drivers/net/ethernet/chelsio/cxgb/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_vlan_mode(struct adapter *adapter, u32 features)
 746 {
 747         struct sge *sge = adapter->sge;
 748
 749         if (features & NETIF_F_HW_VLAN_RX)
 750                 sge->sge_control |= F_VLAN_XTRACT;
 751         else
 752                 sge->sge_control &= ~F_VLAN_XTRACT;
 753         if (adapter->open_device_map) {
 754                 writel(sge->sge_control, adapter->regs + A_SG_CONTROL);
 755                 readl(adapter->regs + A_SG_CONTROL);   /* flush */
 756         }
 757 }
 758
 759 /*
 760  * Programs the various SGE registers. However, the engine is not yet enabled,
 761  * but sge->sge_control is setup and ready to go.
 762  */
 763 static void configure_sge(struct sge *sge, struct sge_params *p)
 764 {
 765         struct adapter *ap = sge->adapter;
 766
 767         writel(0, ap->regs + A_SG_CONTROL);
 768         setup_ring_params(ap, sge->cmdQ[0].dma_addr, sge->cmdQ[0].size,
 769                           A_SG_CMD0BASELWR, A_SG_CMD0BASEUPR, A_SG_CMD0SIZE);
 770         setup_ring_params(ap, sge->cmdQ[1].dma_addr, sge->cmdQ[1].size,
 771                           A_SG_CMD1BASELWR, A_SG_CMD1BASEUPR, A_SG_CMD1SIZE);
 772         setup_ring_params(ap, sge->freelQ[0].dma_addr,
 773                           sge->freelQ[0].size, A_SG_FL0BASELWR,
 774                           A_SG_FL0BASEUPR, A_SG_FL0SIZE);
 775         setup_ring_params(ap, sge->freelQ[1].dma_addr,
 776                           sge->freelQ[1].size, A_SG_FL1BASELWR,
 777                           A_SG_FL1BASEUPR, A_SG_FL1SIZE);
 778
 779         /* The threshold comparison uses <. */
 780         writel(SGE_RX_SM_BUF_SIZE + 1, ap->regs + A_SG_FLTHRESHOLD);
 781
 782         setup_ring_params(ap, sge->respQ.dma_addr, sge->respQ.size,
 783                           A_SG_RSPBASELWR, A_SG_RSPBASEUPR, A_SG_RSPSIZE);
 784         writel((u32)sge->respQ.size - 1, ap->regs + A_SG_RSPQUEUECREDIT);
 785
 786         sge->sge_control = F_CMDQ0_ENABLE | F_CMDQ1_ENABLE | F_FL0_ENABLE |
 787                 F_FL1_ENABLE | F_CPL_ENABLE | F_RESPONSE_QUEUE_ENABLE |
 788                 V_CMDQ_PRIORITY(2) | F_DISABLE_CMDQ1_GTS | F_ISCSI_COALESCE |
 789                 V_RX_PKT_OFFSET(sge->rx_pkt_pad);
 790
 791 #if defined(__BIG_ENDIAN_BITFIELD)
 792         sge->sge_control |= F_ENABLE_BIG_ENDIAN;
 793 #endif
 794
 795         /* Initialize no-resource timer */
 796         sge->intrtimer_nres = SGE_INTRTIMER_NRES * core_ticks_per_usec(ap);
 797
 798         t1_sge_set_coalesce_params(sge, p);
 799 }
 800
 801 /*
 802  * Return the payload capacity of the jumbo free-list buffers.
 803  */
 804 static inline unsigned int jumbo_payload_capacity(const struct sge *sge)
 805 {
 806         return sge->freelQ[sge->jumbo_fl].rx_buffer_size -
 807                 sge->freelQ[sge->jumbo_fl].dma_offset -
 808                 sizeof(struct cpl_rx_data);
 809 }
 810
 811 /*
 812  * Frees all SGE related resources and the sge structure itself
 813  */
 814 void t1_sge_destroy(struct sge *sge)
 815 {
 816         int i;
 817
 818         for_each_port(sge->adapter, i)
 819                 free_percpu(sge->port_stats[i]);
 820
 821         kfree(sge->tx_sched);
 822         free_tx_resources(sge);
 823         free_rx_resources(sge);
 824         kfree(sge);
 825 }
 826
 827 /*
 828  * Allocates new RX buffers on the freelist Q (and tracks them on the freelist
 829  * context Q) until the Q is full or alloc_skb fails.
 830  *
 831  * It is possible that the generation bits already match, indicating that the
 832  * buffer is already valid and nothing needs to be done. This happens when we
 833  * copied a received buffer into a new sk_buff during the interrupt processing.
 834  *
 835  * If the SGE doesn't automatically align packets properly (!sge->rx_pkt_pad),
 836  * we specify a RX_OFFSET in order to make sure that the IP header is 4B
 837  * aligned.
 838  */
 839 static void refill_free_list(struct sge *sge, struct freelQ *q)
 840 {
 841         struct pci_dev *pdev = sge->adapter->pdev;
 842         struct freelQ_ce *ce = &q->centries[q->pidx];
 843         struct freelQ_e *e = &q->entries[q->pidx];
 844         unsigned int dma_len = q->rx_buffer_size - q->dma_offset;
 845
 846         while (q->credits < q->size) {
 847                 struct sk_buff *skb;
 848                 dma_addr_t mapping;
 849
 850                 skb = alloc_skb(q->rx_buffer_size, GFP_ATOMIC);
 851                 if (!skb)
 852                         break;
 853
 854                 skb_reserve(skb, q->dma_offset);
 855                 mapping = pci_map_single(pdev, skb->data, dma_len,
 856                                          PCI_DMA_FROMDEVICE);
 857                 skb_reserve(skb, sge->rx_pkt_pad);
 858
 859                 ce->skb = skb;
 860                 dma_unmap_addr_set(ce, dma_addr, mapping);
 861                 dma_unmap_len_set(ce, dma_len, dma_len);
 862                 e->addr_lo = (u32)mapping;
 863                 e->addr_hi = (u64)mapping >> 32;
 864                 e->len_gen = V_CMD_LEN(dma_len) | V_CMD_GEN1(q->genbit);
 865                 wmb();
 866                 e->gen2 = V_CMD_GEN2(q->genbit);
 867
 868                 e++;
 869                 ce++;
 870                 if (++q->pidx == q->size) {
 871                         q->pidx = 0;
 872                         q->genbit ^= 1;
 873                         ce = q->centries;
 874                         e = q->entries;
 875                 }
 876                 q->credits++;
 877         }
 878 }
 879
 880 /*
 881  * Calls refill_free_list for both free lists. If we cannot fill at least 1/4
 882  * of both rings, we go into 'few interrupt mode' in order to give the system
 883  * time to free up resources.
