Home | History | Annotate | Line # | Download | only in pci
if_bge.c revision 1.98
      1 /*	$NetBSD: if_bge.c,v 1.98 2005/12/07 04:43:05 jonathan Exp $	*/
      2 
      3 /*
      4  * Copyright (c) 2001 Wind River Systems
      5  * Copyright (c) 1997, 1998, 1999, 2001
      6  *	Bill Paul <wpaul (at) windriver.com>.  All rights reserved.
      7  *
      8  * Redistribution and use in source and binary forms, with or without
      9  * modification, are permitted provided that the following conditions
     10  * are met:
     11  * 1. Redistributions of source code must retain the above copyright
     12  *    notice, this list of conditions and the following disclaimer.
     13  * 2. Redistributions in binary form must reproduce the above copyright
     14  *    notice, this list of conditions and the following disclaimer in the
     15  *    documentation and/or other materials provided with the distribution.
     16  * 3. All advertising materials mentioning features or use of this software
     17  *    must display the following acknowledgement:
     18  *	This product includes software developed by Bill Paul.
     19  * 4. Neither the name of the author nor the names of any co-contributors
     20  *    may be used to endorse or promote products derived from this software
     21  *    without specific prior written permission.
     22  *
     23  * THIS SOFTWARE IS PROVIDED BY Bill Paul AND CONTRIBUTORS ``AS IS'' AND
     24  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
     25  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
     26  * ARE DISCLAIMED.  IN NO EVENT SHALL Bill Paul OR THE VOICES IN HIS HEAD
     27  * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
     28  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
     29  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
     30  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
     31  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
     32  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF
     33  * THE POSSIBILITY OF SUCH DAMAGE.
     34  *
     35  * $FreeBSD: if_bge.c,v 1.13 2002/04/04 06:01:31 wpaul Exp $
     36  */
     37 
     38 /*
     39  * Broadcom BCM570x family gigabit ethernet driver for NetBSD.
     40  *
     41  * NetBSD version by:
     42  *
     43  *	Frank van der Linden <fvdl (at) wasabisystems.com>
     44  *	Jason Thorpe <thorpej (at) wasabisystems.com>
     45  *	Jonathan Stone <jonathan (at) dsg.stanford.edu>
     46  *
     47  * Originally written for FreeBSD by Bill Paul <wpaul (at) windriver.com>
     48  * Senior Engineer, Wind River Systems
     49  */
     50 
     51 /*
     52  * The Broadcom BCM5700 is based on technology originally developed by
     53  * Alteon Networks as part of the Tigon I and Tigon II gigabit ethernet
     54  * MAC chips. The BCM5700, sometimes refered to as the Tigon III, has
     55  * two on-board MIPS R4000 CPUs and can have as much as 16MB of external
     56  * SSRAM. The BCM5700 supports TCP, UDP and IP checksum offload, jumbo
     57  * frames, highly configurable RX filtering, and 16 RX and TX queues
     58  * (which, along with RX filter rules, can be used for QOS applications).
     59  * Other features, such as TCP segmentation, may be available as part
     60  * of value-added firmware updates. Unlike the Tigon I and Tigon II,
     61  * firmware images can be stored in hardware and need not be compiled
     62  * into the driver.
     63  *
     64  * The BCM5700 supports the PCI v2.2 and PCI-X v1.0 standards, and will
     65  * function in a 32-bit/64-bit 33/66MHz bus, or a 64-bit/133MHz bus.
     66  *
     67  * The BCM5701 is a single-chip solution incorporating both the BCM5700
     68  * MAC and a BCM5401 10/100/1000 PHY. Unlike the BCM5700, the BCM5701
     69  * does not support external SSRAM.
     70  *
     71  * Broadcom also produces a variation of the BCM5700 under the "Altima"
     72  * brand name, which is functionally similar but lacks PCI-X support.
     73  *
     74  * Without external SSRAM, you can only have at most 4 TX rings,
     75  * and the use of the mini RX ring is disabled. This seems to imply
     76  * that these features are simply not available on the BCM5701. As a
     77  * result, this driver does not implement any support for the mini RX
     78  * ring.
     79  */
     80 
     81 #include <sys/cdefs.h>
     82 __KERNEL_RCSID(0, "$NetBSD: if_bge.c,v 1.98 2005/12/07 04:43:05 jonathan Exp $");
     83 
     84 #include "bpfilter.h"
     85 #include "vlan.h"
     86 
     87 #include <sys/param.h>
     88 #include <sys/systm.h>
     89 #include <sys/callout.h>
     90 #include <sys/sockio.h>
     91 #include <sys/mbuf.h>
     92 #include <sys/malloc.h>
     93 #include <sys/kernel.h>
     94 #include <sys/device.h>
     95 #include <sys/socket.h>
     96 #include <sys/sysctl.h>
     97 
     98 #include <net/if.h>
     99 #include <net/if_dl.h>
    100 #include <net/if_media.h>
    101 #include <net/if_ether.h>
    102 
    103 #ifdef INET
    104 #include <netinet/in.h>
    105 #include <netinet/in_systm.h>
    106 #include <netinet/in_var.h>
    107 #include <netinet/ip.h>
    108 #endif
    109 
    110 /* Headers for TCP  Segmentation Offload (TSO) */
    111 #include <netinet/in_systm.h>		/* n_time for <netinet/ip.h>... */
    112 #include <netinet/in.h>			/* ip_{src,dst}, for <netinet/ip.h> */
    113 #include <netinet/ip.h>			/* for struct ip */
    114 #include <netinet/tcp.h>		/* for struct tcphdr */
    115 
    116 
    117 #if NBPFILTER > 0
    118 #include <net/bpf.h>
    119 #endif
    120 
    121 #include <dev/pci/pcireg.h>
    122 #include <dev/pci/pcivar.h>
    123 #include <dev/pci/pcidevs.h>
    124 
    125 #include <dev/mii/mii.h>
    126 #include <dev/mii/miivar.h>
    127 #include <dev/mii/miidevs.h>
    128 #include <dev/mii/brgphyreg.h>
    129 
    130 #include <dev/pci/if_bgereg.h>
    131 
    132 #include <uvm/uvm_extern.h>
    133 
    134 #define ETHER_MIN_NOPAD (ETHER_MIN_LEN - ETHER_CRC_LEN) /* i.e., 60 */
    135 
    136 
    137 /*
    138  * Tunable thresholds for rx-side bge interrupt mitigation.
    139  */
    140 
    141 /*
    142  * The pairs of values below were obtained from empirical measurement
    143  * on bcm5700 rev B2; they ar designed to give roughly 1 receive
    144  * interrupt for every N packets received, where N is, approximately,
    145  * the second value (rx_max_bds) in each pair.  The values are chosen
    146  * such that moving from one pair to the succeeding pair was observed
    147  * to roughly halve interrupt rate under sustained input packet load.
    148  * The values were empirically chosen to avoid overflowing internal
    149  * limits on the  bcm5700: inreasing rx_ticks much beyond 600
    150  * results in internal wrapping and higher interrupt rates.
    151  * The limit of 46 frames was chosen to match NFS workloads.
    152  *
    153  * These values also work well on bcm5701, bcm5704C, and (less
    154  * tested) bcm5703.  On other chipsets, (including the Altima chip
    155  * family), the larger values may overflow internal chip limits,
    156  * leading to increasing interrupt rates rather than lower interrupt
    157  * rates.
    158  *
    159  * Applications using heavy interrupt mitigation (interrupting every
    160  * 32 or 46 frames) in both directions may need to increase the TCP
    161  * windowsize to above 131072 bytes (e.g., to 199608 bytes) to sustain
    162  * full link bandwidth, due to ACKs and window updates lingering
    163  * in the RX queue during the 30-to-40-frame interrupt-mitigation window.
    164  */
    165 struct bge_load_rx_thresh {
    166 	int rx_ticks;
    167 	int rx_max_bds; }
    168 bge_rx_threshes[] = {
    169 	{ 32,   2 },
    170 	{ 50,   4 },
    171 	{ 100,  8 },
    172 	{ 192, 16 },
    173 	{ 416, 32 },
    174 	{ 598, 46 }
    175 };
    176 #define NBGE_RX_THRESH (sizeof(bge_rx_threshes) / sizeof(bge_rx_threshes[0]))
    177 
    178 /* XXX patchable; should be sysctl'able */
    179 static int	bge_auto_thresh = 1;
    180 static int	bge_rx_thresh_lvl;
    181 
    182 #ifdef __NetBSD__
    183 static int bge_rxthresh_nodenum;
    184 #endif /* __NetBSD__ */
    185 
    186 int bge_probe(struct device *, struct cfdata *, void *);
    187 void bge_attach(struct device *, struct device *, void *);
    188 void bge_powerhook(int, void *);
    189 void bge_release_resources(struct bge_softc *);
    190 void bge_txeof(struct bge_softc *);
    191 void bge_rxeof(struct bge_softc *);
    192 
    193 void bge_tick(void *);
    194 void bge_stats_update(struct bge_softc *);
    195 int bge_encap(struct bge_softc *, struct mbuf *, u_int32_t *);
    196 static __inline int bge_cksum_pad(struct mbuf *pkt);
    197 static __inline int bge_compact_dma_runt(struct mbuf *pkt);
    198 
    199 int bge_intr(void *);
    200 void bge_start(struct ifnet *);
    201 int bge_ioctl(struct ifnet *, u_long, caddr_t);
    202 int bge_init(struct ifnet *);
    203 void bge_stop(struct bge_softc *);
    204 void bge_watchdog(struct ifnet *);
    205 void bge_shutdown(void *);
    206 int bge_ifmedia_upd(struct ifnet *);
    207 void bge_ifmedia_sts(struct ifnet *, struct ifmediareq *);
    208 
    209 u_int8_t bge_eeprom_getbyte(struct bge_softc *, int, u_int8_t *);
    210 int bge_read_eeprom(struct bge_softc *, caddr_t, int, int);
    211 
    212 void bge_setmulti(struct bge_softc *);
    213 
    214 void bge_handle_events(struct bge_softc *);
    215 int bge_alloc_jumbo_mem(struct bge_softc *);
    216 void bge_free_jumbo_mem(struct bge_softc *);
    217 void *bge_jalloc(struct bge_softc *);
    218 void bge_jfree(struct mbuf *, caddr_t, size_t, void *);
    219 int bge_newbuf_std(struct bge_softc *, int, struct mbuf *, bus_dmamap_t);
    220 int bge_newbuf_jumbo(struct bge_softc *, int, struct mbuf *);
    221 int bge_init_rx_ring_std(struct bge_softc *);
    222 void bge_free_rx_ring_std(struct bge_softc *);
    223 int bge_init_rx_ring_jumbo(struct bge_softc *);
    224 void bge_free_rx_ring_jumbo(struct bge_softc *);
    225 void bge_free_tx_ring(struct bge_softc *);
    226 int bge_init_tx_ring(struct bge_softc *);
    227 
    228 int bge_chipinit(struct bge_softc *);
    229 int bge_blockinit(struct bge_softc *);
    230 int bge_setpowerstate(struct bge_softc *, int);
    231 
    232 #ifdef notdef
    233 u_int8_t bge_vpd_readbyte(struct bge_softc *, int);
    234 void bge_vpd_read_res(struct bge_softc *, struct vpd_res *, int);
    235 void bge_vpd_read(struct bge_softc *);
    236 #endif
    237 
    238 u_int32_t bge_readmem_ind(struct bge_softc *, int);
    239 void bge_writemem_ind(struct bge_softc *, int, int);
    240 #ifdef notdef
    241 u_int32_t bge_readreg_ind(struct bge_softc *, int);
    242 #endif
    243 void bge_writereg_ind(struct bge_softc *, int, int);
    244 
    245 int bge_miibus_readreg(struct device *, int, int);
    246 void bge_miibus_writereg(struct device *, int, int, int);
    247 void bge_miibus_statchg(struct device *);
    248 
    249 void bge_reset(struct bge_softc *);
    250 
    251 void	bge_set_thresh(struct ifnet *  /*ifp*/, int /*lvl*/);
    252 void	bge_update_all_threshes(int /*lvl*/);
    253 
    254 void bge_dump_status(struct bge_softc *);
    255 void bge_dump_rxbd(struct bge_rx_bd *);
    256 
    257 
    258 #define BGE_DEBUG
    259 #ifdef BGE_DEBUG
    260 #define DPRINTF(x)	if (bgedebug) printf x
    261 #define DPRINTFN(n,x)	if (bgedebug >= (n)) printf x
    262 #define BGE_TSO_PRINTF(x)  do { if (bge_tso_debug) printf x ;} while (0)
    263 int	bgedebug = 0;
    264 int	bge_tso_debug = 0;
    265 #else
    266 #define DPRINTF(x)
    267 #define DPRINTFN(n,x)
    268 #define BGE_TSO_PRINTF(x)
    269 #endif
    270 
    271 #ifdef BGE_EVENT_COUNTERS
    272 #define	BGE_EVCNT_INCR(ev)	(ev).ev_count++
    273 #define	BGE_EVCNT_ADD(ev, val)	(ev).ev_count += (val)
    274 #define	BGE_EVCNT_UPD(ev, val)	(ev).ev_count = (val)
    275 #else
    276 #define	BGE_EVCNT_INCR(ev)	/* nothing */
    277 #define	BGE_EVCNT_ADD(ev, val)	/* nothing */
    278 #define	BGE_EVCNT_UPD(ev, val)	/* nothing */
    279 #endif
    280 
    281 /* Various chip quirks. */
    282 #define	BGE_QUIRK_LINK_STATE_BROKEN	0x00000001
    283 #define	BGE_QUIRK_CSUM_BROKEN		0x00000002
    284 #define	BGE_QUIRK_ONLY_PHY_1		0x00000004
    285 #define	BGE_QUIRK_5700_SMALLDMA		0x00000008
    286 #define	BGE_QUIRK_5700_PCIX_REG_BUG	0x00000010
    287 #define	BGE_QUIRK_PRODUCER_BUG		0x00000020
    288 #define	BGE_QUIRK_PCIX_DMA_ALIGN_BUG	0x00000040
    289 #define	BGE_QUIRK_5705_CORE		0x00000080
    290 #define	BGE_QUIRK_FEWER_MBUFS		0x00000100
    291 
    292 /*
    293  * XXX: how to handle variants based on 5750 and derivatives:
    294  * 5750 5751, 5721, possibly 5714, 5752, and 5708?, which
    295  * in general behave like a 5705, except with additional quirks.
    296  * This driver's current handling of the 5721 is wrong;
    297  * how we map ASIC revision to "quirks" needs more thought.
    298  * (defined here until the thought is done).
    299  */
    300 #define BGE_IS_5750_OR_BEYOND(sc)  \
    301 	(BGE_ASICREV(sc->bge_chipid) == BGE_ASICREV_BCM5750)
    302 
    303 #define BGE_IS_5705_OR_BEYOND(sc)  \
    304 	( ((sc)->bge_quirks & BGE_QUIRK_5705_CORE) || \
    305 	  BGE_IS_5750_OR_BEYOND(sc) )
    306 
    307 
    308 /* following bugs are common to bcm5700 rev B, all flavours */
    309 #define BGE_QUIRK_5700_COMMON \
    310 	(BGE_QUIRK_5700_SMALLDMA|BGE_QUIRK_PRODUCER_BUG)
    311 
    312 CFATTACH_DECL(bge, sizeof(struct bge_softc),
    313     bge_probe, bge_attach, NULL, NULL);
    314 
    315 u_int32_t
    316 bge_readmem_ind(sc, off)
    317 	struct bge_softc *sc;
    318 	int off;
    319 {
    320 	struct pci_attach_args	*pa = &(sc->bge_pa);
    321 	pcireg_t val;
    322 
    323 	pci_conf_write(pa->pa_pc, pa->pa_tag, BGE_PCI_MEMWIN_BASEADDR, off);
    324 	val = pci_conf_read(pa->pa_pc, pa->pa_tag, BGE_PCI_MEMWIN_DATA);
    325 	return val;
    326 }
    327 
    328 void
    329 bge_writemem_ind(sc, off, val)
    330 	struct bge_softc *sc;
    331 	int off, val;
    332 {
    333 	struct pci_attach_args	*pa = &(sc->bge_pa);
    334 
    335 	pci_conf_write(pa->pa_pc, pa->pa_tag, BGE_PCI_MEMWIN_BASEADDR, off);
    336 	pci_conf_write(pa->pa_pc, pa->pa_tag, BGE_PCI_MEMWIN_DATA, val);
    337 }
    338 
    339 #ifdef notdef
    340 u_int32_t
    341 bge_readreg_ind(sc, off)
    342 	struct bge_softc *sc;
    343 	int off;
    344 {
    345 	struct pci_attach_args	*pa = &(sc->bge_pa);
    346 
    347 	pci_conf_write(pa->pa_pc, pa->pa_tag, BGE_PCI_REG_BASEADDR, off);
    348 	return(pci_conf_read(pa->pa_pc, pa->pa_tag, BGE_PCI_REG_DATA));
    349 }
    350 #endif
    351 
    352 void
    353 bge_writereg_ind(sc, off, val)
    354 	struct bge_softc *sc;
    355 	int off, val;
    356 {
    357 	struct pci_attach_args	*pa = &(sc->bge_pa);
    358 
    359 	pci_conf_write(pa->pa_pc, pa->pa_tag, BGE_PCI_REG_BASEADDR, off);
    360 	pci_conf_write(pa->pa_pc, pa->pa_tag, BGE_PCI_REG_DATA, val);
    361 }
    362 
    363 #ifdef notdef
    364 u_int8_t
    365 bge_vpd_readbyte(sc, addr)
    366 	struct bge_softc *sc;
    367 	int addr;
    368 {
    369 	int i;
    370 	u_int32_t val;
    371 	struct pci_attach_args	*pa = &(sc->bge_pa);
    372 
    373 	pci_conf_write(pa->pa_pc, pa->pa_tag, BGE_PCI_VPD_ADDR, addr);
    374 	for (i = 0; i < BGE_TIMEOUT * 10; i++) {
    375 		DELAY(10);
    376 		if (pci_conf_read(pa->pa_pc, pa->pa_tag, BGE_PCI_VPD_ADDR) &
    377 		    BGE_VPD_FLAG)
    378 			break;
    379 	}
    380 
    381 	if (i == BGE_TIMEOUT) {
    382 		printf("%s: VPD read timed out\n", sc->bge_dev.dv_xname);
    383 		return(0);
    384 	}
    385 
    386 	val = pci_conf_read(pa->pa_pc, pa->pa_tag, BGE_PCI_VPD_DATA);
    387 
    388 	return((val >> ((addr % 4) * 8)) & 0xFF);
    389 }
    390 
    391 void
    392 bge_vpd_read_res(sc, res, addr)
    393 	struct bge_softc *sc;
    394 	struct vpd_res *res;
    395 	int addr;
    396 {
    397 	int i;
    398 	u_int8_t *ptr;
    399 
    400 	ptr = (u_int8_t *)res;
    401 	for (i = 0; i < sizeof(struct vpd_res); i++)
    402 		ptr[i] = bge_vpd_readbyte(sc, i + addr);
    403 }
    404 
    405 void
    406 bge_vpd_read(sc)
    407 	struct bge_softc *sc;
    408 {
    409 	int pos = 0, i;
    410 	struct vpd_res res;
    411 
    412 	if (sc->bge_vpd_prodname != NULL)
    413 		free(sc->bge_vpd_prodname, M_DEVBUF);
    414 	if (sc->bge_vpd_readonly != NULL)
    415 		free(sc->bge_vpd_readonly, M_DEVBUF);
    416 	sc->bge_vpd_prodname = NULL;
    417 	sc->bge_vpd_readonly = NULL;
    418 
    419 	bge_vpd_read_res(sc, &res, pos);
    420 
    421 	if (res.vr_id != VPD_RES_ID) {
    422 		printf("%s: bad VPD resource id: expected %x got %x\n",
    423 			sc->bge_dev.dv_xname, VPD_RES_ID, res.vr_id);
    424 		return;
    425 	}
    426 
    427 	pos += sizeof(res);
    428 	sc->bge_vpd_prodname = malloc(res.vr_len + 1, M_DEVBUF, M_NOWAIT);
    429 	if (sc->bge_vpd_prodname == NULL)
    430 		panic("bge_vpd_read");
    431 	for (i = 0; i < res.vr_len; i++)
    432 		sc->bge_vpd_prodname[i] = bge_vpd_readbyte(sc, i + pos);
    433 	sc->bge_vpd_prodname[i] = '\0';
    434 	pos += i;
    435 
    436 	bge_vpd_read_res(sc, &res, pos);
    437 
    438 	if (res.vr_id != VPD_RES_READ) {
    439 		printf("%s: bad VPD resource id: expected %x got %x\n",
    440 		    sc->bge_dev.dv_xname, VPD_RES_READ, res.vr_id);
    441 		return;
    442 	}
    443 
    444 	pos += sizeof(res);
    445 	sc->bge_vpd_readonly = malloc(res.vr_len, M_DEVBUF, M_NOWAIT);
    446 	if (sc->bge_vpd_readonly == NULL)
    447 		panic("bge_vpd_read");
    448 	for (i = 0; i < res.vr_len + 1; i++)
    449 		sc->bge_vpd_readonly[i] = bge_vpd_readbyte(sc, i + pos);
    450 }
    451 #endif
    452 
    453 /*
    454  * Read a byte of data stored in the EEPROM at address 'addr.' The
    455  * BCM570x supports both the traditional bitbang interface and an
    456  * auto access interface for reading the EEPROM. We use the auto
    457  * access method.
