EVENT_TYPE_PAGE_FAULT:
 157                 return "MLX5_EVENT_TYPE_PAGE_FAULT";
 158         case MLX5_EVENT_TYPE_PPS_EVENT:
 159                 return "MLX5_EVENT_TYPE_PPS_EVENT";
 160         case MLX5_EVENT_TYPE_NIC_VPORT_CHANGE:
 161                 return "MLX5_EVENT_TYPE_NIC_VPORT_CHANGE";
 162         case MLX5_EVENT_TYPE_FPGA_ERROR:
 163                 return "MLX5_EVENT_TYPE_FPGA_ERROR";
 164         default:
 165                 return "Unrecognized event";
 166         }
 167 }
 168
 169 static enum mlx5_dev_event port_subtype_event(u8 subtype)
 170 {
 171         switch (subtype) {
 172         case MLX5_PORT_CHANGE_SUBTYPE_DOWN:
 173                 return MLX5_DEV_EVENT_PORT_DOWN;
 174         case MLX5_PORT_CHANGE_SUBTYPE_ACTIVE:
 175                 return MLX5_DEV_EVENT_PORT_UP;
 176         case MLX5_PORT_CHANGE_SUBTYPE_INITIALIZED:
 177                 return MLX5_DEV_EVENT_PORT_INITIALIZED;
 178         case MLX5_PORT_CHANGE_SUBTYPE_LID:
 179                 return MLX5_DEV_EVENT_LID_CHANGE;
 180         case MLX5_PORT_CHANGE_SUBTYPE_PKEY:
 181                 return MLX5_DEV_EVENT_PKEY_CHANGE;
 182         case MLX5_PORT_CHANGE_SUBTYPE_GUID:
 183                 return MLX5_DEV_EVENT_GUID_CHANGE;
 184         case MLX5_PORT_CHANGE_SUBTYPE_CLIENT_REREG:
 185                 return MLX5_DEV_EVENT_CLIENT_REREG;
 186         }
 187         return -1;
 188 }
 189
 190 static void eq_update_ci(struct mlx5_eq *eq, int arm)
 191 {
 192         __be32 __iomem *addr = eq->doorbell + (arm ? 0 : 2);
 193         u32 val = (eq->cons_index & 0xffffff) | (eq->eqn << 24);
 194         __raw_writel((__force u32)cpu_to_be32(val), addr);
 195         /* We still want ordering, just not swabbing, so add a barrier */
 196         mb();
 197 }
 198
 199 #ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
 200 static void eqe_pf_action(struct work_struct *work)
 201 {
 202         struct mlx5_pagefault *pfault = container_of(work,
 203                                                      struct mlx5_pagefault,
 204                                                      work);
 205         struct mlx5_eq *eq = pfault->eq;
 206
 207         mlx5_core_page_fault(eq->dev, pfault);
 208         mempool_free(pfault, eq->pf_ctx.pool);
 209 }
 210
 211 static void eq_pf_process(struct mlx5_eq *eq)
 212 {
 213         struct mlx5_core_dev *dev = eq->dev;
 214         struct mlx5_eqe_page_fault *pf_eqe;
 215         struct mlx5_pagefault *pfault;
 216         struct mlx5_eqe *eqe;
 217         int set_ci = 0;
 218
 219         while ((eqe = next_eqe_sw(eq))) {
 220                 pfault = mempool_alloc(eq->pf_ctx.pool, GFP_ATOMIC);
 221                 if (!pfault) {
 222                         schedule_work(&eq->pf_ctx.work);
 223                         break;
 224                 }
 225
 226                 dma_rmb();
 227                 pf_eqe = &eqe->data.page_fault;
 228                 pfault->event_subtype = eqe->sub_type;
 229                 pfault->bytes_committed = be32_to_cpu(pf_eqe->bytes_committed);
 230
 231                 mlx5_core_dbg(dev,
 232                               "PAGE_FAULT: subtype: 0x%02x, bytes_committed: 0x%06x\n",
 233                               eqe->sub_type, pfault->bytes_committed);
 234
 235                 switch (eqe->sub_type) {
 236                 case MLX5_PFAULT_SUBTYPE_RDMA:
 237                         /* RDMA based event */
 238                         pfault->type =
 239                                 be32_to_cpu(pf_eqe->rdma.pftype_token) >> 24;
 240                         pfault->token =
 241                                 be32_to_cpu(pf_eqe->rdma.pftype_token) &
 242                                 MLX5_24BIT_MASK;
 243                         pfault->rdma.r_key =
