drivers/infiniband/hw/cxgb4/t4.h

   1 /*
   2  * Copyright (c) 2009-2010 Chelsio, Inc. 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  *      - Redistributions in binary form must reproduce the above
  18  *        copyright notice, this list of conditions and the following
  19  *        disclaimer in the documentation and/or other materials
  20  *        provided with the distribution.
  21  *
  22  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  23  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  24  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  25  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  26  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  27  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  28  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  29  * SOFTWARE.
  30  */
  31 #ifndef __T4_H__
  32 #define __T4_H__
  33
  34 #include "t4_hw.h"
  35 #include "t4_regs.h"
  36 #include "t4_values.h"
  37 #include "t4_msg.h"
  38 #include "t4fw_ri_api.h"
  39
  40 #define T4_MAX_NUM_PD 65536
  41 #define T4_MAX_MR_SIZE (~0ULL)
  42 #define T4_PAGESIZE_MASK 0xffff000  /* 4KB-128MB */
  43 #define T4_STAG_UNSET 0xffffffff
  44 #define T4_FW_MAJ 0
  45 #define PCIE_MA_SYNC_A 0x30b4
  46
  47 struct t4_status_page {
  48         __be32 rsvd1;   /* flit 0 - hw owns */
  49         __be16 rsvd2;
  50         __be16 qid;
  51         __be16 cidx;
  52         __be16 pidx;
  53         u8 qp_err;      /* flit 1 - sw owns */
  54         u8 db_off;
  55         u8 pad;
  56         u16 host_wq_pidx;
  57         u16 host_cidx;
  58         u16 host_pidx;
  59 };
  60
  61 #define T4_EQ_ENTRY_SIZE 64
  62
  63 #define T4_SQ_NUM_SLOTS 5
  64 #define T4_SQ_NUM_BYTES (T4_EQ_ENTRY_SIZE * T4_SQ_NUM_SLOTS)
  65 #define T4_MAX_SEND_SGE ((T4_SQ_NUM_BYTES - sizeof(struct fw_ri_send_wr) - \
  66                         sizeof(struct fw_ri_isgl)) / sizeof(struct fw_ri_sge))
  67 #define T4_MAX_SEND_INLINE ((T4_SQ_NUM_BYTES - sizeof(struct fw_ri_send_wr) - \
  68                         sizeof(struct fw_ri_immd)))
  69 #define T4_MAX_WRITE_INLINE ((T4_SQ_NUM_BYTES - \
  70                         sizeof(struct fw_ri_rdma_write_wr) - \
  71                         sizeof(struct fw_ri_immd)))
  72 #define T4_MAX_WRITE_SGE ((T4_SQ_NUM_BYTES - \
  73                         sizeof(struct fw_ri_rdma_write_wr) - \
  74                         sizeof(struct fw_ri_isgl)) / sizeof(struct fw_ri_sge))
  75 #define T4_MAX_FR_IMMD ((T4_SQ_NUM_BYTES - sizeof(struct fw_ri_fr_nsmr_wr) - \
  76                         sizeof(struct fw_ri_immd)) & ~31UL)
  77 #define T4_MAX_FR_IMMD_DEPTH (T4_MAX_FR_IMMD / sizeof(u64))
  78 #define T4_MAX_FR_DSGL 1024
  79 #define T4_MAX_FR_DSGL_DEPTH (T4_MAX_FR_DSGL / sizeof(u64))
  80
  81 static inline int t4_max_fr_depth(int use_dsgl)
  82 {
  83         return use_dsgl ? T4_MAX_FR_DSGL_DEPTH : T4_MAX_FR_IMMD_DEPTH;
  84 }
  85
  86 #define T4_RQ_NUM_SLOTS 2
  87 #define T4_RQ_NUM_BYTES (T4_EQ_ENTRY_SIZE * T4_RQ_NUM_SLOTS)
  88 #define T4_MAX_RECV_SGE 4
  89
  90 union t4_wr {
  91         struct fw_ri_res_wr res;
  92         struct fw_ri_wr ri;