 884  */
 885 static void freelQs_empty(struct sge *sge)
 886 {
 887         struct adapter *adapter = sge->adapter;
 888         u32 irq_reg = readl(adapter->regs + A_SG_INT_ENABLE);
 889         u32 irqholdoff_reg;
 890
 891         refill_free_list(sge, &sge->freelQ[0]);
 892         refill_free_list(sge, &sge->freelQ[1]);
 893
 894         if (sge->freelQ[0].credits > (sge->freelQ[0].size >> 2) &&
 895             sge->freelQ[1].credits > (sge->freelQ[1].size >> 2)) {
 896                 irq_reg |= F_FL_EXHAUSTED;
 897                 irqholdoff_reg = sge->fixed_intrtimer;
 898         } else {
 899                 /* Clear the F_FL_EXHAUSTED interrupts for now */
 900                 irq_reg &= ~F_FL_EXHAUSTED;
 901                 irqholdoff_reg = sge->intrtimer_nres;
 902         }
 903         writel(irqholdoff_reg, adapter->regs + A_SG_INTRTIMER);
 904         writel(irq_reg, adapter->regs + A_SG_INT_ENABLE);
 905
 906         /* We reenable the Qs to force a freelist GTS interrupt later */
 907         doorbell_pio(adapter, F_FL0_ENABLE | F_FL1_ENABLE);
 908 }
 909
 910 #define SGE_PL_INTR_MASK (F_PL_INTR_SGE_ERR | F_PL_INTR_SGE_DATA)
 911 #define SGE_INT_FATAL (F_RESPQ_OVERFLOW | F_PACKET_TOO_BIG | F_PACKET_MISMATCH)
 912 #define SGE_INT_ENABLE (F_RESPQ_EXHAUSTED | F_RESPQ_OVERFLOW | \
 913                         F_FL_EXHAUSTED | F_PACKET_TOO_BIG | F_PACKET_MISMATCH)
 914
 915 /*
 916  * Disable SGE Interrupts
 917  */
 918 void t1_sge_intr_disable(struct sge *sge)
 919 {
 920         u32 val = readl(sge->adapter->regs + A_PL_ENABLE);
 921
 922         writel(val & ~SGE_PL_INTR_MASK, sge->adapter->regs + A_PL_ENABLE);
 923         writel(0, sge->adapter->regs + A_SG_INT_ENABLE);
 924 }
 925
 926 /*
 927  * Enable SGE interrupts.
 928  */
 929 void t1_sge_intr_enable(struct sge *sge)
 930 {
 931         u32 en = SGE_INT_ENABLE;
 932         u32 val = readl(sge->adapter->regs + A_PL_ENABLE);
 933
 934         if (sge->adapter->port[0].dev->hw_features & NETIF_F_TSO)
 935                 en &= ~F_PACKET_TOO_BIG;
 936         writel(en, sge->adapter->regs + A_SG_INT_ENABLE);
 937         writel(val | SGE_PL_INTR_MASK, sge->adapter->regs + A_PL_ENABLE);
 938 }
 939
 940 /*
 941  * Clear SGE interrupts.
 942  */
 943 void t1_sge_intr_clear(struct sge *sge)
 944 {
 945         writel(SGE_PL_INTR_MASK, sge->adapter->regs + A_PL_CAUSE);
 946         writel(0xffffffff, sge->adapter->regs + A_SG_INT_CAUSE);
 947 }
 948
 949 /*
 950  * SGE 'Error' interrupt handler
 951  */
 952 int t1_sge_intr_error_handler(struct sge *sge)
 953 {
 954         struct adapter *adapter = sge->adapter;
 955         u32 cause = readl(adapter->regs + A_SG_INT_CAUSE);
 956
 957         if (adapter->port[0].dev->hw_features & NETIF_F_TSO)
 958                 cause &= ~F_PACKET_TOO_BIG;
 959         if (cause & F_RESPQ_EXHAUSTED)
 960                 sge->stats.respQ_empty++;
 961         if (cause & F_RESPQ_OVERFLOW) {
 962                 sge->stats.respQ_overflow++;
 963                 pr_alert("%s: SGE response queue overflow\n",
 964                          adapter->name);
 965         }
 966         if (cause & F_FL_EXHAUSTED) {
 967                 sge->stats.freelistQ_empty++;
 968                 freelQs_empty(sge);
 969         }
 970         if (cause & F_PACKET_TOO_BIG) {
 971                 sge->stats.pkt_too_big++;
 972                 pr_alert("%s: SGE max packet size exceeded\n",
 973                          adapter->name);
 974         }
 975         if (cause & F_PACKET_MISMATCH) {
 976                 sge->stats.pkt_mismatch++;
 977                 pr_alert("%s: SGE packet mismatch\n", adapter->name);
 978         }
 979         if (cause & SGE_INT_FATAL)
 980                 t1_fatal_err(adapter);
 981
 982         writel(cause, adapter->regs + A_SG_INT_CAUSE);
 983         return 0;
 984 }
 985
 986 const struct sge_intr_counts *t1_sge_get_intr_counts(const struct sge *sge)
 987 {
 988         return &sge->stats;
 989 }
 990
 991 void t1_sge_get_port_stats(const struct sge *sge, int port,
 992                            struct sge_port_stats *ss)
 993 {
 994         int cpu;
 995
 996         memset(ss, 0, sizeof(*ss));
 997         for_each_possible_cpu(cpu) {
 998                 struct sge_port_stats *st = per_cpu_ptr(sge->port_stats[port], cpu);
 999
1000                 ss->rx_cso_good += st->rx_cso_good;
1001                 ss->tx_cso += st->tx_cso;
1002                 ss->tx_tso += st->tx_tso;
1003                 ss->tx_need_hdrroom += st->tx_need_hdrroom;
1004                 ss->vlan_xtract += st->vlan_xtract;
1005                 ss->vlan_insert += st->vlan_insert;
1006         }
1007 }
1008
1009 /**
1010  *      recycle_fl_buf - recycle a free list buffer
1011  *      @fl: the free list
1012  *      @idx: index of buffer to recycle
1013  *
1014  *      Recycles the specified buffer on the given free list by adding it at
1015  *      the next available slot on the list.
1016  */
1017 static void recycle_fl_buf(struct freelQ *fl, int idx)
1018 {
1019         struct freelQ_e *from = &fl->entries[idx];
1020         struct freelQ_e *to = &fl->entries[fl->pidx];
1021
1022         fl->centries[fl->pidx] = fl->centries[idx];
1023         to->addr_lo = from->addr_lo;
1024         to->addr_hi = from->addr_hi;
1025         to->len_gen = G_CMD_LEN(from->len_gen) | V_CMD_GEN1(fl->genbit);
1026         wmb();
1027         to->gen2 = V_CMD_GEN2(fl->genbit);
1028         fl->credits++;
1029
1030         if (++fl->pidx == fl->size) {
1031                 fl->pidx = 0;
1032                 fl->genbit ^= 1;
1033         }
1034 }
1035
1036 static int copybreak __read_mostly = 256;
1037 module_param(copybreak, int, 0);
1038 MODULE_PARM_DESC(copybreak, "Receive copy threshold");
1039
1040 /**
1041  *      get_packet - return the next ingress packet buffer
1042  *      @pdev: the PCI device that received the packet
1043  *      @fl: the SGE free list holding the packet
1044  *      @len: the actual packet length, excluding any SGE padding
1045  *
1046  *      Get the next packet from a free list and complete setup of the
1047  *      sk_buff.  If the packet is small we make a copy and recycle the
1048  *      original buffer, otherwise we use the original buffer itself.  If a
1049  *      positive drop threshold is supplied packets are dropped and their
1050  *      buffers recycled if (a) the number of remaining buffers is under the
1051  *      threshold and the packet is too big to copy, or (b) the packet should
1052  *      be copied but there is no memory for the copy.