 244                                 be32_to_cpu(pf_eqe->rdma.r_key);
 245                         pfault->rdma.packet_size =
 246                                 be16_to_cpu(pf_eqe->rdma.packet_length);
 247                         pfault->rdma.rdma_op_len =
 248                                 be32_to_cpu(pf_eqe->rdma.rdma_op_len);
 249                         pfault->rdma.rdma_va =
 250                                 be64_to_cpu(pf_eqe->rdma.rdma_va);
 251                         mlx5_core_dbg(dev,
 252                                       "PAGE_FAULT: type:0x%x, token: 0x%06x, r_key: 0x%08x\n",
 253                                       pfault->type, pfault->token,
 254                                       pfault->rdma.r_key);
 255                         mlx5_core_dbg(dev,
 256                                       "PAGE_FAULT: rdma_op_len: 0x%08x, rdma_va: 0x%016llx\n",
 257                                       pfault->rdma.rdma_op_len,
 258                                       pfault->rdma.rdma_va);
 259                         break;
 260
 261                 case MLX5_PFAULT_SUBTYPE_WQE:
 262                         /* WQE based event */
 263                         pfault->type =
 264                                 be32_to_cpu(pf_eqe->wqe.pftype_wq) >> 24;
 265                         pfault->token =
 266                                 be32_to_cpu(pf_eqe->wqe.token);
 267                         pfault->wqe.wq_num =
 268                                 be32_to_cpu(pf_eqe->wqe.pftype_wq) &
 269                                 MLX5_24BIT_MASK;
 270                         pfault->wqe.wqe_index =
 271                                 be16_to_cpu(pf_eqe->wqe.wqe_index);
 272                         pfault->wqe.packet_size =
 273                                 be16_to_cpu(pf_eqe->wqe.packet_length);
 274                         mlx5_core_dbg(dev,
 275                                       "PAGE_FAULT: type:0x%x, token: 0x%06x, wq_num: 0x%06x, wqe_index: 0x%04x\n",
 276                                       pfault->type, pfault->token,
 277                                       pfault->wqe.wq_num,
 278                                       pfault->wqe.wqe_index);
 279                         break;
 280
 281                 default:
 282                         mlx5_core_warn(dev,
 283                                        "Unsupported page fault event sub-type: 0x%02hhx\n",
 284                                        eqe->sub_type);
 285                         /* Unsupported page faults should still be
 286                          * resolved by the page fault handler
 287                          */
 288                 }
 289
 290                 pfault->eq = eq;
 291                 INIT_WORK(&pfault->work, eqe_pf_action);
 292                 queue_work(eq->pf_ctx.wq, &pfault->work);
 293
 294                 ++eq->cons_index;
 295                 ++set_ci;
 296
 297                 if (unlikely(set_ci >= MLX5_NUM_SPARE_EQE)) {
 298                         eq_update_ci(eq, 0);
 299                         set_ci = 0;
 300                 }
 301         }
 302
 303         eq_update_ci(eq, 1);
 304 }
 305
 306 static irqreturn_t mlx5_eq_pf_int(int irq, void *eq_ptr)
 307 {
 308         struct mlx5_eq *eq = eq_ptr;
 309         unsigned long flags;
 310
 311         if (spin_trylock_irqsave(&eq->pf_ctx.lock, flags)) {
 312                 eq_pf_process(eq);
 313                 spin_unlock_irqrestore(&eq->pf_ctx.lock, flags);
 314         } else {
 315                 schedule_work(&eq->pf_ctx.work);
 316         }
 317
 318         return IRQ_HANDLED;
 319 }
 320
 321 /* mempool_refill() was proposed but unfortunately wasn't accepted
 322  * http://lkml.iu.edu/hypermail/linux/kernel/1512.1/05073.html
 323  * Chip workaround.
