drivers/net/chelsio/sge.c

   1 /*****************************************************************************
   2  *                                                                           *
   3  * File: sge.c                                                               *
   4  * $Revision: 1.26 $                                                         *
   5  * $Date: 2005/06/21 18:29:48 $                                              *
   6  * Description:                                                              *
   7  *  DMA engine.                                                              *
   8  *  part of the Chelsio 10Gb Ethernet Driver.                                *
   9  *                                                                           *
  10  * This program is free software; you can redistribute it and/or modify      *
  11  * it under the terms of the GNU General Public License, version 2, as       *
  12  * published by the Free Software Foundation.                                *
  13  *                                                                           *
  14  * You should have received a copy of the GNU General Public License along   *
  15  * with this program; if not, write to the Free Software Foundation, Inc.,   *
  16  * 59 Temple Place - Suite 330, Boston, MA  02111-1307, USA.                 *
  17  *                                                                           *
  18  * THIS SOFTWARE IS PROVIDED ``AS IS'' AND WITHOUT ANY EXPRESS OR IMPLIED    *
  19  * WARRANTIES, INCLUDING, WITHOUT LIMITATION, THE IMPLIED WARRANTIES OF      *
  20  * MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE.                     *
  21  *                                                                           *
  22  * http://www.chelsio.com                                                    *
  23  *                                                                           *
  24  * Copyright (c) 2003 - 2005 Chelsio Communications, Inc.                    *
  25  * All rights reserved.                                                      *
  26  *                                                                           *
  27  * Maintainers: maintainers@chelsio.com                                      *
  28  *                                                                           *
  29  * Authors: Dimitrios Michailidis   <dm@chelsio.com>                         *
  30  *          Tina Yang               <tainay@chelsio.com>                     *
  31  *          Felix Marti             <felix@chelsio.com>                      *
  32  *          Scott Bardone           <sbardone@chelsio.com>                   *
  33  *          Kurt Ottaway            <kottaway@chelsio.com>                   *
  34  *          Frank DiMambro          <frank@chelsio.com>                      *
  35  *                                                                           *
  36  * History:                                                                  *
  37  *                                                                           *
  38  ****************************************************************************/
  39
  40 #include "common.h"
  41
  42 #include <linux/types.h>
  43 #include <linux/errno.h>
  44 #include <linux/pci.h>
  45 #include <linux/ktime.h>
  46 #include <linux/netdevice.h>
  47 #include <linux/etherdevice.h>
  48 #include <linux/if_vlan.h>
  49 #include <linux/skbuff.h>
  50 #include <linux/init.h>
  51 #include <linux/mm.h>
  52 #include <linux/tcp.h>
  53 #include <linux/ip.h>
  54 #include <linux/in.h>
  55 #include <linux/if_arp.h>
  56
  57 #include "cpl5_cmd.h"
  58 #include "sge.h"
  59 #include "regs.h"
  60 #include "espi.h"
  61
  62 /* This belongs in if_ether.h */
  63 #define ETH_P_CPL5 0xf
  64
  65 #define SGE_CMDQ_N              2
  66 #define SGE_FREELQ_N            2
  67 #define SGE_CMDQ0_E_N           1024
  68 #define SGE_CMDQ1_E_N           128
  69 #define SGE_FREEL_SIZE          4096
  70 #define SGE_JUMBO_FREEL_SIZE    512
  71 #define SGE_FREEL_REFILL_THRESH 16
  72 #define SGE_RESPQ_E_N           1024
  73 #define SGE_INTRTIMER_NRES      1000
  74 #define SGE_RX_COPY_THRES       256
  75 #define SGE_RX_SM_BUF_SIZE      1536
  76 #define SGE_TX_DESC_MAX_PLEN    16384
  77
  78 # define SGE_RX_DROP_THRES 2
  79
  80 #define SGE_RESPQ_REPLENISH_THRES (SGE_RESPQ_E_N / 4)
  81
  82 /*
  83  * Period of the TX buffer reclaim timer.  This timer does not need to run
  84  * frequently as TX buffers are usually reclaimed by new TX packets.
  85  */
  86 #define TX_RECLAIM_PERIOD (HZ / 4)
  87
  88 #define M_CMD_LEN       0x7fffffff
  89 #define V_CMD_LEN(v)    (v)
  90 #define G_CMD_LEN(v)    ((v) & M_CMD_LEN)
  91 #define V_CMD_GEN1(v)   ((v) << 31)
  92 #define V_CMD_GEN2(v)   (v)
  93 #define F_CMD_DATAVALID (1 << 1)
  94 #define F_CMD_SOP       (1 << 2)
  95 #define V_CMD_EOP(v)    ((v) << 3)
  96
  97 /*
  98  * Command queue, receive buffer list, and response queue descriptors.
  99  */
 100 #if defined(__BIG_ENDIAN_BITFIELD)
 101 struct cmdQ_e {
 102         u32 addr_lo;
 103         u32 len_gen;
 104         u32 flags;
 105         u32 addr_hi;
 106 };
 107
 108 struct freelQ_e {
 109         u32 addr_lo;
 110         u32 len_gen;
 111         u32 gen2;
 112         u32 addr_hi;
 113 };
 114
 115 struct respQ_e {
 116         u32 Qsleeping           : 4;
 117         u32 Cmdq1CreditReturn   : 5;
 118         u32 Cmdq1DmaComplete    : 5;
 119         u32 Cmdq0CreditReturn   : 5;
 120         u32 Cmdq0DmaComplete    : 5;
 121         u32 FreelistQid         : 2;
 122         u32 CreditValid         : 1;
 123         u32 DataValid           : 1;
 124         u32 Offload             : 1;
 125         u32 Eop                 : 1;
 126         u32 Sop                 : 1;
 127         u32 GenerationBit       : 1;
 128         u32 BufferLength;
 129 };
 130 #elif defined(__LITTLE_ENDIAN_BITFIELD)
 131 struct cmdQ_e {
 132         u32 len_gen;
 133         u32 addr_lo;
 134         u32 addr_hi;
 135         u32 flags;
 136 };
 137
 138 struct freelQ_e {
 139         u32 len_gen;
 140         u32 addr_lo;
 141         u32 addr_hi;
 142         u32 gen2;
 143 };
 144
 145 struct respQ_e {
 146         u32 BufferLength;
 147         u32 GenerationBit       : 1;
 148         u32 Sop                 : 1;
 149         u32 Eop                 : 1;
 150         u32 Offload             : 1;
 151         u32 DataValid           : 1;
 152         u32 CreditValid         : 1;
 153         u32 FreelistQid         : 2;
 154         u32 Cmdq0DmaComplete    : 5;
 155         u32 Cmdq0CreditReturn   : 5;
 156         u32 Cmdq1DmaComplete    : 5;
 157         u32 Cmdq1CreditReturn   : 5;
 158         u32 Qsleeping           : 4;
 159 } ;
 160 #endif
 161
 162 /*
 163  * SW Context Command and Freelist Queue Descriptors
 164  */
 165 struct cmdQ_ce {
 166         struct sk_buff *skb;
 167         DECLARE_PCI_UNMAP_ADDR(dma_addr);
 168         DECLARE_PCI_UNMAP_LEN(dma_len);
 169 };
 170
 171 struct freelQ_ce {
 172         struct sk_buff *skb;
 173         DECLARE_PCI_UNMAP_ADDR(dma_addr);
 174         DECLARE_PCI_UNMAP_LEN(dma_len);
 175 };
 176
 177 /*
 178  * SW command, freelist and response rings
 179  */
 180 struct cmdQ {
 181         unsigned long   status;         /* HW DMA fetch status */
 182         unsigned int    in_use;         /* # of in-use command descriptors */
 183         unsigned int    size;           /* # of descriptors */
 184         unsigned int    processed;      /* total # of descs HW has processed */
 185         unsigned int    cleaned;        /* total # of descs SW has reclaimed */
 186         unsigned int    stop_thres;     /* SW TX queue suspend threshold */
 187         u16             pidx;           /* producer index (SW) */
 188         u16             cidx;           /* consumer index (HW) */
 189         u8              genbit;         /* current generation (=valid) bit */
 190         u8              sop;            /* is next entry start of packet? */
 191         struct cmdQ_e  *entries;        /* HW command descriptor Q */
 192         struct cmdQ_ce *centries;       /* SW command context descriptor Q */
 193         dma_addr_t      dma_addr;       /* DMA addr HW command descriptor Q */
 194         spinlock_t      lock;           /* Lock to protect cmdQ enqueuing */
 195 };
 196
 197 struct freelQ {
 198         unsigned int    credits;        /* # of available RX buffers */
 199         unsigned int    size;           /* free list capacity */
 200         u16             pidx;           /* producer index (SW) */
 201         u16             cidx;           /* consumer index (HW) */
 202         u16             rx_buffer_size; /* Buffer size on this free list */
 203         u16             dma_offset;     /* DMA offset to align IP headers */
 204         u16             recycleq_idx;   /* skb recycle q to use */
 205         u8              genbit;         /* current generation (=valid) bit */
 206         struct freelQ_e *entries;       /* HW freelist descriptor Q */
 207         struct freelQ_ce *centries;     /* SW freelist context descriptor Q */
 208         dma_addr_t      dma_addr;       /* DMA addr HW freelist descriptor Q */
 209 };
 210
 211 struct respQ {
 212         unsigned int    credits;        /* credits to be returned to SGE */
 213         unsigned int    size;           /* # of response Q descriptors */
 214         u16             cidx;           /* consumer index (SW) */
 215         u8              genbit;         /* current generation(=valid) bit */
 216         struct respQ_e *entries;        /* HW response descriptor Q */
 217         dma_addr_t      dma_addr;       /* DMA addr HW response descriptor Q */
 218 };
 219
 220 /* Bit flags for cmdQ.status */
 221 enum {
 222         CMDQ_STAT_RUNNING = 1,          /* fetch engine is running */
 223         CMDQ_STAT_LAST_PKT_DB = 2       /* last packet rung the doorbell */
 224 };
 225
 226 /* T204 TX SW scheduler */
 227
 228 /* Per T204 TX port */
 229 struct sched_port {
 230         unsigned int    avail;          /* available bits - quota */
 231         unsigned int    drain_bits_per_1024ns; /* drain rate */
 232         unsigned int    speed;          /* drain rate, mbps */
 233         unsigned int    mtu;            /* mtu size */
 234         struct sk_buff_head skbq;       /* pending skbs */
 235 };
 236
 237 /* Per T204 device */
 238 struct sched {
 239         ktime_t         last_updated;   /* last time quotas were computed */
 240         unsigned int    max_avail;      /* max bits to be sent to any port */
 241         unsigned int    port;           /* port index (round robin ports) */
 242         unsigned int    num;            /* num skbs in per port queues */
 243         struct sched_port p[MAX_NPORTS];
 244         struct tasklet_struct sched_tsk;/* tasklet used to run scheduler */
 245 };
 246 static void restart_sched(unsigned long);
 247
 248
 249 /*
 250  * Main SGE data structure
 251  *
 252  * Interrupts are handled by a single CPU and it is likely that on a MP system
 253  * the application is migrated to another CPU. In that scenario, we try to
 254  * seperate the RX(in irq context) and TX state in order to decrease memory
 255  * contention.