1053  */
1054 static inline struct sk_buff *get_packet(struct pci_dev *pdev,
1055                                          struct freelQ *fl, unsigned int len)
1056 {
1057         struct sk_buff *skb;
1058         const struct freelQ_ce *ce = &fl->centries[fl->cidx];
1059
1060         if (len < copybreak) {
1061                 skb = alloc_skb(len + 2, GFP_ATOMIC);
1062                 if (!skb)
1063                         goto use_orig_buf;
1064
1065                 skb_reserve(skb, 2);    /* align IP header */
1066                 skb_put(skb, len);
1067                 pci_dma_sync_single_for_cpu(pdev,
1068                                             dma_unmap_addr(ce, dma_addr),
1069                                             dma_unmap_len(ce, dma_len),
1070                                             PCI_DMA_FROMDEVICE);
1071                 skb_copy_from_linear_data(ce->skb, skb->data, len);
1072                 pci_dma_sync_single_for_device(pdev,
1073                                                dma_unmap_addr(ce, dma_addr),
1074                                                dma_unmap_len(ce, dma_len),
1075                                                PCI_DMA_FROMDEVICE);
1076                 recycle_fl_buf(fl, fl->cidx);
1077                 return skb;
1078         }
1079
1080 use_orig_buf:
1081         if (fl->credits < 2) {
1082                 recycle_fl_buf(fl, fl->cidx);
1083                 return NULL;
1084         }
1085
1086         pci_unmap_single(pdev, dma_unmap_addr(ce, dma_addr),
1087                          dma_unmap_len(ce, dma_len), PCI_DMA_FROMDEVICE);
1088         skb = ce->skb;
1089         prefetch(skb->data);
1090
1091         skb_put(skb, len);
1092         return skb;
1093 }
1094
1095 /**
1096  *      unexpected_offload - handle an unexpected offload packet
1097  *      @adapter: the adapter
1098  *      @fl: the free list that received the packet
1099  *
1100  *      Called when we receive an unexpected offload packet (e.g., the TOE
1101  *      function is disabled or the card is a NIC).  Prints a message and
1102  *      recycles the buffer.
1103  */
1104 static void unexpected_offload(struct adapter *adapter, struct freelQ *fl)
1105 {
1106         struct freelQ_ce *ce = &fl->centries[fl->cidx];
1107         struct sk_buff *skb = ce->skb;
1108
1109         pci_dma_sync_single_for_cpu(adapter->pdev, dma_unmap_addr(ce, dma_addr),
1110                             dma_unmap_len(ce, dma_len), PCI_DMA_FROMDEVICE);
1111         pr_err("%s: unexpected offload packet, cmd %u\n",
1112                adapter->name, *skb->data);
1113         recycle_fl_buf(fl, fl->cidx);
1114 }
1115
1116 /*
1117  * T1/T2 SGE limits the maximum DMA size per TX descriptor to
1118  * SGE_TX_DESC_MAX_PLEN (16KB). If the PAGE_SIZE is larger than 16KB, the
1119  * stack might send more than SGE_TX_DESC_MAX_PLEN in a contiguous manner.
1120  * Note that the *_large_page_tx_descs stuff will be optimized out when
1121  * PAGE_SIZE <= SGE_TX_DESC_MAX_PLEN.
1122  *
1123  * compute_large_page_descs() computes how many additional descriptors are
1124  * required to break down the stack's request.
1125  */
1126 static inline unsigned int compute_large_page_tx_descs(struct sk_buff *skb)
1127 {
1128         unsigned int count = 0;
1129
1130         if (PAGE_SIZE > SGE_TX_DESC_MAX_PLEN) {
1131                 unsigned int nfrags = skb_shinfo(skb)->nr_frags;
1132                 unsigned int i, len = skb_headlen(skb);
1133                 while (len > SGE_TX_DESC_MAX_PLEN) {
1134                         count++;
1135                         len -= SGE_TX_DESC_MAX_PLEN;
1136                 }
1137                 for (i = 0; nfrags--; i++) {
1138                         const skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
1139                         len = skb_frag_size(frag);
1140                         while (len > SGE_TX_DESC_MAX_PLEN) {
1141                                 count++;
1142                                 len -= SGE_TX_DESC_MAX_PLEN;
1143                         }
1144                 }
1145         }
1146         return count;
1147 }
1148
1149 /*
1150  * Write a cmdQ entry.
1151  *
1152  * Since this function writes the 'flags' field, it must not be used to
1153  * write the first cmdQ entry.
1154  */
1155 static inline void write_tx_desc(struct cmdQ_e *e, dma_addr_t mapping,
1156                                  unsigned int len, unsigned int gen,
1157                                  unsigned int eop)
1158 {
1159         BUG_ON(len > SGE_TX_DESC_MAX_PLEN);
1160
1161         e->addr_lo = (u32)mapping;
1162         e->addr_hi = (u64)mapping >> 32;
1163         e->len_gen = V_CMD_LEN(len) | V_CMD_GEN1(gen);
1164         e->flags = F_CMD_DATAVALID | V_CMD_EOP(eop) | V_CMD_GEN2(gen);
1165 }
1166
1167 /*
1168  * See comment for previous function.
1169  *
1170  * write_tx_descs_large_page() writes additional SGE tx descriptors if
1171  * *desc_len exceeds HW's capability.
1172  */
1173 static inline unsigned int write_large_page_tx_descs(unsigned int pidx,
1174                                                      struct cmdQ_e **e,
1175                                                      struct cmdQ_ce **ce,
1176                                                      unsigned int *gen,
1177                                                      dma_addr_t *desc_mapping,
1178                                                      unsigned int *desc_len,
1179                                                      unsigned int nfrags,
1180                                                      struct cmdQ *q)
1181 {
1182         if (PAGE_SIZE > SGE_TX_DESC_MAX_PLEN) {
1183                 struct cmdQ_e *e1 = *e;
1184                 struct cmdQ_ce *ce1 = *ce;
1185
1186                 while (*desc_len > SGE_TX_DESC_MAX_PLEN) {
1187                         *desc_len -= SGE_TX_DESC_MAX_PLEN;
1188                         write_tx_desc(e1, *desc_mapping, SGE_TX_DESC_MAX_PLEN,
1189                                       *gen, nfrags == 0 && *desc_len == 0);
1190                         ce1->skb = NULL;
1191                         dma_unmap_len_set(ce1, dma_len, 0);
1192                         *desc_mapping += SGE_TX_DESC_MAX_PLEN;
1193                         if (*desc_len) {
1194                                 ce1++;
1195                                 e1++;
1196                                 if (++pidx == q->size) {
1197                                         pidx = 0;
1198                                         *gen ^= 1;
1199                                         ce1 = q->centries;
1200                                         e1 = q->entries;
1201                                 }
1202                         }
1203                 }
1204                 *e = e1;
1205                 *ce = ce1;
1206         }
1207         return pidx;
1208 }
1209
1210 /*
1211  * Write the command descriptors to transmit the given skb starting at
1212  * descriptor pidx with the given generation.
