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