 256  */
 257 struct sge {
 258         struct adapter *adapter;        /* adapter backpointer */
 259         struct net_device *netdev;      /* netdevice backpointer */
 260         struct freelQ   freelQ[SGE_FREELQ_N]; /* buffer free lists */
 261         struct respQ    respQ;          /* response Q */
 262         unsigned long   stopped_tx_queues; /* bitmap of suspended Tx queues */
 263         unsigned int    rx_pkt_pad;     /* RX padding for L2 packets */
 264         unsigned int    jumbo_fl;       /* jumbo freelist Q index */
 265         unsigned int    intrtimer_nres; /* no-resource interrupt timer */
 266         unsigned int    fixed_intrtimer;/* non-adaptive interrupt timer */
 267         struct timer_list tx_reclaim_timer; /* reclaims TX buffers */
 268         struct timer_list espibug_timer;
 269         unsigned long   espibug_timeout;
 270         struct sk_buff  *espibug_skb[MAX_NPORTS];
 271         u32             sge_control;    /* shadow value of sge control reg */
 272         struct sge_intr_counts stats;
 273         struct sge_port_stats *port_stats[MAX_NPORTS];
 274         struct sched    *tx_sched;
 275         struct cmdQ cmdQ[SGE_CMDQ_N] ____cacheline_aligned_in_smp;
 276 };
 277
 278 /*
 279  * stop tasklet and free all pending skb's
 280  */
 281 static void tx_sched_stop(struct sge *sge)
 282 {
 283         struct sched *s = sge->tx_sched;
 284         int i;
 285
 286         tasklet_kill(&s->sched_tsk);
 287
 288         for (i = 0; i < MAX_NPORTS; i++)
 289                 __skb_queue_purge(&s->p[s->port].skbq);
 290 }
 291
 292 /*
 293  * t1_sched_update_parms() is called when the MTU or link speed changes. It
 294  * re-computes scheduler parameters to scope with the change.
 295  */
 296 unsigned int t1_sched_update_parms(struct sge *sge, unsigned int port,
 297                                    unsigned int mtu, unsigned int speed)
 298 {
 299         struct sched *s = sge->tx_sched;
 300         struct sched_port *p = &s->p[port];
 301         unsigned int max_avail_segs;
 302
 303         pr_debug("t1_sched_update_params mtu=%d speed=%d\n", mtu, speed);
 304         if (speed)
 305                 p->speed = speed;
 306         if (mtu)
 307                 p->mtu = mtu;
 308
 309         if (speed || mtu) {
 310                 unsigned long long drain = 1024ULL * p->speed * (p->mtu - 40);
 311                 do_div(drain, (p->mtu + 50) * 1000);
 312                 p->drain_bits_per_1024ns = (unsigned int) drain;
 313
 314                 if (p->speed < 1000)
 315                         p->drain_bits_per_1024ns =
 316                                 90 * p->drain_bits_per_1024ns / 100;
 317         }
 318
 319         if (board_info(sge->adapter)->board == CHBT_BOARD_CHT204) {
 320                 p->drain_bits_per_1024ns -= 16;
 321                 s->max_avail = max(4096U, p->mtu + 16 + 14 + 4);
 322                 max_avail_segs = max(1U, 4096 / (p->mtu - 40));
 323         } else {
 324                 s->max_avail = 16384;
 325                 max_avail_segs = max(1U, 9000 / (p->mtu - 40));
 326         }
 327
 328         pr_debug("t1_sched_update_parms: mtu %u speed %u max_avail %u "
 329                  "max_avail_segs %u drain_bits_per_1024ns %u\n", p->mtu,
 330                  p->speed, s->max_avail, max_avail_segs,
 331                  p->drain_bits_per_1024ns);
 332
 333         return max_avail_segs * (p->mtu - 40);
 334 }
 335
 336 /*
 337  * t1_sched_max_avail_bytes() tells the scheduler the maximum amount of
 338  * data that can be pushed per port.
 339  */
 340 void t1_sched_set_max_avail_bytes(struct sge *sge, unsigned int val)
 341 {
 342         struct sched *s = sge->tx_sched;
 343         unsigned int i;
 344
 345         s->max_avail = val;
 346         for (i = 0; i < MAX_NPORTS; i++)
 347                 t1_sched_update_parms(sge, i, 0, 0);
 348 }
 349
 350 /*
 351  * t1_sched_set_drain_bits_per_us() tells the scheduler at which rate a port
 352  * is draining.
 353  */
 354 void t1_sched_set_drain_bits_per_us(struct sge *sge, unsigned int port,
 355                                          unsigned int val)
 356 {
 357         struct sched *s = sge->tx_sched;
 358         struct sched_port *p = &s->p[port];
 359         p->drain_bits_per_1024ns = val * 1024 / 1000;
 360         t1_sched_update_parms(sge, port, 0, 0);
 361 }
 362
 363
 364 /*
 365  * get_clock() implements a ns clock (see ktime_get)
 366  */
 367 static inline ktime_t get_clock(void)
 368 {
 369         struct timespec ts;
 370
 371         ktime_get_ts(&ts);
 372         return timespec_to_ktime(ts);
 373 }
 374
 375 /*
 376  * tx_sched_init() allocates resources and does basic initialization.
 377  */
 378 static int tx_sched_init(struct sge *sge)
 379 {
 380         struct sched *s;
 381         int i;
 382
 383         s = kzalloc(sizeof (struct sched), GFP_KERNEL);
 384         if (!s)
 385                 return -ENOMEM;
 386
 387         pr_debug("tx_sched_init\n");
 388         tasklet_init(&s->sched_tsk, restart_sched, (unsigned long) sge);
 389         sge->tx_sched = s;
 390
 391         for (i = 0; i < MAX_NPORTS; i++) {
 392                 skb_queue_head_init(&s->p[i].skbq);
 393                 t1_sched_update_parms(sge, i, 1500, 1000);
 394         }
 395
 396         return 0;
 397 }
 398
 399 /*
 400  * sched_update_avail() computes the delta since the last time it was called
 401  * and updates the per port quota (number of bits that can be sent to the any
 402  * port).
 403  */
 404 static inline int sched_update_avail(struct sge *sge)
 405 {
 406         struct sched *s = sge->tx_sched;
 407         ktime_t now = get_clock();
 408         unsigned int i;
 409         long long delta_time_ns;
 410
 411         delta_time_ns = ktime_to_ns(ktime_sub(now, s->last_updated));
 412
 413         pr_debug("sched_update_avail delta=%lld\n", delta_time_ns);
 414         if (delta_time_ns < 15000)
 415                 return 0;
 416
 417         for (i = 0; i < MAX_NPORTS; i++) {
 418                 struct sched_port *p = &s->p[i];
 419                 unsigned int delta_avail;
 420
 421                 delta_avail = (p->drain_bits_per_1024ns * delta_time_ns) >> 13;
 422                 p->avail = min(p->avail + delta_avail, s->max_avail);
 423         }
 424
 425         s->last_updated = now;
 426
 427         return 1;
 428 }
 429
 430 /*
 431  * sched_skb() is called from two different places. In the tx path, any
 432  * packet generating load on an output port will call sched_skb()
 433  * (skb != NULL). In addition, sched_skb() is called from the irq/soft irq
 434  * context (skb == NULL).
 435  * The scheduler only returns a skb (which will then be sent) if the
 436  * length of the skb is <= the current quota of the output port.
 437  */
 438 static struct sk_buff *sched_skb(struct sge *sge, struct sk_buff *skb,
 439                                 unsigned int credits)
 440 {
 441         struct sched *s = sge->tx_sched;
 442         struct sk_buff_head *skbq;
 443         unsigned int i, len, update = 1;
 444
 445         pr_debug("sched_skb %p\n", skb);
 446         if (!skb) {
 447                 if (!s->num)
 448                         return NULL;
 449         } else {
 450                 skbq = &s->p[skb->dev->if_port].skbq;
 451                 __skb_queue_tail(skbq, skb);
 452                 s->num++;
 453                 skb = NULL;
 454         }
 455
 456         if (credits < MAX_SKB_FRAGS + 1)
 457                 goto out;
 458
 459 again:
 460         for (i = 0; i < MAX_NPORTS; i++) {
 461                 s->port = ++s->port & (MAX_NPORTS - 1);
 462                 skbq = &s->p[s->port].skbq;
 463
 464                 skb = skb_peek(skbq);
 465
 466                 if (!skb)
 467                         continue;
 468
 469                 len = skb->len;
 470                 if (len <= s->p[s->port].avail) {
 471                         s->p[s->port].avail -= len;
 472                         s->num--;
 473                         __skb_unlink(skb, skbq);
 474                         goto out;
 475                 }
 476                 skb = NULL;
 477         }
 478
 479         if (update-- && sched_update_avail(sge))
 480                 goto again;
 481
 482 out:
 483         /* If there are more pending skbs, we use the hardware to schedule us
 484          * again.
 485          */
 486         if (s->num && !skb) {
 487                 struct cmdQ *q = &sge->cmdQ[0];
 488                 clear_bit(CMDQ_STAT_LAST_PKT_DB, &q->status);
 489                 if (test_and_set_bit(CMDQ_STAT_RUNNING, &q->status) == 0) {
 490                         set_bit(CMDQ_STAT_LAST_PKT_DB, &q->status);
 491                         writel(F_CMDQ0_ENABLE, sge->adapter->regs + A_SG_DOORBELL);
 492                 }
 493         }
 494         pr_debug("sched_skb ret %p\n", skb);
 495
 496         return skb;
 497 }
 498
 499 /*
 500  * PIO to indicate that memory mapped Q contains valid descriptor(s).
 501  */
 502 static inline void doorbell_pio(struct adapter *adapter, u32 val)
 503 {
 504         wmb();
 505         writel(val, adapter->regs + A_SG_DOORBELL);
 506 }
 507
 508 /*
 509  * Frees all RX buffers on the freelist Q. The caller must make sure that
 510  * the SGE is turned off before calling this function.
 511  */
 512 static void free_freelQ_buffers(struct pci_dev *pdev, struct freelQ *q)
 513 {
 514         unsigned int cidx = q->cidx;
 515
 516         while (q->credits--) {
 517                 struct freelQ_ce *ce = &q->centries[cidx];
 518
 519                 pci_unmap_single(pdev, pci_unmap_addr(ce, dma_addr),
 520                                  pci_unmap_len(ce, dma_len),
 521                                  PCI_DMA_FROMDEVICE);
 522                 dev_kfree_skb(ce->skb);
 523                 ce->skb = NULL;
 524                 if (++cidx == q->size)
 525                         cidx = 0;
 526         }
 527 }
 528
 529 /*
 530  * Free RX free list and response queue resources.
 531  */
 532 static void free_rx_resources(struct sge *sge)
 533 {
 534         struct pci_dev *pdev = sge->adapter->pdev;
 535         unsigned int size, i;
 536
 537         if (sge->respQ.entries) {
 538                 size = sizeof(struct respQ_e) * sge->respQ.size;
 539                 pci_free_consistent(pdev, size, sge->respQ.entries,
 540                                     sge->respQ.dma_addr);
 541         }
 542
 543         for (i = 0; i < SGE_FREELQ_N; i++) {
 544                 struct freelQ *q = &sge->freelQ[i];
 545
 546                 if (q->centries) {
 547                         free_freelQ_buffers(pdev, q);
 548                         kfree(q->centries);
 549                 }
 550                 if (q->entries) {
 551                         size = sizeof(struct freelQ_e) * q->size;
 552                         pci_free_consistent(pdev, size, q->entries,
 553                                             q->dma_addr);
 554                 }
 555         }
 556 }
 557
 558 /*
 559  * Allocates basic RX resources, consisting of memory mapped freelist Qs and a
 560  * response queue.