1213  */
1214 static inline void write_tx_descs(struct adapter *adapter, struct sk_buff *skb,
1215                                   unsigned int pidx, unsigned int gen,
1216                                   struct cmdQ *q)
1217 {
1218         dma_addr_t mapping, desc_mapping;
1219         struct cmdQ_e *e, *e1;
1220         struct cmdQ_ce *ce;
1221         unsigned int i, flags, first_desc_len, desc_len,
1222             nfrags = skb_shinfo(skb)->nr_frags;
1223
1224         e = e1 = &q->entries[pidx];
1225         ce = &q->centries[pidx];
1226
1227         mapping = pci_map_single(adapter->pdev, skb->data,
1228                                  skb_headlen(skb), PCI_DMA_TODEVICE);
1229
1230         desc_mapping = mapping;
1231         desc_len = skb_headlen(skb);
1232
1233         flags = F_CMD_DATAVALID | F_CMD_SOP |
1234             V_CMD_EOP(nfrags == 0 && desc_len <= SGE_TX_DESC_MAX_PLEN) |
1235             V_CMD_GEN2(gen);
1236         first_desc_len = (desc_len <= SGE_TX_DESC_MAX_PLEN) ?
1237             desc_len : SGE_TX_DESC_MAX_PLEN;
1238         e->addr_lo = (u32)desc_mapping;
1239         e->addr_hi = (u64)desc_mapping >> 32;
1240         e->len_gen = V_CMD_LEN(first_desc_len) | V_CMD_GEN1(gen);
1241         ce->skb = NULL;
1242         dma_unmap_len_set(ce, dma_len, 0);
1243
1244         if (PAGE_SIZE > SGE_TX_DESC_MAX_PLEN &&
1245             desc_len > SGE_TX_DESC_MAX_PLEN) {
1246                 desc_mapping += first_desc_len;
1247                 desc_len -= first_desc_len;
1248                 e1++;
1249                 ce++;
1250                 if (++pidx == q->size) {
1251                         pidx = 0;
1252                         gen ^= 1;
1253                         e1 = q->entries;
1254                         ce = q->centries;
1255                 }
1256                 pidx = write_large_page_tx_descs(pidx, &e1, &ce, &gen,
1257                                                  &desc_mapping, &desc_len,
1258                                                  nfrags, q);
1259
1260                 if (likely(desc_len))
1261                         write_tx_desc(e1, desc_mapping, desc_len, gen,
1262                                       nfrags == 0);
1263         }
1264
1265         ce->skb = NULL;
1266         dma_unmap_addr_set(ce, dma_addr, mapping);
1267         dma_unmap_len_set(ce, dma_len, skb_headlen(skb));
1268
1269         for (i = 0; nfrags--; i++) {
1270                 skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
1271                 e1++;
1272                 ce++;
1273                 if (++pidx == q->size) {
1274                         pidx = 0;
1275                         gen ^= 1;
1276                         e1 = q->entries;
1277                         ce = q->centries;
1278                 }
1279
1280                 mapping = skb_frag_dma_map(&adapter->pdev->dev, frag, 0,
1281                                            skb_frag_size(frag), DMA_TO_DEVICE);
1282                 desc_mapping = mapping;
1283                 desc_len = skb_frag_size(frag);
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, skb_frag_size(frag));
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 (p->vlan_valid) {
1401                 st->vlan_xtract++;
1402                 __vlan_hwaccel_put_tag(skb, ntohs(p->vlan));
1403         }
1404         netif_receive_skb(skb);
1405 }
1406
1407 /*
1408  * Returns true if a command queue has enough available descriptors that
1409  * we can resume Tx operation after temporarily disabling its packet queue.
1410  */
1411 static inline int enough_free_Tx_descs(const struct cmdQ *q)
1412 {
1413         unsigned int r = q->processed - q->cleaned;
1414
1415         return q->in_use - r < (q->size >> 1);
1416 }
1417
1418 /*
1419  * Called when sufficient space has become available in the SGE command queues
1420  * after the Tx packet schedulers have been suspended to restart the Tx path.
1421  */
1422 static void restart_tx_queues(struct sge *sge)
1423 {
1424         struct adapter *adap = sge->adapter;
1425         int i;
1426
1427         if (!enough_free_Tx_descs(&sge->cmdQ[0]))
1428                 return;
1429
1430         for_each_port(adap, i) {
1431                 struct net_device *nd = adap->port[i].dev;
1432
1433                 if (test_and_clear_bit(nd->if_port, &sge->stopped_tx_queues) &&
1434                     netif_running(nd)) {
1435                         sge->stats.cmdQ_restarted[2]++;
1436                         netif_wake_queue(nd);
1437                 }
1438         }
1439 }
1440
1441 /*
1442  * update_tx_info is called from the interrupt handler/NAPI to return cmdQ0
1443  * information.
1444  */
1445 static unsigned int update_tx_info(struct adapter *adapter,
1446                                           unsigned int flags,
1447                                           unsigned int pr0)
1448 {
1449         struct sge *sge = adapter->sge;
1450         struct cmdQ *cmdq = &sge->cmdQ[0];
1451
1452         cmdq->processed += pr0;
1453         if (flags & (F_FL0_ENABLE | F_FL1_ENABLE)) {
1454                 freelQs_empty(sge);
1455                 flags &= ~(F_FL0_ENABLE | F_FL1_ENABLE);
1456         }
1457         if (flags & F_CMDQ0_ENABLE) {
1458                 clear_bit(CMDQ_STAT_RUNNING, &cmdq->status);
1459
1460                 if (cmdq->cleaned + cmdq->in_use != cmdq->processed &&
1461                     !test_and_set_bit(CMDQ_STAT_LAST_PKT_DB, &cmdq->status)) {
1462                         set_bit(CMDQ_STAT_RUNNING, &cmdq->status);
1463                         writel(F_CMDQ0_ENABLE, adapter->regs + A_SG_DOORBELL);
1464                 }
1465                 if (sge->tx_sched)
1466                         tasklet_hi_schedule(&sge->tx_sched->sched_tsk);
1467
1468                 flags &= ~F_CMDQ0_ENABLE;
1469         }
1470
1471         if (unlikely(sge->stopped_tx_queues != 0))
1472                 restart_tx_queues(sge);
1473
1474         return flags;
1475 }
1476
1477 /*
1478  * Process SGE responses, up to the supplied budget.  Returns the number of
1479  * responses processed.  A negative budget is effectively unlimited.