    458  */
    459 u_int8_t
    460 bge_eeprom_getbyte(sc, addr, dest)
    461 	struct bge_softc *sc;
    462 	int addr;
    463 	u_int8_t *dest;
    464 {
    465 	int i;
    466 	u_int32_t byte = 0;
    467 
    468 	/*
    469 	 * Enable use of auto EEPROM access so we can avoid
    470 	 * having to use the bitbang method.
    471 	 */
    472 	BGE_SETBIT(sc, BGE_MISC_LOCAL_CTL, BGE_MLC_AUTO_EEPROM);
    473 
    474 	/* Reset the EEPROM, load the clock period. */
    475 	CSR_WRITE_4(sc, BGE_EE_ADDR,
    476 	    BGE_EEADDR_RESET|BGE_EEHALFCLK(BGE_HALFCLK_384SCL));
    477 	DELAY(20);
    478 
    479 	/* Issue the read EEPROM command. */
    480 	CSR_WRITE_4(sc, BGE_EE_ADDR, BGE_EE_READCMD | addr);
    481 
    482 	/* Wait for completion */
    483 	for(i = 0; i < BGE_TIMEOUT * 10; i++) {
    484 		DELAY(10);
    485 		if (CSR_READ_4(sc, BGE_EE_ADDR) & BGE_EEADDR_DONE)
    486 			break;
    487 	}
    488 
    489 	if (i == BGE_TIMEOUT) {
    490 		printf("%s: eeprom read timed out\n", sc->bge_dev.dv_xname);
    491 		return(0);
    492 	}
    493 
    494 	/* Get result. */
    495 	byte = CSR_READ_4(sc, BGE_EE_DATA);
    496 
    497 	*dest = (byte >> ((addr % 4) * 8)) & 0xFF;
    498 
    499 	return(0);
    500 }
    501 
    502 /*
    503  * Read a sequence of bytes from the EEPROM.
    504  */
    505 int
    506 bge_read_eeprom(sc, dest, off, cnt)
    507 	struct bge_softc *sc;
    508 	caddr_t dest;
    509 	int off;
    510 	int cnt;
    511 {
    512 	int err = 0, i;
    513 	u_int8_t byte = 0;
    514 
    515 	for (i = 0; i < cnt; i++) {
    516 		err = bge_eeprom_getbyte(sc, off + i, &byte);
    517 		if (err)
    518 			break;
    519 		*(dest + i) = byte;
    520 	}
    521 
    522 	return(err ? 1 : 0);
    523 }
    524 
    525 int
    526 bge_miibus_readreg(dev, phy, reg)
    527 	struct device *dev;
    528 	int phy, reg;
    529 {
    530 	struct bge_softc *sc = (struct bge_softc *)dev;
    531 	u_int32_t val;
    532 	u_int32_t saved_autopoll;
    533 	int i;
    534 
    535 	/*
    536 	 * Several chips with builtin PHYs will incorrectly answer to
    537 	 * other PHY instances than the builtin PHY at id 1.
    538 	 */
    539 	if (phy != 1 && (sc->bge_quirks & BGE_QUIRK_ONLY_PHY_1))
    540 		return(0);
    541 
    542 	/* Reading with autopolling on may trigger PCI errors */
    543 	saved_autopoll = CSR_READ_4(sc, BGE_MI_MODE);
    544 	if (saved_autopoll & BGE_MIMODE_AUTOPOLL) {
    545 		CSR_WRITE_4(sc, BGE_MI_MODE,
    546 		    saved_autopoll &~ BGE_MIMODE_AUTOPOLL);
    547 		DELAY(40);
    548 	}
    549 
    550 	CSR_WRITE_4(sc, BGE_MI_COMM, BGE_MICMD_READ|BGE_MICOMM_BUSY|
    551 	    BGE_MIPHY(phy)|BGE_MIREG(reg));
    552 
    553 	for (i = 0; i < BGE_TIMEOUT; i++) {
    554 		val = CSR_READ_4(sc, BGE_MI_COMM);
    555 		if (!(val & BGE_MICOMM_BUSY))
    556 			break;
    557 		delay(10);
    558 	}
    559 
    560 	if (i == BGE_TIMEOUT) {
    561 		printf("%s: PHY read timed out\n", sc->bge_dev.dv_xname);
    562 		val = 0;
    563 		goto done;
    564 	}
    565 
    566 	val = CSR_READ_4(sc, BGE_MI_COMM);
    567 
    568 done:
    569 	if (saved_autopoll & BGE_MIMODE_AUTOPOLL) {
    570 		CSR_WRITE_4(sc, BGE_MI_MODE, saved_autopoll);
    571 		DELAY(40);
    572 	}
    573 
    574 	if (val & BGE_MICOMM_READFAIL)
    575 		return(0);
    576 
    577 	return(val & 0xFFFF);
    578 }
    579 
    580 void
    581 bge_miibus_writereg(dev, phy, reg, val)
    582 	struct device *dev;
    583 	int phy, reg, val;
    584 {
    585 	struct bge_softc *sc = (struct bge_softc *)dev;
    586 	u_int32_t saved_autopoll;
    587 	int i;
    588 
    589 	/* Touching the PHY while autopolling is on may trigger PCI errors */
    590 	saved_autopoll = CSR_READ_4(sc, BGE_MI_MODE);
    591 	if (saved_autopoll & BGE_MIMODE_AUTOPOLL) {
    592 		delay(40);
    593 		CSR_WRITE_4(sc, BGE_MI_MODE,
    594 		    saved_autopoll & (~BGE_MIMODE_AUTOPOLL));
    595 		delay(10); /* 40 usec is supposed to be adequate */
    596 	}
    597 
    598 	CSR_WRITE_4(sc, BGE_MI_COMM, BGE_MICMD_WRITE|BGE_MICOMM_BUSY|
    599 	    BGE_MIPHY(phy)|BGE_MIREG(reg)|val);
    600 
    601 	for (i = 0; i < BGE_TIMEOUT; i++) {
    602 		if (!(CSR_READ_4(sc, BGE_MI_COMM) & BGE_MICOMM_BUSY))
    603 			break;
    604 		delay(10);
    605 	}
    606 
    607 	if (saved_autopoll & BGE_MIMODE_AUTOPOLL) {
    608 		CSR_WRITE_4(sc, BGE_MI_MODE, saved_autopoll);
    609 		delay(40);
    610 	}
    611 
    612 	if (i == BGE_TIMEOUT) {
    613 		printf("%s: PHY read timed out\n", sc->bge_dev.dv_xname);
    614 	}
    615 }
    616 
    617 void
    618 bge_miibus_statchg(dev)
    619 	struct device *dev;
    620 {
    621 	struct bge_softc *sc = (struct bge_softc *)dev;
    622 	struct mii_data *mii = &sc->bge_mii;
    623 
    624 	/*
    625 	 * Get flow control negotiation result.
    626 	 */
    627 	if (IFM_SUBTYPE(mii->mii_media.ifm_cur->ifm_media) == IFM_AUTO &&
    628 	    (mii->mii_media_active & IFM_ETH_FMASK) != sc->bge_flowflags) {
    629 		sc->bge_flowflags = mii->mii_media_active & IFM_ETH_FMASK;
    630 		mii->mii_media_active &= ~IFM_ETH_FMASK;
    631 	}
    632 
    633 	BGE_CLRBIT(sc, BGE_MAC_MODE, BGE_MACMODE_PORTMODE);
    634 	if (IFM_SUBTYPE(mii->mii_media_active) == IFM_1000_T) {
    635 		BGE_SETBIT(sc, BGE_MAC_MODE, BGE_PORTMODE_GMII);
    636 	} else {
    637 		BGE_SETBIT(sc, BGE_MAC_MODE, BGE_PORTMODE_MII);
    638 	}
    639 
    640 	if ((mii->mii_media_active & IFM_GMASK) == IFM_FDX) {
    641 		BGE_CLRBIT(sc, BGE_MAC_MODE, BGE_MACMODE_HALF_DUPLEX);
    642 	} else {
    643 		BGE_SETBIT(sc, BGE_MAC_MODE, BGE_MACMODE_HALF_DUPLEX);
    644 	}
    645 
    646 	/*
    647 	 * 802.3x flow control
    648 	 */
    649 	if (sc->bge_flowflags & IFM_ETH_RXPAUSE) {
    650 		BGE_SETBIT(sc, BGE_RX_MODE, BGE_RXMODE_FLOWCTL_ENABLE);
    651 	} else {
    652 		BGE_CLRBIT(sc, BGE_RX_MODE, BGE_RXMODE_FLOWCTL_ENABLE);
    653 	}
    654 	if (sc->bge_flowflags & IFM_ETH_TXPAUSE) {
    655 		BGE_SETBIT(sc, BGE_TX_MODE, BGE_TXMODE_FLOWCTL_ENABLE);
    656 	} else {
    657 		BGE_CLRBIT(sc, BGE_TX_MODE, BGE_TXMODE_FLOWCTL_ENABLE);
    658 	}
    659 }
    660 
    661 /*
    662  * Update rx threshold levels to values in a particular slot
    663  * of the interrupt-mitigation table bge_rx_threshes.
    664  */
    665 void
    666 bge_set_thresh(struct ifnet *ifp, int lvl)
    667 {
    668 	struct bge_softc *sc = ifp->if_softc;
    669 	int s;
    670 
    671 	/* For now, just save the new Rx-intr thresholds and record
    672 	 * that a threshold update is pending.  Updating the hardware
    673 	 * registers here (even at splhigh()) is observed to
    674 	 * occasionaly cause glitches where Rx-interrupts are not
    675 	 * honoured for up to 10 seconds. jonathan (at) NetBSD.org, 2003-04-05
    676 	 */
    677 	s = splnet();
    678 	sc->bge_rx_coal_ticks = bge_rx_threshes[lvl].rx_ticks;
    679 	sc->bge_rx_max_coal_bds = bge_rx_threshes[lvl].rx_max_bds;
    680 	sc->bge_pending_rxintr_change = 1;
    681 	splx(s);
    682 
    683 	 return;
    684 }
    685 
    686 
    687 /*
    688  * Update Rx thresholds of all bge devices
    689  */
    690 void
    691 bge_update_all_threshes(int lvl)
    692 {
    693 	struct ifnet *ifp;
    694 	const char * const namebuf = "bge";
    695 	int namelen;
    696 
    697 	if (lvl < 0)
    698 		lvl = 0;
    699 	else if( lvl >= NBGE_RX_THRESH)
    700 		lvl = NBGE_RX_THRESH - 1;
    701 
    702 	namelen = strlen(namebuf);
    703 	/*
    704 	 * Now search all the interfaces for this name/number
    705 	 */
    706 	IFNET_FOREACH(ifp) {
    707 		if (strncmp(ifp->if_xname, namebuf, namelen) != 0)
    708 		      continue;
    709 		/* We got a match: update if doing auto-threshold-tuning */
    710 		if (bge_auto_thresh)
    711 			bge_set_thresh(ifp, lvl);
    712 	}
    713 }
    714 
    715 /*
    716  * Handle events that have triggered interrupts.
    717  */
    718 void
    719 bge_handle_events(sc)
    720 	struct bge_softc		*sc;
    721 {
    722 
    723 	return;
    724 }
    725 
    726 /*
    727  * Memory management for jumbo frames.
    728  */
    729 
    730 int
    731 bge_alloc_jumbo_mem(sc)
    732 	struct bge_softc		*sc;
    733 {
    734 	caddr_t			ptr, kva;
    735 	bus_dma_segment_t	seg;
    736 	int		i, rseg, state, error;
    737 	struct bge_jpool_entry   *entry;
    738 
    739 	state = error = 0;
    740 
    741 	/* Grab a big chunk o' storage. */
    742 	if (bus_dmamem_alloc(sc->bge_dmatag, BGE_JMEM, PAGE_SIZE, 0,
    743 	     &seg, 1, &rseg, BUS_DMA_NOWAIT)) {
    744 		printf("%s: can't alloc rx buffers\n", sc->bge_dev.dv_xname);
    745 		return ENOBUFS;
    746 	}
    747 
    748 	state = 1;
    749 	if (bus_dmamem_map(sc->bge_dmatag, &seg, rseg, BGE_JMEM, &kva,
    750 	    BUS_DMA_NOWAIT)) {
    751 		printf("%s: can't map DMA buffers (%d bytes)\n",
    752 		    sc->bge_dev.dv_xname, (int)BGE_JMEM);
    753 		error = ENOBUFS;
    754 		goto out;
    755 	}
    756 
    757 	state = 2;
    758 	if (bus_dmamap_create(sc->bge_dmatag, BGE_JMEM, 1, BGE_JMEM, 0,
    759 	    BUS_DMA_NOWAIT, &sc->bge_cdata.bge_rx_jumbo_map)) {
    760 		printf("%s: can't create DMA map\n", sc->bge_dev.dv_xname);
    761 		error = ENOBUFS;
    762 		goto out;
    763 	}
    764 
    765 	state = 3;
    766 	if (bus_dmamap_load(sc->bge_dmatag, sc->bge_cdata.bge_rx_jumbo_map,
    767 	    kva, BGE_JMEM, NULL, BUS_DMA_NOWAIT)) {
    768 		printf("%s: can't load DMA map\n", sc->bge_dev.dv_xname);
    769 		error = ENOBUFS;
    770 		goto out;
    771 	}
    772 
    773 	state = 4;
    774 	sc->bge_cdata.bge_jumbo_buf = (caddr_t)kva;
    775 	DPRINTFN(1,("bge_jumbo_buf = %p\n", sc->bge_cdata.bge_jumbo_buf));
    776 
    777 	SLIST_INIT(&sc->bge_jfree_listhead);
    778 	SLIST_INIT(&sc->bge_jinuse_listhead);
    779 
    780 	/*
    781 	 * Now divide it up into 9K pieces and save the addresses
    782 	 * in an array.
    783 	 */
    784 	ptr = sc->bge_cdata.bge_jumbo_buf;
    785 	for (i = 0; i < BGE_JSLOTS; i++) {
    786 		sc->bge_cdata.bge_jslots[i] = ptr;
    787 		ptr += BGE_JLEN;
    788 		entry = malloc(sizeof(struct bge_jpool_entry),
    789 		    M_DEVBUF, M_NOWAIT);
    790 		if (entry == NULL) {
    791 			printf("%s: no memory for jumbo buffer queue!\n",
    792 			    sc->bge_dev.dv_xname);
    793 			error = ENOBUFS;
    794 			goto out;
    795 		}
    796 		entry->slot = i;
    797 		SLIST_INSERT_HEAD(&sc->bge_jfree_listhead,
    798 				 entry, jpool_entries);
    799 	}
    800 out:
    801 	if (error != 0) {
    802 		switch (state) {
    803 		case 4:
    804 			bus_dmamap_unload(sc->bge_dmatag,
    805 			    sc->bge_cdata.bge_rx_jumbo_map);
    806 		case 3:
    807 			bus_dmamap_destroy(sc->bge_dmatag,
    808 			    sc->bge_cdata.bge_rx_jumbo_map);
    809 		case 2:
    810 			bus_dmamem_unmap(sc->bge_dmatag, kva, BGE_JMEM);
    811 		case 1:
    812 			bus_dmamem_free(sc->bge_dmatag, &seg, rseg);
    813 			break;
    814 		default:
    815 			break;
    816 		}
    817 	}
    818 
    819 	return error;
    820 }
    821 
    822 /*
    823  * Allocate a jumbo buffer.
    824  */
    825 void *
    826 bge_jalloc(sc)
    827 	struct bge_softc		*sc;
    828 {
    829 	struct bge_jpool_entry   *entry;
    830 
    831 	entry = SLIST_FIRST(&sc->bge_jfree_listhead);
    832 
    833 	if (entry == NULL) {
    834 		printf("%s: no free jumbo buffers\n", sc->bge_dev.dv_xname);
    835 		return(NULL);
    836 	}
    837 
    838 	SLIST_REMOVE_HEAD(&sc->bge_jfree_listhead, jpool_entries);
    839 	SLIST_INSERT_HEAD(&sc->bge_jinuse_listhead, entry, jpool_entries);
    840 	return(sc->bge_cdata.bge_jslots[entry->slot]);
    841 }
    842 
    843 /*
    844  * Release a jumbo buffer.
    845  */
    846 void
    847 bge_jfree(m, buf, size, arg)
    848 	struct mbuf	*m;
    849 	caddr_t		buf;
    850 	size_t		size;
    851 	void		*arg;
    852 {
    853 	struct bge_jpool_entry *entry;
    854 	struct bge_softc *sc;
    855 	int i, s;
    856 
    857 	/* Extract the softc struct pointer. */
    858 	sc = (struct bge_softc *)arg;
    859 
    860 	if (sc == NULL)
    861 		panic("bge_jfree: can't find softc pointer!");
    862 
    863 	/* calculate the slot this buffer belongs to */
    864 
    865 	i = ((caddr_t)buf
    866 	     - (caddr_t)sc->bge_cdata.bge_jumbo_buf) / BGE_JLEN;
    867 
    868 	if ((i < 0) || (i >= BGE_JSLOTS))
    869 		panic("bge_jfree: asked to free buffer that we don't manage!");
    870 
    871 	s = splvm();
    872 	entry = SLIST_FIRST(&sc->bge_jinuse_listhead);
    873 	if (entry == NULL)
    874 		panic("bge_jfree: buffer not in use!");
    875 	entry->slot = i;
    876 	SLIST_REMOVE_HEAD(&sc->bge_jinuse_listhead, jpool_entries);
    877 	SLIST_INSERT_HEAD(&sc->bge_jfree_listhead, entry, jpool_entries);
    878 
    879 	if (__predict_true(m != NULL))
    880   		pool_cache_put(&mbpool_cache, m);
    881 	splx(s);
    882 }
    883 
    884 
    885 /*
    886  * Intialize a standard receive ring descriptor.
    887  */
    888 int
    889 bge_newbuf_std(sc, i, m, dmamap)
    890 	struct bge_softc	*sc;
    891 	int			i;
    892 	struct mbuf		*m;
    893 	bus_dmamap_t dmamap;
    894 {
    895 	struct mbuf		*m_new = NULL;
    896 	struct bge_rx_bd	*r;
    897 	int			error;
    898 
    899 	if (dmamap == NULL) {
    900 		error = bus_dmamap_create(sc->bge_dmatag, MCLBYTES, 1,
    901 		    MCLBYTES, 0, BUS_DMA_NOWAIT, &dmamap);
    902 		if (error != 0)
    903 			return error;
    904 	}
    905 
    906 	sc->bge_cdata.bge_rx_std_map[i] = dmamap;
    907 
    908 	if (m == NULL) {
    909 		MGETHDR(m_new, M_DONTWAIT, MT_DATA);
    910 		if (m_new == NULL) {
    911 			return(ENOBUFS);
    912 		}
    913 
    914 		MCLGET(m_new, M_DONTWAIT);
    915 		if (!(m_new->m_flags & M_EXT)) {
    916 			m_freem(m_new);
    917 			return(ENOBUFS);
    918 		}
    919 		m_new->m_len = m_new->m_pkthdr.len = MCLBYTES;
    920 		if (!sc->bge_rx_alignment_bug)
    921 		    m_adj(m_new, ETHER_ALIGN);
    922 
    923 		if (bus_dmamap_load_mbuf(sc->bge_dmatag, dmamap, m_new,
    924 		    BUS_DMA_READ|BUS_DMA_NOWAIT))
    925 			return(ENOBUFS);
    926 	} else {
    927 		m_new = m;
    928 		m_new->m_len = m_new->m_pkthdr.len = MCLBYTES;
    929 		m_new->m_data = m_new->m_ext.ext_buf;
    930 		if (!sc->bge_rx_alignment_bug)
    931 		    m_adj(m_new, ETHER_ALIGN);
    932 	}
    933 
    934 	sc->bge_cdata.bge_rx_std_chain[i] = m_new;
    935 	r = &sc->bge_rdata->bge_rx_std_ring[i];
    936 	bge_set_hostaddr(&r->bge_addr,
    937 	    dmamap->dm_segs[0].ds_addr);
    938 	r->bge_flags = BGE_RXBDFLAG_END;
    939 	r->bge_len = m_new->m_len;
    940 	r->bge_idx = i;
    941 
    942 	bus_dmamap_sync(sc->bge_dmatag, sc->bge_ring_map,
    943 	    offsetof(struct bge_ring_data, bge_rx_std_ring) +
    944 		i * sizeof (struct bge_rx_bd),
    945 	    sizeof (struct bge_rx_bd),
    946 	    BUS_DMASYNC_PREWRITE|BUS_DMASYNC_PREREAD);
    947 
    948 	return(0);
    949 }
    950 
    951 /*
    952  * Initialize a jumbo receive ring descriptor. This allocates
    953  * a jumbo buffer from the pool managed internally by the driver.