 561  */
 562 static int alloc_rx_resources(struct sge *sge, struct sge_params *p)
 563 {
 564         struct pci_dev *pdev = sge->adapter->pdev;
 565         unsigned int size, i;
 566
 567         for (i = 0; i < SGE_FREELQ_N; i++) {
 568                 struct freelQ *q = &sge->freelQ[i];
 569
 570                 q->genbit = 1;
 571                 q->size = p->freelQ_size[i];
 572                 q->dma_offset = sge->rx_pkt_pad ? 0 : NET_IP_ALIGN;
 573                 size = sizeof(struct freelQ_e) * q->size;
 574                 q->entries = pci_alloc_consistent(pdev, size, &q->dma_addr);
 575                 if (!q->entries)
 576                         goto err_no_mem;
 577
 578                 size = sizeof(struct freelQ_ce) * q->size;
 579                 q->centries = kzalloc(size, GFP_KERNEL);
 580                 if (!q->centries)
 581                         goto err_no_mem;
 582         }
 583
 584         /*
 585          * Calculate the buffer sizes for the two free lists.  FL0 accommodates
 586          * regular sized Ethernet frames, FL1 is sized not to exceed 16K,
 587          * including all the sk_buff overhead.
 588          *
 589          * Note: For T2 FL0 and FL1 are reversed.
 590          */
 591         sge->freelQ[!sge->jumbo_fl].rx_buffer_size = SGE_RX_SM_BUF_SIZE +
 592                 sizeof(struct cpl_rx_data) +
 593                 sge->freelQ[!sge->jumbo_fl].dma_offset;
 594
 595                 size = (16 * 1024) -
 596                     SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
 597
 598         sge->freelQ[sge->jumbo_fl].rx_buffer_size = size;
 599
 600         /*
 601          * Setup which skb recycle Q should be used when recycling buffers from
 602          * each free list.
 603          */
 604         sge->freelQ[!sge->jumbo_fl].recycleq_idx = 0;
 605         sge->freelQ[sge->jumbo_fl].recycleq_idx = 1;
 606
 607         sge->respQ.genbit = 1;
 608         sge->respQ.size = SGE_RESPQ_E_N;
 609         sge->respQ.credits = 0;
 610         size = sizeof(struct respQ_e) * sge->respQ.size;
 611         sge->respQ.entries =
 612                 pci_alloc_consistent(pdev, size, &sge->respQ.dma_addr);
 613         if (!sge->respQ.entries)
 614                 goto err_no_mem;
 615         return 0;
 616
 617 err_no_mem:
 618         free_rx_resources(sge);
 619         return -ENOMEM;
 620 }
 621
 622 /*
 623  * Reclaims n TX descriptors and frees the buffers associated with them.
 624  */
 625 static void free_cmdQ_buffers(struct sge *sge, struct cmdQ *q, unsigned int n)
 626 {
 627         struct cmdQ_ce *ce;
 628         struct pci_dev *pdev = sge->adapter->pdev;
 629         unsigned int cidx = q->cidx;
 630
 631         q->in_use -= n;
 632         ce = &q->centries[cidx];
 633         while (n--) {
 634                 if (likely(pci_unmap_len(ce, dma_len))) {
 635                         pci_unmap_single(pdev, pci_unmap_addr(ce, dma_addr),
 636                                          pci_unmap_len(ce, dma_len),
 637                                          PCI_DMA_TODEVICE);
 638                         if (q->sop)
 639                                 q->sop = 0;
 640                 }
 641                 if (ce->skb) {
 642                         dev_kfree_skb_any(ce->skb);
 643                         q->sop = 1;
 644                 }
 645                 ce++;
 646                 if (++cidx == q->size) {
 647                         cidx = 0;
 648                         ce = q->centries;
 649                 }
 650         }
 651         q->cidx = cidx;
 652 }
 653
 654 /*
 655  * Free TX resources.
 656  *
 657  * Assumes that SGE is stopped and all interrupts are disabled.
 658  */
 659 static void free_tx_resources(struct sge *sge)
 660 {
 661         struct pci_dev *pdev = sge->adapter->pdev;
 662         unsigned int size, i;
 663
 664         for (i = 0; i < SGE_CMDQ_N; i++) {
 665                 struct cmdQ *q = &sge->cmdQ[i];
 666
 667                 if (q->centries) {
 668                         if (q->in_use)
 669                                 free_cmdQ_buffers(sge, q, q->in_use);
 670                         kfree(q->centries);
 671                 }
 672                 if (q->entries) {
 673                         size = sizeof(struct cmdQ_e) * q->size;
 674                         pci_free_consistent(pdev, size, q->entries,
 675                                             q->dma_addr);
 676                 }
 677         }
 678 }
 679
 680 /*
 681  * Allocates basic TX resources, consisting of memory mapped command Qs.
 682  */
 683 static int alloc_tx_resources(struct sge *sge, struct sge_params *p)
 684 {
 685         struct pci_dev *pdev = sge->adapter->pdev;
 686         unsigned int size, i;
 687
 688         for (i = 0; i < SGE_CMDQ_N; i++) {
 689                 struct cmdQ *q = &sge->cmdQ[i];
 690
 691                 q->genbit = 1;
 692                 q->sop = 1;
 693                 q->size = p->cmdQ_size[i];
 694                 q->in_use = 0;
 695                 q->status = 0;
 696                 q->processed = q->cleaned = 0;
 697                 q->stop_thres = 0;
 698                 spin_lock_init(&q->lock);
 699                 size = sizeof(struct cmdQ_e) * q->size;
 700                 q->entries = pci_alloc_consistent(pdev, size, &q->dma_addr);
 701                 if (!q->entries)
 702                         goto err_no_mem;
 703
 704                 size = sizeof(struct cmdQ_ce) * q->size;
 705                 q->centries = kzalloc(size, GFP_KERNEL);
 706                 if (!q->centries)
 707                         goto err_no_mem;
 708         }
 709
 710         /*
 711          * CommandQ 0 handles Ethernet and TOE packets, while queue 1 is TOE
 712          * only.  For queue 0 set the stop threshold so we can handle one more
 713          * packet from each port, plus reserve an additional 24 entries for
 714          * Ethernet packets only.  Queue 1 never suspends nor do we reserve
 715          * space for Ethernet packets.
 716          */
 717         sge->cmdQ[0].stop_thres = sge->adapter->params.nports *
 718                 (MAX_SKB_FRAGS + 1);
 719         return 0;
 720
 721 err_no_mem:
 722         free_tx_resources(sge);
 723         return -ENOMEM;
 724 }
 725
 726 static inline void setup_ring_params(struct adapter *adapter, u64 addr,
 727                                      u32 size, int base_reg_lo,
 728                                      int base_reg_hi, int size_reg)
 729 {
 730         writel((u32)addr, adapter->regs + base_reg_lo);
 731         writel(addr >> 32, adapter->regs + base_reg_hi);
 732         writel(size, adapter->regs + size_reg);
 733 }
 734
 735 /*
 736  * Enable/disable VLAN acceleration.
 737  */
 738 void t1_set_vlan_accel(struct adapter *adapter, int on_off)
 739 {
 740         struct sge *sge = adapter->sge;
 741
 742         sge->sge_control &= ~F_VLAN_XTRACT;
 743         if (on_off)
 744                 sge->sge_control |= F_VLAN_XTRACT;
 745         if (adapter->open_device_map) {
 746                 writel(sge->sge_control, adapter->regs + A_SG_CONTROL);
 747                 readl(adapter->regs + A_SG_CONTROL);   /* flush */
 748         }
 749 }
 750
 751 /*
 752  * Programs the various SGE registers. However, the engine is not yet enabled,
 753  * but sge->sge_control is setup and ready to go.
 754  */
 755 static void configure_sge(struct sge *sge, struct sge_params *p)
 756 {
 757         struct adapter *ap = sge->adapter;
 758
 759         writel(0, ap->regs + A_SG_CONTROL);
 760         setup_ring_params(ap, sge->cmdQ[0].dma_addr, sge->cmdQ[0].size,
 761                           A_SG_CMD0BASELWR, A_SG_CMD0BASEUPR, A_SG_CMD0SIZE);
 762         setup_ring_params(ap, sge->cmdQ[1].dma_addr, sge->cmdQ[1].size,
 763                           A_SG_CMD1BASELWR, A_SG_CMD1BASEUPR, A_SG_CMD1SIZE);
 764         setup_ring_params(ap, sge->freelQ[0].dma_addr,
 765                           sge->freelQ[0].size, A_SG_FL0BASELWR,
 766                           A_SG_FL0BASEUPR, A_SG_FL0SIZE);
 767         setup_ring_params(ap, sge->freelQ[1].dma_addr,
 768                           sge->freelQ[1].size, A_SG_FL1BASELWR,
 769                           A_SG_FL1BASEUPR, A_SG_FL1SIZE);
 770
 771         /* The threshold comparison uses <. */
 772         writel(SGE_RX_SM_BUF_SIZE + 1, ap->regs + A_SG_FLTHRESHOLD);
 773
 774         setup_ring_params(ap, sge->respQ.dma_addr, sge->respQ.size,
 775                           A_SG_RSPBASELWR, A_SG_RSPBASEUPR, A_SG_RSPSIZE);
 776         writel((u32)sge->respQ.size - 1, ap->regs + A_SG_RSPQUEUECREDIT);
 777
 778         sge->sge_control = F_CMDQ0_ENABLE | F_CMDQ1_ENABLE | F_FL0_ENABLE |
 779                 F_FL1_ENABLE | F_CPL_ENABLE | F_RESPONSE_QUEUE_ENABLE |
 780                 V_CMDQ_PRIORITY(2) | F_DISABLE_CMDQ1_GTS | F_ISCSI_COALESCE |
 781                 V_RX_PKT_OFFSET(sge->rx_pkt_pad);
 782
 783 #if defined(__BIG_ENDIAN_BITFIELD)
 784         sge->sge_control |= F_ENABLE_BIG_ENDIAN;
 785 #endif
 786
 787         /* Initialize no-resource timer */
 788         sge->intrtimer_nres = SGE_INTRTIMER_NRES * core_ticks_per_usec(ap);
 789
 790         t1_sge_set_coalesce_params(sge, p);
 791 }
 792
 793 /*
 794  * Return the payload capacity of the jumbo free-list buffers.
 795  */
 796 static inline unsigned int jumbo_payload_capacity(const struct sge *sge)
 797 {
 798         return sge->freelQ[sge->jumbo_fl].rx_buffer_size -
 799                 sge->freelQ[sge->jumbo_fl].dma_offset -
 800                 sizeof(struct cpl_rx_data);
 801 }
 802
 803 /*
 804  * Frees all SGE related resources and the sge structure itself
 805  */
 806 void t1_sge_destroy(struct sge *sge)
 807 {
 808         int i;
 809
 810         for_each_port(sge->adapter, i)
 811                 free_percpu(sge->port_stats[i]);
 812
 813         kfree(sge->tx_sched);
 814         free_tx_resources(sge);
 815         free_rx_resources(sge);
 816         kfree(sge);
 817 }
 818
 819 /*
 820  * Allocates new RX buffers on the freelist Q (and tracks them on the freelist
 821  * context Q) until the Q is full or alloc_skb fails.
 822  *
 823  * It is possible that the generation bits already match, indicating that the
 824  * buffer is already valid and nothing needs to be done. This happens when we
 825  * copied a received buffer into a new sk_buff during the interrupt processing.
 826  *
 827  * If the SGE doesn't automatically align packets properly (!sge->rx_pkt_pad),
 828  * we specify a RX_OFFSET in order to make sure that the IP header is 4B
 829  * aligned.