1480  */
1481 static int process_responses(struct adapter *adapter, int budget)
1482 {
1483         struct sge *sge = adapter->sge;
1484         struct respQ *q = &sge->respQ;
1485         struct respQ_e *e = &q->entries[q->cidx];
1486         int done = 0;
1487         unsigned int flags = 0;
1488         unsigned int cmdq_processed[SGE_CMDQ_N] = {0, 0};
1489
1490         while (done < budget && e->GenerationBit == q->genbit) {
1491                 flags |= e->Qsleeping;
1492
1493                 cmdq_processed[0] += e->Cmdq0CreditReturn;
1494                 cmdq_processed[1] += e->Cmdq1CreditReturn;
1495
1496                 /* We batch updates to the TX side to avoid cacheline
1497                  * ping-pong of TX state information on MP where the sender
1498                  * might run on a different CPU than this function...
1499                  */
1500                 if (unlikely((flags & F_CMDQ0_ENABLE) || cmdq_processed[0] > 64)) {
1501                         flags = update_tx_info(adapter, flags, cmdq_processed[0]);
1502                         cmdq_processed[0] = 0;
1503                 }
1504
1505                 if (unlikely(cmdq_processed[1] > 16)) {
1506                         sge->cmdQ[1].processed += cmdq_processed[1];
1507                         cmdq_processed[1] = 0;
1508                 }
1509
1510                 if (likely(e->DataValid)) {
1511                         struct freelQ *fl = &sge->freelQ[e->FreelistQid];
1512
1513                         BUG_ON(!e->Sop || !e->Eop);
1514                         if (unlikely(e->Offload))
1515                                 unexpected_offload(adapter, fl);
1516                         else
1517                                 sge_rx(sge, fl, e->BufferLength);
1518
1519                         ++done;
1520
1521                         /*
1522                          * Note: this depends on each packet consuming a
1523                          * single free-list buffer; cf. the BUG above.
1524                          */
1525                         if (++fl->cidx == fl->size)
1526                                 fl->cidx = 0;
1527                         prefetch(fl->centries[fl->cidx].skb);
1528
1529                         if (unlikely(--fl->credits <
1530                                      fl->size - SGE_FREEL_REFILL_THRESH))
1531                                 refill_free_list(sge, fl);
1532                 } else
1533                         sge->stats.pure_rsps++;
1534
1535                 e++;
1536                 if (unlikely(++q->cidx == q->size)) {
1537                         q->cidx = 0;
1538                         q->genbit ^= 1;
1539                         e = q->entries;
1540                 }
1541                 prefetch(e);
1542
1543                 if (++q->credits > SGE_RESPQ_REPLENISH_THRES) {
1544                         writel(q->credits, adapter->regs + A_SG_RSPQUEUECREDIT);
1545                         q->credits = 0;
1546                 }
1547         }
1548
1549         flags = update_tx_info(adapter, flags, cmdq_processed[0]);
1550         sge->cmdQ[1].processed += cmdq_processed[1];
1551
1552         return done;
1553 }
1554
1555 static inline int responses_pending(const struct adapter *adapter)
1556 {
1557         const struct respQ *Q = &adapter->sge->respQ;
1558         const struct respQ_e *e = &Q->entries[Q->cidx];
1559
1560         return e->GenerationBit == Q->genbit;
1561 }
1562
1563 /*
1564  * A simpler version of process_responses() that handles only pure (i.e.,
1565  * non data-carrying) responses.  Such respones are too light-weight to justify
1566  * calling a softirq when using NAPI, so we handle them specially in hard
1567  * interrupt context.  The function is called with a pointer to a response,
1568  * which the caller must ensure is a valid pure response.  Returns 1 if it
1569  * encounters a valid data-carrying response, 0 otherwise.
1570  */
1571 static int process_pure_responses(struct adapter *adapter)
1572 {
1573         struct sge *sge = adapter->sge;
1574         struct respQ *q = &sge->respQ;
1575         struct respQ_e *e = &q->entries[q->cidx];
1576         const struct freelQ *fl = &sge->freelQ[e->FreelistQid];
1577         unsigned int flags = 0;
1578         unsigned int cmdq_processed[SGE_CMDQ_N] = {0, 0};
1579
1580         prefetch(fl->centries[fl->cidx].skb);
1581         if (e->DataValid)
1582                 return 1;
1583
1584         do {
1585                 flags |= e->Qsleeping;
1586
1587                 cmdq_processed[0] += e->Cmdq0CreditReturn;
1588                 cmdq_processed[1] += e->Cmdq1CreditReturn;
1589
1590                 e++;
1591                 if (unlikely(++q->cidx == q->size)) {
1592                         q->cidx = 0;
1593                         q->genbit ^= 1;
1594                         e = q->entries;
1595                 }
1596                 prefetch(e);
1597
1598                 if (++q->credits > SGE_RESPQ_REPLENISH_THRES) {
1599                         writel(q->credits, adapter->regs + A_SG_RSPQUEUECREDIT);
1600                         q->credits = 0;
1601                 }
1602                 sge->stats.pure_rsps++;
1603         } while (e->GenerationBit == q->genbit && !e->DataValid);
1604
1605         flags = update_tx_info(adapter, flags, cmdq_processed[0]);
1606         sge->cmdQ[1].processed += cmdq_processed[1];
1607
1608         return e->GenerationBit == q->genbit;
1609 }
1610
1611 /*
1612  * Handler for new data events when using NAPI.  This does not need any locking
1613  * or protection from interrupts as data interrupts are off at this point and
1614  * other adapter interrupts do not interfere.
1615  */
1616 int t1_poll(struct napi_struct *napi, int budget)
1617 {
1618         struct adapter *adapter = container_of(napi, struct adapter, napi);
1619         int work_done = process_responses(adapter, budget);
1620
1621         if (likely(work_done < budget)) {
1622                 napi_complete(napi);
1623                 writel(adapter->sge->respQ.cidx,
1624                        adapter->regs + A_SG_SLEEPING);
1625         }
1626         return work_done;
1627 }
1628
1629 irqreturn_t t1_interrupt(int irq, void *data)
1630 {
1631         struct adapter *adapter = data;
1632         struct sge *sge = adapter->sge;
1633         int handled;
1634
1635         if (likely(responses_pending(adapter))) {
1636                 writel(F_PL_INTR_SGE_DATA, adapter->regs + A_PL_CAUSE);
1637
1638                 if (napi_schedule_prep(&adapter->napi)) {
1639                         if (process_pure_responses(adapter))
1640                                 __napi_schedule(&adapter->napi);
1641                         else {
1642                                 /* no data, no NAPI needed */
1643                                 writel(sge->respQ.cidx, adapter->regs + A_SG_SLEEPING);
1644                                 /* undo schedule_prep */
1645                                 napi_enable(&adapter->napi);
1646                         }
1647                 }
1648                 return IRQ_HANDLED;
1649         }
1650
1651         spin_lock(&adapter->async_lock);
1652         handled = t1_slow_intr_handler(adapter);
1653         spin_unlock(&adapter->async_lock);
1654
1655         if (!handled)
1656                 sge->stats.unhandled_irqs++;
1657
1658         return IRQ_RETVAL(handled != 0);
1659 }
1660
1661 /*
1662  * Enqueues the sk_buff onto the cmdQ[qid] and has hardware fetch it.