  93         struct fw_ri_rdma_write_wr write;
  94         struct fw_ri_send_wr send;
  95         struct fw_ri_rdma_read_wr read;
  96         struct fw_ri_bind_mw_wr bind;
  97         struct fw_ri_fr_nsmr_wr fr;
  98         struct fw_ri_inv_lstag_wr inv;
  99         struct t4_status_page status;
 100         __be64 flits[T4_EQ_ENTRY_SIZE / sizeof(__be64) * T4_SQ_NUM_SLOTS];
 101 };
 102
 103 union t4_recv_wr {
 104         struct fw_ri_recv_wr recv;
 105         struct t4_status_page status;
 106         __be64 flits[T4_EQ_ENTRY_SIZE / sizeof(__be64) * T4_RQ_NUM_SLOTS];
 107 };
 108
 109 static inline void init_wr_hdr(union t4_wr *wqe, u16 wrid,
 110                                enum fw_wr_opcodes opcode, u8 flags, u8 len16)
 111 {
 112         wqe->send.opcode = (u8)opcode;
 113         wqe->send.flags = flags;
 114         wqe->send.wrid = wrid;
 115         wqe->send.r1[0] = 0;
 116         wqe->send.r1[1] = 0;
 117         wqe->send.r1[2] = 0;
 118         wqe->send.len16 = len16;
 119 }
 120
 121 /* CQE/AE status codes */
 122 #define T4_ERR_SUCCESS                     0x0
 123 #define T4_ERR_STAG                        0x1  /* STAG invalid: either the */
 124                                                 /* STAG is offlimt, being 0, */
 125                                                 /* or STAG_key mismatch */
 126 #define T4_ERR_PDID                        0x2  /* PDID mismatch */
 127 #define T4_ERR_QPID                        0x3  /* QPID mismatch */
 128 #define T4_ERR_ACCESS                      0x4  /* Invalid access right */
 129 #define T4_ERR_WRAP                        0x5  /* Wrap error */
 130 #define T4_ERR_BOUND                       0x6  /* base and bounds voilation */
 131 #define T4_ERR_INVALIDATE_SHARED_MR        0x7  /* attempt to invalidate a  */
 132                                                 /* shared memory region */
 133 #define T4_ERR_INVALIDATE_MR_WITH_MW_BOUND 0x8  /* attempt to invalidate a  */
 134                                                 /* shared memory region */
 135 #define T4_ERR_ECC                         0x9  /* ECC error detected */
 136 #define T4_ERR_ECC_PSTAG                   0xA  /* ECC error detected when  */
 137                                                 /* reading PSTAG for a MW  */
 138                                                 /* Invalidate */
 139 #define T4_ERR_PBL_ADDR_BOUND              0xB  /* pbl addr out of bounds:  */
 140                                                 /* software error */
 141 #define T4_ERR_SWFLUSH                     0xC  /* SW FLUSHED */
 142 #define T4_ERR_CRC                         0x10 /* CRC error */
 143 #define T4_ERR_MARKER                      0x11 /* Marker error */
 144 #define T4_ERR_PDU_LEN_ERR                 0x12 /* invalid PDU length */
 145 #define T4_ERR_OUT_OF_RQE                  0x13 /* out of RQE */
 146 #define T4_ERR_DDP_VERSION                 0x14 /* wrong DDP version */
 147 #define T4_ERR_RDMA_VERSION                0x15 /* wrong RDMA version */
 148 #define T4_ERR_OPCODE                      0x16 /* invalid rdma opcode */