 830  */
 831 static void refill_free_list(struct sge *sge, struct freelQ *q)
 832 {
 833         struct pci_dev *pdev = sge->adapter->pdev;
 834         struct freelQ_ce *ce = &q->centries[q->pidx];
 835         struct freelQ_e *e = &q->entries[q->pidx];
 836         unsigned int dma_len = q->rx_buffer_size - q->dma_offset;
 837
 838         while (q->credits < q->size) {
 839                 struct sk_buff *skb;
 840                 dma_addr_t mapping;
 841
 842                 skb = alloc_skb(q->rx_buffer_size, GFP_ATOMIC);
 843                 if (!skb)
 844                         break;
 845
 846                 skb_reserve(skb, q->dma_offset);
 847                 mapping = pci_map_single(pdev, skb->data, dma_len,
 848                                          PCI_DMA_FROMDEVICE);
 849                 ce->skb = skb;
 850                 pci_unmap_addr_set(ce, dma_addr, mapping);
 851                 pci_unmap_len_set(ce, dma_len, dma_len);
 852                 e->addr_lo = (u32)mapping;
 853                 e->addr_hi = (u64)mapping >> 32;
 854                 e->len_gen = V_CMD_LEN(dma_len) | V_CMD_GEN1(q->genbit);
 855                 wmb();
 856                 e->gen2 = V_CMD_GEN2(q->genbit);
 857
 858                 e++;
 859                 ce++;
 860                 if (++q->pidx == q->size) {
 861                         q->pidx = 0;
 862                         q->genbit ^= 1;
 863                         ce = q->centries;
 864                         e = q->entries;
 865                 }
 866                 q->credits++;
 867         }
 868 }
 869
 870 /*
 871  * Calls refill_free_list for both free lists. If we cannot fill at least 1/4
 872  * of both rings, we go into 'few interrupt mode' in order to give the system
 873  * time to free up resources.
 874  */
 875 static void freelQs_empty(struct sge *sge)
 876 {
 877         struct adapter *adapter = sge->adapter;
 878         u32 irq_reg = readl(adapter->regs + A_SG_INT_ENABLE);
 879         u32 irqholdoff_reg;
 880
 881         refill_free_list(sge, &sge->freelQ[0]);
 882         refill_free_list(sge, &sge->freelQ[1]);
 883
 884         if (sge->freelQ[0].credits > (sge->freelQ[0].size >> 2) &&
 885             sge->freelQ[1].credits > (sge->freelQ[1].size >> 2)) {
 886                 irq_reg |= F_FL_EXHAUSTED;
 887                 irqholdoff_reg = sge->fixed_intrtimer;
 888         } else {
 889                 /* Clear the F_FL_EXHAUSTED interrupts for now */
 890                 irq_reg &= ~F_FL_EXHAUSTED;
 891                 irqholdoff_reg = sge->intrtimer_nres;
 892         }
 893         writel(irqholdoff_reg, adapter->regs + A_SG_INTRTIMER);
 894         writel(irq_reg, adapter->regs + A_SG_INT_ENABLE);
 895
 896         /* We reenable the Qs to force a freelist GTS interrupt later */
 897         doorbell_pio(adapter, F_FL0_ENABLE | F_FL1_ENABLE);
 898 }
 899
 900 #define SGE_PL_INTR_MASK (F_PL_INTR_SGE_ERR | F_PL_INTR_SGE_DATA)
 901 #define SGE_INT_FATAL (F_RESPQ_OVERFLOW | F_PACKET_TOO_BIG | F_PACKET_MISMATCH)
 902 #define SGE_INT_ENABLE (F_RESPQ_EXHAUSTED | F_RESPQ_OVERFLOW | \
 903                         F_FL_EXHAUSTED | F_PACKET_TOO_BIG | F_PACKET_MISMATCH)
 904
 905 /*
 906  * Disable SGE Interrupts
 907  */
 908 void t1_sge_intr_disable(struct sge *sge)
 909 {
 910         u32 val = readl(sge->adapter->regs + A_PL_ENABLE);
 911
 912         writel(val & ~SGE_PL_INTR_MASK, sge->adapter->regs + A_PL_ENABLE);
 913         writel(0, sge->adapter->regs + A_SG_INT_ENABLE);
 914 }
 915
 916 /*
 917  * Enable SGE interrupts.
 918  */
 919 void t1_sge_intr_enable(struct sge *sge)
 920 {
 921         u32 en = SGE_INT_ENABLE;
 922         u32 val = readl(sge->adapter->regs + A_PL_ENABLE);
 923
 924         if (sge->adapter->flags & TSO_CAPABLE)
 925                 en &= ~F_PACKET_TOO_BIG;
 926         writel(en, sge->adapter->regs + A_SG_INT_ENABLE);
 927         writel(val | SGE_PL_INTR_MASK, sge->adapter->regs + A_PL_ENABLE);
 928 }
 929
 930 /*
 931  * Clear SGE interrupts.
 932  */
 933 void t1_sge_intr_clear(struct sge *sge)
 934 {
 935         writel(SGE_PL_INTR_MASK, sge->adapter->regs + A_PL_CAUSE);
 936         writel(0xffffffff, sge->adapter->regs + A_SG_INT_CAUSE);
 937 }
 938
 939 /*
 940  * SGE 'Error' interrupt handler
 941  */
 942 int t1_sge_intr_error_handler(struct sge *sge)
 943 {
 944         struct adapter *adapter = sge->adapter;
 945         u32 cause = readl(adapter->regs + A_SG_INT_CAUSE);
 946
 947         if (adapter->flags & TSO_CAPABLE)
 948                 cause &= ~F_PACKET_TOO_BIG;
 949         if (cause & F_RESPQ_EXHAUSTED)
 950                 sge->stats.respQ_empty++;
 951         if (cause & F_RESPQ_OVERFLOW) {
 952                 sge->stats.respQ_overflow++;
 953                 CH_ALERT("%s: SGE response queue overflow\n",
 954                          adapter->name);
 955         }
 956         if (cause & F_FL_EXHAUSTED) {
 957                 sge->stats.freelistQ_empty++;
 958                 freelQs_empty(sge);
 959         }
 960         if (cause & F_PACKET_TOO_BIG) {
 961                 sge->stats.pkt_too_big++;
 962                 CH_ALERT("%s: SGE max packet size exceeded\n",
 963                          adapter->name);
 964         }
 965         if (cause & F_PACKET_MISMATCH) {
 966                 sge->stats.pkt_mismatch++;
 967                 CH_ALERT("%s: SGE packet mismatch\n", adapter->name);
 968         }
 969         if (cause & SGE_INT_FATAL)
 970                 t1_fatal_err(adapter);
 971
 972         writel(cause, adapter->regs + A_SG_INT_CAUSE);
 973         return 0;
 974 }
 975
 976 const struct sge_intr_counts *t1_sge_get_intr_counts(const struct sge *sge)
 977 {
 978         return &sge->stats;
 979 }
 980
 981 void t1_sge_get_port_stats(const struct sge *sge, int port,
 982                            struct sge_port_stats *ss)
 983 {
 984         int cpu;
 985
 986         memset(ss, 0, sizeof(*ss));
 987         for_each_possible_cpu(cpu) {
 988                 struct sge_port_stats *st = per_cpu_ptr(sge->port_stats[port], cpu);
 989
 990                 ss->rx_packets += st->rx_packets;
 991                 ss->rx_cso_good += st->rx_cso_good;
 992                 ss->tx_packets += st->tx_packets;
 993                 ss->tx_cso += st->tx_cso;
 994                 ss->tx_tso += st->tx_tso;
 995                 ss->vlan_xtract += st->vlan_xtract;
 996                 ss->vlan_insert += st->vlan_insert;
 997         }
 998 }
 999
1000 /**
1001  *      recycle_fl_buf - recycle a free list buffer
1002  *      @fl: the free list
1003  *      @idx: index of buffer to recycle
1004  *
1005  *      Recycles the specified buffer on the given free list by adding it at
1006  *      the next available slot on the list.
1007  */
1008 static void recycle_fl_buf(struct freelQ *fl, int idx)
1009 {
1010         struct freelQ_e *from = &fl->entries[idx];
1011         struct freelQ_e *to = &fl->entries[fl->pidx];
1012
1013         fl->centries[fl->pidx] = fl->centries[idx];
1014         to->addr_lo = from->addr_lo;
1015         to->addr_hi = from->addr_hi;
1016         to->len_gen = G_CMD_LEN(from->len_gen) | V_CMD_GEN1(fl->genbit);
1017         wmb();
1018         to->gen2 = V_CMD_GEN2(fl->genbit);
1019         fl->credits++;
1020
1021         if (++fl->pidx == fl->size) {
1022                 fl->pidx = 0;
1023                 fl->genbit ^= 1;
1024         }
1025 }
1026
1027 /**
1028  *      get_packet - return the next ingress packet buffer
1029  *      @pdev: the PCI device that received the packet
1030  *      @fl: the SGE free list holding the packet
1031  *      @len: the actual packet length, excluding any SGE padding
1032  *      @dma_pad: padding at beginning of buffer left by SGE DMA
1033  *      @skb_pad: padding to be used if the packet is copied
1034  *      @copy_thres: length threshold under which a packet should be copied
1035  *      @drop_thres: # of remaining buffers before we start dropping packets
1036  *
1037  *      Get the next packet from a free list and complete setup of the
1038  *      sk_buff.  If the packet is small we make a copy and recycle the
1039  *      original buffer, otherwise we use the original buffer itself.  If a
1040  *      positive drop threshold is supplied packets are dropped and their
1041  *      buffers recycled if (a) the number of remaining buffers is under the
1042  *      threshold and the packet is too big to copy, or (b) the packet should
1043  *      be copied but there is no memory for the copy.
1044  */
1045 static inline struct sk_buff *get_packet(struct pci_dev *pdev,
1046                                          struct freelQ *fl, unsigned int len,
1047                                          int dma_pad, int skb_pad,
1048                                          unsigned int copy_thres,
1049                                          unsigned int drop_thres)
1050 {
1051         struct sk_buff *skb;
1052         struct freelQ_ce *ce = &fl->centries[fl->cidx];
1053
1054         if (len < copy_thres) {
1055                 skb = alloc_skb(len + skb_pad, GFP_ATOMIC);
1056                 if (likely(skb != NULL)) {
1057                         skb_reserve(skb, skb_pad);
1058                         skb_put(skb, len);
1059                         pci_dma_sync_single_for_cpu(pdev,
1060                                             pci_unmap_addr(ce, dma_addr),
1061                                             pci_unmap_len(ce, dma_len),
1062                                             PCI_DMA_FROMDEVICE);
1063                         memcpy(skb->data, ce->skb->data + dma_pad, len);
1064                         pci_dma_sync_single_for_device(pdev,
1065                                             pci_unmap_addr(ce, dma_addr),
1066                                             pci_unmap_len(ce, dma_len),
1067                                             PCI_DMA_FROMDEVICE);
1068                 } else if (!drop_thres)
1069                         goto use_orig_buf;
1070
1071                 recycle_fl_buf(fl, fl->cidx);
1072                 return skb;
1073         }
1074
1075         if (fl->credits < drop_thres) {
1076                 recycle_fl_buf(fl, fl->cidx);
1077                 return NULL;
1078         }
1079
1080 use_orig_buf:
1081         pci_unmap_single(pdev, pci_unmap_addr(ce, dma_addr),
1082                          pci_unmap_len(ce, dma_len), PCI_DMA_FROMDEVICE);
1083         skb = ce->skb;
1084         skb_reserve(skb, dma_pad);
1085         skb_put(skb, len);
1086         return skb;
1087 }
1088
1089 /**
1090  *      unexpected_offload - handle an unexpected offload packet
1091  *      @adapter: the adapter
1092  *      @fl: the free list that received the packet
1093  *
1094  *      Called when we receive an unexpected offload packet (e.g., the TOE
1095  *      function is disabled or the card is a NIC).  Prints a message and
1096  *      recycles the buffer.