1663  *
1664  * The code figures out how many entries the sk_buff will require in the
1665  * cmdQ and updates the cmdQ data structure with the state once the enqueue
1666  * has complete. Then, it doesn't access the global structure anymore, but
1667  * uses the corresponding fields on the stack. In conjunction with a spinlock
1668  * around that code, we can make the function reentrant without holding the
1669  * lock when we actually enqueue (which might be expensive, especially on
1670  * architectures with IO MMUs).
1671  *
1672  * This runs with softirqs disabled.
1673  */
1674 static int t1_sge_tx(struct sk_buff *skb, struct adapter *adapter,
1675                      unsigned int qid, struct net_device *dev)
1676 {
1677         struct sge *sge = adapter->sge;
1678         struct cmdQ *q = &sge->cmdQ[qid];
1679         unsigned int credits, pidx, genbit, count, use_sched_skb = 0;
1680
1681         if (!spin_trylock(&q->lock))
1682                 return NETDEV_TX_LOCKED;
1683
1684         reclaim_completed_tx(sge, q);
1685
1686         pidx = q->pidx;
1687         credits = q->size - q->in_use;
1688         count = 1 + skb_shinfo(skb)->nr_frags;
1689         count += compute_large_page_tx_descs(skb);
1690
1691         /* Ethernet packet */
1692         if (unlikely(credits < count)) {
1693                 if (!netif_queue_stopped(dev)) {
1694                         netif_stop_queue(dev);
1695                         set_bit(dev->if_port, &sge->stopped_tx_queues);
1696                         sge->stats.cmdQ_full[2]++;
1697                         pr_err("%s: Tx ring full while queue awake!\n",
1698                                adapter->name);
1699                 }
1700                 spin_unlock(&q->lock);
1701                 return NETDEV_TX_BUSY;
1702         }
1703
1704         if (unlikely(credits - count < q->stop_thres)) {
1705                 netif_stop_queue(dev);
1706                 set_bit(dev->if_port, &sge->stopped_tx_queues);
1707                 sge->stats.cmdQ_full[2]++;
1708         }
1709
1710         /* T204 cmdQ0 skbs that are destined for a certain port have to go
1711          * through the scheduler.
1712          */
1713         if (sge->tx_sched && !qid && skb->dev) {
1714 use_sched:
1715                 use_sched_skb = 1;
1716                 /* Note that the scheduler might return a different skb than
1717                  * the one passed in.
1718                  */
1719                 skb = sched_skb(sge, skb, credits);
1720                 if (!skb) {
1721                         spin_unlock(&q->lock);
1722                         return NETDEV_TX_OK;
1723                 }
1724                 pidx = q->pidx;
1725                 count = 1 + skb_shinfo(skb)->nr_frags;
1726                 count += compute_large_page_tx_descs(skb);
1727         }
1728
1729         q->in_use += count;
1730         genbit = q->genbit;
1731         pidx = q->pidx;
1732         q->pidx += count;
1733         if (q->pidx >= q->size) {
1734                 q->pidx -= q->size;
1735                 q->genbit ^= 1;
1736         }
1737         spin_unlock(&q->lock);
1738
1739         write_tx_descs(adapter, skb, pidx, genbit, q);
1740
1741         /*
1742          * We always ring the doorbell for cmdQ1.  For cmdQ0, we only ring
1743          * the doorbell if the Q is asleep. There is a natural race, where
1744          * the hardware is going to sleep just after we checked, however,
1745          * then the interrupt handler will detect the outstanding TX packet
1746          * and ring the doorbell for us.
1747          */
1748         if (qid)
1749                 doorbell_pio(adapter, F_CMDQ1_ENABLE);
1750         else {
1751                 clear_bit(CMDQ_STAT_LAST_PKT_DB, &q->status);
1752                 if (test_and_set_bit(CMDQ_STAT_RUNNING, &q->status) == 0) {
1753                         set_bit(CMDQ_STAT_LAST_PKT_DB, &q->status);
1754                         writel(F_CMDQ0_ENABLE, adapter->regs + A_SG_DOORBELL);
1755                 }
1756         }
1757
1758         if (use_sched_skb) {
1759                 if (spin_trylock(&q->lock)) {
1760                         credits = q->size - q->in_use;
1761                         skb = NULL;
1762                         goto use_sched;
1763                 }
1764         }
1765         return NETDEV_TX_OK;
1766 }
1767
1768 #define MK_ETH_TYPE_MSS(type, mss) (((mss) & 0x3FFF) | ((type) << 14))
1769
1770 /*
1771  *      eth_hdr_len - return the length of an Ethernet header
1772  *      @data: pointer to the start of the Ethernet header
1773  *
1774  *      Returns the length of an Ethernet header, including optional VLAN tag.
1775  */
1776 static inline int eth_hdr_len(const void *data)
1777 {
1778         const struct ethhdr *e = data;
1779
1780         return e->h_proto == htons(ETH_P_8021Q) ? VLAN_ETH_HLEN : ETH_HLEN;
1781 }
1782
1783 /*
1784  * Adds the CPL header to the sk_buff and passes it to t1_sge_tx.
1785  */
1786 netdev_tx_t t1_start_xmit(struct sk_buff *skb, struct net_device *dev)
1787 {
1788         struct adapter *adapter = dev->ml_priv;
1789         struct sge *sge = adapter->sge;
1790         struct sge_port_stats *st = this_cpu_ptr(sge->port_stats[dev->if_port]);
1791         struct cpl_tx_pkt *cpl;
1792         struct sk_buff *orig_skb = skb;
1793         int ret;
1794
1795         if (skb->protocol == htons(ETH_P_CPL5))
1796                 goto send;
1797
1798         /*
1799          * We are using a non-standard hard_header_len.
1800          * Allocate more header room in the rare cases it is not big enough.
1801          */
1802         if (unlikely(skb_headroom(skb) < dev->hard_header_len - ETH_HLEN)) {
1803                 skb = skb_realloc_headroom(skb, sizeof(struct cpl_tx_pkt_lso));
1804                 ++st->tx_need_hdrroom;
1805                 dev_kfree_skb_any(orig_skb);
1806                 if (!skb)
1807                         return NETDEV_TX_OK;
1808         }
1809
1810         if (skb_shinfo(skb)->gso_size) {
1811                 int eth_type;
1812                 struct cpl_tx_pkt_lso *hdr;
1813
1814                 ++st->tx_tso;
1815
1816                 eth_type = skb_network_offset(skb) == ETH_HLEN ?