    954  */
    955 int
    956 bge_newbuf_jumbo(sc, i, m)
    957 	struct bge_softc *sc;
    958 	int i;
    959 	struct mbuf *m;
    960 {
    961 	struct mbuf *m_new = NULL;
    962 	struct bge_rx_bd *r;
    963 
    964 	if (m == NULL) {
    965 		caddr_t			buf = NULL;
    966 
    967 		/* Allocate the mbuf. */
    968 		MGETHDR(m_new, M_DONTWAIT, MT_DATA);
    969 		if (m_new == NULL) {
    970 			return(ENOBUFS);
    971 		}
    972 
    973 		/* Allocate the jumbo buffer */
    974 		buf = bge_jalloc(sc);
    975 		if (buf == NULL) {
    976 			m_freem(m_new);
    977 			printf("%s: jumbo allocation failed "
    978 			    "-- packet dropped!\n", sc->bge_dev.dv_xname);
    979 			return(ENOBUFS);
    980 		}
    981 
    982 		/* Attach the buffer to the mbuf. */
    983 		m_new->m_len = m_new->m_pkthdr.len = BGE_JUMBO_FRAMELEN;
    984 		MEXTADD(m_new, buf, BGE_JUMBO_FRAMELEN, M_DEVBUF,
    985 		    bge_jfree, sc);
    986 		m_new->m_flags |= M_EXT_RW;
    987 	} else {
    988 		m_new = m;
    989 		m_new->m_data = m_new->m_ext.ext_buf;
    990 		m_new->m_ext.ext_size = BGE_JUMBO_FRAMELEN;
    991 	}
    992 
    993 	if (!sc->bge_rx_alignment_bug)
    994 	    m_adj(m_new, ETHER_ALIGN);
    995 	/* Set up the descriptor. */
    996 	r = &sc->bge_rdata->bge_rx_jumbo_ring[i];
    997 	sc->bge_cdata.bge_rx_jumbo_chain[i] = m_new;
    998 	bge_set_hostaddr(&r->bge_addr, BGE_JUMBO_DMA_ADDR(sc, m_new));
    999 	r->bge_flags = BGE_RXBDFLAG_END|BGE_RXBDFLAG_JUMBO_RING;
   1000 	r->bge_len = m_new->m_len;
   1001 	r->bge_idx = i;
   1002 
   1003 	bus_dmamap_sync(sc->bge_dmatag, sc->bge_ring_map,
   1004 	    offsetof(struct bge_ring_data, bge_rx_jumbo_ring) +
   1005 		i * sizeof (struct bge_rx_bd),
   1006 	    sizeof (struct bge_rx_bd),
   1007 	    BUS_DMASYNC_PREWRITE|BUS_DMASYNC_PREREAD);
   1008 
   1009 	return(0);
   1010 }
   1011 
   1012 /*
   1013  * The standard receive ring has 512 entries in it. At 2K per mbuf cluster,
   1014  * that's 1MB or memory, which is a lot. For now, we fill only the first
   1015  * 256 ring entries and hope that our CPU is fast enough to keep up with
   1016  * the NIC.
   1017  */
   1018 int
   1019 bge_init_rx_ring_std(sc)
   1020 	struct bge_softc *sc;
   1021 {
   1022 	int i;
   1023 
   1024 	if (sc->bge_flags & BGE_RXRING_VALID)
   1025 		return 0;
   1026 
   1027 	for (i = 0; i < BGE_SSLOTS; i++) {
   1028 		if (bge_newbuf_std(sc, i, NULL, 0) == ENOBUFS)
   1029 			return(ENOBUFS);
   1030 	}
   1031 
   1032 	sc->bge_std = i - 1;
   1033 	CSR_WRITE_4(sc, BGE_MBX_RX_STD_PROD_LO, sc->bge_std);
   1034 
   1035 	sc->bge_flags |= BGE_RXRING_VALID;
   1036 
   1037 	return(0);
   1038 }
   1039 
   1040 void
   1041 bge_free_rx_ring_std(sc)
   1042 	struct bge_softc *sc;
   1043 {
   1044 	int i;
   1045 
   1046 	if (!(sc->bge_flags & BGE_RXRING_VALID))
   1047 		return;
   1048 
   1049 	for (i = 0; i < BGE_STD_RX_RING_CNT; i++) {
   1050 		if (sc->bge_cdata.bge_rx_std_chain[i] != NULL) {
   1051 			m_freem(sc->bge_cdata.bge_rx_std_chain[i]);
   1052 			sc->bge_cdata.bge_rx_std_chain[i] = NULL;
   1053 			bus_dmamap_destroy(sc->bge_dmatag,
   1054 			    sc->bge_cdata.bge_rx_std_map[i]);
   1055 		}
   1056 		memset((char *)&sc->bge_rdata->bge_rx_std_ring[i], 0,
   1057 		    sizeof(struct bge_rx_bd));
   1058 	}
   1059 
   1060 	sc->bge_flags &= ~BGE_RXRING_VALID;
   1061 }
   1062 
   1063 int
   1064 bge_init_rx_ring_jumbo(sc)
   1065 	struct bge_softc *sc;
   1066 {
   1067 	int i;
   1068 	volatile struct bge_rcb *rcb;
   1069 
   1070 	if (sc->bge_flags & BGE_JUMBO_RXRING_VALID)
   1071 		return 0;
   1072 
   1073 	for (i = 0; i < BGE_JUMBO_RX_RING_CNT; i++) {
   1074 		if (bge_newbuf_jumbo(sc, i, NULL) == ENOBUFS)
   1075 			return(ENOBUFS);
   1076 	};
   1077 
   1078 	sc->bge_jumbo = i - 1;
   1079 	sc->bge_flags |= BGE_JUMBO_RXRING_VALID;
   1080 
   1081 	rcb = &sc->bge_rdata->bge_info.bge_jumbo_rx_rcb;
   1082 	rcb->bge_maxlen_flags = 0;
   1083 	CSR_WRITE_4(sc, BGE_RX_JUMBO_RCB_MAXLEN_FLAGS, rcb->bge_maxlen_flags);
   1084 
   1085 	CSR_WRITE_4(sc, BGE_MBX_RX_JUMBO_PROD_LO, sc->bge_jumbo);
   1086 
   1087 	return(0);
   1088 }
   1089 
   1090 void
   1091 bge_free_rx_ring_jumbo(sc)
   1092 	struct bge_softc *sc;
   1093 {
   1094 	int i;
   1095 
   1096 	if (!(sc->bge_flags & BGE_JUMBO_RXRING_VALID))
   1097 		return;
   1098 
   1099 	for (i = 0; i < BGE_JUMBO_RX_RING_CNT; i++) {
   1100 		if (sc->bge_cdata.bge_rx_jumbo_chain[i] != NULL) {
   1101 			m_freem(sc->bge_cdata.bge_rx_jumbo_chain[i]);
   1102 			sc->bge_cdata.bge_rx_jumbo_chain[i] = NULL;
   1103 		}
   1104 		memset((char *)&sc->bge_rdata->bge_rx_jumbo_ring[i], 0,
   1105 		    sizeof(struct bge_rx_bd));
   1106 	}
   1107 
   1108 	sc->bge_flags &= ~BGE_JUMBO_RXRING_VALID;
   1109 }
   1110 
   1111 void
   1112 bge_free_tx_ring(sc)
   1113 	struct bge_softc *sc;
   1114 {
   1115 	int i, freed;
   1116 	struct txdmamap_pool_entry *dma;
   1117 
   1118 	if (!(sc->bge_flags & BGE_TXRING_VALID))
   1119 		return;
   1120 
   1121 	freed = 0;
   1122 
   1123 	for (i = 0; i < BGE_TX_RING_CNT; i++) {
   1124 		if (sc->bge_cdata.bge_tx_chain[i] != NULL) {
   1125 			freed++;
   1126 			m_freem(sc->bge_cdata.bge_tx_chain[i]);
   1127 			sc->bge_cdata.bge_tx_chain[i] = NULL;
   1128 			SLIST_INSERT_HEAD(&sc->txdma_list, sc->txdma[i],
   1129 					    link);
   1130 			sc->txdma[i] = 0;
   1131 		}
   1132 		memset((char *)&sc->bge_rdata->bge_tx_ring[i], 0,
   1133 		    sizeof(struct bge_tx_bd));
   1134 	}
   1135 
   1136 	while ((dma = SLIST_FIRST(&sc->txdma_list))) {
   1137 		SLIST_REMOVE_HEAD(&sc->txdma_list, link);
   1138 		bus_dmamap_destroy(sc->bge_dmatag, dma->dmamap);
   1139 		free(dma, M_DEVBUF);
   1140 	}
   1141 
   1142 	sc->bge_flags &= ~BGE_TXRING_VALID;
   1143 }
   1144 
   1145 int
   1146 bge_init_tx_ring(sc)
   1147 	struct bge_softc *sc;
   1148 {
   1149 	int i;
   1150 	bus_dmamap_t dmamap;
   1151 	struct txdmamap_pool_entry *dma;
   1152 
   1153 	if (sc->bge_flags & BGE_TXRING_VALID)
   1154 		return 0;
   1155 
   1156 	sc->bge_txcnt = 0;
   1157 	sc->bge_tx_saved_considx = 0;
   1158 
   1159 	/* Initialize transmit producer index for host-memory send ring. */
   1160 	sc->bge_tx_prodidx = 0;
   1161 	CSR_WRITE_4(sc, BGE_MBX_TX_HOST_PROD0_LO, sc->bge_tx_prodidx);
   1162 	if (sc->bge_quirks & BGE_QUIRK_PRODUCER_BUG)	/* 5700 b2 errata */
   1163 		CSR_WRITE_4(sc, BGE_MBX_TX_HOST_PROD0_LO, sc->bge_tx_prodidx);
   1164 
   1165 	/* NIC-memory send ring  not used; initialize to zero. */
   1166 	CSR_WRITE_4(sc, BGE_MBX_TX_NIC_PROD0_LO, 0);
   1167 	if (sc->bge_quirks & BGE_QUIRK_PRODUCER_BUG)	/* 5700 b2 errata */
   1168 		CSR_WRITE_4(sc, BGE_MBX_TX_HOST_PROD0_LO, 0);
   1169 
   1170 	SLIST_INIT(&sc->txdma_list);
   1171 	for (i = 0; i < BGE_RSLOTS; i++) {
   1172 		if (bus_dmamap_create(sc->bge_dmatag, BGE_TXDMA_MAX,
   1173 		    BGE_NTXSEG, ETHER_MAX_LEN_JUMBO, 0, BUS_DMA_NOWAIT,
   1174 		    &dmamap))
   1175 			return(ENOBUFS);
   1176 		if (dmamap == NULL)
   1177 			panic("dmamap NULL in bge_init_tx_ring");
   1178 		dma = malloc(sizeof(*dma), M_DEVBUF, M_NOWAIT);
   1179 		if (dma == NULL) {
   1180 			printf("%s: can't alloc txdmamap_pool_entry\n",
   1181 			    sc->bge_dev.dv_xname);
   1182 			bus_dmamap_destroy(sc->bge_dmatag, dmamap);
   1183 			return (ENOMEM);
   1184 		}
   1185 		dma->dmamap = dmamap;
   1186 		SLIST_INSERT_HEAD(&sc->txdma_list, dma, link);
   1187 	}
   1188 
   1189 	sc->bge_flags |= BGE_TXRING_VALID;
   1190 
   1191 	return(0);
   1192 }
   1193 
   1194 void
   1195 bge_setmulti(sc)
   1196 	struct bge_softc *sc;
   1197 {
   1198 	struct ethercom		*ac = &sc->ethercom;
   1199 	struct ifnet		*ifp = &ac->ec_if;
   1200 	struct ether_multi	*enm;
   1201 	struct ether_multistep  step;
   1202 	u_int32_t		hashes[4] = { 0, 0, 0, 0 };
   1203 	u_int32_t		h;
   1204 	int			i;
   1205 
   1206 	if (ifp->if_flags & IFF_PROMISC)
   1207 		goto allmulti;
   1208 
   1209 	/* Now program new ones. */
   1210 	ETHER_FIRST_MULTI(step, ac, enm);
   1211 	while (enm != NULL) {
   1212 		if (memcmp(enm->enm_addrlo, enm->enm_addrhi, ETHER_ADDR_LEN)) {
   1213 			/*
   1214 			 * We must listen to a range of multicast addresses.
   1215 			 * For now, just accept all multicasts, rather than
   1216 			 * trying to set only those filter bits needed to match
   1217 			 * the range.  (At this time, the only use of address
   1218 			 * ranges is for IP multicast routing, for which the
   1219 			 * range is big enough to require all bits set.)
   1220 			 */
   1221 			goto allmulti;
   1222 		}
   1223 
   1224 		h = ether_crc32_le(enm->enm_addrlo, ETHER_ADDR_LEN);
   1225 
   1226 		/* Just want the 7 least-significant bits. */
   1227 		h &= 0x7f;
   1228 
   1229 		hashes[(h & 0x60) >> 5] |= 1 << (h & 0x1F);
   1230 		ETHER_NEXT_MULTI(step, enm);
   1231 	}
   1232 
   1233 	ifp->if_flags &= ~IFF_ALLMULTI;
   1234 	goto setit;
   1235 
   1236  allmulti:
   1237 	ifp->if_flags |= IFF_ALLMULTI;
   1238 	hashes[0] = hashes[1] = hashes[2] = hashes[3] = 0xffffffff;
   1239 
   1240  setit:
   1241 	for (i = 0; i < 4; i++)
   1242 		CSR_WRITE_4(sc, BGE_MAR0 + (i * 4), hashes[i]);
   1243 }
   1244 
   1245 const int bge_swapbits[] = {
   1246 	0,
   1247 	BGE_MODECTL_BYTESWAP_DATA,
   1248 	BGE_MODECTL_WORDSWAP_DATA,
   1249 	BGE_MODECTL_BYTESWAP_NONFRAME,
   1250 	BGE_MODECTL_WORDSWAP_NONFRAME,
   1251 
   1252 	BGE_MODECTL_BYTESWAP_DATA|BGE_MODECTL_WORDSWAP_DATA,
   1253 	BGE_MODECTL_BYTESWAP_DATA|BGE_MODECTL_BYTESWAP_NONFRAME,
   1254 	BGE_MODECTL_BYTESWAP_DATA|BGE_MODECTL_WORDSWAP_NONFRAME,
   1255 
   1256 	BGE_MODECTL_WORDSWAP_DATA|BGE_MODECTL_BYTESWAP_NONFRAME,
   1257 	BGE_MODECTL_WORDSWAP_DATA|BGE_MODECTL_WORDSWAP_NONFRAME,
   1258 
   1259 	BGE_MODECTL_BYTESWAP_NONFRAME|BGE_MODECTL_WORDSWAP_NONFRAME,
   1260 
   1261 	BGE_MODECTL_BYTESWAP_DATA|BGE_MODECTL_WORDSWAP_DATA|
   1262 	    BGE_MODECTL_BYTESWAP_NONFRAME,
   1263 	BGE_MODECTL_BYTESWAP_DATA|BGE_MODECTL_WORDSWAP_DATA|
   1264 	    BGE_MODECTL_WORDSWAP_NONFRAME,
   1265 	BGE_MODECTL_BYTESWAP_DATA|BGE_MODECTL_BYTESWAP_NONFRAME|
   1266 	    BGE_MODECTL_WORDSWAP_NONFRAME,
   1267 	BGE_MODECTL_WORDSWAP_DATA|BGE_MODECTL_BYTESWAP_NONFRAME|
   1268 	    BGE_MODECTL_WORDSWAP_NONFRAME,
   1269 
   1270 	BGE_MODECTL_BYTESWAP_DATA|BGE_MODECTL_WORDSWAP_DATA|
   1271 	    BGE_MODECTL_BYTESWAP_NONFRAME|BGE_MODECTL_WORDSWAP_NONFRAME,
   1272 };
   1273 
   1274 int bge_swapindex = 0;
   1275 
   1276 /*
   1277  * Do endian, PCI and DMA initialization. Also check the on-board ROM
   1278  * self-test results.
   1279  */
   1280 int
   1281 bge_chipinit(sc)
   1282 	struct bge_softc *sc;
   1283 {
   1284 	u_int32_t		cachesize;
   1285 	int			i;
   1286 	u_int32_t		dma_rw_ctl;
   1287 	struct pci_attach_args	*pa = &(sc->bge_pa);
   1288 
   1289 
   1290 	/* Set endianness before we access any non-PCI registers. */
   1291 	pci_conf_write(pa->pa_pc, pa->pa_tag, BGE_PCI_MISC_CTL,
   1292 	    BGE_INIT);
   1293 
   1294 	/* Set power state to D0. */
   1295 	bge_setpowerstate(sc, 0);
   1296 
   1297 	/*
   1298 	 * Check the 'ROM failed' bit on the RX CPU to see if
   1299 	 * self-tests passed.
   1300 	 */
   1301 	if (CSR_READ_4(sc, BGE_RXCPU_MODE) & BGE_RXCPUMODE_ROMFAIL) {
   1302 		printf("%s: RX CPU self-diagnostics failed!\n",
   1303 		    sc->bge_dev.dv_xname);
   1304 		return(ENODEV);
   1305 	}
   1306 
   1307 	/* Clear the MAC control register */
   1308 	CSR_WRITE_4(sc, BGE_MAC_MODE, 0);
   1309 
   1310 	/*
   1311 	 * Clear the MAC statistics block in the NIC's
   1312 	 * internal memory.
   1313 	 */
   1314 	for (i = BGE_STATS_BLOCK;
   1315 	    i < BGE_STATS_BLOCK_END + 1; i += sizeof(u_int32_t))
   1316 		BGE_MEMWIN_WRITE(pa->pa_pc, pa->pa_tag, i, 0);
   1317 
   1318 	for (i = BGE_STATUS_BLOCK;
   1319 	    i < BGE_STATUS_BLOCK_END + 1; i += sizeof(u_int32_t))
   1320 		BGE_MEMWIN_WRITE(pa->pa_pc, pa->pa_tag, i, 0);
   1321 
   1322 	/* Set up the PCI DMA control register. */
   1323 	if (sc->bge_pcie) {
   1324 	  u_int32_t device_ctl;
   1325 
   1326 		/* From FreeBSD */
   1327 		DPRINTFN(4, ("(%s: PCI-Express DMA setting)\n",
   1328 		    sc->bge_dev.dv_xname));
   1329 		dma_rw_ctl = (BGE_PCI_READ_CMD | BGE_PCI_WRITE_CMD |
   1330 		    (0xf << BGE_PCIDMARWCTL_RD_WAT_SHIFT) |
   1331 		    (0x2 << BGE_PCIDMARWCTL_WR_WAT_SHIFT));
   1332 
   1333 		/* jonathan: alternative from Linux driver */
   1334 #define DMA_CTRL_WRITE_PCIE_H20MARK_128         0x00180000
   1335 #define DMA_CTRL_WRITE_PCIE_H20MARK_256         0x00380000
   1336 
   1337 		dma_rw_ctl =   0x76000000; /* XXX XXX XXX */;
   1338 		device_ctl = pci_conf_read(pa->pa_pc, pa->pa_tag,
   1339 					   BGE_PCI_CONF_DEV_CTRL);
   1340 		printf("%s: pcie mode=0x%x\n", sc->bge_dev.dv_xname, device_ctl);
   1341 
   1342 		if ((device_ctl & 0x00e0) && 0) {
   1343 			/*
   1344 			 * XXX jonathan (at) NetBSD.org:
   1345 			 * This clause is exactly what the Broadcom-supplied
   1346 			 * Linux does; but given overall register programming
   1347 			 * by if_bge(4), this larger DMA-write watermark
   1348 			 * value causes bcm5721 chips to totally wedge.
   1349 			 */
   1350 			dma_rw_ctl |= BGE_PCIDMA_RWCTL_PCIE_WRITE_WATRMARK_256;
   1351 		} else {
   1352 			dma_rw_ctl |= BGE_PCIDMA_RWCTL_PCIE_WRITE_WATRMARK_128;
   1353 		}
   1354 	} else if (pci_conf_read(pa->pa_pc, pa->pa_tag,BGE_PCI_PCISTATE) &
   1355 	    BGE_PCISTATE_PCI_BUSMODE) {
   1356 		/* Conventional PCI bus */
   1357 	  	DPRINTFN(4, ("(%s: PCI 2.2 DMA setting)\n", sc->bge_dev.dv_xname));
   1358 		dma_rw_ctl = (BGE_PCI_READ_CMD | BGE_PCI_WRITE_CMD |
   1359 		   (0x7 << BGE_PCIDMARWCTL_RD_WAT_SHIFT) |
   1360 		   (0x7 << BGE_PCIDMARWCTL_WR_WAT_SHIFT));
   1361 		if ((sc->bge_quirks & BGE_QUIRK_5705_CORE) == 0) {
   1362 			dma_rw_ctl |= 0x0F;
   1363 		}
   1364 	} else {
   1365 	  	DPRINTFN(4, ("(:%s: PCI-X DMA setting)\n", sc->bge_dev.dv_xname));
   1366 		/* PCI-X bus */
   1367 		dma_rw_ctl = BGE_PCI_READ_CMD|BGE_PCI_WRITE_CMD |
   1368 		    (0x3 << BGE_PCIDMARWCTL_RD_WAT_SHIFT) |
   1369 		    (0x3 << BGE_PCIDMARWCTL_WR_WAT_SHIFT) |
   1370 		    (0x0F);
   1371 		/*
   1372 		 * 5703 and 5704 need ONEDMA_AT_ONCE as a workaround
   1373 		 * for hardware bugs, which means we should also clear
   1374 		 * the low-order MINDMA bits.  In addition, the 5704
   1375 		 * uses a different encoding of read/write watermarks.