1097  */
1098 static void unexpected_offload(struct adapter *adapter, struct freelQ *fl)
1099 {
1100         struct freelQ_ce *ce = &fl->centries[fl->cidx];
1101         struct sk_buff *skb = ce->skb;
1102
1103         pci_dma_sync_single_for_cpu(adapter->pdev, pci_unmap_addr(ce, dma_addr),
1104                             pci_unmap_len(ce, dma_len), PCI_DMA_FROMDEVICE);
1105         CH_ERR("%s: unexpected offload packet, cmd %u\n",
1106                adapter->name, *skb->data);
1107         recycle_fl_buf(fl, fl->cidx);
1108 }
1109
1110 /*
1111  * T1/T2 SGE limits the maximum DMA size per TX descriptor to
1112  * SGE_TX_DESC_MAX_PLEN (16KB). If the PAGE_SIZE is larger than 16KB, the
1113  * stack might send more than SGE_TX_DESC_MAX_PLEN in a contiguous manner.
1114  * Note that the *_large_page_tx_descs stuff will be optimized out when
1115  * PAGE_SIZE <= SGE_TX_DESC_MAX_PLEN.
1116  *
1117  * compute_large_page_descs() computes how many additional descriptors are
1118  * required to break down the stack's request.
1119  */
1120 static inline unsigned int compute_large_page_tx_descs(struct sk_buff *skb)
1121 {
1122         unsigned int count = 0;
1123
1124         if (PAGE_SIZE > SGE_TX_DESC_MAX_PLEN) {
1125                 unsigned int nfrags = skb_shinfo(skb)->nr_frags;
1126                 unsigned int i, len = skb->len - skb->data_len;
1127                 while (len > SGE_TX_DESC_MAX_PLEN) {
1128                         count++;
1129                         len -= SGE_TX_DESC_MAX_PLEN;
1130                 }
1131                 for (i = 0; nfrags--; i++) {
1132                         skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
1133                         len = frag->size;
1134                         while (len > SGE_TX_DESC_MAX_PLEN) {
1135                                 count++;
1136                                 len -= SGE_TX_DESC_MAX_PLEN;
1137                         }
1138                 }
1139         }
1140         return count;
1141 }
1142
1143 /*
1144  * Write a cmdQ entry.
1145  *
1146  * Since this function writes the 'flags' field, it must not be used to
1147  * write the first cmdQ entry.
1148  */
1149 static inline void write_tx_desc(struct cmdQ_e *e, dma_addr_t mapping,
1150                                  unsigned int len, unsigned int gen,
1151                                  unsigned int eop)
1152 {
1153         if (unlikely(len > SGE_TX_DESC_MAX_PLEN))
1154                 BUG();
1155         e->addr_lo = (u32)mapping;
1156         e->addr_hi = (u64)mapping >> 32;
1157         e->len_gen = V_CMD_LEN(len) | V_CMD_GEN1(gen);
1158         e->flags = F_CMD_DATAVALID | V_CMD_EOP(eop) | V_CMD_GEN2(gen);
1159 }
1160
1161 /*
1162  * See comment for previous function.
1163  *
1164  * write_tx_descs_large_page() writes additional SGE tx descriptors if
1165  * *desc_len exceeds HW's capability.
1166  */
1167 static inline unsigned int write_large_page_tx_descs(unsigned int pidx,
1168                                                      struct cmdQ_e **e,
1169                                                      struct cmdQ_ce **ce,
1170                                                      unsigned int *gen,
1171                                                      dma_addr_t *desc_mapping,
1172                                                      unsigned int *desc_len,
1173                                                      unsigned int nfrags,
1174                                                      struct cmdQ *q)
1175 {
1176         if (PAGE_SIZE > SGE_TX_DESC_MAX_PLEN) {
1177                 struct cmdQ_e *e1 = *e;
1178                 struct cmdQ_ce *ce1 = *ce;
1179
1180                 while (*desc_len > SGE_TX_DESC_MAX_PLEN) {
1181                         *desc_len -= SGE_TX_DESC_MAX_PLEN;
1182                         write_tx_desc(e1, *desc_mapping, SGE_TX_DESC_MAX_PLEN,
1183                                       *gen, nfrags == 0 && *desc_len == 0);
1184                         ce1->skb = NULL;
1185                         pci_unmap_len_set(ce1, dma_len, 0);
1186                         *desc_mapping += SGE_TX_DESC_MAX_PLEN;
1187                         if (*desc_len) {
1188                                 ce1++;
1189                                 e1++;
1190                                 if (++pidx == q->size) {
1191                                         pidx = 0;
1192                                         *gen ^= 1;
1193                                         ce1 = q->centries;
1194                                         e1 = q->entries;
1195                                 }
1196                         }
1197                 }
1198                 *e = e1;
1199                 *ce = ce1;
1200         }
1201         return pidx;
1202 }
1203
1204 /*
1205  * Write the command descriptors to transmit the given skb starting at
1206  * descriptor pidx with the given generation.
1207  */
1208 static inline void write_tx_descs(struct adapter *adapter, struct sk_buff *skb,
1209                                   unsigned int pidx, unsigned int gen,
1210                                   struct cmdQ *q)
1211 {
1212         dma_addr_t mapping, desc_mapping;
1213         struct cmdQ_e *e, *e1;
1214         struct cmdQ_ce *ce;
1215         unsigned int i, flags, first_desc_len, desc_len,
1216             nfrags = skb_shinfo(skb)->nr_frags;
1217
1218         e = e1 = &q->entries[pidx];
1219         ce = &q->centries[pidx];
1220
1221         mapping = pci_map_single(adapter->pdev, skb->data,
1222                                 skb->len - skb->data_len, PCI_DMA_TODEVICE);
1223
1224         desc_mapping = mapping;
1225         desc_len = skb->len - skb->data_len;
1226
1227         flags = F_CMD_DATAVALID | F_CMD_SOP |
1228             V_CMD_EOP(nfrags == 0 && desc_len <= SGE_TX_DESC_MAX_PLEN) |
1229             V_CMD_GEN2(gen);
1230         first_desc_len = (desc_len <= SGE_TX_DESC_MAX_PLEN) ?
1231             desc_len : SGE_TX_DESC_MAX_PLEN;
1232         e->addr_lo = (u32)desc_mapping;
1233         e->addr_hi = (u64)desc_mapping >> 32;
1234         e->len_gen = V_CMD_LEN(first_desc_len) | V_CMD_GEN1(gen);
1235         ce->skb = NULL;
1236         pci_unmap_len_set(ce, dma_len, 0);
1237
1238         if (PAGE_SIZE > SGE_TX_DESC_MAX_PLEN &&
1239             desc_len > SGE_TX_DESC_MAX_PLEN) {
1240                 desc_mapping += first_desc_len;
1241                 desc_len -= first_desc_len;
1242                 e1++;
1243                 ce++;
1244                 if (++pidx == q->size) {
1245                         pidx = 0;
1246                         gen ^= 1;
1247                         e1 = q->entries;
1248                         ce = q->centries;
1249                 }
1250                 pidx = write_large_page_tx_descs(pidx, &e1, &ce, &gen,
1251                                                  &desc_mapping, &desc_len,
1252                                                  nfrags, q);
1253
1254                 if (likely(desc_len))
1255                         write_tx_desc(e1, desc_mapping, desc_len, gen,
1256                                       nfrags == 0);
1257         }
1258
1259         ce->skb = NULL;
1260         pci_unmap_addr_set(ce, dma_addr, mapping);
1261         pci_unmap_len_set(ce, dma_len, skb->len - skb->data_len);
1262
1263         for (i = 0; nfrags--; i++) {
1264                 skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
1265                 e1++;
1266                 ce++;
1267                 if (++pidx == q->size) {
1268                         pidx = 0;
1269                         gen ^= 1;
1270                         e1 = q->entries;
1271                         ce = q->centries;
1272                 }
1273
1274                 mapping = pci_map_page(adapter->pdev, frag->page,
1275                                        frag->page_offset, frag->size,
1276                                        PCI_DMA_TODEVICE);
1277                 desc_mapping = mapping;
1278                 desc_len = frag->size;
1279
1280                 pidx = write_large_page_tx_descs(pidx, &e1, &ce, &gen,
1281                                                  &desc_mapping, &desc_len,
1282                                                  nfrags, q);
1283                 if (likely(desc_len))
1284                         write_tx_desc(e1, desc_mapping, desc_len, gen,
1285                                       nfrags == 0);
1286                 ce->skb = NULL;
1287                 pci_unmap_addr_set(ce, dma_addr, mapping);
1288                 pci_unmap_len_set(ce, dma_len, frag->size);
1289         }
1290         ce->skb = skb;
1291         wmb();
1292         e->flags = flags;
1293 }
1294
1295 /*
1296  * Clean up completed Tx buffers.
1297  */
1298 static inline void reclaim_completed_tx(struct sge *sge, struct cmdQ *q)
1299 {
1300         unsigned int reclaim = q->processed - q->cleaned;
1301
1302         if (reclaim) {
1303                 pr_debug("reclaim_completed_tx processed:%d cleaned:%d\n",
1304                          q->processed, q->cleaned);
1305                 free_cmdQ_buffers(sge, q, reclaim);
1306                 q->cleaned += reclaim;
1307         }
1308 }
1309
1310 /*
1311  * Called from tasklet. Checks the scheduler for any
1312  * pending skbs that can be sent.
1313  */
1314 static void restart_sched(unsigned long arg)
1315 {
1316         struct sge *sge = (struct sge *) arg;
1317         struct adapter *adapter = sge->adapter;
1318         struct cmdQ *q = &sge->cmdQ[0];
1319         struct sk_buff *skb;
1320         unsigned int credits, queued_skb = 0;
1321
1322         spin_lock(&q->lock);
1323         reclaim_completed_tx(sge, q);
1324
1325         credits = q->size - q->in_use;
1326         pr_debug("restart_sched credits=%d\n", credits);
1327         while ((skb = sched_skb(sge, NULL, credits)) != NULL) {
1328                 unsigned int genbit, pidx, count;
1329                 count = 1 + skb_shinfo(skb)->nr_frags;
1330                 count += compute_large_page_tx_descs(skb);
1331                 q->in_use += count;
1332                 genbit = q->genbit;
1333                 pidx = q->pidx;
1334                 q->pidx += count;
1335                 if (q->pidx >= q->size) {
1336                         q->pidx -= q->size;
1337                         q->genbit ^= 1;
1338                 }
1339                 write_tx_descs(adapter, skb, pidx, genbit, q);
1340                 credits = q->size - q->in_use;
1341                 queued_skb = 1;
1342         }
1343
1344         if (queued_skb) {
1345                 clear_bit(CMDQ_STAT_LAST_PKT_DB, &q->status);
1346                 if (test_and_set_bit(CMDQ_STAT_RUNNING, &q->status) == 0) {
1347                         set_bit(CMDQ_STAT_LAST_PKT_DB, &q->status);
1348                         writel(F_CMDQ0_ENABLE, adapter->regs + A_SG_DOORBELL);
1349                 }
1350         }
1351         spin_unlock(&q->lock);
1352 }
1353
1354 /**
1355  *      sge_rx - process an ingress ethernet packet
1356  *      @sge: the sge structure
1357  *      @fl: the free list that contains the packet buffer
1358  *      @len: the packet length
1359  *
1360  *      Process an ingress ethernet pakcet and deliver it to the stack.
