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