   1376 		 */
   1377 		if (BGE_ASICREV(sc->bge_chipid) == BGE_ASICREV_BCM5704) {
   1378 			dma_rw_ctl = BGE_PCI_READ_CMD|BGE_PCI_WRITE_CMD |
   1379 			  /* should be 0x1f0000 */
   1380 			  (0x7 << BGE_PCIDMARWCTL_RD_WAT_SHIFT) |
   1381 			  (0x3 << BGE_PCIDMARWCTL_WR_WAT_SHIFT);
   1382 			dma_rw_ctl |= BGE_PCIDMARWCTL_ONEDMA_ATONCE;
   1383 		}
   1384 		else if (BGE_ASICREV(sc->bge_chipid) == BGE_ASICREV_BCM5703) {
   1385 			dma_rw_ctl &=  0xfffffff0;
   1386 			dma_rw_ctl |= BGE_PCIDMARWCTL_ONEDMA_ATONCE;
   1387 		}
   1388 	}
   1389 
   1390 	pci_conf_write(pa->pa_pc, pa->pa_tag, BGE_PCI_DMA_RW_CTL, dma_rw_ctl);
   1391 
   1392 	/*
   1393 	 * Set up general mode register.
   1394 	 */
   1395 	CSR_WRITE_4(sc, BGE_MODE_CTL, BGE_DMA_SWAP_OPTIONS|
   1396 		    BGE_MODECTL_MAC_ATTN_INTR|BGE_MODECTL_HOST_SEND_BDS|
   1397 		    BGE_MODECTL_TX_NO_PHDR_CSUM|BGE_MODECTL_RX_NO_PHDR_CSUM);
   1398 
   1399 	/* Get cache line size. */
   1400 	cachesize = pci_conf_read(pa->pa_pc, pa->pa_tag, BGE_PCI_CACHESZ);
   1401 
   1402 	/*
   1403 	 * Avoid violating PCI spec on certain chip revs.
   1404 	 */
   1405 	if (pci_conf_read(pa->pa_pc, pa->pa_tag, BGE_PCI_CMD) &
   1406 	    PCIM_CMD_MWIEN) {
   1407 		switch(cachesize) {
   1408 		case 1:
   1409 			PCI_SETBIT(pa->pa_pc, pa->pa_tag, BGE_PCI_DMA_RW_CTL,
   1410 				   BGE_PCI_WRITE_BNDRY_16BYTES);
   1411 			break;
   1412 		case 2:
   1413 			PCI_SETBIT(pa->pa_pc, pa->pa_tag, BGE_PCI_DMA_RW_CTL,
   1414 				   BGE_PCI_WRITE_BNDRY_32BYTES);
   1415 			break;
   1416 		case 4:
   1417 			PCI_SETBIT(pa->pa_pc, pa->pa_tag, BGE_PCI_DMA_RW_CTL,
   1418 				   BGE_PCI_WRITE_BNDRY_64BYTES);
   1419 			break;
   1420 		case 8:
   1421 			PCI_SETBIT(pa->pa_pc, pa->pa_tag, BGE_PCI_DMA_RW_CTL,
   1422 				   BGE_PCI_WRITE_BNDRY_128BYTES);
   1423 			break;
   1424 		case 16:
   1425 			PCI_SETBIT(pa->pa_pc, pa->pa_tag, BGE_PCI_DMA_RW_CTL,
   1426 				   BGE_PCI_WRITE_BNDRY_256BYTES);
   1427 			break;
   1428 		case 32:
   1429 			PCI_SETBIT(pa->pa_pc, pa->pa_tag, BGE_PCI_DMA_RW_CTL,
   1430 				   BGE_PCI_WRITE_BNDRY_512BYTES);
   1431 			break;
   1432 		case 64:
   1433 			PCI_SETBIT(pa->pa_pc, pa->pa_tag, BGE_PCI_DMA_RW_CTL,
   1434 				   BGE_PCI_WRITE_BNDRY_1024BYTES);
   1435 			break;
   1436 		default:
   1437 		/* Disable PCI memory write and invalidate. */
   1438 #if 0
   1439 			if (bootverbose)
   1440 				printf("%s: cache line size %d not "
   1441 				    "supported; disabling PCI MWI\n",
   1442 				    sc->bge_dev.dv_xname, cachesize);
   1443 #endif
   1444 			PCI_CLRBIT(pa->pa_pc, pa->pa_tag, BGE_PCI_CMD,
   1445 			    PCIM_CMD_MWIEN);
   1446 			break;
   1447 		}
   1448 	}
   1449 
   1450 	/*
   1451 	 * Disable memory write invalidate.  Apparently it is not supported
   1452 	 * properly by these devices.
   1453 	 */
   1454 	PCI_CLRBIT(pa->pa_pc, pa->pa_tag, BGE_PCI_CMD, PCIM_CMD_MWIEN);
   1455 
   1456 
   1457 #ifdef __brokenalpha__
   1458 	/*
   1459 	 * Must insure that we do not cross an 8K (bytes) boundary
   1460 	 * for DMA reads.  Our highest limit is 1K bytes.  This is a
   1461 	 * restriction on some ALPHA platforms with early revision
   1462 	 * 21174 PCI chipsets, such as the AlphaPC 164lx
   1463 	 */
   1464 	PCI_SETBIT(sc, BGE_PCI_DMA_RW_CTL, BGE_PCI_READ_BNDRY_1024, 4);
   1465 #endif
   1466 
   1467 	/* Set the timer prescaler (always 66MHz) */
   1468 	CSR_WRITE_4(sc, BGE_MISC_CFG, 65 << 1/*BGE_32BITTIME_66MHZ*/);
   1469 
   1470 	return(0);
   1471 }
   1472 
   1473 int
   1474 bge_blockinit(sc)
   1475 	struct bge_softc *sc;
   1476 {
   1477 	volatile struct bge_rcb		*rcb;
   1478 	bus_size_t		rcb_addr;
   1479 	int			i;
   1480 	struct ifnet		*ifp = &sc->ethercom.ec_if;
   1481 	bge_hostaddr		taddr;
   1482 
   1483 	/*
   1484 	 * Initialize the memory window pointer register so that
   1485 	 * we can access the first 32K of internal NIC RAM. This will
   1486 	 * allow us to set up the TX send ring RCBs and the RX return
   1487 	 * ring RCBs, plus other things which live in NIC memory.
   1488 	 */
   1489 
   1490 	pci_conf_write(sc->bge_pa.pa_pc, sc->bge_pa.pa_tag,
   1491 	    BGE_PCI_MEMWIN_BASEADDR, 0);
   1492 
   1493 	/* Configure mbuf memory pool */
   1494 	if ((sc->bge_quirks & BGE_QUIRK_5705_CORE) == 0) {
   1495 		if (sc->bge_extram) {
   1496 			CSR_WRITE_4(sc, BGE_BMAN_MBUFPOOL_BASEADDR,
   1497 			    BGE_EXT_SSRAM);
   1498 			if ((sc->bge_quirks & BGE_QUIRK_FEWER_MBUFS) != 0)
   1499 				CSR_WRITE_4(sc, BGE_BMAN_MBUFPOOL_LEN, 0x10000);
   1500 			else
   1501 				CSR_WRITE_4(sc, BGE_BMAN_MBUFPOOL_LEN, 0x18000);
   1502 		} else {
   1503 			CSR_WRITE_4(sc, BGE_BMAN_MBUFPOOL_BASEADDR,
   1504 			    BGE_BUFFPOOL_1);
   1505 			if ((sc->bge_quirks & BGE_QUIRK_FEWER_MBUFS) != 0)
   1506 				CSR_WRITE_4(sc, BGE_BMAN_MBUFPOOL_LEN, 0x10000);
   1507 			else
   1508 				CSR_WRITE_4(sc, BGE_BMAN_MBUFPOOL_LEN, 0x18000);
   1509 		}
   1510 
   1511 		/* Configure DMA resource pool */
   1512 		CSR_WRITE_4(sc, BGE_BMAN_DMA_DESCPOOL_BASEADDR,
   1513 		    BGE_DMA_DESCRIPTORS);
   1514 		CSR_WRITE_4(sc, BGE_BMAN_DMA_DESCPOOL_LEN, 0x2000);
   1515 	}
   1516 
   1517 	/* Configure mbuf pool watermarks */
   1518 #ifdef ORIG_WPAUL_VALUES
   1519 	CSR_WRITE_4(sc, BGE_BMAN_MBUFPOOL_READDMA_LOWAT, 24);
   1520 	CSR_WRITE_4(sc, BGE_BMAN_MBUFPOOL_MACRX_LOWAT, 24);
   1521 	CSR_WRITE_4(sc, BGE_BMAN_MBUFPOOL_HIWAT, 48);
   1522 #else
   1523 	/* new broadcom docs strongly recommend these: */
   1524 	if ((sc->bge_quirks & BGE_QUIRK_5705_CORE) == 0) {
   1525 		if (ifp->if_mtu > ETHER_MAX_LEN) {
   1526 			CSR_WRITE_4(sc, BGE_BMAN_MBUFPOOL_READDMA_LOWAT, 0x50);
   1527 			CSR_WRITE_4(sc, BGE_BMAN_MBUFPOOL_MACRX_LOWAT, 0x20);
   1528 			CSR_WRITE_4(sc, BGE_BMAN_MBUFPOOL_HIWAT, 0x60);
   1529 		} else {
   1530 			/* Values from Linux driver... */
   1531 			CSR_WRITE_4(sc, BGE_BMAN_MBUFPOOL_READDMA_LOWAT, 304);
   1532 			CSR_WRITE_4(sc, BGE_BMAN_MBUFPOOL_MACRX_LOWAT, 152);
   1533 			CSR_WRITE_4(sc, BGE_BMAN_MBUFPOOL_HIWAT, 380);
   1534 		}
   1535 	} else {
   1536 		CSR_WRITE_4(sc, BGE_BMAN_MBUFPOOL_READDMA_LOWAT, 0x0);
   1537 		CSR_WRITE_4(sc, BGE_BMAN_MBUFPOOL_MACRX_LOWAT, 0x10);
   1538 		CSR_WRITE_4(sc, BGE_BMAN_MBUFPOOL_HIWAT, 0x60);
   1539 	}
   1540 #endif
   1541 
   1542 	/* Configure DMA resource watermarks */
   1543 	CSR_WRITE_4(sc, BGE_BMAN_DMA_DESCPOOL_LOWAT, 5);
   1544 	CSR_WRITE_4(sc, BGE_BMAN_DMA_DESCPOOL_HIWAT, 10);
   1545 
   1546 	/* Enable buffer manager */
   1547 	if ((sc->bge_quirks & BGE_QUIRK_5705_CORE) == 0) {
   1548 		CSR_WRITE_4(sc, BGE_BMAN_MODE,
   1549 		    BGE_BMANMODE_ENABLE|BGE_BMANMODE_LOMBUF_ATTN);
   1550 
   1551 		/* Poll for buffer manager start indication */
   1552 		for (i = 0; i < BGE_TIMEOUT; i++) {
   1553 			if (CSR_READ_4(sc, BGE_BMAN_MODE) & BGE_BMANMODE_ENABLE)
   1554 				break;
   1555 			DELAY(10);
   1556 		}
   1557 
   1558 		if (i == BGE_TIMEOUT) {
   1559 			printf("%s: buffer manager failed to start\n",
   1560 			    sc->bge_dev.dv_xname);
   1561 			return(ENXIO);
   1562 		}
   1563 	}
   1564 
   1565 	/* Enable flow-through queues */
   1566 	CSR_WRITE_4(sc, BGE_FTQ_RESET, 0xFFFFFFFF);
   1567 	CSR_WRITE_4(sc, BGE_FTQ_RESET, 0);
   1568 
   1569 	/* Wait until queue initialization is complete */
   1570 	for (i = 0; i < BGE_TIMEOUT; i++) {
   1571 		if (CSR_READ_4(sc, BGE_FTQ_RESET) == 0)
   1572 			break;
   1573 		DELAY(10);
   1574 	}
   1575 
   1576 	if (i == BGE_TIMEOUT) {
   1577 		printf("%s: flow-through queue init failed\n",
   1578 		    sc->bge_dev.dv_xname);
   1579 		return(ENXIO);
   1580 	}
   1581 
   1582 	/* Initialize the standard RX ring control block */
   1583 	rcb = &sc->bge_rdata->bge_info.bge_std_rx_rcb;
   1584 	bge_set_hostaddr(&rcb->bge_hostaddr,
   1585 	    BGE_RING_DMA_ADDR(sc, bge_rx_std_ring));
   1586 	if ((sc->bge_quirks & BGE_QUIRK_5705_CORE) == 0) {
   1587 		rcb->bge_maxlen_flags =
   1588 		    BGE_RCB_MAXLEN_FLAGS(BGE_MAX_FRAMELEN, 0);
   1589 	} else {
   1590 		rcb->bge_maxlen_flags = BGE_RCB_MAXLEN_FLAGS(512, 0);
   1591 	}
   1592 	if (sc->bge_extram)
   1593 		rcb->bge_nicaddr = BGE_EXT_STD_RX_RINGS;
   1594 	else
   1595 		rcb->bge_nicaddr = BGE_STD_RX_RINGS;
   1596 	CSR_WRITE_4(sc, BGE_RX_STD_RCB_HADDR_HI, rcb->bge_hostaddr.bge_addr_hi);
   1597 	CSR_WRITE_4(sc, BGE_RX_STD_RCB_HADDR_LO, rcb->bge_hostaddr.bge_addr_lo);
   1598 	CSR_WRITE_4(sc, BGE_RX_STD_RCB_MAXLEN_FLAGS, rcb->bge_maxlen_flags);
   1599 	CSR_WRITE_4(sc, BGE_RX_STD_RCB_NICADDR, rcb->bge_nicaddr);
   1600 
   1601 	if ((sc->bge_quirks & BGE_QUIRK_5705_CORE) == 0) {
   1602 		sc->bge_return_ring_cnt = BGE_RETURN_RING_CNT;
   1603 	} else {
   1604 		sc->bge_return_ring_cnt = BGE_RETURN_RING_CNT_5705;
   1605 	}
   1606 
   1607 	/*
   1608 	 * Initialize the jumbo RX ring control block
   1609 	 * We set the 'ring disabled' bit in the flags
   1610 	 * field until we're actually ready to start
   1611 	 * using this ring (i.e. once we set the MTU
   1612 	 * high enough to require it).
   1613 	 */
   1614 	if ((sc->bge_quirks & BGE_QUIRK_5705_CORE) == 0) {
   1615 		rcb = &sc->bge_rdata->bge_info.bge_jumbo_rx_rcb;
   1616 		bge_set_hostaddr(&rcb->bge_hostaddr,
   1617 		    BGE_RING_DMA_ADDR(sc, bge_rx_jumbo_ring));
   1618 		rcb->bge_maxlen_flags =
   1619 		    BGE_RCB_MAXLEN_FLAGS(BGE_MAX_FRAMELEN,
   1620 			BGE_RCB_FLAG_RING_DISABLED);
   1621 		if (sc->bge_extram)
   1622 			rcb->bge_nicaddr = BGE_EXT_JUMBO_RX_RINGS;
   1623 		else
   1624 			rcb->bge_nicaddr = BGE_JUMBO_RX_RINGS;
   1625 
   1626 		CSR_WRITE_4(sc, BGE_RX_JUMBO_RCB_HADDR_HI,
   1627 		    rcb->bge_hostaddr.bge_addr_hi);
   1628 		CSR_WRITE_4(sc, BGE_RX_JUMBO_RCB_HADDR_LO,
   1629 		    rcb->bge_hostaddr.bge_addr_lo);
   1630 		CSR_WRITE_4(sc, BGE_RX_JUMBO_RCB_MAXLEN_FLAGS,
   1631 		    rcb->bge_maxlen_flags);
   1632 		CSR_WRITE_4(sc, BGE_RX_JUMBO_RCB_NICADDR, rcb->bge_nicaddr);
   1633 
   1634 		/* Set up dummy disabled mini ring RCB */
   1635 		rcb = &sc->bge_rdata->bge_info.bge_mini_rx_rcb;
   1636 		rcb->bge_maxlen_flags = BGE_RCB_MAXLEN_FLAGS(0,
   1637 		    BGE_RCB_FLAG_RING_DISABLED);
   1638 		CSR_WRITE_4(sc, BGE_RX_MINI_RCB_MAXLEN_FLAGS,
   1639 		    rcb->bge_maxlen_flags);
   1640 
   1641 		bus_dmamap_sync(sc->bge_dmatag, sc->bge_ring_map,
   1642 		    offsetof(struct bge_ring_data, bge_info),
   1643 		    sizeof (struct bge_gib),
   1644 		    BUS_DMASYNC_PREREAD|BUS_DMASYNC_PREWRITE);
   1645 	}
   1646 
   1647 	/*
   1648 	 * Set the BD ring replentish thresholds. The recommended
   1649 	 * values are 1/8th the number of descriptors allocated to
   1650 	 * each ring.
   1651 	 */
   1652 	CSR_WRITE_4(sc, BGE_RBDI_STD_REPL_THRESH, BGE_STD_RX_RING_CNT/8);
   1653 	CSR_WRITE_4(sc, BGE_RBDI_JUMBO_REPL_THRESH, BGE_JUMBO_RX_RING_CNT/8);
   1654 
   1655 	/*
   1656 	 * Disable all unused send rings by setting the 'ring disabled'
   1657 	 * bit in the flags field of all the TX send ring control blocks.
   1658 	 * These are located in NIC memory.
   1659 	 */
   1660 	rcb_addr = BGE_MEMWIN_START + BGE_SEND_RING_RCB;
   1661 	for (i = 0; i < BGE_TX_RINGS_EXTSSRAM_MAX; i++) {
   1662 		RCB_WRITE_4(sc, rcb_addr, bge_maxlen_flags,
   1663 		    BGE_RCB_MAXLEN_FLAGS(0,BGE_RCB_FLAG_RING_DISABLED));
   1664 		RCB_WRITE_4(sc, rcb_addr, bge_nicaddr, 0);
   1665 		rcb_addr += sizeof(struct bge_rcb);
   1666 	}
   1667 
   1668 	/* Configure TX RCB 0 (we use only the first ring) */
   1669 	rcb_addr = BGE_MEMWIN_START + BGE_SEND_RING_RCB;
   1670 	bge_set_hostaddr(&taddr, BGE_RING_DMA_ADDR(sc, bge_tx_ring));
   1671 	RCB_WRITE_4(sc, rcb_addr, bge_hostaddr.bge_addr_hi, taddr.bge_addr_hi);
   1672 	RCB_WRITE_4(sc, rcb_addr, bge_hostaddr.bge_addr_lo, taddr.bge_addr_lo);
   1673 	RCB_WRITE_4(sc, rcb_addr, bge_nicaddr,
   1674 		    BGE_NIC_TXRING_ADDR(0, BGE_TX_RING_CNT));
   1675 	if ((sc->bge_quirks & BGE_QUIRK_5705_CORE) == 0) {
   1676 		RCB_WRITE_4(sc, rcb_addr, bge_maxlen_flags,
   1677 		    BGE_RCB_MAXLEN_FLAGS(BGE_TX_RING_CNT, 0));
   1678 	}
   1679 
   1680 	/* Disable all unused RX return rings */
   1681 	rcb_addr = BGE_MEMWIN_START + BGE_RX_RETURN_RING_RCB;
   1682 	for (i = 0; i < BGE_RX_RINGS_MAX; i++) {
   1683 		RCB_WRITE_4(sc, rcb_addr, bge_hostaddr.bge_addr_hi, 0);
   1684 		RCB_WRITE_4(sc, rcb_addr, bge_hostaddr.bge_addr_lo, 0);
   1685 		RCB_WRITE_4(sc, rcb_addr, bge_maxlen_flags,
   1686 			    BGE_RCB_MAXLEN_FLAGS(sc->bge_return_ring_cnt,
   1687                                      BGE_RCB_FLAG_RING_DISABLED));
   1688 		RCB_WRITE_4(sc, rcb_addr, bge_nicaddr, 0);
   1689 		CSR_WRITE_4(sc, BGE_MBX_RX_CONS0_LO +
   1690 		    (i * (sizeof(u_int64_t))), 0);
   1691 		rcb_addr += sizeof(struct bge_rcb);
   1692 	}
   1693 
   1694 	/* Initialize RX ring indexes */
   1695 	CSR_WRITE_4(sc, BGE_MBX_RX_STD_PROD_LO, 0);
   1696 	CSR_WRITE_4(sc, BGE_MBX_RX_JUMBO_PROD_LO, 0);
   1697 	CSR_WRITE_4(sc, BGE_MBX_RX_MINI_PROD_LO, 0);
   1698 
   1699 	/*
   1700 	 * Set up RX return ring 0
   1701 	 * Note that the NIC address for RX return rings is 0x00000000.