 149 #define T4_ERR_DDP_QUEUE_NUM               0x17 /* invalid ddp queue number */
 150 #define T4_ERR_MSN                         0x18 /* MSN error */
 151 #define T4_ERR_TBIT                        0x19 /* tag bit not set correctly */
 152 #define T4_ERR_MO                          0x1A /* MO not 0 for TERMINATE  */
 153                                                 /* or READ_REQ */
 154 #define T4_ERR_MSN_GAP                     0x1B
 155 #define T4_ERR_MSN_RANGE                   0x1C
 156 #define T4_ERR_IRD_OVERFLOW                0x1D
 157 #define T4_ERR_RQE_ADDR_BOUND              0x1E /* RQE addr out of bounds:  */
 158                                                 /* software error */
 159 #define T4_ERR_INTERNAL_ERR                0x1F /* internal error (opcode  */
 160                                                 /* mismatch) */
 161 /*
 162  * CQE defs
 163  */
 164 struct t4_cqe {
 165         __be32 header;
 166         __be32 len;
 167         union {
 168                 struct {
 169                         __be32 stag;
 170                         __be32 msn;
 171                 } rcqe;
 172                 struct {
 173                         u32 nada1;
 174                         u16 nada2;
 175                         u16 cidx;
 176                 } scqe;
 177                 struct {
 178                         __be32 wrid_hi;
 179                         __be32 wrid_low;
 180                 } gen;
 181         } u;
 182         __be64 reserved;
 183         __be64 bits_type_ts;
 184 };
 185
 186 /* macros for flit 0 of the cqe */
 187
 188 #define CQE_QPID_S        12
 189 #define CQE_QPID_M        0xFFFFF
 190 #define CQE_QPID_G(x)     ((((x) >> CQE_QPID_S)) & CQE_QPID_M)
 191 #define CQE_QPID_V(x)     ((x)<<CQE_QPID_S)
 192
 193 #define CQE_SWCQE_S       11
 194 #define CQE_SWCQE_M       0x1
 195 #define CQE_SWCQE_G(x)    ((((x) >> CQE_SWCQE_S)) & CQE_SWCQE_M)
 196 #define CQE_SWCQE_V(x)    ((x)<<CQE_SWCQE_S)
 197
 198 #define CQE_STATUS_S      5
 199 #define CQE_STATUS_M      0x1F
 200 #define CQE_STATUS_G(x)   ((((x) >> CQE_STATUS_S)) & CQE_STATUS_M)
 201 #define CQE_STATUS_V(x)   ((x)<<CQE_STATUS_S)
 202
 203 #define CQE_TYPE_S        4
 204 #define CQE_TYPE_M        0x1
 205 #define CQE_TYPE_G(x)     ((((x) >> CQE_TYPE_S)) & CQE_TYPE_M)
 206 #define CQE_TYPE_V(x)     ((x)<<CQE_TYPE_S)
 207
 208 #define CQE_OPCODE_S      0
 209 #define CQE_OPCODE_M      0xF
 210 #define CQE_OPCODE_G(x)   ((((x) >> CQE_OPCODE_S)) & CQE_OPCODE_M)
 211 #define CQE_OPCODE_V(x)   ((x)<<CQE_OPCODE_S)
 212
 213 #define SW_CQE(x)         (CQE_SWCQE_G(be32_to_cpu((x)->header)))
 214 #define CQE_QPID(x)       (CQE_QPID_G(be32_to_cpu((x)->header)))
 215 #define CQE_TYPE(x)       (CQE_TYPE_G(be32_to_cpu((x)->header)))
 216 #define SQ_TYPE(x)        (CQE_TYPE((x)))
 217 #define RQ_TYPE(x)        (!CQE_TYPE((x)))
 218 #define CQE_STATUS(x)     (CQE_STATUS_G(be32_to_cpu((x)->header)))
 219 #define CQE_OPCODE(x)     (CQE_OPCODE_G(be32_to_cpu((x)->header)))
 220
 221 #define CQE_SEND_OPCODE(x)( \
 222         (CQE_OPCODE_G(be32_to_cpu((x)->header)) == FW_RI_SEND) || \