1361  */
1362 static int sge_rx(struct sge *sge, struct freelQ *fl, unsigned int len)
1363 {
1364         struct sk_buff *skb;
1365         struct cpl_rx_pkt *p;
1366         struct adapter *adapter = sge->adapter;
1367         struct sge_port_stats *st;
1368
1369         skb = get_packet(adapter->pdev, fl, len - sge->rx_pkt_pad,
1370                          sge->rx_pkt_pad, 2, SGE_RX_COPY_THRES,
1371                          SGE_RX_DROP_THRES);
1372         if (unlikely(!skb)) {
1373                 sge->stats.rx_drops++;
1374                 return 0;
1375         }
1376
1377         p = (struct cpl_rx_pkt *)skb->data;
1378         skb_pull(skb, sizeof(*p));
1379         if (p->iff >= adapter->params.nports) {
1380                 kfree_skb(skb);
1381                 return 0;
1382         }
1383
1384         skb->dev = adapter->port[p->iff].dev;
1385         skb->dev->last_rx = jiffies;
1386         st = per_cpu_ptr(sge->port_stats[p->iff], smp_processor_id());
1387         st->rx_packets++;
1388
1389         skb->protocol = eth_type_trans(skb, skb->dev);
1390         if ((adapter->flags & RX_CSUM_ENABLED) && p->csum == 0xffff &&
1391             skb->protocol == htons(ETH_P_IP) &&
1392             (skb->data[9] == IPPROTO_TCP || skb->data[9] == IPPROTO_UDP)) {
1393                 ++st->rx_cso_good;
1394                 skb->ip_summed = CHECKSUM_UNNECESSARY;
1395         } else
1396                 skb->ip_summed = CHECKSUM_NONE;
1397
1398         if (unlikely(adapter->vlan_grp && p->vlan_valid)) {
1399                 st->vlan_xtract++;
1400 #ifdef CONFIG_CHELSIO_T1_NAPI
1401                         vlan_hwaccel_receive_skb(skb, adapter->vlan_grp,
1402                                                  ntohs(p->vlan));
1403 #else
1404                         vlan_hwaccel_rx(skb, adapter->vlan_grp,
1405                                         ntohs(p->vlan));
1406 #endif
1407         } else {
1408 #ifdef CONFIG_CHELSIO_T1_NAPI
1409                 netif_receive_skb(skb);
1410 #else
1411                 netif_rx(skb);
1412 #endif
1413         }
1414         return 0;
1415 }
1416
1417 /*
1418  * Returns true if a command queue has enough available descriptors that
1419  * we can resume Tx operation after temporarily disabling its packet queue.
1420  */
1421 static inline int enough_free_Tx_descs(const struct cmdQ *q)
1422 {
1423         unsigned int r = q->processed - q->cleaned;
1424
1425         return q->in_use - r < (q->size >> 1);
1426 }
1427
1428 /*
1429  * Called when sufficient space has become available in the SGE command queues
1430  * after the Tx packet schedulers have been suspended to restart the Tx path.
1431  */
1432 static void restart_tx_queues(struct sge *sge)
1433 {
1434         struct adapter *adap = sge->adapter;
1435         int i;
1436
1437         if (!enough_free_Tx_descs(&sge->cmdQ[0]))
1438                 return;
1439
1440         for_each_port(adap, i) {
1441                 struct net_device *nd = adap->port[i].dev;
1442
1443                 if (test_and_clear_bit(nd->if_port, &sge->stopped_tx_queues) &&
1444                     netif_running(nd)) {
1445                         sge->stats.cmdQ_restarted[2]++;
1446                         netif_wake_queue(nd);
1447                 }
1448         }
1449 }
1450
1451 /*
1452  * update_tx_info is called from the interrupt handler/NAPI to return cmdQ0
1453  * information.
1454  */
1455 static unsigned int update_tx_info(struct adapter *adapter,
1456                                           unsigned int flags,
1457                                           unsigned int pr0)
1458 {
1459         struct sge *sge = adapter->sge;
1460         struct cmdQ *cmdq = &sge->cmdQ[0];
1461
1462         cmdq->processed += pr0;
1463         if (flags & (F_FL0_ENABLE | F_FL1_ENABLE)) {
1464                 freelQs_empty(sge);
1465                 flags &= ~(F_FL0_ENABLE | F_FL1_ENABLE);
1466         }
1467         if (flags & F_CMDQ0_ENABLE) {
1468                 clear_bit(CMDQ_STAT_RUNNING, &cmdq->status);
1469
1470                 if (cmdq->cleaned + cmdq->in_use != cmdq->processed &&
1471                     !test_and_set_bit(CMDQ_STAT_LAST_PKT_DB, &cmdq->status)) {
1472                         set_bit(CMDQ_STAT_RUNNING, &cmdq->status);
1473                         writel(F_CMDQ0_ENABLE, adapter->regs + A_SG_DOORBELL);
1474                 }
1475                 if (sge->tx_sched)
1476                         tasklet_hi_schedule(&sge->tx_sched->sched_tsk);
1477
1478                 flags &= ~F_CMDQ0_ENABLE;
1479         }
1480
1481         if (unlikely(sge->stopped_tx_queues != 0))
1482                 restart_tx_queues(sge);
1483
1484         return flags;
1485 }
1486
1487 /*
1488  * Process SGE responses, up to the supplied budget.  Returns the number of
1489  * responses processed.  A negative budget is effectively unlimited.
1490  */
1491 static int process_responses(struct adapter *adapter, int budget)
1492 {
1493         struct sge *sge = adapter->sge;
1494         struct respQ *q = &sge->respQ;
1495         struct respQ_e *e = &q->entries[q->cidx];
1496         int budget_left = budget;
1497         unsigned int flags = 0;
1498         unsigned int cmdq_processed[SGE_CMDQ_N] = {0, 0};
1499
1500
1501         while (likely(budget_left && e->GenerationBit == q->genbit)) {
1502                 flags |= e->Qsleeping;
1503
1504                 cmdq_processed[0] += e->Cmdq0CreditReturn;
1505                 cmdq_processed[1] += e->Cmdq1CreditReturn;
1506
1507                 /* We batch updates to the TX side to avoid cacheline
1508                  * ping-pong of TX state information on MP where the sender
1509                  * might run on a different CPU than this function...
1510                  */
1511                 if (unlikely(flags & F_CMDQ0_ENABLE || cmdq_processed[0] > 64)) {
1512                         flags = update_tx_info(adapter, flags, cmdq_processed[0]);
1513                         cmdq_processed[0] = 0;
1514                 }
1515                 if (unlikely(cmdq_processed[1] > 16)) {
1516                         sge->cmdQ[1].processed += cmdq_processed[1];
1517                         cmdq_processed[1] = 0;
1518                 }
1519                 if (likely(e->DataValid)) {
1520                         struct freelQ *fl = &sge->freelQ[e->FreelistQid];
1521
1522                         BUG_ON(!e->Sop || !e->Eop);
1523                         if (unlikely(e->Offload))
1524                                 unexpected_offload(adapter, fl);
1525                         else
1526                                 sge_rx(sge, fl, e->BufferLength);
1527
1528                         /*
1529                          * Note: this depends on each packet consuming a
1530                          * single free-list buffer; cf. the BUG above.
1531                          */
1532                         if (++fl->cidx == fl->size)
1533                                 fl->cidx = 0;
1534                         if (unlikely(--fl->credits <
1535                                      fl->size - SGE_FREEL_REFILL_THRESH))
1536                                 refill_free_list(sge, fl);
1537                 } else
1538                         sge->stats.pure_rsps++;
1539
1540                 e++;
1541                 if (unlikely(++q->cidx == q->size)) {
1542                         q->cidx = 0;
1543                         q->genbit ^= 1;
1544                         e = q->entries;
1545                 }
1546                 prefetch(e);
1547
1548                 if (++q->credits > SGE_RESPQ_REPLENISH_THRES) {
1549                         writel(q->credits, adapter->regs + A_SG_RSPQUEUECREDIT);
1550                         q->credits = 0;
1551                 }
1552                 --budget_left;
1553         }
1554
1555         flags = update_tx_info(adapter, flags, cmdq_processed[0]);
1556         sge->cmdQ[1].processed += cmdq_processed[1];
1557
1558         budget -= budget_left;
1559         return budget;
1560 }
1561
1562 static inline int responses_pending(const struct adapter *adapter)
1563 {
1564         const struct respQ *Q = &adapter->sge->respQ;
1565         const struct respQ_e *e = &Q->entries[Q->cidx];
1566
1567         return (e->GenerationBit == Q->genbit);
1568 }
1569
1570 #ifdef CONFIG_CHELSIO_T1_NAPI
1571 /*
1572  * A simpler version of process_responses() that handles only pure (i.e.,
1573  * non data-carrying) responses.  Such respones are too light-weight to justify
1574  * calling a softirq when using NAPI, so we handle them specially in hard
1575  * interrupt context.  The function is called with a pointer to a response,
1576  * which the caller must ensure is a valid pure response.  Returns 1 if it
1577  * encounters a valid data-carrying response, 0 otherwise.
1578  */
1579 static int process_pure_responses(struct adapter *adapter)
1580 {
1581         struct sge *sge = adapter->sge;
1582         struct respQ *q = &sge->respQ;
1583         struct respQ_e *e = &q->entries[q->cidx];
1584         unsigned int flags = 0;
1585         unsigned int cmdq_processed[SGE_CMDQ_N] = {0, 0};
1586
1587         if (e->DataValid)
1588                 return 1;
1589         do {
1590                 flags |= e->Qsleeping;
1591
1592                 cmdq_processed[0] += e->Cmdq0CreditReturn;
1593                 cmdq_processed[1] += e->Cmdq1CreditReturn;
1594
1595                 e++;
1596                 if (unlikely(++q->cidx == q->size)) {
1597                         q->cidx = 0;
1598                         q->genbit ^= 1;
1599                         e = q->entries;
1600                 }
1601                 prefetch(e);
1602
1603                 if (++q->credits > SGE_RESPQ_REPLENISH_THRES) {
1604                         writel(q->credits, adapter->regs + A_SG_RSPQUEUECREDIT);
1605                         q->credits = 0;
1606                 }
1607                 sge->stats.pure_rsps++;
1608         } while (e->GenerationBit == q->genbit && !e->DataValid);
1609
1610         flags = update_tx_info(adapter, flags, cmdq_processed[0]);
1611         sge->cmdQ[1].processed += cmdq_processed[1];
1612
1613         return e->GenerationBit == q->genbit;
1614 }
1615
1616 /*
1617  * Handler for new data events when using NAPI.  This does not need any locking
1618  * or protection from interrupts as data interrupts are off at this point and
1619  * other adapter interrupts do not interfere.
1620  */
1621 int t1_poll(struct net_device *dev, int *budget)
1622 {
1623         struct adapter *adapter = dev->priv;
1624         int work_done;
1625
1626         work_done = process_responses(adapter, min(*budget, dev->quota));
1627         *budget -= work_done;
1628         dev->quota -= work_done;
1629
1630         if (unlikely(responses_pending(adapter)))
1631                 return 1;
1632
1633         netif_rx_complete(dev);
1634         writel(adapter->sge->respQ.cidx, adapter->regs + A_SG_SLEEPING);
1635
1636         return 0;
1637
1638 }
1639
1640 /*
1641  * NAPI version of the main interrupt handler.
1642  */
1643 irqreturn_t t1_interrupt(int irq, void *data)
1644 {
1645         struct adapter *adapter = data;
1646         struct sge *sge = adapter->sge;
1647         int handled;
1648
1649         if (likely(responses_pending(adapter))) {
1650                 struct net_device *dev = sge->netdev;
1651
1652                 writel(F_PL_INTR_SGE_DATA, adapter->regs + A_PL_CAUSE);
1653
1654                 if (__netif_rx_schedule_prep(dev)) {
1655                         if (process_pure_responses(adapter))
1656                                 __netif_rx_schedule(dev);
1657                         else {
1658                                 /* no data, no NAPI needed */
1659                                 writel(sge->respQ.cidx, adapter->regs + A_SG_SLEEPING);
1660                                 netif_poll_enable(dev); /* undo schedule_prep */
1661                         }
1662                 }
1663                 return IRQ_HANDLED;
1664         }
1665
1666         spin_lock(&adapter->async_lock);
1667         handled = t1_slow_intr_handler(adapter);
1668         spin_unlock(&adapter->async_lock);
1669
1670         if (!handled)
1671                 sge->stats.unhandled_irqs++;
1672
1673         return IRQ_RETVAL(handled != 0);
1674 }
1675
1676 #else
1677 /*
1678  * Main interrupt handler, optimized assuming that we took a 'DATA'
1679  * interrupt.