1817                         CPL_ETH_II : CPL_ETH_II_VLAN;
1818
1819                 hdr = (struct cpl_tx_pkt_lso *)skb_push(skb, sizeof(*hdr));
1820                 hdr->opcode = CPL_TX_PKT_LSO;
1821                 hdr->ip_csum_dis = hdr->l4_csum_dis = 0;
1822                 hdr->ip_hdr_words = ip_hdr(skb)->ihl;
1823                 hdr->tcp_hdr_words = tcp_hdr(skb)->doff;
1824                 hdr->eth_type_mss = htons(MK_ETH_TYPE_MSS(eth_type,
1825                                                           skb_shinfo(skb)->gso_size));
1826                 hdr->len = htonl(skb->len - sizeof(*hdr));
1827                 cpl = (struct cpl_tx_pkt *)hdr;
1828         } else {
1829                 /*
1830                  * Packets shorter than ETH_HLEN can break the MAC, drop them
1831                  * early.  Also, we may get oversized packets because some
1832                  * parts of the kernel don't handle our unusual hard_header_len
1833                  * right, drop those too.
1834                  */
1835                 if (unlikely(skb->len < ETH_HLEN ||
1836                              skb->len > dev->mtu + eth_hdr_len(skb->data))) {
1837                         pr_debug("%s: packet size %d hdr %d mtu%d\n", dev->name,
1838                                  skb->len, eth_hdr_len(skb->data), dev->mtu);
1839                         dev_kfree_skb_any(skb);
1840                         return NETDEV_TX_OK;
1841                 }
1842
1843                 if (skb->ip_summed == CHECKSUM_PARTIAL &&
1844                     ip_hdr(skb)->protocol == IPPROTO_UDP) {
1845                         if (unlikely(skb_checksum_help(skb))) {
1846                                 pr_debug("%s: unable to do udp checksum\n", dev->name);
1847                                 dev_kfree_skb_any(skb);
1848                                 return NETDEV_TX_OK;
1849                         }
1850                 }
1851
1852                 /* Hmmm, assuming to catch the gratious arp... and we'll use
1853                  * it to flush out stuck espi packets...
1854                  */
1855                 if ((unlikely(!adapter->sge->espibug_skb[dev->if_port]))) {
1856                         if (skb->protocol == htons(ETH_P_ARP) &&
1857                             arp_hdr(skb)->ar_op == htons(ARPOP_REQUEST)) {
1858                                 adapter->sge->espibug_skb[dev->if_port] = skb;
1859                                 /* We want to re-use this skb later. We
1860                                  * simply bump the reference count and it
1861                                  * will not be freed...
1862                                  */
1863                                 skb = skb_get(skb);
1864                         }
1865                 }
1866
1867                 cpl = (struct cpl_tx_pkt *)__skb_push(skb, sizeof(*cpl));
1868                 cpl->opcode = CPL_TX_PKT;
1869                 cpl->ip_csum_dis = 1;    /* SW calculates IP csum */
1870                 cpl->l4_csum_dis = skb->ip_summed == CHECKSUM_PARTIAL ? 0 : 1;
1871                 /* the length field isn't used so don't bother setting it */
1872
1873                 st->tx_cso += (skb->ip_summed == CHECKSUM_PARTIAL);
1874         }
1875         cpl->iff = dev->if_port;
1876
1877         if (vlan_tx_tag_present(skb)) {
1878                 cpl->vlan_valid = 1;
1879                 cpl->vlan = htons(vlan_tx_tag_get(skb));
1880                 st->vlan_insert++;
1881         } else
1882                 cpl->vlan_valid = 0;
1883
1884 send:
1885         ret = t1_sge_tx(skb, adapter, 0, dev);
1886
1887         /* If transmit busy, and we reallocated skb's due to headroom limit,
1888          * then silently discard to avoid leak.
1889          */
1890         if (unlikely(ret != NETDEV_TX_OK && skb != orig_skb)) {
1891                 dev_kfree_skb_any(skb);
1892                 ret = NETDEV_TX_OK;
1893         }
1894         return ret;
1895 }
1896
1897 /*
1898  * Callback for the Tx buffer reclaim timer.  Runs with softirqs disabled.
1899  */
1900 static void sge_tx_reclaim_cb(unsigned long data)
1901 {
1902         int i;
1903         struct sge *sge = (struct sge *)data;
1904
1905         for (i = 0; i < SGE_CMDQ_N; ++i) {
1906                 struct cmdQ *q = &sge->cmdQ[i];
1907
1908                 if (!spin_trylock(&q->lock))
1909                         continue;
1910
1911                 reclaim_completed_tx(sge, q);
1912                 if (i == 0 && q->in_use) {    /* flush pending credits */
1913                         writel(F_CMDQ0_ENABLE, sge->adapter->regs + A_SG_DOORBELL);
1914                 }
1915                 spin_unlock(&q->lock);
1916         }
1917         mod_timer(&sge->tx_reclaim_timer, jiffies + TX_RECLAIM_PERIOD);
1918 }
1919
1920 /*
1921  * Propagate changes of the SGE coalescing parameters to the HW.
1922  */
1923 int t1_sge_set_coalesce_params(struct sge *sge, struct sge_params *p)
1924 {
1925         sge->fixed_intrtimer = p->rx_coalesce_usecs *
1926                 core_ticks_per_usec(sge->adapter);
1927         writel(sge->fixed_intrtimer, sge->adapter->regs + A_SG_INTRTIMER);
1928         return 0;
1929 }
1930
1931 /*
1932  * Allocates both RX and TX resources and configures the SGE. However,
1933  * the hardware is not enabled yet.
1934  */
1935 int t1_sge_configure(struct sge *sge, struct sge_params *p)
1936 {
1937         if (alloc_rx_resources(sge, p))
1938                 return -ENOMEM;
1939         if (alloc_tx_resources(sge, p)) {
1940                 free_rx_resources(sge);
1941                 return -ENOMEM;
1942         }
1943         configure_sge(sge, p);
1944
1945         /*
1946          * Now that we have sized the free lists calculate the payload
1947          * capacity of the large buffers.  Other parts of the driver use
1948          * this to set the max offload coalescing size so that RX packets
1949          * do not overflow our large buffers.
1950          */
1951         p->large_buf_capacity = jumbo_payload_capacity(sge);
1952         return 0;
1953 }
1954
1955 /*
1956  * Disables the DMA engine.
1957  */
1958 void t1_sge_stop(struct sge *sge)
1959 {
1960         int i;
1961         writel(0, sge->adapter->regs + A_SG_CONTROL);
1962         readl(sge->adapter->regs + A_SG_CONTROL); /* flush */
1963
1964         if (is_T2(sge->adapter))
1965                 del_timer_sync(&sge->espibug_timer);
1966
1967         del_timer_sync(&sge->tx_reclaim_timer);
1968         if (sge->tx_sched)
1969                 tx_sched_stop(sge);
1970
1971         for (i = 0; i < MAX_NPORTS; i++)
1972                 kfree_skb(sge->espibug_skb[i]);
1973 }
1974
1975 /*
1976  * Enables the DMA engine.