 223         (CQE_OPCODE_G(be32_to_cpu((x)->header)) == FW_RI_SEND_WITH_SE) || \
 224         (CQE_OPCODE_G(be32_to_cpu((x)->header)) == FW_RI_SEND_WITH_INV) || \
 225         (CQE_OPCODE_G(be32_to_cpu((x)->header)) == FW_RI_SEND_WITH_SE_INV))
 226
 227 #define CQE_LEN(x)        (be32_to_cpu((x)->len))
 228
 229 /* used for RQ completion processing */
 230 #define CQE_WRID_STAG(x)  (be32_to_cpu((x)->u.rcqe.stag))
 231 #define CQE_WRID_MSN(x)   (be32_to_cpu((x)->u.rcqe.msn))
 232
 233 /* used for SQ completion processing */
 234 #define CQE_WRID_SQ_IDX(x)      ((x)->u.scqe.cidx)
 235
 236 /* generic accessor macros */
 237 #define CQE_WRID_HI(x)          (be32_to_cpu((x)->u.gen.wrid_hi))
 238 #define CQE_WRID_LOW(x)         (be32_to_cpu((x)->u.gen.wrid_low))
 239
 240 /* macros for flit 3 of the cqe */
 241 #define CQE_GENBIT_S    63
 242 #define CQE_GENBIT_M    0x1
 243 #define CQE_GENBIT_G(x) (((x) >> CQE_GENBIT_S) & CQE_GENBIT_M)
 244 #define CQE_GENBIT_V(x) ((x)<<CQE_GENBIT_S)
 245
 246 #define CQE_OVFBIT_S    62
 247 #define CQE_OVFBIT_M    0x1
 248 #define CQE_OVFBIT_G(x) ((((x) >> CQE_OVFBIT_S)) & CQE_OVFBIT_M)
 249
 250 #define CQE_IQTYPE_S    60
 251 #define CQE_IQTYPE_M    0x3
 252 #define CQE_IQTYPE_G(x) ((((x) >> CQE_IQTYPE_S)) & CQE_IQTYPE_M)
 253
 254 #define CQE_TS_M        0x0fffffffffffffffULL
 255 #define CQE_TS_G(x)     ((x) & CQE_TS_M)
 256
 257 #define CQE_OVFBIT(x)   ((unsigned)CQE_OVFBIT_G(be64_to_cpu((x)->bits_type_ts)))
 258 #define CQE_GENBIT(x)   ((unsigned)CQE_GENBIT_G(be64_to_cpu((x)->bits_type_ts)))
 259 #define CQE_TS(x)       (CQE_TS_G(be64_to_cpu((x)->bits_type_ts)))
 260
 261 struct t4_swsqe {
 262         u64                     wr_id;
 263         struct t4_cqe           cqe;
 264         int                     read_len;
 265         int                     opcode;
 266         int                     complete;
 267         int                     signaled;
 268         u16                     idx;
 269         int                     flushed;
 270         struct timespec         host_ts;
 271         u64                     sge_ts;
 272 };
 273
 274 static inline pgprot_t t4_pgprot_wc(pgprot_t prot)
 275 {
 276 #if defined(__i386__) || defined(__x86_64__) || defined(CONFIG_PPC64)
 277         return pgprot_writecombine(prot);
 278 #else
 279         return pgprot_noncached(prot);
 280 #endif
 281 }
 282
 283 enum {
 284         T4_SQ_ONCHIP = (1<<0),
 285 };
 286
 287 struct t4_sq {
 288         union t4_wr *queue;
 289         dma_addr_t dma_addr;
 290         DEFINE_DMA_UNMAP_ADDR(mapping);
 291         unsigned long phys_addr;
 292         struct t4_swsqe *sw_sq;
 293         struct t4_swsqe *oldest_read;
 294         void __iomem *bar2_va;
 295         u64 bar2_pa;
 296         size_t memsize;
 297         u32 bar2_qid;
 298         u32 qid;
 299         u16 in_use;
 300         u16 size;
 301         u16 cidx;
 302         u16 pidx;
 303         u16 wq_pidx;
 304         u16 wq_pidx_inc;
 305         u16 flags;
 306         short flush_cidx;
 307 };
 308
 309 struct t4_swrqe {
 310         u64 wr_id;
 311         struct timespec host_ts;
 312         u64 sge_ts;
 313 };
 314