 324  */
 325 static void mempool_refill(mempool_t *pool)
 326 {
 327         while (pool->curr_nr < pool->min_nr)
 328                 mempool_free(mempool_alloc(pool, GFP_KERNEL), pool);
 329 }
 330
 331 static void eq_pf_action(struct work_struct *work)
 332 {
 333         struct mlx5_eq *eq = container_of(work, struct mlx5_eq, pf_ctx.work);
 334
 335         mempool_refill(eq->pf_ctx.pool);
 336
 337         spin_lock_irq(&eq->pf_ctx.lock);
 338         eq_pf_process(eq);
 339         spin_unlock_irq(&eq->pf_ctx.lock);
 340 }
 341
 342 static int init_pf_ctx(struct mlx5_eq_pagefault *pf_ctx, const char *name)
 343 {
 344         spin_lock_init(&pf_ctx->lock);
 345         INIT_WORK(&pf_ctx->work, eq_pf_action);
 346
 347         pf_ctx->wq = alloc_ordered_workqueue(name,
 348                                              WQ_MEM_RECLAIM);
 349         if (!pf_ctx->wq)
 350                 return -ENOMEM;
 351
 352         pf_ctx->pool = mempool_create_kmalloc_pool
 353                 (MLX5_NUM_PF_DRAIN, sizeof(struct mlx5_pagefault));
 354         if (!pf_ctx->pool)
 355                 goto err_wq;
 356
 357         return 0;
 358 err_wq:
 359         destroy_workqueue(pf_ctx->wq);
 360         return -ENOMEM;
 361 }
 362
 363 int mlx5_core_page_fault_resume(struct mlx5_core_dev *dev, u32 token,
 364                                 u32 wq_num, u8 type, int error)
 365 {
 366         u32 out[MLX5_ST_SZ_DW(page_fault_resume_out)] = {0};
 367         u32 in[MLX5_ST_SZ_DW(page_fault_resume_in)]   = {0};
 368
 369         MLX5_SET(page_fault_resume_in, in, opcode,
 370                  MLX5_CMD_OP_PAGE_FAULT_RESUME);
 371         MLX5_SET(page_fault_resume_in, in, error, !!error);
 372         MLX5_SET(page_fault_resume_in, in, page_fault_type, type);
 373         MLX5_SET(page_fault_resume_in, in, wq_number, wq_num);
 374         MLX5_SET(page_fault_resume_in, in, token, token);
 375
 376         return mlx5_cmd_exec(dev, in, sizeof(in), out, sizeof(out));
 377 }
 378 EXPORT_SYMBOL_GPL(mlx5_core_page_fault_resume);
 379 #endif
 380
 381 static irqreturn_t mlx5_eq_int(int irq, void *eq_ptr)
 382 {
 383         struct mlx5_eq *eq = eq_ptr;
 384         struct mlx5_core_dev *dev = eq->dev;
 385         struct mlx5_eqe *eqe;
 386         int set_ci = 0;
 387         u32 cqn = -1;
 388         u32 rsn;
 389         u8 port;
 390
 391         while ((eqe = next_eqe_sw(eq))) {
 392                 /*
 393                  * Make sure we read EQ entry contents after we've
 394                  * checked the ownership bit.
 395                  */
 396                 dma_rmb();
 397
 398                 mlx5_core_dbg(eq->dev, "eqn %d, eqe type %s\n",
 399                               eq->eqn, eqe_type_str(eqe->type));
 400                 switch (eqe->type) {
 401                 case MLX5_EVENT_TYPE_COMP:
 402                         cqn = be32_to_cpu(eqe->data.comp.cqn) & 0xffffff;
 403                         mlx5_cq_completion(dev, cqn);
 404                         break;
 405
 406                 case MLX5_EVENT_TYPE_PATH_MIG:
 407                 case MLX5_EVENT_TYPE_COMM_EST:
 408                 case MLX5_EVENT_TYPE_SQ_DRAINED:
 409                 case MLX5_EVENT_TYPE_SRQ_LAST_WQE:
 410                 case MLX5_EVENT_TYPE_WQ_CATAS_ERROR:
 411                 case MLX5_EVENT_TYPE_PATH_MIG_FAILED:
 412                 case MLX5_EVENT_TYPE_WQ_INVAL_REQ_ERROR:
 413                 case MLX5_EVENT_TYPE_WQ_ACCESS_ERROR:
 414                         rsn = be32_to_cpu(eqe->data.qp_srq.qp_srq_n) & 0xffffff;
 415                         rsn |= (eqe->data.qp_srq.type << MLX5_USER_INDEX_LEN);
 416                         mlx5_core_dbg(dev, "event %s(%d) arrived on resource 0x%x\n",
 417                                       eqe_type_str(eqe->type), eqe->type, rsn);
 418                         mlx5_rsc_event(dev, rsn, eqe->type);
 419                         break;
 420
 421                 case MLX5_EVENT_TYPE_SRQ_RQ_LIMIT:
 422                 case MLX5_EVENT_TYPE_SRQ_CATAS_ERROR:
 423                         rsn = be32_to_cpu(eqe->data.qp_srq.qp_srq_n) & 0xffffff;