   1702 	 * The return rings live entirely within the host, so the
   1703 	 * nicaddr field in the RCB isn't used.
   1704 	 */
   1705 	rcb_addr = BGE_MEMWIN_START + BGE_RX_RETURN_RING_RCB;
   1706 	bge_set_hostaddr(&taddr, BGE_RING_DMA_ADDR(sc, bge_rx_return_ring));
   1707 	RCB_WRITE_4(sc, rcb_addr, bge_hostaddr.bge_addr_hi, taddr.bge_addr_hi);
   1708 	RCB_WRITE_4(sc, rcb_addr, bge_hostaddr.bge_addr_lo, taddr.bge_addr_lo);
   1709 	RCB_WRITE_4(sc, rcb_addr, bge_nicaddr, 0x00000000);
   1710 	RCB_WRITE_4(sc, rcb_addr, bge_maxlen_flags,
   1711 	    BGE_RCB_MAXLEN_FLAGS(sc->bge_return_ring_cnt, 0));
   1712 
   1713 	/* Set random backoff seed for TX */
   1714 	CSR_WRITE_4(sc, BGE_TX_RANDOM_BACKOFF,
   1715 	    LLADDR(ifp->if_sadl)[0] + LLADDR(ifp->if_sadl)[1] +
   1716 	    LLADDR(ifp->if_sadl)[2] + LLADDR(ifp->if_sadl)[3] +
   1717 	    LLADDR(ifp->if_sadl)[4] + LLADDR(ifp->if_sadl)[5] +
   1718 	    BGE_TX_BACKOFF_SEED_MASK);
   1719 
   1720 	/* Set inter-packet gap */
   1721 	CSR_WRITE_4(sc, BGE_TX_LENGTHS, 0x2620);
   1722 
   1723 	/*
   1724 	 * Specify which ring to use for packets that don't match
   1725 	 * any RX rules.
   1726 	 */
   1727 	CSR_WRITE_4(sc, BGE_RX_RULES_CFG, 0x08);
   1728 
   1729 	/*
   1730 	 * Configure number of RX lists. One interrupt distribution
   1731 	 * list, sixteen active lists, one bad frames class.
   1732 	 */
   1733 	CSR_WRITE_4(sc, BGE_RXLP_CFG, 0x181);
   1734 
   1735 	/* Inialize RX list placement stats mask. */
   1736 	CSR_WRITE_4(sc, BGE_RXLP_STATS_ENABLE_MASK, 0x007FFFFF);
   1737 	CSR_WRITE_4(sc, BGE_RXLP_STATS_CTL, 0x1);
   1738 
   1739 	/* Disable host coalescing until we get it set up */
   1740 	CSR_WRITE_4(sc, BGE_HCC_MODE, 0x00000000);
   1741 
   1742 	/* Poll to make sure it's shut down. */
   1743 	for (i = 0; i < BGE_TIMEOUT; i++) {
   1744 		if (!(CSR_READ_4(sc, BGE_HCC_MODE) & BGE_HCCMODE_ENABLE))
   1745 			break;
   1746 		DELAY(10);
   1747 	}
   1748 
   1749 	if (i == BGE_TIMEOUT) {
   1750 		printf("%s: host coalescing engine failed to idle\n",
   1751 		    sc->bge_dev.dv_xname);
   1752 		return(ENXIO);
   1753 	}
   1754 
   1755 	/* Set up host coalescing defaults */
   1756 	CSR_WRITE_4(sc, BGE_HCC_RX_COAL_TICKS, sc->bge_rx_coal_ticks);
   1757 	CSR_WRITE_4(sc, BGE_HCC_TX_COAL_TICKS, sc->bge_tx_coal_ticks);
   1758 	CSR_WRITE_4(sc, BGE_HCC_RX_MAX_COAL_BDS, sc->bge_rx_max_coal_bds);
   1759 	CSR_WRITE_4(sc, BGE_HCC_TX_MAX_COAL_BDS, sc->bge_tx_max_coal_bds);
   1760 	if ((sc->bge_quirks & BGE_QUIRK_5705_CORE) == 0) {
   1761 		CSR_WRITE_4(sc, BGE_HCC_RX_COAL_TICKS_INT, 0);
   1762 		CSR_WRITE_4(sc, BGE_HCC_TX_COAL_TICKS_INT, 0);
   1763 	}
   1764 	CSR_WRITE_4(sc, BGE_HCC_RX_MAX_COAL_BDS_INT, 0);
   1765 	CSR_WRITE_4(sc, BGE_HCC_TX_MAX_COAL_BDS_INT, 0);
   1766 
   1767 	/* Set up address of statistics block */
   1768 	if ((sc->bge_quirks & BGE_QUIRK_5705_CORE) == 0) {
   1769 		bge_set_hostaddr(&taddr,
   1770 		    BGE_RING_DMA_ADDR(sc, bge_info.bge_stats));
   1771 		CSR_WRITE_4(sc, BGE_HCC_STATS_TICKS, sc->bge_stat_ticks);
   1772 		CSR_WRITE_4(sc, BGE_HCC_STATS_BASEADDR, BGE_STATS_BLOCK);
   1773 		CSR_WRITE_4(sc, BGE_HCC_STATS_ADDR_HI, taddr.bge_addr_hi);
   1774 		CSR_WRITE_4(sc, BGE_HCC_STATS_ADDR_LO, taddr.bge_addr_lo);
   1775 	}
   1776 
   1777 	/* Set up address of status block */
   1778 	bge_set_hostaddr(&taddr, BGE_RING_DMA_ADDR(sc, bge_status_block));
   1779 	CSR_WRITE_4(sc, BGE_HCC_STATUSBLK_BASEADDR, BGE_STATUS_BLOCK);
   1780 	CSR_WRITE_4(sc, BGE_HCC_STATUSBLK_ADDR_HI, taddr.bge_addr_hi);
   1781 	CSR_WRITE_4(sc, BGE_HCC_STATUSBLK_ADDR_LO, taddr.bge_addr_lo);
   1782 	sc->bge_rdata->bge_status_block.bge_idx[0].bge_rx_prod_idx = 0;
   1783 	sc->bge_rdata->bge_status_block.bge_idx[0].bge_tx_cons_idx = 0;
   1784 
   1785 	/* Turn on host coalescing state machine */
   1786 	CSR_WRITE_4(sc, BGE_HCC_MODE, BGE_HCCMODE_ENABLE);
   1787 
   1788 	/* Turn on RX BD completion state machine and enable attentions */
   1789 	CSR_WRITE_4(sc, BGE_RBDC_MODE,
   1790 	    BGE_RBDCMODE_ENABLE|BGE_RBDCMODE_ATTN);
   1791 
   1792 	/* Turn on RX list placement state machine */
   1793 	CSR_WRITE_4(sc, BGE_RXLP_MODE, BGE_RXLPMODE_ENABLE);
   1794 
   1795 	/* Turn on RX list selector state machine. */
   1796 	if ((sc->bge_quirks & BGE_QUIRK_5705_CORE) == 0) {
   1797 		CSR_WRITE_4(sc, BGE_RXLS_MODE, BGE_RXLSMODE_ENABLE);
   1798 	}
   1799 
   1800 	/* Turn on DMA, clear stats */
   1801 	CSR_WRITE_4(sc, BGE_MAC_MODE, BGE_MACMODE_TXDMA_ENB|
   1802 	    BGE_MACMODE_RXDMA_ENB|BGE_MACMODE_RX_STATS_CLEAR|
   1803 	    BGE_MACMODE_TX_STATS_CLEAR|BGE_MACMODE_RX_STATS_ENB|
   1804 	    BGE_MACMODE_TX_STATS_ENB|BGE_MACMODE_FRMHDR_DMA_ENB|
   1805 	    (sc->bge_tbi ? BGE_PORTMODE_TBI : BGE_PORTMODE_MII));
   1806 
   1807 	/* Set misc. local control, enable interrupts on attentions */
   1808 	sc->bge_local_ctrl_reg = BGE_MLC_INTR_ONATTN | BGE_MLC_AUTO_EEPROM;
   1809 
   1810 #ifdef notdef
   1811 	/* Assert GPIO pins for PHY reset */
   1812 	BGE_SETBIT(sc, BGE_MISC_LOCAL_CTL, BGE_MLC_MISCIO_OUT0|
   1813 	    BGE_MLC_MISCIO_OUT1|BGE_MLC_MISCIO_OUT2);
   1814 	BGE_SETBIT(sc, BGE_MISC_LOCAL_CTL, BGE_MLC_MISCIO_OUTEN0|
   1815 	    BGE_MLC_MISCIO_OUTEN1|BGE_MLC_MISCIO_OUTEN2);
   1816 #endif
   1817 
   1818 #if defined(not_quite_yet)
   1819 	/* Linux driver enables enable gpio pin #1 on 5700s */
   1820 	if (sc->bge_chipid == BGE_CHIPID_BCM5700) {
   1821 		sc->bge_local_ctrl_reg |=
   1822 		  (BGE_MLC_MISCIO_OUT1|BGE_MLC_MISCIO_OUTEN1);
   1823 	}
   1824 #endif
   1825 	CSR_WRITE_4(sc, BGE_MISC_LOCAL_CTL, sc->bge_local_ctrl_reg);
   1826 
   1827 	/* Turn on DMA completion state machine */
   1828 	if ((sc->bge_quirks & BGE_QUIRK_5705_CORE) == 0) {
   1829 		CSR_WRITE_4(sc, BGE_DMAC_MODE, BGE_DMACMODE_ENABLE);
   1830 	}
   1831 
   1832 	/* Turn on write DMA state machine */
   1833 	CSR_WRITE_4(sc, BGE_WDMA_MODE,
   1834 	    BGE_WDMAMODE_ENABLE|BGE_WDMAMODE_ALL_ATTNS);
   1835 
   1836 	/* Turn on read DMA state machine */
   1837 	{
   1838 		uint32_t dma_read_modebits;
   1839 
   1840 		dma_read_modebits =
   1841 		  BGE_RDMAMODE_ENABLE | BGE_RDMAMODE_ALL_ATTNS;
   1842 
   1843 		if (sc->bge_pcie && 0) {
   1844 			dma_read_modebits |= BGE_RDMA_MODE_FIFO_LONG_BURST;
   1845 		} else if ((sc->bge_quirks & BGE_QUIRK_5705_CORE)) {
   1846 			dma_read_modebits |= BGE_RDMA_MODE_FIFO_SIZE_128;
   1847 		}
   1848 
   1849 		/* XXX broadcom-supplied linux driver; undocumented */
   1850 		if (BGE_IS_5750_OR_BEYOND(sc)) {
   1851  			/*
   1852 			 * XXX: magic values.
   1853 			 * From Broadcom-supplied Linux driver;  apparently
   1854 			 * required to workaround a DMA bug affecting TSO
   1855 			 * on bcm575x/bcm5721?
   1856 			 */
   1857 			dma_read_modebits |= (1 << 27);
   1858 		}
   1859 		CSR_WRITE_4(sc, BGE_RDMA_MODE, dma_read_modebits);
   1860 	}
   1861 
   1862 	/* Turn on RX data completion state machine */
   1863 	CSR_WRITE_4(sc, BGE_RDC_MODE, BGE_RDCMODE_ENABLE);
   1864 
   1865 	/* Turn on RX BD initiator state machine */
   1866 	CSR_WRITE_4(sc, BGE_RBDI_MODE, BGE_RBDIMODE_ENABLE);
   1867 
   1868 	/* Turn on RX data and RX BD initiator state machine */
   1869 	CSR_WRITE_4(sc, BGE_RDBDI_MODE, BGE_RDBDIMODE_ENABLE);
   1870 
   1871 	/* Turn on Mbuf cluster free state machine */
   1872 	if ((sc->bge_quirks & BGE_QUIRK_5705_CORE) == 0) {
   1873 		CSR_WRITE_4(sc, BGE_MBCF_MODE, BGE_MBCFMODE_ENABLE);
   1874 	}
   1875 
   1876 	/* Turn on send BD completion state machine */
   1877 	CSR_WRITE_4(sc, BGE_SBDC_MODE, BGE_SBDCMODE_ENABLE);
   1878 
   1879 	/* Turn on send data completion state machine */
   1880 	CSR_WRITE_4(sc, BGE_SDC_MODE, BGE_SDCMODE_ENABLE);
   1881 
   1882 	/* Turn on send data initiator state machine */
   1883 	if (BGE_IS_5750_OR_BEYOND(sc)) {
   1884 		/* XXX: magic value from Linux driver */
   1885 		CSR_WRITE_4(sc, BGE_SDI_MODE, BGE_SDIMODE_ENABLE | 0x08);
   1886 	} else {
   1887 		CSR_WRITE_4(sc, BGE_SDI_MODE, BGE_SDIMODE_ENABLE);
   1888 	}
   1889 
   1890 	/* Turn on send BD initiator state machine */
   1891 	CSR_WRITE_4(sc, BGE_SBDI_MODE, BGE_SBDIMODE_ENABLE);
   1892 
   1893 	/* Turn on send BD selector state machine */
   1894 	CSR_WRITE_4(sc, BGE_SRS_MODE, BGE_SRSMODE_ENABLE);
   1895 
   1896 	CSR_WRITE_4(sc, BGE_SDI_STATS_ENABLE_MASK, 0x007FFFFF);
   1897 	CSR_WRITE_4(sc, BGE_SDI_STATS_CTL,
   1898 	    BGE_SDISTATSCTL_ENABLE|BGE_SDISTATSCTL_FASTER);
   1899 
   1900 	/* ack/clear link change events */
   1901 	CSR_WRITE_4(sc, BGE_MAC_STS, BGE_MACSTAT_SYNC_CHANGED|
   1902 	    BGE_MACSTAT_CFG_CHANGED);
   1903 	CSR_WRITE_4(sc, BGE_MI_STS, 0);
   1904 
   1905 	/* Enable PHY auto polling (for MII/GMII only) */
   1906 	if (sc->bge_tbi) {
   1907 		CSR_WRITE_4(sc, BGE_MI_STS, BGE_MISTS_LINK);
   1908  	} else {
   1909 		BGE_SETBIT(sc, BGE_MI_MODE, BGE_MIMODE_AUTOPOLL|10<<16);
   1910 		if (sc->bge_quirks & BGE_QUIRK_LINK_STATE_BROKEN)
   1911 			CSR_WRITE_4(sc, BGE_MAC_EVT_ENB,
   1912 			    BGE_EVTENB_MI_INTERRUPT);
   1913 	}
   1914 
   1915 	/* Enable link state change attentions. */
   1916 	BGE_SETBIT(sc, BGE_MAC_EVT_ENB, BGE_EVTENB_LINK_CHANGED);
   1917 
   1918 	return(0);
   1919 }
   1920 
   1921 static const struct bge_revision {
   1922 	uint32_t		br_chipid;
   1923 	uint32_t		br_quirks;
   1924 	const char		*br_name;
   1925 } bge_revisions[] = {
   1926 	{ BGE_CHIPID_BCM5700_A0,
   1927 	  BGE_QUIRK_LINK_STATE_BROKEN,
   1928 	  "BCM5700 A0" },
   1929 
   1930 	{ BGE_CHIPID_BCM5700_A1,
   1931 	  BGE_QUIRK_LINK_STATE_BROKEN,
   1932 	  "BCM5700 A1" },
   1933 
   1934 	{ BGE_CHIPID_BCM5700_B0,
   1935 	  BGE_QUIRK_LINK_STATE_BROKEN|BGE_QUIRK_CSUM_BROKEN|BGE_QUIRK_5700_COMMON,
   1936 	  "BCM5700 B0" },
   1937 
   1938 	{ BGE_CHIPID_BCM5700_B1,
   1939 	  BGE_QUIRK_LINK_STATE_BROKEN|BGE_QUIRK_5700_COMMON,
   1940 	  "BCM5700 B1" },
   1941 
   1942 	{ BGE_CHIPID_BCM5700_B2,
   1943 	  BGE_QUIRK_LINK_STATE_BROKEN|BGE_QUIRK_5700_COMMON,
   1944 	  "BCM5700 B2" },
   1945 
   1946 	/* This is treated like a BCM5700 Bx */
   1947 	{ BGE_CHIPID_BCM5700_ALTIMA,
   1948 	  BGE_QUIRK_LINK_STATE_BROKEN|BGE_QUIRK_5700_COMMON,
   1949 	  "BCM5700 Altima" },
   1950 
   1951 	{ BGE_CHIPID_BCM5700_C0,
   1952 	  0,
   1953 	  "BCM5700 C0" },
   1954 
   1955 	{ BGE_CHIPID_BCM5701_A0,
   1956 	  0, /*XXX really, just not known */
   1957 	  "BCM5701 A0" },
   1958 
   1959 	{ BGE_CHIPID_BCM5701_B0,
   1960 	  BGE_QUIRK_PCIX_DMA_ALIGN_BUG,
   1961 	  "BCM5701 B0" },
   1962 
   1963 	{ BGE_CHIPID_BCM5701_B2,
   1964 	  BGE_QUIRK_PCIX_DMA_ALIGN_BUG,
   1965 	  "BCM5701 B2" },
   1966 
   1967 	{ BGE_CHIPID_BCM5701_B5,
   1968 	  BGE_QUIRK_ONLY_PHY_1|BGE_QUIRK_PCIX_DMA_ALIGN_BUG,
   1969 	  "BCM5701 B5" },
   1970 
   1971 	{ BGE_CHIPID_BCM5703_A0,
   1972 	  0,
   1973 	  "BCM5703 A0" },
   1974 
   1975 	{ BGE_CHIPID_BCM5703_A1,
   1976 	  0,
   1977 	  "BCM5703 A1" },
   1978 
   1979 	{ BGE_CHIPID_BCM5703_A2,
   1980 	  BGE_QUIRK_ONLY_PHY_1,
   1981 	  "BCM5703 A2" },
   1982 
   1983 	{ BGE_CHIPID_BCM5703_A3,
   1984 	  BGE_QUIRK_ONLY_PHY_1,
   1985 	  "BCM5703 A3" },
   1986 
   1987 	{ BGE_CHIPID_BCM5704_A0,
   1988   	  BGE_QUIRK_ONLY_PHY_1|BGE_QUIRK_FEWER_MBUFS,
   1989 	  "BCM5704 A0" },
   1990 
   1991 	{ BGE_CHIPID_BCM5704_A1,
   1992   	  BGE_QUIRK_ONLY_PHY_1|BGE_QUIRK_FEWER_MBUFS,
   1993 	  "BCM5704 A1" },
   1994 
   1995 	{ BGE_CHIPID_BCM5704_A2,
   1996   	  BGE_QUIRK_ONLY_PHY_1|BGE_QUIRK_FEWER_MBUFS,
   1997 	  "BCM5704 A2" },
   1998 
   1999 	{ BGE_CHIPID_BCM5704_A3,
   2000   	  BGE_QUIRK_ONLY_PHY_1|BGE_QUIRK_FEWER_MBUFS,
   2001 	  "BCM5704 A3" },
   2002 
   2003 	{ BGE_CHIPID_BCM5705_A0,
   2004 	  BGE_QUIRK_ONLY_PHY_1|BGE_QUIRK_5705_CORE,
   2005 	  "BCM5705 A0" },
   2006 
   2007 	{ BGE_CHIPID_BCM5705_A1,
   2008 	  BGE_QUIRK_ONLY_PHY_1|BGE_QUIRK_5705_CORE,
   2009 	  "BCM5705 A1" },
   2010 
   2011 	{ BGE_CHIPID_BCM5705_A2,
   2012 	  BGE_QUIRK_ONLY_PHY_1|BGE_QUIRK_5705_CORE,
   2013 	  "BCM5705 A2" },
   2014 
   2015 	{ BGE_CHIPID_BCM5705_A3,
   2016 	  BGE_QUIRK_ONLY_PHY_1|BGE_QUIRK_5705_CORE,
   2017 	  "BCM5705 A3" },
   2018 
   2019 	{ BGE_CHIPID_BCM5750_A0,
   2020 	  BGE_QUIRK_ONLY_PHY_1|BGE_QUIRK_5705_CORE,
   2021 	  "BCM5750 A1" },
   2022 
   2023 	{ BGE_CHIPID_BCM5750_A1,
   2024 	  BGE_QUIRK_ONLY_PHY_1|BGE_QUIRK_5705_CORE,
   2025 	  "BCM5750 A1" },
   2026 
   2027 	{ BGE_CHIPID_BCM5751_A1,
   2028 	  BGE_QUIRK_ONLY_PHY_1|BGE_QUIRK_5705_CORE,
   2029 	  "BCM5751 A1" },
   2030 
   2031 	{ 0, 0, NULL }
   2032 };
   2033 
   2034 /*
   2035  * Some defaults for major revisions, so that newer steppings
   2036  * that we don't know about have a shot at working.