1680  *
1681  * 1. Clear the interrupt
1682  * 2. Loop while we find valid descriptors and process them; accumulate
1683  *      information that can be processed after the loop
1684  * 3. Tell the SGE at which index we stopped processing descriptors
1685  * 4. Bookkeeping; free TX buffers, ring doorbell if there are any
1686  *      outstanding TX buffers waiting, replenish RX buffers, potentially
1687  *      reenable upper layers if they were turned off due to lack of TX
1688  *      resources which are available again.
1689  * 5. If we took an interrupt, but no valid respQ descriptors was found we
1690  *      let the slow_intr_handler run and do error handling.
1691  */
1692 irqreturn_t t1_interrupt(int irq, void *cookie)
1693 {
1694         int work_done;
1695         struct adapter *adapter = cookie;
1696
1697         spin_lock(&adapter->async_lock);
1698
1699         writel(F_PL_INTR_SGE_DATA, adapter->regs + A_PL_CAUSE);
1700
1701         if (likely(responses_pending(adapter))
1702                 work_done = process_responses(adapter, -1);
1703         else
1704                 work_done = t1_slow_intr_handler(adapter);
1705
1706         /*
1707          * The unconditional clearing of the PL_CAUSE above may have raced
1708          * with DMA completion and the corresponding generation of a response
1709          * to cause us to miss the resulting data interrupt.  The next write
1710          * is also unconditional to recover the missed interrupt and render
1711          * this race harmless.
1712          */
1713         writel(Q->cidx, adapter->regs + A_SG_SLEEPING);
1714
1715         if (!work_done)
1716                 adapter->sge->stats.unhandled_irqs++;
1717         spin_unlock(&adapter->async_lock);
1718         return IRQ_RETVAL(work_done != 0);
1719 }
1720 #endif
1721
1722 /*
1723  * Enqueues the sk_buff onto the cmdQ[qid] and has hardware fetch it.
1724  *
1725  * The code figures out how many entries the sk_buff will require in the
1726  * cmdQ and updates the cmdQ data structure with the state once the enqueue
1727  * has complete. Then, it doesn't access the global structure anymore, but
1728  * uses the corresponding fields on the stack. In conjuction with a spinlock
1729  * around that code, we can make the function reentrant without holding the
1730  * lock when we actually enqueue (which might be expensive, especially on
1731  * architectures with IO MMUs).
1732  *
1733  * This runs with softirqs disabled.
1734  */
1735 static int t1_sge_tx(struct sk_buff *skb, struct adapter *adapter,
1736                      unsigned int qid, struct net_device *dev)
1737 {
1738         struct sge *sge = adapter->sge;
1739         struct cmdQ *q = &sge->cmdQ[qid];
1740         unsigned int credits, pidx, genbit, count, use_sched_skb = 0;
1741
1742         if (!spin_trylock(&q->lock))
1743                 return NETDEV_TX_LOCKED;
1744
1745         reclaim_completed_tx(sge, q);
1746
1747         pidx = q->pidx;
1748         credits = q->size - q->in_use;
1749         count = 1 + skb_shinfo(skb)->nr_frags;
1750         count += compute_large_page_tx_descs(skb);
1751
1752         /* Ethernet packet */
1753         if (unlikely(credits < count)) {
1754                 if (!netif_queue_stopped(dev)) {
1755                         netif_stop_queue(dev);
1756                         set_bit(dev->if_port, &sge->stopped_tx_queues);
1757                         sge->stats.cmdQ_full[2]++;
1758                         CH_ERR("%s: Tx ring full while queue awake!\n",
1759                                adapter->name);
1760                 }
1761                 spin_unlock(&q->lock);
1762                 return NETDEV_TX_BUSY;
1763         }
1764
1765         if (unlikely(credits - count < q->stop_thres)) {
1766                 netif_stop_queue(dev);
1767                 set_bit(dev->if_port, &sge->stopped_tx_queues);
1768                 sge->stats.cmdQ_full[2]++;
1769         }
1770
1771         /* T204 cmdQ0 skbs that are destined for a certain port have to go
1772          * through the scheduler.
1773          */
1774         if (sge->tx_sched && !qid && skb->dev) {
1775 use_sched:
1776                 use_sched_skb = 1;
1777                 /* Note that the scheduler might return a different skb than
1778                  * the one passed in.
1779                  */
1780                 skb = sched_skb(sge, skb, credits);
1781                 if (!skb) {
1782                         spin_unlock(&q->lock);
1783                         return NETDEV_TX_OK;
1784                 }
1785                 pidx = q->pidx;
1786                 count = 1 + skb_shinfo(skb)->nr_frags;
1787                 count += compute_large_page_tx_descs(skb);
1788         }
1789
1790         q->in_use += count;
1791         genbit = q->genbit;
1792         pidx = q->pidx;
1793         q->pidx += count;
1794         if (q->pidx >= q->size) {
1795                 q->pidx -= q->size;
1796                 q->genbit ^= 1;
1797         }
1798         spin_unlock(&q->lock);
1799
1800         write_tx_descs(adapter, skb, pidx, genbit, q);
1801
1802         /*
1803          * We always ring the doorbell for cmdQ1.  For cmdQ0, we only ring
1804          * the doorbell if the Q is asleep. There is a natural race, where
1805          * the hardware is going to sleep just after we checked, however,
1806          * then the interrupt handler will detect the outstanding TX packet
1807          * and ring the doorbell for us.
1808          */
1809         if (qid)
1810                 doorbell_pio(adapter, F_CMDQ1_ENABLE);
1811         else {
1812                 clear_bit(CMDQ_STAT_LAST_PKT_DB, &q->status);
1813                 if (test_and_set_bit(CMDQ_STAT_RUNNING, &q->status) == 0) {
1814                         set_bit(CMDQ_STAT_LAST_PKT_DB, &q->status);
1815                         writel(F_CMDQ0_ENABLE, adapter->regs + A_SG_DOORBELL);
1816                 }
1817         }
1818
1819         if (use_sched_skb) {
1820                 if (spin_trylock(&q->lock)) {
1821                         credits = q->size - q->in_use;
1822                         skb = NULL;
1823                         goto use_sched;
1824                 }
1825         }
1826         return NETDEV_TX_OK;
1827 }
1828
1829 #define MK_ETH_TYPE_MSS(type, mss) (((mss) & 0x3FFF) | ((type) << 14))
1830
1831 /*
1832  *      eth_hdr_len - return the length of an Ethernet header
1833  *      @data: pointer to the start of the Ethernet header
1834  *
1835  *      Returns the length of an Ethernet header, including optional VLAN tag.
1836  */
1837 static inline int eth_hdr_len(const void *data)
1838 {
1839         const struct ethhdr *e = data;
1840
1841         return e->h_proto == htons(ETH_P_8021Q) ? VLAN_ETH_HLEN : ETH_HLEN;
1842 }
1843
1844 /*
1845  * Adds the CPL header to the sk_buff and passes it to t1_sge_tx.
1846  */
1847 int t1_start_xmit(struct sk_buff *skb, struct net_device *dev)
1848 {
1849         struct adapter *adapter = dev->priv;
1850         struct sge *sge = adapter->sge;
1851         struct sge_port_stats *st = per_cpu_ptr(sge->port_stats[dev->if_port], smp_processor_id());
1852         struct cpl_tx_pkt *cpl;
1853         struct sk_buff *orig_skb = skb;
1854         int ret;
1855
1856         if (skb->protocol == htons(ETH_P_CPL5))
1857                 goto send;
1858
1859         if (skb_shinfo(skb)->gso_size) {
1860                 int eth_type;
1861                 struct cpl_tx_pkt_lso *hdr;
1862
1863                 ++st->tx_tso;
1864
1865                 eth_type = skb->nh.raw - skb->data == ETH_HLEN ?
1866                         CPL_ETH_II : CPL_ETH_II_VLAN;
1867
1868                 hdr = (struct cpl_tx_pkt_lso *)skb_push(skb, sizeof(*hdr));
1869                 hdr->opcode = CPL_TX_PKT_LSO;
1870                 hdr->ip_csum_dis = hdr->l4_csum_dis = 0;
1871                 hdr->ip_hdr_words = skb->nh.iph->ihl;
1872                 hdr->tcp_hdr_words = skb->h.th->doff;
1873                 hdr->eth_type_mss = htons(MK_ETH_TYPE_MSS(eth_type,
1874                                                           skb_shinfo(skb)->gso_size));
1875                 hdr->len = htonl(skb->len - sizeof(*hdr));
1876                 cpl = (struct cpl_tx_pkt *)hdr;
1877         } else {
1878                 /*
1879                  * Packets shorter than ETH_HLEN can break the MAC, drop them
1880                  * early.  Also, we may get oversized packets because some
1881                  * parts of the kernel don't handle our unusual hard_header_len
1882                  * right, drop those too.
1883                  */
1884                 if (unlikely(skb->len < ETH_HLEN ||
1885                              skb->len > dev->mtu + eth_hdr_len(skb->data))) {
1886                         pr_debug("%s: packet size %d hdr %d mtu%d\n", dev->name,
1887                                  skb->len, eth_hdr_len(skb->data), dev->mtu);
1888                         dev_kfree_skb_any(skb);
1889                         return NETDEV_TX_OK;
1890                 }
1891
1892                 /*
1893                  * We are using a non-standard hard_header_len and some kernel
1894                  * components, such as pktgen, do not handle it right.
1895                  * Complain when this happens but try to fix things up.
1896                  */
1897                 if (unlikely(skb_headroom(skb) < dev->hard_header_len - ETH_HLEN)) {
1898                         pr_debug("%s: headroom %d header_len %d\n", dev->name,
1899                                  skb_headroom(skb), dev->hard_header_len);
1900
1901                         if (net_ratelimit())
1902                                 printk(KERN_ERR "%s: inadequate headroom in "
1903                                        "Tx packet\n", dev->name);
1904                         skb = skb_realloc_headroom(skb, sizeof(*cpl));
1905                         dev_kfree_skb_any(orig_skb);
1906                         if (!skb)
1907                                 return NETDEV_TX_OK;
1908                 }
1909
1910                 if (!(adapter->flags & UDP_CSUM_CAPABLE) &&
1911                     skb->ip_summed == CHECKSUM_PARTIAL &&
1912                     skb->nh.iph->protocol == IPPROTO_UDP) {
1913                         if (unlikely(skb_checksum_help(skb))) {
1914                                 pr_debug("%s: unable to do udp checksum\n", dev->name);
1915                                 dev_kfree_skb_any(skb);
1916                                 return NETDEV_TX_OK;
1917                         }
1918                 }
1919
1920                 /* Hmmm, assuming to catch the gratious arp... and we'll use
1921                  * it to flush out stuck espi packets...
1922                  */
1923                 if ((unlikely(!adapter->sge->espibug_skb[dev->if_port]))) {
1924                         if (skb->protocol == htons(ETH_P_ARP) &&
1925                             skb->nh.arph->ar_op == htons(ARPOP_REQUEST)) {
1926                                 adapter->sge->espibug_skb[dev->if_port] = skb;
1927                                 /* We want to re-use this skb later. We
1928                                  * simply bump the reference count and it
1929                                  * will not be freed...