 424                         mlx5_core_dbg(dev, "SRQ event %s(%d): srqn 0x%x\n",
 425                                       eqe_type_str(eqe->type), eqe->type, rsn);
 426                         mlx5_srq_event(dev, rsn, eqe->type);
 427                         break;
 428
 429                 case MLX5_EVENT_TYPE_CMD:
 430                         mlx5_cmd_comp_handler(dev, be32_to_cpu(eqe->data.cmd.vector), false);
 431                         break;
 432
 433                 case MLX5_EVENT_TYPE_PORT_CHANGE:
 434                         port = (eqe->data.port.port >> 4) & 0xf;
 435                         switch (eqe->sub_type) {
 436                         case MLX5_PORT_CHANGE_SUBTYPE_DOWN:
 437                         case MLX5_PORT_CHANGE_SUBTYPE_ACTIVE:
 438                         case MLX5_PORT_CHANGE_SUBTYPE_LID:
 439                         case MLX5_PORT_CHANGE_SUBTYPE_PKEY:
 440                         case MLX5_PORT_CHANGE_SUBTYPE_GUID:
 441                         case MLX5_PORT_CHANGE_SUBTYPE_CLIENT_REREG:
 442                         case MLX5_PORT_CHANGE_SUBTYPE_INITIALIZED:
 443                                 if (dev->event)
 444                                         dev->event(dev, port_subtype_event(eqe->sub_type),
 445                                                    (unsigned long)port);
 446                                 break;
 447                         default:
 448                                 mlx5_core_warn(dev, "Port event with unrecognized subtype: port %d, sub_type %d\n",
 449                                                port, eqe->sub_type);
 450                         }
 451                         break;
 452                 case MLX5_EVENT_TYPE_CQ_ERROR:
 453                         cqn = be32_to_cpu(eqe->data.cq_err.cqn) & 0xffffff;
 454                         mlx5_core_warn(dev, "CQ error on CQN 0x%x, syndrom 0x%x\n",
 455                                        cqn, eqe->data.cq_err.syndrome);
 456                         mlx5_cq_event(dev, cqn, eqe->type);
 457                         break;
 458
 459                 case MLX5_EVENT_TYPE_PAGE_REQUEST:
 460                         {
 461                                 u16 func_id = be16_to_cpu(eqe->data.req_pages.func_id);
 462                                 s32 npages = be32_to_cpu(eqe->data.req_pages.num_pages);
 463
 464                                 mlx5_core_dbg(dev, "page request for func 0x%x, npages %d\n",
 465                                               func_id, npages);
 466                                 mlx5_core_req_pages_handler(dev, func_id, npages);
 467                         }
 468                         break;
 469
 470 #ifdef CONFIG_MLX5_CORE_EN
 471                 case MLX5_EVENT_TYPE_NIC_VPORT_CHANGE:
 472                         mlx5_eswitch_vport_event(dev->priv.eswitch, eqe);
 473                         break;
 474 #endif
 475
 476                 case MLX5_EVENT_TYPE_PORT_MODULE_EVENT:
 477                         mlx5_port_module_event(dev, eqe);
 478                         break;
 479
 480                 case MLX5_EVENT_TYPE_PPS_EVENT:
 481                         if (dev->event)
 482                                 dev->event(dev, MLX5_DEV_EVENT_PPS, (unsigned long)eqe);
 483                         break;
 484
 485                 case MLX5_EVENT_TYPE_FPGA_ERROR:
 486                         mlx5_fpga_event(dev, eqe->type, &eqe->data.raw);
 487                         break;
 488
 489                 default:
 490                         mlx5_core_warn(dev, "Unhandled event 0x%x on EQ 0x%x\n",
 491                                        eqe->type, eq->eqn);
 492                         break;
 493                 }
 494
 495                 ++eq->cons_index;
 496                 ++set_ci;
 497
 498                 /* The HCA will think the queue has overflowed if we
 499                  * don't tell it we've been processing events.  We
 500                  * create our EQs with MLX5_NUM_SPARE_EQE extra
 501                  * entries, so we must update our consumer index at
 502                  * least that often.