1977  */
1978 void t1_sge_start(struct sge *sge)
1979 {
1980         refill_free_list(sge, &sge->freelQ[0]);
1981         refill_free_list(sge, &sge->freelQ[1]);
1982
1983         writel(sge->sge_control, sge->adapter->regs + A_SG_CONTROL);
1984         doorbell_pio(sge->adapter, F_FL0_ENABLE | F_FL1_ENABLE);
1985         readl(sge->adapter->regs + A_SG_CONTROL); /* flush */
1986
1987         mod_timer(&sge->tx_reclaim_timer, jiffies + TX_RECLAIM_PERIOD);
1988
1989         if (is_T2(sge->adapter))
1990                 mod_timer(&sge->espibug_timer, jiffies + sge->espibug_timeout);
1991 }
1992
1993 /*
1994  * Callback for the T2 ESPI 'stuck packet feature' workaorund
1995  */
1996 static void espibug_workaround_t204(unsigned long data)
1997 {
1998         struct adapter *adapter = (struct adapter *)data;
1999         struct sge *sge = adapter->sge;
2000         unsigned int nports = adapter->params.nports;
2001         u32 seop[MAX_NPORTS];
2002
2003         if (adapter->open_device_map & PORT_MASK) {
2004                 int i;
2005
2006                 if (t1_espi_get_mon_t204(adapter, &(seop[0]), 0) < 0)
2007                         return;
2008
2009                 for (i = 0; i < nports; i++) {
2010                         struct sk_buff *skb = sge->espibug_skb[i];
2011
2012                         if (!netif_running(adapter->port[i].dev) ||
2013                             netif_queue_stopped(adapter->port[i].dev) ||
2014                             !seop[i] || ((seop[i] & 0xfff) != 0) || !skb)
2015                                 continue;
2016
2017                         if (!skb->cb[0]) {
2018                                 skb_copy_to_linear_data_offset(skb,
2019                                                     sizeof(struct cpl_tx_pkt),
2020                                                                ch_mac_addr,
2021                                                                ETH_ALEN);
2022                                 skb_copy_to_linear_data_offset(skb,
2023                                                                skb->len - 10,
2024                                                                ch_mac_addr,
2025                                                                ETH_ALEN);
2026                                 skb->cb[0] = 0xff;
2027                         }
2028
2029                         /* bump the reference count to avoid freeing of
2030                          * the skb once the DMA has completed.
2031                          */
2032                         skb = skb_get(skb);
2033                         t1_sge_tx(skb, adapter, 0, adapter->port[i].dev);
2034                 }
2035         }
2036         mod_timer(&sge->espibug_timer, jiffies + sge->espibug_timeout);
2037 }
2038
2039 static void espibug_workaround(unsigned long data)
2040 {
2041         struct adapter *adapter = (struct adapter *)data;
2042         struct sge *sge = adapter->sge;
2043
2044         if (netif_running(adapter->port[0].dev)) {
2045                 struct sk_buff *skb = sge->espibug_skb[0];
2046                 u32 seop = t1_espi_get_mon(adapter, 0x930, 0);
2047
2048                 if ((seop & 0xfff0fff) == 0xfff && skb) {
2049                         if (!skb->cb[0]) {
2050                                 skb_copy_to_linear_data_offset(skb,
2051                                                      sizeof(struct cpl_tx_pkt),
2052                                                                ch_mac_addr,
2053                                                                ETH_ALEN);
2054                                 skb_copy_to_linear_data_offset(skb,
2055                                                                skb->len - 10,
2056                                                                ch_mac_addr,
2057                                                                ETH_ALEN);
2058                                 skb->cb[0] = 0xff;
2059                         }
2060
2061                         /* bump the reference count to avoid freeing of the
2062                          * skb once the DMA has completed.
2063                          */
2064                         skb = skb_get(skb);
2065                         t1_sge_tx(skb, adapter, 0, adapter->port[0].dev);
2066                 }
2067         }
2068         mod_timer(&sge->espibug_timer, jiffies + sge->espibug_timeout);
2069 }
2070
2071 /*
2072  * Creates a t1_sge structure and returns suggested resource parameters.
2073  */
2074 struct sge * __devinit t1_sge_create(struct adapter *adapter,
2075                                      struct sge_params *p)
2076 {
2077         struct sge *sge = kzalloc(sizeof(*sge), GFP_KERNEL);
2078         int i;
2079
2080         if (!sge)
2081                 return NULL;
2082
2083         sge->adapter = adapter;
2084         sge->netdev = adapter->port[0].dev;
2085         sge->rx_pkt_pad = t1_is_T1B(adapter) ? 0 : 2;
2086         sge->jumbo_fl = t1_is_T1B(adapter) ? 1 : 0;
2087
2088         for_each_port(adapter, i) {
2089                 sge->port_stats[i] = alloc_percpu(struct sge_port_stats);
2090                 if (!sge->port_stats[i])
2091                         goto nomem_port;
2092         }
2093
2094         init_timer(&sge->tx_reclaim_timer);
2095         sge->tx_reclaim_timer.data = (unsigned long)sge;
2096         sge->tx_reclaim_timer.function = sge_tx_reclaim_cb;
2097
2098         if (is_T2(sge->adapter)) {
2099                 init_timer(&sge->espibug_timer);
2100
2101                 if (adapter->params.nports > 1) {
2102                         tx_sched_init(sge);
2103                         sge->espibug_timer.function = espibug_workaround_t204;
2104                 } else
2105                         sge->espibug_timer.function = espibug_workaround;
2106                 sge->espibug_timer.data = (unsigned long)sge->adapter;
2107
2108                 sge->espibug_timeout = 1;
2109                 /* for T204, every 10ms */
2110                 if (adapter->params.nports > 1)
2111                         sge->espibug_timeout = HZ/100;
2112         }
2113
2114
2115         p->cmdQ_size[0] = SGE_CMDQ0_E_N;
2116         p->cmdQ_size[1] = SGE_CMDQ1_E_N;
2117         p->freelQ_size[!sge->jumbo_fl] = SGE_FREEL_SIZE;
2118         p->freelQ_size[sge->jumbo_fl] = SGE_JUMBO_FREEL_SIZE;
2119         if (sge->tx_sched) {
2120                 if (board_info(sge->adapter)->board == CHBT_BOARD_CHT204)
2121                         p->rx_coalesce_usecs = 15;
2122                 else
2123                         p->rx_coalesce_usecs = 50;
2124         } else
2125                 p->rx_coalesce_usecs = 50;
2126
2127         p->coalesce_enable = 0;
2128         p->sample_interval_usecs = 0;
2129
2130         return sge;
2131 nomem_port:
2132         while (i >= 0) {
2133                 free_percpu(sge->port_stats[i]);
2134                 --i;
2135         }
2136         kfree(sge);
2137         return NULL;
2138
2139 }