 315 struct t4_rq {
 316         union  t4_recv_wr *queue;
 317         dma_addr_t dma_addr;
 318         DEFINE_DMA_UNMAP_ADDR(mapping);
 319         struct t4_swrqe *sw_rq;
 320         void __iomem *bar2_va;
 321         u64 bar2_pa;
 322         size_t memsize;
 323         u32 bar2_qid;
 324         u32 qid;
 325         u32 msn;
 326         u32 rqt_hwaddr;
 327         u16 rqt_size;
 328         u16 in_use;
 329         u16 size;
 330         u16 cidx;
 331         u16 pidx;
 332         u16 wq_pidx;
 333         u16 wq_pidx_inc;
 334 };
 335
 336 struct t4_wq {
 337         struct t4_sq sq;
 338         struct t4_rq rq;
 339         void __iomem *db;
 340         struct c4iw_rdev *rdev;
 341         int flushed;
 342 };
 343
 344 static inline int t4_rqes_posted(struct t4_wq *wq)
 345 {
 346         return wq->rq.in_use;
 347 }
 348
 349 static inline int t4_rq_empty(struct t4_wq *wq)
 350 {
 351         return wq->rq.in_use == 0;
 352 }
 353
 354 static inline int t4_rq_full(struct t4_wq *wq)
 355 {
 356         return wq->rq.in_use == (wq->rq.size - 1);
 357 }
 358
 359 static inline u32 t4_rq_avail(struct t4_wq *wq)
 360 {
 361         return wq->rq.size - 1 - wq->rq.in_use;
 362 }
 363
 364 static inline void t4_rq_produce(struct t4_wq *wq, u8 len16)
 365 {
 366         wq->rq.in_use++;
 367         if (++wq->rq.pidx == wq->rq.size)
 368                 wq->rq.pidx = 0;
 369         wq->rq.wq_pidx += DIV_ROUND_UP(len16*16, T4_EQ_ENTRY_SIZE);
 370         if (wq->rq.wq_pidx >= wq->rq.size * T4_RQ_NUM_SLOTS)
 371                 wq->rq.wq_pidx %= wq->rq.size * T4_RQ_NUM_SLOTS;
 372 }
 373
 374 static inline void t4_rq_consume(struct t4_wq *wq)
 375 {
 376         wq->rq.in_use--;
 377         wq->rq.msn++;
 378         if (++wq->rq.cidx == wq->rq.size)
 379                 wq->rq.cidx = 0;
 380 }
 381
 382 static inline u16 t4_rq_host_wq_pidx(struct t4_wq *wq)
 383 {
 384         return wq->rq.queue[wq->rq.size].status.host_wq_pidx;
 385 }
 386
 387 static inline u16 t4_rq_wq_size(struct t4_wq *wq)
 388 {
 389                 return wq->rq.size * T4_RQ_NUM_SLOTS;
 390 }
 391
 392 static inline int t4_sq_onchip(struct t4_sq *sq)
 393 {
 394         return sq->flags & T4_SQ_ONCHIP;
 395 }
 396
 397 static inline int t4_sq_empty(struct t4_wq *wq)
 398 {
 399         return wq->sq.in_use == 0;
 400 }
 401
 402 static inline int t4_sq_full(struct t4_wq *wq)
 403 {
 404         return wq->sq.in_use == (wq->sq.size - 1);
 405 }
 406
 407 static inline u32 t4_sq_avail(struct t4_wq *wq)
 408 {
 409         return wq->sq.size - 1 - wq->sq.in_use;
 410 }
 411
 412 static inline void t4_sq_produce(struct t4_wq *wq, u8 len16)
 413 {
 414         wq->sq.in_use++;
 415         if (++wq->sq.pidx == wq->sq.size)
 416                 wq->sq.pidx = 0;
 417         wq->sq.wq_pidx += DIV_ROUND_UP(len16*16, T4_EQ_ENTRY_SIZE);
 418         if (wq->sq.wq_pidx >= wq->sq.size * T4_SQ_NUM_SLOTS)
 419                 wq->sq.wq_pidx %= wq->sq.size * T4_SQ_NUM_SLOTS;
 420 }
 421
 422 static inline void t4_sq_consume(struct t4_wq *wq)
 423 {
 424         BUG_ON(wq->sq.in_use < 1);
 425         if (wq->sq.cidx == wq->sq.flush_cidx)