1930                                  */
1931                                 skb = skb_get(skb);
1932                         }
1933                 }
1934
1935                 cpl = (struct cpl_tx_pkt *)__skb_push(skb, sizeof(*cpl));
1936                 cpl->opcode = CPL_TX_PKT;
1937                 cpl->ip_csum_dis = 1;    /* SW calculates IP csum */
1938                 cpl->l4_csum_dis = skb->ip_summed == CHECKSUM_PARTIAL ? 0 : 1;
1939                 /* the length field isn't used so don't bother setting it */
1940
1941                 st->tx_cso += (skb->ip_summed == CHECKSUM_PARTIAL);
1942         }
1943         cpl->iff = dev->if_port;
1944
1945 #if defined(CONFIG_VLAN_8021Q) || defined(CONFIG_VLAN_8021Q_MODULE)
1946         if (adapter->vlan_grp && vlan_tx_tag_present(skb)) {
1947                 cpl->vlan_valid = 1;
1948                 cpl->vlan = htons(vlan_tx_tag_get(skb));
1949                 st->vlan_insert++;
1950         } else
1951 #endif
1952                 cpl->vlan_valid = 0;
1953
1954 send:
1955         st->tx_packets++;
1956         dev->trans_start = jiffies;
1957         ret = t1_sge_tx(skb, adapter, 0, dev);
1958
1959         /* If transmit busy, and we reallocated skb's due to headroom limit,
1960          * then silently discard to avoid leak.
1961          */
1962         if (unlikely(ret != NETDEV_TX_OK && skb != orig_skb)) {
1963                 dev_kfree_skb_any(skb);
1964                 ret = NETDEV_TX_OK;
1965         }
1966         return ret;
1967 }
1968
1969 /*
1970  * Callback for the Tx buffer reclaim timer.  Runs with softirqs disabled.
1971  */
1972 static void sge_tx_reclaim_cb(unsigned long data)
1973 {
1974         int i;
1975         struct sge *sge = (struct sge *)data;
1976
1977         for (i = 0; i < SGE_CMDQ_N; ++i) {
1978                 struct cmdQ *q = &sge->cmdQ[i];
1979
1980                 if (!spin_trylock(&q->lock))
1981                         continue;
1982
1983                 reclaim_completed_tx(sge, q);
1984                 if (i == 0 && q->in_use) {    /* flush pending credits */
1985                         writel(F_CMDQ0_ENABLE, sge->adapter->regs + A_SG_DOORBELL);
1986                 }
1987                 spin_unlock(&q->lock);
1988         }
1989         mod_timer(&sge->tx_reclaim_timer, jiffies + TX_RECLAIM_PERIOD);
1990 }
1991
1992 /*
1993  * Propagate changes of the SGE coalescing parameters to the HW.
1994  */
1995 int t1_sge_set_coalesce_params(struct sge *sge, struct sge_params *p)
1996 {
1997         sge->fixed_intrtimer = p->rx_coalesce_usecs *
1998                 core_ticks_per_usec(sge->adapter);
1999         writel(sge->fixed_intrtimer, sge->adapter->regs + A_SG_INTRTIMER);
2000         return 0;
2001 }
2002
2003 /*
2004  * Allocates both RX and TX resources and configures the SGE. However,
2005  * the hardware is not enabled yet.
2006  */
2007 int t1_sge_configure(struct sge *sge, struct sge_params *p)
2008 {
2009         if (alloc_rx_resources(sge, p))
2010                 return -ENOMEM;
2011         if (alloc_tx_resources(sge, p)) {
2012                 free_rx_resources(sge);
2013                 return -ENOMEM;
2014         }
2015         configure_sge(sge, p);
2016
2017         /*
2018          * Now that we have sized the free lists calculate the payload
2019          * capacity of the large buffers.  Other parts of the driver use
2020          * this to set the max offload coalescing size so that RX packets
2021          * do not overflow our large buffers.
2022          */
2023         p->large_buf_capacity = jumbo_payload_capacity(sge);
2024         return 0;
2025 }
2026
2027 /*
2028  * Disables the DMA engine.
2029  */
2030 void t1_sge_stop(struct sge *sge)
2031 {
2032         int i;
2033         writel(0, sge->adapter->regs + A_SG_CONTROL);
2034         readl(sge->adapter->regs + A_SG_CONTROL); /* flush */
2035
2036         if (is_T2(sge->adapter))
2037                 del_timer_sync(&sge->espibug_timer);
2038
2039         del_timer_sync(&sge->tx_reclaim_timer);
2040         if (sge->tx_sched)
2041                 tx_sched_stop(sge);
2042
2043         for (i = 0; i < MAX_NPORTS; i++)
2044                 if (sge->espibug_skb[i])
2045                         kfree_skb(sge->espibug_skb[i]);
2046 }
2047
2048 /*
2049  * Enables the DMA engine.
2050  */
2051 void t1_sge_start(struct sge *sge)
2052 {
2053         refill_free_list(sge, &sge->freelQ[0]);
2054         refill_free_list(sge, &sge->freelQ[1]);
2055
2056         writel(sge->sge_control, sge->adapter->regs + A_SG_CONTROL);
2057         doorbell_pio(sge->adapter, F_FL0_ENABLE | F_FL1_ENABLE);
2058         readl(sge->adapter->regs + A_SG_CONTROL); /* flush */
2059
2060         mod_timer(&sge->tx_reclaim_timer, jiffies + TX_RECLAIM_PERIOD);
2061
2062         if (is_T2(sge->adapter))
2063                 mod_timer(&sge->espibug_timer, jiffies + sge->espibug_timeout);
2064 }
2065
2066 /*
2067  * Callback for the T2 ESPI 'stuck packet feature' workaorund
2068  */
2069 static void espibug_workaround_t204(unsigned long data)
2070 {
2071         struct adapter *adapter = (struct adapter *)data;
2072         struct sge *sge = adapter->sge;
2073         unsigned int nports = adapter->params.nports;
2074         u32 seop[MAX_NPORTS];
2075
2076         if (adapter->open_device_map & PORT_MASK) {
2077                 int i;
2078
2079                 if (t1_espi_get_mon_t204(adapter, &(seop[0]), 0) < 0)
2080                         return;
2081
2082                 for (i = 0; i < nports; i++) {
2083                         struct sk_buff *skb = sge->espibug_skb[i];
2084
2085                         if (!netif_running(adapter->port[i].dev) ||
2086                             netif_queue_stopped(adapter->port[i].dev) ||
2087                             !seop[i] || ((seop[i] & 0xfff) != 0) || !skb)
2088                                 continue;
2089
2090                         if (!skb->cb[0]) {
2091                                 u8 ch_mac_addr[ETH_ALEN] = {
2092                                         0x0, 0x7, 0x43, 0x0, 0x0, 0x0
2093                                 };
2094
2095                                 memcpy(skb->data + sizeof(struct cpl_tx_pkt),
2096                                         ch_mac_addr, ETH_ALEN);
2097                                 memcpy(skb->data + skb->len - 10,
2098                                         ch_mac_addr, ETH_ALEN);
2099                                 skb->cb[0] = 0xff;
2100                         }
2101
2102                         /* bump the reference count to avoid freeing of
2103                          * the skb once the DMA has completed.
2104                          */
2105                         skb = skb_get(skb);
2106                         t1_sge_tx(skb, adapter, 0, adapter->port[i].dev);
2107                 }
2108         }
2109         mod_timer(&sge->espibug_timer, jiffies + sge->espibug_timeout);
2110 }
2111
2112 static void espibug_workaround(unsigned long data)
2113 {
2114         struct adapter *adapter = (struct adapter *)data;
2115         struct sge *sge = adapter->sge;
2116
2117         if (netif_running(adapter->port[0].dev)) {
2118                 struct sk_buff *skb = sge->espibug_skb[0];
2119                 u32 seop = t1_espi_get_mon(adapter, 0x930, 0);
2120
2121                 if ((seop & 0xfff0fff) == 0xfff && skb) {
2122                         if (!skb->cb[0]) {
2123                                 u8 ch_mac_addr[ETH_ALEN] =
2124                                     {0x0, 0x7, 0x43, 0x0, 0x0, 0x0};
2125                                 memcpy(skb->data + sizeof(struct cpl_tx_pkt),
2126                                        ch_mac_addr, ETH_ALEN);
2127                                 memcpy(skb->data + skb->len - 10, ch_mac_addr,
2128                                        ETH_ALEN);
2129                                 skb->cb[0] = 0xff;
2130                         }
2131
2132                         /* bump the reference count to avoid freeing of the
2133                          * skb once the DMA has completed.
2134                          */
2135                         skb = skb_get(skb);
2136                         t1_sge_tx(skb, adapter, 0, adapter->port[0].dev);
2137                 }
2138         }
2139         mod_timer(&sge->espibug_timer, jiffies + sge->espibug_timeout);
2140 }
2141
2142 /*
2143  * Creates a t1_sge structure and returns suggested resource parameters.
2144  */
2145 struct sge * __devinit t1_sge_create(struct adapter *adapter,
2146                                      struct sge_params *p)
2147 {
2148         struct sge *sge = kzalloc(sizeof(*sge), GFP_KERNEL);
2149         int i;
2150
2151         if (!sge)
2152                 return NULL;
2153
2154         sge->adapter = adapter;
2155         sge->netdev = adapter->port[0].dev;
2156         sge->rx_pkt_pad = t1_is_T1B(adapter) ? 0 : 2;
2157         sge->jumbo_fl = t1_is_T1B(adapter) ? 1 : 0;
2158
2159         for_each_port(adapter, i) {
2160                 sge->port_stats[i] = alloc_percpu(struct sge_port_stats);
2161                 if (!sge->port_stats[i])
2162                         goto nomem_port;
2163         }
2164
2165         init_timer(&sge->tx_reclaim_timer);
2166         sge->tx_reclaim_timer.data = (unsigned long)sge;
2167         sge->tx_reclaim_timer.function = sge_tx_reclaim_cb;
2168
2169         if (is_T2(sge->adapter)) {
2170                 init_timer(&sge->espibug_timer);
2171
2172                 if (adapter->params.nports > 1) {
2173                         tx_sched_init(sge);
2174                         sge->espibug_timer.function = espibug_workaround_t204;
2175                 } else
2176                         sge->espibug_timer.function = espibug_workaround;
2177                 sge->espibug_timer.data = (unsigned long)sge->adapter;
2178
2179                 sge->espibug_timeout = 1;
2180                 /* for T204, every 10ms */
2181                 if (adapter->params.nports > 1)
2182                         sge->espibug_timeout = HZ/100;
2183         }
2184
2185
2186         p->cmdQ_size[0] = SGE_CMDQ0_E_N;
2187         p->cmdQ_size[1] = SGE_CMDQ1_E_N;
2188         p->freelQ_size[!sge->jumbo_fl] = SGE_FREEL_SIZE;
2189         p->freelQ_size[sge->jumbo_fl] = SGE_JUMBO_FREEL_SIZE;
2190         if (sge->tx_sched) {
2191                 if (board_info(sge->adapter)->board == CHBT_BOARD_CHT204)
2192                         p->rx_coalesce_usecs = 15;
2193                 else
2194                         p->rx_coalesce_usecs = 50;
2195         } else
2196                 p->rx_coalesce_usecs = 50;
2197
2198         p->coalesce_enable = 0;
2199         p->sample_interval_usecs = 0;
2200
2201         return sge;
2202 nomem_port:
2203         while (i >= 0) {
2204                 free_percpu(sge->port_stats[i]);
2205                 --i;
2206         }
2207         kfree(sge);
2208         return NULL;
2209
2210 }