   2037  */
   2038 static const struct bge_revision bge_majorrevs[] = {
   2039 	{ BGE_ASICREV_BCM5700,
   2040 	  BGE_QUIRK_LINK_STATE_BROKEN,
   2041 	  "unknown BCM5700" },
   2042 
   2043 	{ BGE_ASICREV_BCM5701,
   2044 	  BGE_QUIRK_PCIX_DMA_ALIGN_BUG,
   2045 	  "unknown BCM5701" },
   2046 
   2047 	{ BGE_ASICREV_BCM5703,
   2048 	  0,
   2049 	  "unknown BCM5703" },
   2050 
   2051 	{ BGE_ASICREV_BCM5704,
   2052 	  BGE_QUIRK_ONLY_PHY_1,
   2053 	  "unknown BCM5704" },
   2054 
   2055 	{ BGE_ASICREV_BCM5705,
   2056 	  BGE_QUIRK_ONLY_PHY_1|BGE_QUIRK_5705_CORE,
   2057 	  "unknown BCM5705" },
   2058 
   2059 	{ BGE_ASICREV_BCM5750,
   2060 	  BGE_QUIRK_ONLY_PHY_1|BGE_QUIRK_5705_CORE,
   2061 	  "unknown BCM575x family" },
   2062 
   2063 	{ BGE_ASICREV_BCM5714,
   2064 	  BGE_QUIRK_ONLY_PHY_1|BGE_QUIRK_5705_CORE,
   2065 	  "unknown BCM5714" },
   2066 
   2067 	{ BGE_ASICREV_BCM5752,
   2068 	  BGE_QUIRK_ONLY_PHY_1|BGE_QUIRK_5705_CORE,
   2069 	  "unknown BCM5752 family" },
   2070 
   2071 
   2072 	{ BGE_ASICREV_BCM5715,
   2073 	  BGE_QUIRK_ONLY_PHY_1|BGE_QUIRK_5705_CORE,
   2074 	  "unknown BCM5715" },
   2075 
   2076 	{ 0,
   2077 	  0,
   2078 	  NULL }
   2079 };
   2080 
   2081 
   2082 static const struct bge_revision *
   2083 bge_lookup_rev(uint32_t chipid)
   2084 {
   2085 	const struct bge_revision *br;
   2086 
   2087 	for (br = bge_revisions; br->br_name != NULL; br++) {
   2088 		if (br->br_chipid == chipid)
   2089 			return (br);
   2090 	}
   2091 
   2092 	for (br = bge_majorrevs; br->br_name != NULL; br++) {
   2093 		if (br->br_chipid == BGE_ASICREV(chipid))
   2094 			return (br);
   2095 	}
   2096 
   2097 	return (NULL);
   2098 }
   2099 
   2100 static const struct bge_product {
   2101 	pci_vendor_id_t		bp_vendor;
   2102 	pci_product_id_t	bp_product;
   2103 	const char		*bp_name;
   2104 } bge_products[] = {
   2105 	/*
   2106 	 * The BCM5700 documentation seems to indicate that the hardware
   2107 	 * still has the Alteon vendor ID burned into it, though it
   2108 	 * should always be overridden by the value in the EEPROM.  We'll
   2109 	 * check for it anyway.
   2110 	 */
   2111 	{ PCI_VENDOR_ALTEON,
   2112 	  PCI_PRODUCT_ALTEON_BCM5700,
   2113 	  "Broadcom BCM5700 Gigabit Ethernet",
   2114 	  },
   2115 	{ PCI_VENDOR_ALTEON,
   2116 	  PCI_PRODUCT_ALTEON_BCM5701,
   2117 	  "Broadcom BCM5701 Gigabit Ethernet",
   2118 	  },
   2119 
   2120 	{ PCI_VENDOR_ALTIMA,
   2121 	  PCI_PRODUCT_ALTIMA_AC1000,
   2122 	  "Altima AC1000 Gigabit Ethernet",
   2123 	  },
   2124 	{ PCI_VENDOR_ALTIMA,
   2125 	  PCI_PRODUCT_ALTIMA_AC1001,
   2126 	  "Altima AC1001 Gigabit Ethernet",
   2127 	   },
   2128 	{ PCI_VENDOR_ALTIMA,
   2129 	  PCI_PRODUCT_ALTIMA_AC9100,
   2130 	  "Altima AC9100 Gigabit Ethernet",
   2131 	  },
   2132 
   2133 	{ PCI_VENDOR_BROADCOM,
   2134 	  PCI_PRODUCT_BROADCOM_BCM5700,
   2135 	  "Broadcom BCM5700 Gigabit Ethernet",
   2136 	  },
   2137 	{ PCI_VENDOR_BROADCOM,
   2138 	  PCI_PRODUCT_BROADCOM_BCM5701,
   2139 	  "Broadcom BCM5701 Gigabit Ethernet",
   2140 	  },
   2141 	{ PCI_VENDOR_BROADCOM,
   2142 	  PCI_PRODUCT_BROADCOM_BCM5702,
   2143 	  "Broadcom BCM5702 Gigabit Ethernet",
   2144 	  },
   2145 	{ PCI_VENDOR_BROADCOM,
   2146 	  PCI_PRODUCT_BROADCOM_BCM5702X,
   2147 	  "Broadcom BCM5702X Gigabit Ethernet" },
   2148 
   2149 	{ PCI_VENDOR_BROADCOM,
   2150 	  PCI_PRODUCT_BROADCOM_BCM5703,
   2151 	  "Broadcom BCM5703 Gigabit Ethernet",
   2152 	  },
   2153 	{ PCI_VENDOR_BROADCOM,
   2154 	  PCI_PRODUCT_BROADCOM_BCM5703X,
   2155 	  "Broadcom BCM5703X Gigabit Ethernet",
   2156 	  },
   2157 	{ PCI_VENDOR_BROADCOM,
   2158 	  PCI_PRODUCT_BROADCOM_BCM5703A3,
   2159 	  "Broadcom BCM5703A3 Gigabit Ethernet",
   2160 	  },
   2161 
   2162    	{ PCI_VENDOR_BROADCOM,
   2163 	  PCI_PRODUCT_BROADCOM_BCM5704C,
   2164 	  "Broadcom BCM5704C Dual Gigabit Ethernet",
   2165 	  },
   2166    	{ PCI_VENDOR_BROADCOM,
   2167 	  PCI_PRODUCT_BROADCOM_BCM5704S,
   2168 	  "Broadcom BCM5704S Dual Gigabit Ethernet",
   2169 	  },
   2170 
   2171    	{ PCI_VENDOR_BROADCOM,
   2172 	  PCI_PRODUCT_BROADCOM_BCM5705,
   2173 	  "Broadcom BCM5705 Gigabit Ethernet",
   2174 	  },
   2175    	{ PCI_VENDOR_BROADCOM,
   2176 	  PCI_PRODUCT_BROADCOM_BCM5705K,
   2177 	  "Broadcom BCM5705K Gigabit Ethernet",
   2178 	  },
   2179    	{ PCI_VENDOR_BROADCOM,
   2180 	  PCI_PRODUCT_BROADCOM_BCM5705_ALT,
   2181 	  "Broadcom BCM5705 Gigabit Ethernet",
   2182 	  },
   2183    	{ PCI_VENDOR_BROADCOM,
   2184 	  PCI_PRODUCT_BROADCOM_BCM5705M,
   2185 	  "Broadcom BCM5705M Gigabit Ethernet",
   2186 	  },
   2187 
   2188 	{ PCI_VENDOR_BROADCOM,
   2189 	  PCI_PRODUCT_BROADCOM_BCM5714,
   2190 	  "Broadcom BCM5714/5715 Gigabit Ethernet",
   2191 	  },
   2192 
   2193 	{ PCI_VENDOR_BROADCOM,
   2194 	  PCI_PRODUCT_BROADCOM_BCM5721,
   2195 	  "Broadcom BCM5721 Gigabit Ethernet",
   2196 	  },
   2197 
   2198 	{ PCI_VENDOR_BROADCOM,
   2199 	  PCI_PRODUCT_BROADCOM_BCM5750,
   2200 	  "Broadcom BCM5750 Gigabit Ethernet",
   2201 	  },
   2202 
   2203 	{ PCI_VENDOR_BROADCOM,
   2204 	  PCI_PRODUCT_BROADCOM_BCM5750M,
   2205 	  "Broadcom BCM5750M Gigabit Ethernet",
   2206 	  },
   2207 
   2208 	{ PCI_VENDOR_BROADCOM,
   2209 	  PCI_PRODUCT_BROADCOM_BCM5751,
   2210 	  "Broadcom BCM5751 Gigabit Ethernet",
   2211 	  },
   2212 
   2213 	{ PCI_VENDOR_BROADCOM,
   2214 	  PCI_PRODUCT_BROADCOM_BCM5751M,
   2215 	  "Broadcom BCM5751M Gigabit Ethernet",
   2216 	  },
   2217 
   2218 	{ PCI_VENDOR_BROADCOM,
   2219 	  PCI_PRODUCT_BROADCOM_BCM5752,
   2220 	  "Broadcom BCM5752 Gigabit Ethernet",
   2221 	  },
   2222 
   2223    	{ PCI_VENDOR_BROADCOM,
   2224 	  PCI_PRODUCT_BROADCOM_BCM5782,
   2225 	  "Broadcom BCM5782 Gigabit Ethernet",
   2226 	  },
   2227    	{ PCI_VENDOR_BROADCOM,
   2228 	  PCI_PRODUCT_BROADCOM_BCM5788,
   2229 	  "Broadcom BCM5788 Gigabit Ethernet",
   2230 	  },
   2231    	{ PCI_VENDOR_BROADCOM,
   2232 	  PCI_PRODUCT_BROADCOM_BCM5789,
   2233 	  "Broadcom BCM5789 Gigabit Ethernet",
   2234 	  },
   2235 
   2236    	{ PCI_VENDOR_BROADCOM,
   2237 	  PCI_PRODUCT_BROADCOM_BCM5901,
   2238 	  "Broadcom BCM5901 Fast Ethernet",
   2239 	  },
   2240    	{ PCI_VENDOR_BROADCOM,
   2241 	  PCI_PRODUCT_BROADCOM_BCM5901A2,
   2242 	  "Broadcom BCM5901A2 Fast Ethernet",
   2243 	  },
   2244 
   2245 	{ PCI_VENDOR_SCHNEIDERKOCH,
   2246 	  PCI_PRODUCT_SCHNEIDERKOCH_SK_9DX1,
   2247 	  "SysKonnect SK-9Dx1 Gigabit Ethernet",
   2248 	  },
   2249 
   2250 	{ PCI_VENDOR_3COM,
   2251 	  PCI_PRODUCT_3COM_3C996,
   2252 	  "3Com 3c996 Gigabit Ethernet",
   2253 	  },
   2254 
   2255 	{ 0,
   2256 	  0,
   2257 	  NULL },
   2258 };
   2259 
   2260 static const struct bge_product *
   2261 bge_lookup(const struct pci_attach_args *pa)
   2262 {
   2263 	const struct bge_product *bp;
   2264 
   2265 	for (bp = bge_products; bp->bp_name != NULL; bp++) {
   2266 		if (PCI_VENDOR(pa->pa_id) == bp->bp_vendor &&
   2267 		    PCI_PRODUCT(pa->pa_id) == bp->bp_product)
   2268 			return (bp);
   2269 	}
   2270 
   2271 	return (NULL);
   2272 }
   2273 
   2274 int
   2275 bge_setpowerstate(sc, powerlevel)
   2276 	struct bge_softc *sc;
   2277 	int powerlevel;
   2278 {
   2279 #ifdef NOTYET
   2280 	u_int32_t pm_ctl = 0;
   2281 
   2282 	/* XXX FIXME: make sure indirect accesses enabled? */
   2283 	pm_ctl = pci_conf_read(sc->bge_dev, BGE_PCI_MISC_CTL, 4);
   2284 	pm_ctl |= BGE_PCIMISCCTL_INDIRECT_ACCESS;
   2285 	pci_write_config(sc->bge_dev, BGE_PCI_MISC_CTL, pm_ctl, 4);
   2286 
   2287 	/* clear the PME_assert bit and power state bits, enable PME */
   2288 	pm_ctl = pci_conf_read(sc->bge_dev, BGE_PCI_PWRMGMT_CMD, 2);
   2289 	pm_ctl &= ~PCIM_PSTAT_DMASK;
   2290 	pm_ctl |= (1 << 8);
   2291 
   2292 	if (powerlevel == 0) {
   2293 		pm_ctl |= PCIM_PSTAT_D0;
   2294 		pci_write_config(sc->bge_dev, BGE_PCI_PWRMGMT_CMD,
   2295 		    pm_ctl, 2);
   2296 		DELAY(10000);
   2297 		CSR_WRITE_4(sc, BGE_MISC_LOCAL_CTL, sc->bge_local_ctrl_reg);
   2298 		DELAY(10000);
   2299 
   2300 #ifdef NOTYET
   2301 		/* XXX FIXME: write 0x02 to phy aux_Ctrl reg */
   2302 		bge_miibus_writereg(sc->bge_dev, 1, 0x18, 0x02);
   2303 #endif
   2304 		DELAY(40); DELAY(40); DELAY(40);
   2305 		DELAY(10000);	/* above not quite adequate on 5700 */
   2306 		return 0;
   2307 	}
   2308 
   2309 
   2310 	/*
   2311 	 * Entering ACPI power states D1-D3 is achieved by wiggling
   2312 	 * GMII gpio pins. Example code assumes all hardware vendors
   2313 	 * followed Broadom's sample pcb layout. Until we verify that
   2314 	 * for all supported OEM cards, states D1-D3 are  unsupported.
   2315 	 */
   2316 	printf("%s: power state %d unimplemented; check GPIO pins\n",
   2317 	       sc->bge_dev.dv_xname, powerlevel);
   2318 #endif
   2319 	return EOPNOTSUPP;
   2320 }
   2321 
   2322 
   2323 /*
   2324  * Probe for a Broadcom chip. Check the PCI vendor and device IDs
   2325  * against our list and return its name if we find a match. Note
   2326  * that since the Broadcom controller contains VPD support, we
   2327  * can get the device name string from the controller itself instead
   2328  * of the compiled-in string. This is a little slow, but it guarantees
   2329  * we'll always announce the right product name.
   2330  */
   2331 int
   2332 bge_probe(parent, match, aux)
   2333 	struct device *parent;
   2334 	struct cfdata *match;
   2335 	void *aux;
   2336 {
   2337 	struct pci_attach_args *pa = (struct pci_attach_args *)aux;
   2338 
   2339 	if (bge_lookup(pa) != NULL)
   2340 		return (1);
   2341 
   2342 	return (0);
   2343 }
   2344 
   2345 void
   2346 bge_attach(parent, self, aux)
   2347 	struct device *parent, *self;
   2348 	void *aux;
   2349 {
   2350 	struct bge_softc	*sc = (struct bge_softc *)self;
   2351 	struct pci_attach_args	*pa = aux;
   2352 	const struct bge_product *bp;
   2353 	const struct bge_revision *br;
   2354 	pci_chipset_tag_t	pc = pa->pa_pc;
   2355 	pci_intr_handle_t	ih;
   2356 	const char		*intrstr = NULL;
   2357 	bus_dma_segment_t	seg;
   2358 	int			rseg;
   2359 	u_int32_t		hwcfg = 0;
   2360 	u_int32_t		mac_addr = 0;
   2361 	u_int32_t		command;
   2362 	struct ifnet		*ifp;
   2363 	caddr_t			kva;
   2364 	u_char			eaddr[ETHER_ADDR_LEN];
   2365 	pcireg_t		memtype;
   2366 	bus_addr_t		memaddr;
   2367 	bus_size_t		memsize;
   2368 	u_int32_t		pm_ctl;
   2369 
   2370 	bp = bge_lookup(pa);
   2371 	KASSERT(bp != NULL);
   2372 
   2373 	sc->bge_pa = *pa;
   2374 
   2375 	aprint_naive(": Ethernet controller\n");
   2376 	aprint_normal(": %s\n", bp->bp_name);
   2377 
   2378 	/*
   2379 	 * Map control/status registers.
   2380 	 */
   2381 	DPRINTFN(5, ("Map control/status regs\n"));
   2382 	command = pci_conf_read(pc, pa->pa_tag, PCI_COMMAND_STATUS_REG);
   2383 	command |= PCI_COMMAND_MEM_ENABLE | PCI_COMMAND_MASTER_ENABLE;
   2384 	pci_conf_write(pc, pa->pa_tag, PCI_COMMAND_STATUS_REG, command);
   2385 	command = pci_conf_read(pc, pa->pa_tag, PCI_COMMAND_STATUS_REG);
   2386 
   2387 	if (!(command & PCI_COMMAND_MEM_ENABLE)) {
   2388 		aprint_error("%s: failed to enable memory mapping!\n",
   2389 		    sc->bge_dev.dv_xname);
   2390 		return;
   2391 	}
   2392 
   2393 	DPRINTFN(5, ("pci_mem_find\n"));
   2394 	memtype = pci_mapreg_type(pa->pa_pc, pa->pa_tag, BGE_PCI_BAR0);
   2395  	switch (memtype) {
   2396 	case PCI_MAPREG_TYPE_MEM | PCI_MAPREG_MEM_TYPE_32BIT:
   2397 	case PCI_MAPREG_TYPE_MEM | PCI_MAPREG_MEM_TYPE_64BIT:
   2398 		if (pci_mapreg_map(pa, BGE_PCI_BAR0,
   2399 		    memtype, 0, &sc->bge_btag, &sc->bge_bhandle,
   2400 		    &memaddr, &memsize) == 0)
   2401 			break;
   2402 	default:
   2403 		aprint_error("%s: can't find mem space\n",
   2404 		    sc->bge_dev.dv_xname);
   2405 		return;
   2406 	}
   2407 
   2408 	DPRINTFN(5, ("pci_intr_map\n"));
   2409 	if (pci_intr_map(pa, &ih)) {
   2410 		aprint_error("%s: couldn't map interrupt\n",
   2411 		    sc->bge_dev.dv_xname);
   2412 		return;
   2413 	}
   2414 
   2415 	DPRINTFN(5, ("pci_intr_string\n"));
   2416 	intrstr = pci_intr_string(pc, ih);
   2417 
   2418 	DPRINTFN(5, ("pci_intr_establish\n"));
   2419 	sc->bge_intrhand = pci_intr_establish(pc, ih, IPL_NET, bge_intr, sc);
   2420 
   2421 	if (sc->bge_intrhand == NULL) {
   2422 		aprint_error("%s: couldn't establish interrupt",
   2423 		    sc->bge_dev.dv_xname);
   2424 		if (intrstr != NULL)
   2425 			aprint_normal(" at %s", intrstr);
   2426 		aprint_normal("\n");
   2427 		return;
   2428 	}
   2429 	aprint_normal("%s: interrupting at %s\n",
   2430 	    sc->bge_dev.dv_xname, intrstr);
   2431 
   2432 	/*
   2433 	 * Kludge for 5700 Bx bug: a hardware bug (PCIX byte enable?)
   2434 	 * can clobber the chip's PCI config-space power control registers,
   2435 	 * leaving the card in D3 powersave state.
   2436 	 * We do not have memory-mapped registers in this state,
   2437 	 * so force device into D0 state before starting initialization.
   2438 	 */
   2439 	pm_ctl = pci_conf_read(pc, pa->pa_tag, BGE_PCI_PWRMGMT_CMD);
   2440 	pm_ctl &= ~(PCI_PWR_D0|PCI_PWR_D1|PCI_PWR_D2|PCI_PWR_D3);
   2441 	pm_ctl |= (1 << 8) | PCI_PWR_D0 ; /* D0 state */
   2442 	pci_conf_write(pc, pa->pa_tag, BGE_PCI_PWRMGMT_CMD, pm_ctl);
   2443 	DELAY(1000);	/* 27 usec is allegedly sufficent */
   2444 
   2445 	/*
   2446 	 * Save ASIC rev.  Look up any quirks associated with this
   2447 	 * ASIC.
   2448 	 */
   2449 	sc->bge_chipid =
   2450 	    pci_conf_read(pa->pa_pc, pa->pa_tag, BGE_PCI_MISC_CTL) &
   2451 	    BGE_PCIMISCCTL_ASICREV;
   2452 
   2453 	/*
   2454 	 * Detect PCI-Express devices
   2455 	 * XXX: guessed from Linux/FreeBSD; no documentation
   2456 	 */
   2457 	if (BGE_ASICREV(sc->bge_chipid) == BGE_ASICREV_BCM5750 &&
   2458 	    pci_get_capability(pa->pa_pc, pa->pa_tag, PCI_CAP_PCIEXPRESS,
   2459 	    NULL, NULL) != 0)
   2460 		sc->bge_pcie = 1;
   2461 	else
   2462 		sc->bge_pcie = 0;
   2463 
   2464 	/* Try to reset the chip. */
   2465 	DPRINTFN(5, ("bge_reset\n"));
   2466 	bge_reset(sc);
   2467 
   2468 	if (bge_chipinit(sc)) {
   2469 		aprint_error("%s: chip initialization failed\n",
   2470 		    sc->bge_dev.dv_xname);
   2471 		bge_release_resources(sc);
   2472 		return;
   2473 	}
   2474 
   2475 	/*
   2476 	 * Get station address from the EEPROM.