 426                 wq->sq.flush_cidx = -1;
 427         wq->sq.in_use--;
 428         if (++wq->sq.cidx == wq->sq.size)
 429                 wq->sq.cidx = 0;
 430 }
 431
 432 static inline u16 t4_sq_host_wq_pidx(struct t4_wq *wq)
 433 {
 434         return wq->sq.queue[wq->sq.size].status.host_wq_pidx;
 435 }
 436
 437 static inline u16 t4_sq_wq_size(struct t4_wq *wq)
 438 {
 439                 return wq->sq.size * T4_SQ_NUM_SLOTS;
 440 }
 441
 442 /* This function copies 64 byte coalesced work request to memory
 443  * mapped BAR2 space. For coalesced WRs, the SGE fetches data
 444  * from the FIFO instead of from Host.
 445  */
 446 static inline void pio_copy(u64 __iomem *dst, u64 *src)
 447 {
 448         int count = 8;
 449
 450         while (count) {
 451                 writeq(*src, dst);
 452                 src++;
 453                 dst++;
 454                 count--;
 455         }
 456 }
 457
 458 static inline void t4_ring_sq_db(struct t4_wq *wq, u16 inc, union t4_wr *wqe)
 459 {
 460
 461         /* Flush host queue memory writes. */
 462         wmb();
 463         if (wq->sq.bar2_va) {
 464                 if (inc == 1 && wq->sq.bar2_qid == 0 && wqe) {
 465                         PDBG("%s: WC wq->sq.pidx = %d\n",
 466                              __func__, wq->sq.pidx);
 467                         pio_copy((u64 __iomem *)
 468                                  (wq->sq.bar2_va + SGE_UDB_WCDOORBELL),
 469                                  (u64 *)wqe);
 470                 } else {
 471                         PDBG("%s: DB wq->sq.pidx = %d\n",
 472                              __func__, wq->sq.pidx);
 473                         writel(PIDX_T5_V(inc) | QID_V(wq->sq.bar2_qid),
 474                                wq->sq.bar2_va + SGE_UDB_KDOORBELL);
 475                 }
 476
 477                 /* Flush user doorbell area writes. */
 478                 wmb();
 479                 return;
 480         }
 481         writel(QID_V(wq->sq.qid) | PIDX_V(inc), wq->db);
 482 }
 483
 484 static inline void t4_ring_rq_db(struct t4_wq *wq, u16 inc,
 485                                  union t4_recv_wr *wqe)
 486 {
 487
 488         /* Flush host queue memory writes. */
 489         wmb();
 490         if (wq->rq.bar2_va) {
 491                 if (inc == 1 && wq->rq.bar2_qid == 0 && wqe) {
 492                         PDBG("%s: WC wq->rq.pidx = %d\n",
 493                              __func__, wq->rq.pidx);
 494                         pio_copy((u64 __iomem *)
 495                                  (wq->rq.bar2_va + SGE_UDB_WCDOORBELL),
 496                                  (void *)wqe);
 497                 } else {
 498                         PDBG("%s: DB wq->rq.pidx = %d\n",
 499                              __func__, wq->rq.pidx);
 500                         writel(PIDX_T5_V(inc) | QID_V(wq->rq.bar2_qid),
 501                                wq->rq.bar2_va + SGE_UDB_KDOORBELL);
 502                 }
 503
 504                 /* Flush user doorbell area writes. */
 505                 wmb();
 506                 return;
 507         }