 503                  */
 504                 if (unlikely(set_ci >= MLX5_NUM_SPARE_EQE)) {
 505                         eq_update_ci(eq, 0);
 506                         set_ci = 0;
 507                 }
 508         }
 509
 510         eq_update_ci(eq, 1);
 511
 512         if (cqn != -1)
 513                 tasklet_schedule(&eq->tasklet_ctx.task);
 514
 515         return IRQ_HANDLED;
 516 }
 517
 518 static void init_eq_buf(struct mlx5_eq *eq)
 519 {
 520         struct mlx5_eqe *eqe;
 521         int i;
 522
 523         for (i = 0; i < eq->nent; i++) {
 524                 eqe = get_eqe(eq, i);
 525                 eqe->owner = MLX5_EQE_OWNER_INIT_VAL;
 526         }
 527 }
 528
 529 int mlx5_create_map_eq(struct mlx5_core_dev *dev, struct mlx5_eq *eq, u8 vecidx,
 530                        int nent, u64 mask, const char *name,
 531                        enum mlx5_eq_type type)
 532 {
 533         u32 out[MLX5_ST_SZ_DW(create_eq_out)] = {0};
 534         struct mlx5_priv *priv = &dev->priv;
 535         irq_handler_t handler;
 536         __be64 *pas;
 537         void *eqc;
 538         int inlen;
 539         u32 *in;
 540         int err;
 541
 542         eq->type = type;
 543         eq->nent = roundup_pow_of_two(nent + MLX5_NUM_SPARE_EQE);
 544         eq->cons_index = 0;
 545         err = mlx5_buf_alloc(dev, eq->nent * MLX5_EQE_SIZE, &eq->buf);
 546         if (err)
 547                 return err;
 548
 549 #ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
 550         if (type == MLX5_EQ_TYPE_PF)
 551                 handler = mlx5_eq_pf_int;
 552         else
 553 #endif
 554                 handler = mlx5_eq_int;
 555
 556         init_eq_buf(eq);
 557
 558         inlen = MLX5_ST_SZ_BYTES(create_eq_in) +
 559                 MLX5_FLD_SZ_BYTES(create_eq_in, pas[0]) * eq->buf.npages;
 560
 561         in = kvzalloc(inlen, GFP_KERNEL);
 562         if (!in) {
 563                 err = -ENOMEM;
 564                 goto err_buf;
 565         }
 566
 567         pas = (__be64 *)MLX5_ADDR_OF(create_eq_in, in, pas);
 568         mlx5_fill_page_array(&eq->buf, pas);
 569
 570         MLX5_SET(create_eq_in, in, opcode, MLX5_CMD_OP_CREATE_EQ);
 571         MLX5_SET64(create_eq_in, in, event_bitmask, mask);
 572
 573         eqc = MLX5_ADDR_OF(create_eq_in, in, eq_context_entry);
 574         MLX5_SET(eqc, eqc, log_eq_size, ilog2(eq->nent));
 575         MLX5_SET(eqc, eqc, uar_page, priv->uar->index);
 576         MLX5_SET(eqc, eqc, intr, vecidx);
 577         MLX5_SET(eqc, eqc, log_page_size,
 578                  eq->buf.page_shift - MLX5_ADAPTER_PAGE_SHIFT);
 579
 580         err = mlx5_cmd_exec(dev, in, inlen, out, sizeof(out));
 581         if (err)
 582                 goto err_in;
 583
 584         snprintf(priv->irq_info[vecidx].name, MLX5_MAX_IRQ_NAME, "%s@pci:%s",
 585                  name, pci_name(dev->pdev));
 586
 587         eq->eqn = MLX5_GET(create_eq_out, out, eq_number);
 588         eq->irqn = priv->msix_arr[vecidx].vector;
 589         eq->dev = dev;
 590         eq->doorbell = priv->uar->map + MLX5_EQ_DOORBEL_OFFSET;
 591         err = request_irq(eq->irqn, handler, 0,
 592                           priv->irq_info[vecidx].name, eq);
 593         if (err)
 594                 goto err_eq;
 595
 596         err = mlx5_debug_eq_add(dev, eq);
 597         if (err)
 598                 goto err_irq;
 599
 600 #ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
 601         if (type == MLX5_EQ_TYPE_PF) {
 602                 err = init_pf_ctx(&eq->pf_ctx, name);