   2477 	 */
   2478 	mac_addr = bge_readmem_ind(sc, 0x0c14);
   2479 	if ((mac_addr >> 16) == 0x484b) {
   2480 		eaddr[0] = (u_char)(mac_addr >> 8);
   2481 		eaddr[1] = (u_char)(mac_addr >> 0);
   2482 		mac_addr = bge_readmem_ind(sc, 0x0c18);
   2483 		eaddr[2] = (u_char)(mac_addr >> 24);
   2484 		eaddr[3] = (u_char)(mac_addr >> 16);
   2485 		eaddr[4] = (u_char)(mac_addr >> 8);
   2486 		eaddr[5] = (u_char)(mac_addr >> 0);
   2487 	} else if (bge_read_eeprom(sc, (caddr_t)eaddr,
   2488 	    BGE_EE_MAC_OFFSET + 2, ETHER_ADDR_LEN)) {
   2489 		aprint_error("%s: failed to read station address\n",
   2490 		    sc->bge_dev.dv_xname);
   2491 		bge_release_resources(sc);
   2492 		return;
   2493 	}
   2494 
   2495 	br = bge_lookup_rev(sc->bge_chipid);
   2496 	aprint_normal("%s: ", sc->bge_dev.dv_xname);
   2497 
   2498 	if (br == NULL) {
   2499 		aprint_normal("unknown ASIC (0x%04x)", sc->bge_chipid >> 16);
   2500 		sc->bge_quirks = 0;
   2501 	} else {
   2502 		aprint_normal("ASIC %s (0x%04x)",
   2503 		    br->br_name, sc->bge_chipid >> 16);
   2504 		sc->bge_quirks |= br->br_quirks;
   2505 	}
   2506 	aprint_normal(", Ethernet address %s\n", ether_sprintf(eaddr));
   2507 
   2508 	/* Allocate the general information block and ring buffers. */
   2509 	if (pci_dma64_available(pa))
   2510 		sc->bge_dmatag = pa->pa_dmat64;
   2511 	else
   2512 		sc->bge_dmatag = pa->pa_dmat;
   2513 	DPRINTFN(5, ("bus_dmamem_alloc\n"));
   2514 	if (bus_dmamem_alloc(sc->bge_dmatag, sizeof(struct bge_ring_data),
   2515 			     PAGE_SIZE, 0, &seg, 1, &rseg, BUS_DMA_NOWAIT)) {
   2516 		aprint_error("%s: can't alloc rx buffers\n",
   2517 		    sc->bge_dev.dv_xname);
   2518 		return;
   2519 	}
   2520 	DPRINTFN(5, ("bus_dmamem_map\n"));
   2521 	if (bus_dmamem_map(sc->bge_dmatag, &seg, rseg,
   2522 			   sizeof(struct bge_ring_data), &kva,
   2523 			   BUS_DMA_NOWAIT)) {
   2524 		aprint_error("%s: can't map DMA buffers (%d bytes)\n",
   2525 		    sc->bge_dev.dv_xname, (int)sizeof(struct bge_ring_data));
   2526 		bus_dmamem_free(sc->bge_dmatag, &seg, rseg);
   2527 		return;
   2528 	}
   2529 	DPRINTFN(5, ("bus_dmamem_create\n"));
   2530 	if (bus_dmamap_create(sc->bge_dmatag, sizeof(struct bge_ring_data), 1,
   2531 	    sizeof(struct bge_ring_data), 0,
   2532 	    BUS_DMA_NOWAIT, &sc->bge_ring_map)) {
   2533 		aprint_error("%s: can't create DMA map\n",
   2534 		    sc->bge_dev.dv_xname);
   2535 		bus_dmamem_unmap(sc->bge_dmatag, kva,
   2536 				 sizeof(struct bge_ring_data));
   2537 		bus_dmamem_free(sc->bge_dmatag, &seg, rseg);
   2538 		return;
   2539 	}
   2540 	DPRINTFN(5, ("bus_dmamem_load\n"));
   2541 	if (bus_dmamap_load(sc->bge_dmatag, sc->bge_ring_map, kva,
   2542 			    sizeof(struct bge_ring_data), NULL,
   2543 			    BUS_DMA_NOWAIT)) {
   2544 		bus_dmamap_destroy(sc->bge_dmatag, sc->bge_ring_map);
   2545 		bus_dmamem_unmap(sc->bge_dmatag, kva,
   2546 				 sizeof(struct bge_ring_data));
   2547 		bus_dmamem_free(sc->bge_dmatag, &seg, rseg);
   2548 		return;
   2549 	}
   2550 
   2551 	DPRINTFN(5, ("bzero\n"));
   2552 	sc->bge_rdata = (struct bge_ring_data *)kva;
   2553 
   2554 	memset(sc->bge_rdata, 0, sizeof(struct bge_ring_data));
   2555 
   2556 	/* Try to allocate memory for jumbo buffers. */
   2557 	if ((sc->bge_quirks & BGE_QUIRK_5705_CORE) == 0) {
   2558 		if (bge_alloc_jumbo_mem(sc)) {
   2559 			aprint_error("%s: jumbo buffer allocation failed\n",
   2560 			    sc->bge_dev.dv_xname);
   2561 		} else
   2562 			sc->ethercom.ec_capabilities |= ETHERCAP_JUMBO_MTU;
   2563 	}
   2564 
   2565 	/* Set default tuneable values. */
   2566 	sc->bge_stat_ticks = BGE_TICKS_PER_SEC;
   2567 	sc->bge_rx_coal_ticks = 150;
   2568 	sc->bge_rx_max_coal_bds = 64;
   2569 #ifdef ORIG_WPAUL_VALUES
   2570 	sc->bge_tx_coal_ticks = 150;
   2571 	sc->bge_tx_max_coal_bds = 128;
   2572 #else
   2573 	sc->bge_tx_coal_ticks = 300;
   2574 	sc->bge_tx_max_coal_bds = 400;
   2575 #endif
   2576 	if (sc->bge_quirks & BGE_QUIRK_5705_CORE) {
   2577 		sc->bge_tx_coal_ticks = (12 * 5);
   2578 		sc->bge_rx_max_coal_bds = (12 * 5);
   2579 			aprint_error("%s: setting short Tx thresholds\n",
   2580 			    sc->bge_dev.dv_xname);
   2581 	}
   2582 
   2583 	/* Set up ifnet structure */
   2584 	ifp = &sc->ethercom.ec_if;
   2585 	ifp->if_softc = sc;
   2586 	ifp->if_flags = IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST;
   2587 	ifp->if_ioctl = bge_ioctl;
   2588 	ifp->if_start = bge_start;
   2589 	ifp->if_init = bge_init;
   2590 	ifp->if_watchdog = bge_watchdog;
   2591 	IFQ_SET_MAXLEN(&ifp->if_snd, max(BGE_TX_RING_CNT - 1, IFQ_MAXLEN));
   2592 	IFQ_SET_READY(&ifp->if_snd);
   2593 	DPRINTFN(5, ("bcopy\n"));
   2594 	strcpy(ifp->if_xname, sc->bge_dev.dv_xname);
   2595 
   2596 	if ((sc->bge_quirks & BGE_QUIRK_CSUM_BROKEN) == 0)
   2597 		sc->ethercom.ec_if.if_capabilities |=
   2598 		    IFCAP_CSUM_IPv4_Tx | IFCAP_CSUM_IPv4_Rx |
   2599 		    IFCAP_CSUM_TCPv4_Tx | IFCAP_CSUM_TCPv4_Rx |
   2600 		    IFCAP_CSUM_UDPv4_Tx | IFCAP_CSUM_UDPv4_Rx;
   2601 	sc->ethercom.ec_capabilities |=
   2602 	    ETHERCAP_VLAN_HWTAGGING | ETHERCAP_VLAN_MTU;
   2603 
   2604 	if (sc->bge_pcie)
   2605 		sc->ethercom.ec_if.if_capabilities |= IFCAP_TSOv4;
   2606 
   2607 	/*
   2608 	 * Do MII setup.
   2609 	 */
   2610 	DPRINTFN(5, ("mii setup\n"));
   2611 	sc->bge_mii.mii_ifp = ifp;
   2612 	sc->bge_mii.mii_readreg = bge_miibus_readreg;
   2613 	sc->bge_mii.mii_writereg = bge_miibus_writereg;
   2614 	sc->bge_mii.mii_statchg = bge_miibus_statchg;
   2615 
   2616 	/*
   2617 	 * Figure out what sort of media we have by checking the
   2618 	 * hardware config word in the first 32k of NIC internal memory,
   2619 	 * or fall back to the config word in the EEPROM. Note: on some BCM5700
   2620 	 * cards, this value appears to be unset. If that's the
   2621 	 * case, we have to rely on identifying the NIC by its PCI
   2622 	 * subsystem ID, as we do below for the SysKonnect SK-9D41.
   2623 	 */
   2624 	if (bge_readmem_ind(sc, BGE_SOFTWARE_GENCOMM_SIG) == BGE_MAGIC_NUMBER) {
   2625 		hwcfg = bge_readmem_ind(sc, BGE_SOFTWARE_GENCOMM_NICCFG);
   2626 	} else {
   2627 		bge_read_eeprom(sc, (caddr_t)&hwcfg,
   2628 		    BGE_EE_HWCFG_OFFSET, sizeof(hwcfg));
   2629 		hwcfg = be32toh(hwcfg);
   2630 	}
   2631 	if ((hwcfg & BGE_HWCFG_MEDIA) == BGE_MEDIA_FIBER)
   2632 		sc->bge_tbi = 1;
   2633 
   2634 	/* The SysKonnect SK-9D41 is a 1000baseSX card. */
   2635 	if ((pci_conf_read(pa->pa_pc, pa->pa_tag, BGE_PCI_SUBSYS) >> 16) ==
   2636 	    SK_SUBSYSID_9D41)
   2637 		sc->bge_tbi = 1;
   2638 
   2639 	if (sc->bge_tbi) {
   2640 		ifmedia_init(&sc->bge_ifmedia, IFM_IMASK, bge_ifmedia_upd,
   2641 		    bge_ifmedia_sts);
   2642 		ifmedia_add(&sc->bge_ifmedia, IFM_ETHER|IFM_1000_SX, 0, NULL);
   2643 		ifmedia_add(&sc->bge_ifmedia, IFM_ETHER|IFM_1000_SX|IFM_FDX,
   2644 			    0, NULL);
   2645 		ifmedia_add(&sc->bge_ifmedia, IFM_ETHER|IFM_AUTO, 0, NULL);
   2646 		ifmedia_set(&sc->bge_ifmedia, IFM_ETHER|IFM_AUTO);
   2647 	} else {
   2648 		/*
   2649 		 * Do transceiver setup.
   2650 		 */
   2651 		ifmedia_init(&sc->bge_mii.mii_media, 0, bge_ifmedia_upd,
   2652 			     bge_ifmedia_sts);
   2653 		mii_attach(&sc->bge_dev, &sc->bge_mii, 0xffffffff,
   2654 			   MII_PHY_ANY, MII_OFFSET_ANY,
   2655 			   MIIF_FORCEANEG|MIIF_DOPAUSE);
   2656 
   2657 		if (LIST_FIRST(&sc->bge_mii.mii_phys) == NULL) {
   2658 			printf("%s: no PHY found!\n", sc->bge_dev.dv_xname);
   2659 			ifmedia_add(&sc->bge_mii.mii_media,
   2660 				    IFM_ETHER|IFM_MANUAL, 0, NULL);
   2661 			ifmedia_set(&sc->bge_mii.mii_media,
   2662 				    IFM_ETHER|IFM_MANUAL);
   2663 		} else
   2664 			ifmedia_set(&sc->bge_mii.mii_media,
   2665 				    IFM_ETHER|IFM_AUTO);
   2666 	}
   2667 
   2668 	/*
   2669 	 * When using the BCM5701 in PCI-X mode, data corruption has
   2670 	 * been observed in the first few bytes of some received packets.
   2671 	 * Aligning the packet buffer in memory eliminates the corruption.
   2672 	 * Unfortunately, this misaligns the packet payloads.  On platforms
   2673 	 * which do not support unaligned accesses, we will realign the
   2674 	 * payloads by copying the received packets.
   2675 	 */
   2676 	if (sc->bge_quirks & BGE_QUIRK_PCIX_DMA_ALIGN_BUG) {
   2677 		/* If in PCI-X mode, work around the alignment bug. */
   2678 		if ((pci_conf_read(pa->pa_pc, pa->pa_tag, BGE_PCI_PCISTATE) &
   2679                     (BGE_PCISTATE_PCI_BUSMODE | BGE_PCISTATE_PCI_BUSSPEED)) ==
   2680                          BGE_PCISTATE_PCI_BUSSPEED)
   2681 		sc->bge_rx_alignment_bug = 1;
   2682         }
   2683 
   2684 	/*
   2685 	 * Call MI attach routine.
   2686 	 */
   2687 	DPRINTFN(5, ("if_attach\n"));
   2688 	if_attach(ifp);
   2689 	DPRINTFN(5, ("ether_ifattach\n"));
   2690 	ether_ifattach(ifp, eaddr);
   2691 #ifdef BGE_EVENT_COUNTERS
   2692 	/*
   2693 	 * Attach event counters.
   2694 	 */
   2695 	evcnt_attach_dynamic(&sc->bge_ev_intr, EVCNT_TYPE_INTR,
   2696 	    NULL, sc->bge_dev.dv_xname, "intr");
   2697 	evcnt_attach_dynamic(&sc->bge_ev_tx_xoff, EVCNT_TYPE_MISC,
   2698 	    NULL, sc->bge_dev.dv_xname, "tx_xoff");
   2699 	evcnt_attach_dynamic(&sc->bge_ev_tx_xon, EVCNT_TYPE_MISC,
   2700 	    NULL, sc->bge_dev.dv_xname, "tx_xon");
   2701 	evcnt_attach_dynamic(&sc->bge_ev_rx_xoff, EVCNT_TYPE_MISC,
   2702 	    NULL, sc->bge_dev.dv_xname, "rx_xoff");
   2703 	evcnt_attach_dynamic(&sc->bge_ev_rx_xon, EVCNT_TYPE_MISC,
   2704 	    NULL, sc->bge_dev.dv_xname, "rx_xon");
   2705 	evcnt_attach_dynamic(&sc->bge_ev_rx_macctl, EVCNT_TYPE_MISC,
   2706 	    NULL, sc->bge_dev.dv_xname, "rx_macctl");
   2707 	evcnt_attach_dynamic(&sc->bge_ev_xoffentered, EVCNT_TYPE_MISC,
   2708 	    NULL, sc->bge_dev.dv_xname, "xoffentered");
   2709 #endif /* BGE_EVENT_COUNTERS */
   2710 	DPRINTFN(5, ("callout_init\n"));
   2711 	callout_init(&sc->bge_timeout);
   2712 
   2713 	sc->bge_powerhook = powerhook_establish(bge_powerhook, sc);
   2714 	if (sc->bge_powerhook == NULL)
   2715 		printf("%s: WARNING: unable to establish PCI power hook\n",
   2716 		    sc->bge_dev.dv_xname);
   2717 }
   2718 
   2719 void
   2720 bge_release_resources(sc)
   2721 	struct bge_softc *sc;
   2722 {
   2723 	if (sc->bge_vpd_prodname != NULL)
   2724 		free(sc->bge_vpd_prodname, M_DEVBUF);
   2725 
   2726 	if (sc->bge_vpd_readonly != NULL)
   2727 		free(sc->bge_vpd_readonly, M_DEVBUF);
   2728 }
   2729 
   2730 void
   2731 bge_reset(sc)
   2732 	struct bge_softc *sc;
   2733 {
   2734 	struct pci_attach_args *pa = &sc->bge_pa;
   2735 	u_int32_t cachesize, command, pcistate, new_pcistate;
   2736 	int i, val;
   2737 
   2738 	/* Save some important PCI state. */
   2739 	cachesize = pci_conf_read(pa->pa_pc, pa->pa_tag, BGE_PCI_CACHESZ);
   2740 	command = pci_conf_read(pa->pa_pc, pa->pa_tag, BGE_PCI_CMD);
   2741 	pcistate = pci_conf_read(pa->pa_pc, pa->pa_tag, BGE_PCI_PCISTATE);
   2742 
   2743 	pci_conf_write(pa->pa_pc, pa->pa_tag, BGE_PCI_MISC_CTL,
   2744 	    BGE_PCIMISCCTL_INDIRECT_ACCESS|BGE_PCIMISCCTL_MASK_PCI_INTR|
   2745 	    BGE_HIF_SWAP_OPTIONS|BGE_PCIMISCCTL_PCISTATE_RW);
   2746 
   2747 	val = BGE_MISCCFG_RESET_CORE_CLOCKS | (65<<1);
   2748 	/*
   2749 	 * XXX: from FreeBSD/Linux; no documentation
   2750 	 */
   2751 	if (sc->bge_pcie) {
   2752 		if (CSR_READ_4(sc, BGE_PCIE_CTL1) == 0x60)
   2753 			CSR_WRITE_4(sc, BGE_PCIE_CTL1, 0x20);
   2754 		if (sc->bge_chipid != BGE_CHIPID_BCM5750_A0) {
   2755 			/* No idea what that actually means */
   2756 			CSR_WRITE_4(sc, BGE_MISC_CFG, 1 << 29);
   2757 			val |= (1<<29);
   2758 		}
   2759 	}
   2760 
   2761 	/* Issue global reset */
   2762 	bge_writereg_ind(sc, BGE_MISC_CFG, val);
   2763 
   2764 	DELAY(1000);
   2765 
   2766 	/*
   2767 	 * XXX: from FreeBSD/Linux; no documentation
   2768 	 */
   2769 	if (sc->bge_pcie) {
   2770 		if (sc->bge_chipid == BGE_CHIPID_BCM5750_A0) {
   2771 			pcireg_t reg;
   2772 
   2773 			DELAY(500000);
   2774 			/* XXX: Magic Numbers */
   2775 			reg = pci_conf_read(pa->pa_pc, pa->pa_tag, BGE_PCI_UNKNOWN0);
   2776 			pci_conf_write(pa->pa_pc, pa->pa_tag, BGE_PCI_UNKNOWN0,
   2777 			    reg | (1 << 15));
   2778 		}
   2779 		/*
   2780 		 * XXX: Magic Numbers.
   2781 		 * Sets maximal PCI-e payload and clears any PCI-e errors.
   2782 		 * Should be replaced with references to PCI config-space
   2783 		 * capability block for PCI-Express.
   2784 		 */
   2785 		pci_conf_write(pa->pa_pc, pa->pa_tag,
   2786 		    BGE_PCI_CONF_DEV_CTRL, 0xf5000);
   2787 
   2788 	}
   2789 
   2790 	/* Reset some of the PCI state that got zapped by reset */
   2791 	pci_conf_write(pa->pa_pc, pa->pa_tag, BGE_PCI_MISC_CTL,
   2792 	    BGE_PCIMISCCTL_INDIRECT_ACCESS|BGE_PCIMISCCTL_MASK_PCI_INTR|
   2793 	    BGE_HIF_SWAP_OPTIONS|BGE_PCIMISCCTL_PCISTATE_RW);
   2794 	pci_conf_write(pa->pa_pc, pa->pa_tag, BGE_PCI_CMD, command);
   2795 	pci_conf_write(pa->pa_pc, pa->pa_tag, BGE_PCI_CACHESZ, cachesize);
   2796 	bge_writereg_ind(sc, BGE_MISC_CFG, (65 << 1));
   2797 
   2798 	/* Enable memory arbiter. */
   2799 	if ((sc->bge_quirks & BGE_QUIRK_5705_CORE) == 0) {
   2800 		CSR_WRITE_4(sc, BGE_MARB_MODE, BGE_MARBMODE_ENABLE);
   2801 	}
   2802 
   2803 	/*
   2804 	 * Prevent PXE restart: write a magic number to the
   2805 	 * general communications memory at 0xB50.
   2806 	 */
   2807 	bge_writemem_ind(sc, BGE_SOFTWARE_GENCOMM, BGE_MAGIC_NUMBER);
   2808 
   2809 	/*
   2810 	 * Poll the value location we just wrote until
   2811 	 * we see the 1's complement of the magic number.
   2812 	 * This indicates that the firmware initialization
   2813 	 * is complete.