 508         writel(QID_V(wq->rq.qid) | PIDX_V(inc), wq->db);
 509 }
 510
 511 static inline int t4_wq_in_error(struct t4_wq *wq)
 512 {
 513         return wq->rq.queue[wq->rq.size].status.qp_err;
 514 }
 515
 516 static inline void t4_set_wq_in_error(struct t4_wq *wq)
 517 {
 518         wq->rq.queue[wq->rq.size].status.qp_err = 1;
 519 }
 520
 521 static inline void t4_disable_wq_db(struct t4_wq *wq)
 522 {
 523         wq->rq.queue[wq->rq.size].status.db_off = 1;
 524 }
 525
 526 static inline void t4_enable_wq_db(struct t4_wq *wq)
 527 {
 528         wq->rq.queue[wq->rq.size].status.db_off = 0;
 529 }
 530
 531 static inline int t4_wq_db_enabled(struct t4_wq *wq)
 532 {
 533         return !wq->rq.queue[wq->rq.size].status.db_off;
 534 }
 535
 536 enum t4_cq_flags {
 537         CQ_ARMED        = 1,
 538 };
 539
 540 struct t4_cq {
 541         struct t4_cqe *queue;
 542         dma_addr_t dma_addr;
 543         DEFINE_DMA_UNMAP_ADDR(mapping);
 544         struct t4_cqe *sw_queue;
 545         void __iomem *gts;
 546         void __iomem *bar2_va;
 547         u64 bar2_pa;
 548         u32 bar2_qid;
 549         struct c4iw_rdev *rdev;
 550         size_t memsize;
 551         __be64 bits_type_ts;
 552         u32 cqid;
 553         u32 qid_mask;
 554         int vector;
 555         u16 size; /* including status page */
 556         u16 cidx;
 557         u16 sw_pidx;
 558         u16 sw_cidx;
 559         u16 sw_in_use;
 560         u16 cidx_inc;
 561         u8 gen;
 562         u8 error;
 563         unsigned long flags;
 564 };
 565
 566 static inline void write_gts(struct t4_cq *cq, u32 val)
 567 {
 568         if (cq->bar2_va)
 569                 writel(val | INGRESSQID_V(cq->bar2_qid),
 570                        cq->bar2_va + SGE_UDB_GTS);
 571         else
 572                 writel(val | INGRESSQID_V(cq->cqid), cq->gts);
 573 }
 574
 575 static inline int t4_clear_cq_armed(struct t4_cq *cq)
 576 {
 577         return test_and_clear_bit(CQ_ARMED, &cq->flags);
 578 }
 579
 580 static inline int t4_arm_cq(struct t4_cq *cq, int se)
 581 {
 582         u32 val;
 583
 584         set_bit(CQ_ARMED, &cq->flags);
 585         while (cq->cidx_inc > CIDXINC_M) {
 586                 val = SEINTARM_V(0) | CIDXINC_V(CIDXINC_M) | TIMERREG_V(7);
 587                 write_gts(cq, val);
 588                 cq->cidx_inc -= CIDXINC_M;
 589         }
 590         val = SEINTARM_V(se) | CIDXINC_V(cq->cidx_inc) | TIMERREG_V(6);
 591         write_gts(cq, val);
 592         cq->cidx_inc = 0;
 593         return 0;
 594 }
 595
 596 static inline void t4_swcq_produce(struct t4_cq *cq)
 597 {
 598         cq->sw_in_use++;
 599         if (cq->sw_in_use == cq->size) {
 600                 PDBG("%s cxgb4 sw cq overflow cqid %u\n", __func__, cq->cqid);
 601                 cq->error = 1;
 602                 BUG_ON(1);
 603         }
 604         if (++cq->sw_pidx == cq->size)
 605                 cq->sw_pidx = 0;
 606 }
 607
 608 static inline void t4_swcq_consume(struct t4_cq *cq)
 609 {
 610         BUG_ON(cq->sw_in_use < 1);
 611         cq->sw_in_use--;
 612         if (++cq->sw_cidx == cq->size)
 613                 cq->sw_cidx = 0;