 603                 if (err)
 604                         goto err_irq;
 605         } else
 606 #endif
 607         {
 608                 INIT_LIST_HEAD(&eq->tasklet_ctx.list);
 609                 INIT_LIST_HEAD(&eq->tasklet_ctx.process_list);
 610                 spin_lock_init(&eq->tasklet_ctx.lock);
 611                 tasklet_init(&eq->tasklet_ctx.task, mlx5_cq_tasklet_cb,
 612                              (unsigned long)&eq->tasklet_ctx);
 613         }
 614
 615         /* EQs are created in ARMED state
 616          */
 617         eq_update_ci(eq, 1);
 618
 619         kvfree(in);
 620         return 0;
 621
 622 err_irq:
 623         free_irq(priv->msix_arr[vecidx].vector, eq);
 624
 625 err_eq:
 626         mlx5_cmd_destroy_eq(dev, eq->eqn);
 627
 628 err_in:
 629         kvfree(in);
 630
 631 err_buf:
 632         mlx5_buf_free(dev, &eq->buf);
 633         return err;
 634 }
 635 EXPORT_SYMBOL_GPL(mlx5_create_map_eq);
 636
 637 int mlx5_destroy_unmap_eq(struct mlx5_core_dev *dev, struct mlx5_eq *eq)
 638 {
 639         int err;
 640
 641         mlx5_debug_eq_remove(dev, eq);
 642         free_irq(eq->irqn, eq);
 643         err = mlx5_cmd_destroy_eq(dev, eq->eqn);
 644         if (err)
 645                 mlx5_core_warn(dev, "failed to destroy a previously created eq: eqn %d\n",
 646                                eq->eqn);
 647         synchronize_irq(eq->irqn);
 648
 649         if (eq->type == MLX5_EQ_TYPE_COMP) {
 650                 tasklet_disable(&eq->tasklet_ctx.task);
 651 #ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
 652         } else if (eq->type == MLX5_EQ_TYPE_PF) {
 653                 cancel_work_sync(&eq->pf_ctx.work);
 654                 destroy_workqueue(eq->pf_ctx.wq);
 655                 mempool_destroy(eq->pf_ctx.pool);
 656 #endif
 657         }
 658         mlx5_buf_free(dev, &eq->buf);
 659
 660         return err;
 661 }
 662 EXPORT_SYMBOL_GPL(mlx5_destroy_unmap_eq);
 663
 664 u32 mlx5_get_msix_vec(struct mlx5_core_dev *dev, int vecidx)
 665 {
 666         return dev->priv.msix_arr[MLX5_EQ_VEC_ASYNC].vector;
 667 }
 668
 669 int mlx5_eq_init(struct mlx5_core_dev *dev)
 670 {
 671         int err;
 672
 673         spin_lock_init(&dev->priv.eq_table.lock);
 674
 675         err = mlx5_eq_debugfs_init(dev);
 676
 677         return err;
 678 }
 679
 680 void mlx5_eq_cleanup(struct mlx5_core_dev *dev)
 681 {
 682         mlx5_eq_debugfs_cleanup(dev);
 683 }
 684
 685 int mlx5_start_eqs(struct mlx5_core_dev *dev)
 686 {
 687         struct mlx5_eq_table *table = &dev->priv.eq_table;
 688         u64 async_event_mask = MLX5_ASYNC_EVENT_MASK;
 689         int err;
 690
 691         if (MLX5_CAP_GEN(dev, port_type) == MLX5_CAP_PORT_TYPE_ETH &&
 692             MLX5_CAP_GEN(dev, vport_group_manager) &&
 693             mlx5_core_is_pf(dev))
 694                 async_event_mask |= (1ull << MLX5_EVENT_TYPE_NIC_VPORT_CHANGE);
 695
 696         if (MLX5_CAP_GEN(dev, port_module_event))
 697                 async_event_mask |= (1ull << MLX5_EVENT_TYPE_PORT_MODULE_EVENT);
 698         else
 699                 mlx5_core_dbg(dev, "port_module_event is not set\n");
 700
 701         if (MLX5_PPS_CAP(dev))
 702                 async_event_mask |= (1ull << MLX5_EVENT_TYPE_PPS_EVENT);
 703
 704         if (MLX5_CAP_GEN(dev, fpga))
 705                 async_event_mask |= (1ull << MLX5_EVENT_TYPE_FPGA_ERROR);
 706
 707         err = mlx5_create_map_eq(dev, &table->cmd_eq, MLX5_EQ_VEC_CMD,
 708                                  MLX5_NUM_CMD_EQE, 1ull << MLX5_EVENT_TYPE_CMD,