 614 }
 615
 616 static inline void t4_hwcq_consume(struct t4_cq *cq)
 617 {
 618         cq->bits_type_ts = cq->queue[cq->cidx].bits_type_ts;
 619         if (++cq->cidx_inc == (cq->size >> 4) || cq->cidx_inc == CIDXINC_M) {
 620                 u32 val;
 621
 622                 val = SEINTARM_V(0) | CIDXINC_V(cq->cidx_inc) | TIMERREG_V(7);
 623                 write_gts(cq, val);
 624                 cq->cidx_inc = 0;
 625         }
 626         if (++cq->cidx == cq->size) {
 627                 cq->cidx = 0;
 628                 cq->gen ^= 1;
 629         }
 630 }
 631
 632 static inline int t4_valid_cqe(struct t4_cq *cq, struct t4_cqe *cqe)
 633 {
 634         return (CQE_GENBIT(cqe) == cq->gen);
 635 }
 636
 637 static inline int t4_cq_notempty(struct t4_cq *cq)
 638 {
 639         return cq->sw_in_use || t4_valid_cqe(cq, &cq->queue[cq->cidx]);
 640 }
 641
 642 static inline int t4_next_hw_cqe(struct t4_cq *cq, struct t4_cqe **cqe)
 643 {
 644         int ret;
 645         u16 prev_cidx;
 646
 647         if (cq->cidx == 0)
 648                 prev_cidx = cq->size - 1;
 649         else
 650                 prev_cidx = cq->cidx - 1;
 651
 652         if (cq->queue[prev_cidx].bits_type_ts != cq->bits_type_ts) {
 653                 ret = -EOVERFLOW;
 654                 cq->error = 1;
 655                 printk(KERN_ERR MOD "cq overflow cqid %u\n", cq->cqid);
 656                 BUG_ON(1);
 657         } else if (t4_valid_cqe(cq, &cq->queue[cq->cidx])) {
 658
 659                 /* Ensure CQE is flushed to memory */
 660                 rmb();
 661                 *cqe = &cq->queue[cq->cidx];
 662                 ret = 0;
 663         } else
 664                 ret = -ENODATA;
 665         return ret;
 666 }
 667
 668 static inline struct t4_cqe *t4_next_sw_cqe(struct t4_cq *cq)
 669 {
 670         if (cq->sw_in_use == cq->size) {
 671                 PDBG("%s cxgb4 sw cq overflow cqid %u\n", __func__, cq->cqid);
 672                 cq->error = 1;
 673                 BUG_ON(1);
 674                 return NULL;
 675         }
 676         if (cq->sw_in_use)
 677                 return &cq->sw_queue[cq->sw_cidx];
 678         return NULL;
 679 }
 680
 681 static inline int t4_next_cqe(struct t4_cq *cq, struct t4_cqe **cqe)
 682 {
 683         int ret = 0;
 684
 685         if (cq->error)
 686                 ret = -ENODATA;
 687         else if (cq->sw_in_use)
 688                 *cqe = &cq->sw_queue[cq->sw_cidx];
 689         else
 690                 ret = t4_next_hw_cqe(cq, cqe);
 691         return ret;
 692 }
 693
 694 static inline int t4_cq_in_error(struct t4_cq *cq)
 695 {
 696         return ((struct t4_status_page *)&cq->queue[cq->size])->qp_err;
 697 }
 698
 699 static inline void t4_set_cq_in_error(struct t4_cq *cq)
 700 {
 701         ((struct t4_status_page *)&cq->queue[cq->size])->qp_err = 1;
 702 }
 703 #endif
 704
 705 struct t4_dev_status_page {
 706         u8 db_off;
 707         u8 pad1;
 708         u16 pad2;
 709         u32 pad3;
 710         u64 qp_start;
 711         u64 qp_size;
 712         u64 cq_start;
 713         u64 cq_size;
 714 };