 709                                  "mlx5_cmd_eq", MLX5_EQ_TYPE_ASYNC);
 710         if (err) {
 711                 mlx5_core_warn(dev, "failed to create cmd EQ %d\n", err);
 712                 return err;
 713         }
 714
 715         mlx5_cmd_use_events(dev);
 716
 717         err = mlx5_create_map_eq(dev, &table->async_eq, MLX5_EQ_VEC_ASYNC,
 718                                  MLX5_NUM_ASYNC_EQE, async_event_mask,
 719                                  "mlx5_async_eq", MLX5_EQ_TYPE_ASYNC);
 720         if (err) {
 721                 mlx5_core_warn(dev, "failed to create async EQ %d\n", err);
 722                 goto err1;
 723         }
 724
 725         err = mlx5_create_map_eq(dev, &table->pages_eq,
 726                                  MLX5_EQ_VEC_PAGES,
 727                                  /* TODO: sriov max_vf + */ 1,
 728                                  1 << MLX5_EVENT_TYPE_PAGE_REQUEST, "mlx5_pages_eq",
 729                                  MLX5_EQ_TYPE_ASYNC);
 730         if (err) {
 731                 mlx5_core_warn(dev, "failed to create pages EQ %d\n", err);
 732                 goto err2;
 733         }
 734
 735 #ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
 736         if (MLX5_CAP_GEN(dev, pg)) {
 737                 err = mlx5_create_map_eq(dev, &table->pfault_eq,
 738                                          MLX5_EQ_VEC_PFAULT,
 739                                          MLX5_NUM_ASYNC_EQE,
 740                                          1 << MLX5_EVENT_TYPE_PAGE_FAULT,
 741                                          "mlx5_page_fault_eq",
 742                                          MLX5_EQ_TYPE_PF);
 743                 if (err) {
 744                         mlx5_core_warn(dev, "failed to create page fault EQ %d\n",
 745                                        err);
 746                         goto err3;
 747                 }
 748         }
 749
 750         return err;
 751 err3:
 752         mlx5_destroy_unmap_eq(dev, &table->pages_eq);
 753 #else
 754         return err;
 755 #endif
 756
 757 err2:
 758         mlx5_destroy_unmap_eq(dev, &table->async_eq);
 759
 760 err1:
 761         mlx5_cmd_use_polling(dev);
 762         mlx5_destroy_unmap_eq(dev, &table->cmd_eq);
 763         return err;
 764 }
 765
 766 int mlx5_stop_eqs(struct mlx5_core_dev *dev)
 767 {
 768         struct mlx5_eq_table *table = &dev->priv.eq_table;
 769         int err;
 770
 771 #ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
 772         if (MLX5_CAP_GEN(dev, pg)) {
 773                 err = mlx5_destroy_unmap_eq(dev, &table->pfault_eq);
 774                 if (err)
 775                         return err;
 776         }
 777 #endif
 778
 779         err = mlx5_destroy_unmap_eq(dev, &table->pages_eq);
 780         if (err)
 781                 return err;
 782
 783         mlx5_destroy_unmap_eq(dev, &table->async_eq);
 784         mlx5_cmd_use_polling(dev);
 785
 786         err = mlx5_destroy_unmap_eq(dev, &table->cmd_eq);
 787         if (err)
 788                 mlx5_cmd_use_events(dev);
 789
 790         return err;
 791 }
 792
 793 int mlx5_core_eq_query(struct mlx5_core_dev *dev, struct mlx5_eq *eq,
 794                        u32 *out, int outlen)
 795 {
 796         u32 in[MLX5_ST_SZ_DW(query_eq_in)] = {0};
 797
 798         MLX5_SET(query_eq_in, in, opcode, MLX5_CMD_OP_QUERY_EQ);
 799         MLX5_SET(query_eq_in, in, eq_number, eq->eqn);
 800         return mlx5_cmd_exec(dev, in, sizeof(in), out, outlen);
 801 }
 802 EXPORT_SYMBOL_GPL(mlx5_core_eq_query);