   2814 	 */
   2815 	for (i = 0; i < BGE_TIMEOUT; i++) {
   2816 		val = bge_readmem_ind(sc, BGE_SOFTWARE_GENCOMM);
   2817 		if (val == ~BGE_MAGIC_NUMBER)
   2818 			break;
   2819 		DELAY(1000);
   2820 	}
   2821 
   2822 	if (i >= BGE_TIMEOUT) {
   2823 		printf("%s: firmware handshake timed out, val = %x\n",
   2824 		    sc->bge_dev.dv_xname, val);
   2825 		/*
   2826 		 * XXX: occasionally fired on bcm5721, but without
   2827 		 * apparent harm.  For now, keep going if we timeout
   2828 		 * against PCI-E devices.
   2829 		 */
   2830 		 if (!sc->bge_pcie)
   2831 		  return;
   2832 	}
   2833 
   2834 	/*
   2835 	 * XXX Wait for the value of the PCISTATE register to
   2836 	 * return to its original pre-reset state. This is a
   2837 	 * fairly good indicator of reset completion. If we don't
   2838 	 * wait for the reset to fully complete, trying to read
   2839 	 * from the device's non-PCI registers may yield garbage
   2840 	 * results.
   2841 	 */
   2842 	for (i = 0; i < BGE_TIMEOUT; i++) {
   2843 		new_pcistate = pci_conf_read(pa->pa_pc, pa->pa_tag,
   2844 		    BGE_PCI_PCISTATE);
   2845 		if ((new_pcistate & ~BGE_PCISTATE_RESERVED) ==
   2846 		    (pcistate & ~BGE_PCISTATE_RESERVED))
   2847 			break;
   2848 		DELAY(10);
   2849 	}
   2850 	if ((new_pcistate & ~BGE_PCISTATE_RESERVED) !=
   2851 	    (pcistate & ~BGE_PCISTATE_RESERVED)) {
   2852 		printf("%s: pcistate failed to revert\n",
   2853 		    sc->bge_dev.dv_xname);
   2854 	}
   2855 
   2856 	/* XXX: from FreeBSD/Linux; no documentation */
   2857 	if (sc->bge_pcie && sc->bge_chipid != BGE_CHIPID_BCM5750_A0)
   2858 		CSR_WRITE_4(sc, BGE_PCIE_CTL0, CSR_READ_4(sc, BGE_PCIE_CTL0) | (1<<25));
   2859 
   2860 	/* Enable memory arbiter. */
   2861 	if ((sc->bge_quirks & BGE_QUIRK_5705_CORE) == 0) {
   2862 		CSR_WRITE_4(sc, BGE_MARB_MODE, BGE_MARBMODE_ENABLE);
   2863 	}
   2864 
   2865 	/* Fix up byte swapping */
   2866 	CSR_WRITE_4(sc, BGE_MODE_CTL, BGE_DMA_SWAP_OPTIONS);
   2867 
   2868 	CSR_WRITE_4(sc, BGE_MAC_MODE, 0);
   2869 
   2870 	DELAY(10000);
   2871 }
   2872 
   2873 /*
   2874  * Frame reception handling. This is called if there's a frame
   2875  * on the receive return list.
   2876  *
   2877  * Note: we have to be able to handle two possibilities here:
   2878  * 1) the frame is from the jumbo recieve ring
   2879  * 2) the frame is from the standard receive ring
   2880  */
   2881 
   2882 void
   2883 bge_rxeof(sc)
   2884 	struct bge_softc *sc;
   2885 {
   2886 	struct ifnet *ifp;
   2887 	int stdcnt = 0, jumbocnt = 0;
   2888 	bus_dmamap_t dmamap;
   2889 	bus_addr_t offset, toff;
   2890 	bus_size_t tlen;
   2891 	int tosync;
   2892 
   2893 	ifp = &sc->ethercom.ec_if;
   2894 
   2895 	bus_dmamap_sync(sc->bge_dmatag, sc->bge_ring_map,
   2896 	    offsetof(struct bge_ring_data, bge_status_block),
   2897 	    sizeof (struct bge_status_block),
   2898 	    BUS_DMASYNC_POSTREAD);
   2899 
   2900 	offset = offsetof(struct bge_ring_data, bge_rx_return_ring);
   2901 	tosync = sc->bge_rdata->bge_status_block.bge_idx[0].bge_rx_prod_idx -
   2902 	    sc->bge_rx_saved_considx;
   2903 
   2904 	toff = offset + (sc->bge_rx_saved_considx * sizeof (struct bge_rx_bd));
   2905 
   2906 	if (tosync < 0) {
   2907 		tlen = (sc->bge_return_ring_cnt - sc->bge_rx_saved_considx) *
   2908 		    sizeof (struct bge_rx_bd);
   2909 		bus_dmamap_sync(sc->bge_dmatag, sc->bge_ring_map,
   2910 		    toff, tlen, BUS_DMASYNC_POSTREAD);
   2911 		tosync = -tosync;
   2912 	}
   2913 
   2914 	bus_dmamap_sync(sc->bge_dmatag, sc->bge_ring_map,
   2915 	    offset, tosync * sizeof (struct bge_rx_bd),
   2916 	    BUS_DMASYNC_POSTREAD);
   2917 
   2918 	while(sc->bge_rx_saved_considx !=
   2919 	    sc->bge_rdata->bge_status_block.bge_idx[0].bge_rx_prod_idx) {
   2920 		struct bge_rx_bd	*cur_rx;
   2921 		u_int32_t		rxidx;
   2922 		struct mbuf		*m = NULL;
   2923 
   2924 		cur_rx = &sc->bge_rdata->
   2925 			bge_rx_return_ring[sc->bge_rx_saved_considx];
   2926 
   2927 		rxidx = cur_rx->bge_idx;
   2928 		BGE_INC(sc->bge_rx_saved_considx, sc->bge_return_ring_cnt);
   2929 
   2930 		if (cur_rx->bge_flags & BGE_RXBDFLAG_JUMBO_RING) {
   2931 			BGE_INC(sc->bge_jumbo, BGE_JUMBO_RX_RING_CNT);
   2932 			m = sc->bge_cdata.bge_rx_jumbo_chain[rxidx];
   2933 			sc->bge_cdata.bge_rx_jumbo_chain[rxidx] = NULL;
   2934 			jumbocnt++;
   2935 			if (cur_rx->bge_flags & BGE_RXBDFLAG_ERROR) {
   2936 				ifp->if_ierrors++;
   2937 				bge_newbuf_jumbo(sc, sc->bge_jumbo, m);
   2938 				continue;
   2939 			}
   2940 			if (bge_newbuf_jumbo(sc, sc->bge_jumbo,
   2941 					     NULL)== ENOBUFS) {
   2942 				ifp->if_ierrors++;
   2943 				bge_newbuf_jumbo(sc, sc->bge_jumbo, m);
   2944 				continue;
   2945 			}
   2946 		} else {
   2947 			BGE_INC(sc->bge_std, BGE_STD_RX_RING_CNT);
   2948 			m = sc->bge_cdata.bge_rx_std_chain[rxidx];
   2949 			sc->bge_cdata.bge_rx_std_chain[rxidx] = NULL;
   2950 			stdcnt++;
   2951 			dmamap = sc->bge_cdata.bge_rx_std_map[rxidx];
   2952 			sc->bge_cdata.bge_rx_std_map[rxidx] = 0;
   2953 			if (cur_rx->bge_flags & BGE_RXBDFLAG_ERROR) {
   2954 				ifp->if_ierrors++;
   2955 				bge_newbuf_std(sc, sc->bge_std, m, dmamap);
   2956 				continue;
   2957 			}
   2958 			if (bge_newbuf_std(sc, sc->bge_std,
   2959 			    NULL, dmamap) == ENOBUFS) {
   2960 				ifp->if_ierrors++;
   2961 				bge_newbuf_std(sc, sc->bge_std, m, dmamap);
   2962 				continue;
   2963 			}
   2964 		}
   2965 
   2966 		ifp->if_ipackets++;
   2967 #ifndef __NO_STRICT_ALIGNMENT
   2968                 /*
   2969                  * XXX: if the 5701 PCIX-Rx-DMA workaround is in effect,
   2970                  * the Rx buffer has the layer-2 header unaligned.
   2971                  * If our CPU requires alignment, re-align by copying.
   2972                  */
   2973 		if (sc->bge_rx_alignment_bug) {
   2974 			memmove(mtod(m, caddr_t) + ETHER_ALIGN, m->m_data,
   2975                                 cur_rx->bge_len);
   2976 			m->m_data += ETHER_ALIGN;
   2977 		}
   2978 #endif
   2979 
   2980 		m->m_pkthdr.len = m->m_len = cur_rx->bge_len - ETHER_CRC_LEN;
   2981 		m->m_pkthdr.rcvif = ifp;
   2982 
   2983 #if NBPFILTER > 0
   2984 		/*
   2985 		 * Handle BPF listeners. Let the BPF user see the packet.
   2986 		 */
   2987 		if (ifp->if_bpf)
   2988 			bpf_mtap(ifp->if_bpf, m);
   2989 #endif
   2990 
   2991 		m->m_pkthdr.csum_flags = M_CSUM_IPv4;
   2992 
   2993 		if ((cur_rx->bge_ip_csum ^ 0xffff) != 0)
   2994 			m->m_pkthdr.csum_flags |= M_CSUM_IPv4_BAD;
   2995 		/*
   2996 		 * Rx transport checksum-offload may also
   2997 		 * have bugs with packets which, when transmitted,
   2998 		 * were `runts' requiring padding.
   2999 		 */
   3000 		if (cur_rx->bge_flags & BGE_RXBDFLAG_TCP_UDP_CSUM &&
   3001 		    (/* (sc->_bge_quirks & BGE_QUIRK_SHORT_CKSUM_BUG) == 0 ||*/
   3002 		     m->m_pkthdr.len >= ETHER_MIN_NOPAD)) {
   3003 			m->m_pkthdr.csum_data =
   3004 			    cur_rx->bge_tcp_udp_csum;
   3005 			m->m_pkthdr.csum_flags |=
   3006 			    (M_CSUM_TCPv4|M_CSUM_UDPv4|
   3007 			     M_CSUM_DATA|M_CSUM_NO_PSEUDOHDR);
   3008 		}
   3009 
   3010 		/*
   3011 		 * If we received a packet with a vlan tag, pass it
   3012 		 * to vlan_input() instead of ether_input().
   3013 		 */
   3014 		if (cur_rx->bge_flags & BGE_RXBDFLAG_VLAN_TAG)
   3015 			VLAN_INPUT_TAG(ifp, m, cur_rx->bge_vlan_tag, continue);
   3016 
   3017 		(*ifp->if_input)(ifp, m);
   3018 	}
   3019 
   3020 	CSR_WRITE_4(sc, BGE_MBX_RX_CONS0_LO, sc->bge_rx_saved_considx);
   3021 	if (stdcnt)
   3022 		CSR_WRITE_4(sc, BGE_MBX_RX_STD_PROD_LO, sc->bge_std);
   3023 	if (jumbocnt)
   3024 		CSR_WRITE_4(sc, BGE_MBX_RX_JUMBO_PROD_LO, sc->bge_jumbo);
   3025 }
   3026 
   3027 void
   3028 bge_txeof(sc)
   3029 	struct bge_softc *sc;
   3030 {
   3031 	struct bge_tx_bd *cur_tx = NULL;
   3032 	struct ifnet *ifp;
   3033 	struct txdmamap_pool_entry *dma;
   3034 	bus_addr_t offset, toff;
   3035 	bus_size_t tlen;
   3036 	int tosync;
   3037 	struct mbuf *m;
   3038 
   3039 	ifp = &sc->ethercom.ec_if;
   3040 
   3041 	bus_dmamap_sync(sc->bge_dmatag, sc->bge_ring_map,
   3042 	    offsetof(struct bge_ring_data, bge_status_block),
   3043 	    sizeof (struct bge_status_block),
   3044 	    BUS_DMASYNC_POSTREAD);
   3045 
   3046 	offset = offsetof(struct bge_ring_data, bge_tx_ring);
   3047 	tosync = sc->bge_rdata->bge_status_block.bge_idx[0].bge_tx_cons_idx -
   3048 	    sc->bge_tx_saved_considx;
   3049 
   3050 	toff = offset + (sc->bge_tx_saved_considx * sizeof (struct bge_tx_bd));
   3051 
   3052 	if (tosync < 0) {
   3053 		tlen = (BGE_TX_RING_CNT - sc->bge_tx_saved_considx) *
   3054 		    sizeof (struct bge_tx_bd);
   3055 		bus_dmamap_sync(sc->bge_dmatag, sc->bge_ring_map,
   3056 		    toff, tlen, BUS_DMASYNC_POSTREAD|BUS_DMASYNC_POSTWRITE);
   3057 		tosync = -tosync;
   3058 	}
   3059 
   3060 	bus_dmamap_sync(sc->bge_dmatag, sc->bge_ring_map,
   3061 	    offset, tosync * sizeof (struct bge_tx_bd),
   3062 	    BUS_DMASYNC_POSTREAD|BUS_DMASYNC_POSTWRITE);
   3063 
   3064 	/*
   3065 	 * Go through our tx ring and free mbufs for those
   3066 	 * frames that have been sent.
   3067 	 */
   3068 	while (sc->bge_tx_saved_considx !=
   3069 	    sc->bge_rdata->bge_status_block.bge_idx[0].bge_tx_cons_idx) {
   3070 		u_int32_t		idx = 0;
   3071 
   3072 		idx = sc->bge_tx_saved_considx;
   3073 		cur_tx = &sc->bge_rdata->bge_tx_ring[idx];
   3074 		if (cur_tx->bge_flags & BGE_TXBDFLAG_END)
   3075 			ifp->if_opackets++;
   3076 		m = sc->bge_cdata.bge_tx_chain[idx];
   3077 		if (m != NULL) {
   3078 			sc->bge_cdata.bge_tx_chain[idx] = NULL;
   3079 			dma = sc->txdma[idx];
   3080 			bus_dmamap_sync(sc->bge_dmatag, dma->dmamap, 0,
   3081 			    dma->dmamap->dm_mapsize, BUS_DMASYNC_POSTWRITE);
   3082 			bus_dmamap_unload(sc->bge_dmatag, dma->dmamap);
   3083 			SLIST_INSERT_HEAD(&sc->txdma_list, dma, link);
   3084 			sc->txdma[idx] = NULL;
   3085 
   3086 			m_freem(m);
   3087 		}
   3088 		sc->bge_txcnt--;
   3089 		BGE_INC(sc->bge_tx_saved_considx, BGE_TX_RING_CNT);
   3090 		ifp->if_timer = 0;
   3091 	}
   3092 
   3093 	if (cur_tx != NULL)
   3094 		ifp->if_flags &= ~IFF_OACTIVE;
   3095 }
   3096 
   3097 int
   3098 bge_intr(xsc)
   3099 	void *xsc;
   3100 {
   3101 	struct bge_softc *sc;
   3102 	struct ifnet *ifp;
   3103 
   3104 	sc = xsc;
   3105 	ifp = &sc->ethercom.ec_if;
   3106 
   3107 #ifdef notdef
   3108 	/* Avoid this for now -- checking this register is expensive. */
   3109 	/* Make sure this is really our interrupt. */
   3110 	if (!(CSR_READ_4(sc, BGE_MISC_LOCAL_CTL) & BGE_MLC_INTR_STATE))
   3111 		return (0);
   3112 #endif
   3113 	/* Ack interrupt and stop others from occuring. */
   3114 	CSR_WRITE_4(sc, BGE_MBX_IRQ0_LO, 1);
   3115 
   3116 	BGE_EVCNT_INCR(sc->bge_ev_intr);
   3117 
   3118 	/*
   3119 	 * Process link state changes.
   3120 	 * Grrr. The link status word in the status block does
   3121 	 * not work correctly on the BCM5700 rev AX and BX chips,
   3122 	 * according to all avaibable information. Hence, we have
   3123 	 * to enable MII interrupts in order to properly obtain
   3124 	 * async link changes. Unfortunately, this also means that
   3125 	 * we have to read the MAC status register to detect link
   3126 	 * changes, thereby adding an additional register access to
   3127 	 * the interrupt handler.
   3128 	 */
   3129 
   3130 	if (sc->bge_quirks & BGE_QUIRK_LINK_STATE_BROKEN) {
   3131 		u_int32_t		status;
   3132 
   3133 		status = CSR_READ_4(sc, BGE_MAC_STS);
   3134 		if (status & BGE_MACSTAT_MI_INTERRUPT) {
   3135 			sc->bge_link = 0;
   3136 			callout_stop(&sc->bge_timeout);
   3137 			bge_tick(sc);
   3138 			/* Clear the interrupt */
   3139 			CSR_WRITE_4(sc, BGE_MAC_EVT_ENB,
   3140 			    BGE_EVTENB_MI_INTERRUPT);
   3141 			bge_miibus_readreg(&sc->bge_dev, 1, BRGPHY_MII_ISR);
   3142 			bge_miibus_writereg(&sc->bge_dev, 1, BRGPHY_MII_IMR,
   3143 			    BRGPHY_INTRS);
   3144 		}
   3145 	} else {
   3146 		if (sc->bge_rdata->bge_status_block.bge_status &
   3147 		    BGE_STATFLAG_LINKSTATE_CHANGED) {
   3148 			sc->bge_link = 0;
   3149 			callout_stop(&sc->bge_timeout);
   3150 			bge_tick(sc);
   3151 			/* Clear the interrupt */
   3152 			CSR_WRITE_4(sc, BGE_MAC_STS, BGE_MACSTAT_SYNC_CHANGED|
   3153 			    BGE_MACSTAT_CFG_CHANGED|BGE_MACSTAT_MI_COMPLETE|
   3154 			    BGE_MACSTAT_LINK_CHANGED);
   3155 		}
   3156 	}
   3157 
   3158 	if (ifp->if_flags & IFF_RUNNING) {
   3159 		/* Check RX return ring producer/consumer */
   3160 		bge_rxeof(sc);
   3161 
   3162 		/* Check TX ring producer/consumer */
   3163 		bge_txeof(sc);
   3164 	}
   3165 
   3166 	if (sc->bge_pending_rxintr_change) {
   3167 		uint32_t rx_ticks = sc->bge_rx_coal_ticks;
   3168 		uint32_t rx_bds = sc->bge_rx_max_coal_bds;
   3169 		uint32_t junk;
   3170 
   3171 		CSR_WRITE_4(sc, BGE_HCC_RX_COAL_TICKS, rx_ticks);
   3172 		DELAY(10);
   3173 		junk = CSR_READ_4(sc, BGE_HCC_RX_COAL_TICKS);
   3174 
   3175 		CSR_WRITE_4(sc, BGE_HCC_RX_MAX_COAL_BDS, rx_bds);
   3176 		DELAY(10);
   3177 		junk = CSR_READ_4(sc, BGE_HCC_RX_MAX_COAL_BDS);
   3178 
   3179 		sc->bge_pending_rxintr_change = 0;
   3180 	}
   3181 	bge_handle_events(sc);
   3182 
   3183 	/* Re-enable interrupts. */
   3184 	CSR_WRITE_4(sc, BGE_MBX_IRQ0_LO, 0);
   3185 
   3186 	if (ifp->if_flags & IFF_RUNNING && !IFQ_IS_EMPTY(&ifp->if_snd))
   3187 		bge_start(ifp);
   3188 
   3189 	return (1);
   3190 }
   3191 
   3192 void
   3193 bge_tick(xsc)
   3194 	void *xsc;
   3195 {
   3196 	struct bge_softc *sc = xsc;
   3197 	struct mii_data *mii = &sc->bge_mii;
   3198 	struct ifmedia *ifm = NULL;
   3199 	struct ifnet *ifp = &sc->ethercom.ec_if;
   3200 	int s;
   3201 
   3202 	s = splnet();
   3203 
   3204 	bge_stats_update(sc);
   3205 	callout_reset(&sc->bge_timeout, hz, bge_tick, sc);
   3206 	if (sc->bge_link) {
   3207 		splx(s);
   3208 		return;
   3209 	}
   3210 
   3211 	if (sc->bge_tbi) {
   3212 		ifm = &sc->bge_ifmedia;
   3213 		if (CSR_READ_4(sc, BGE_MAC_STS) &
   3214 		    BGE_MACSTAT_TBI_PCS_SYNCHED) {
   3215 			sc->bge_link++;
   3216 			CSR_WRITE_4(sc, BGE_MAC_STS, 0xFFFFFFFF);
   3217 			if (!IFQ_IS_EMPTY(&ifp->if_snd))
   3218 				bge_start(ifp);
   3219 		}
   3220 		splx(s);
   3221 		return;
   3222 	}
   3223 
   3224 	mii_tick(mii);
   3225 
   3226 	if (!sc->bge_link && mii->mii_media_status & IFM_ACTIVE &&
   3227 	    IFM_SUBTYPE(mii->mii_media_active) != IFM_NONE) {
   3228 		sc->bge_link++;
   3229 		if (!IFQ_IS_EMPTY(&ifp->if_snd))
   3230 			bge_start(ifp);
   3231 	}
   3232 
   3233 	splx(s);
   3234 }
   3235 
   3236 void
   3237 bge_stats_update(sc)
   3238 	struct bge_softc *sc;
   3239 {
   3240 	struct ifnet *ifp = &sc->ethercom.ec_if;
   3241 	bus_size_t stats = BGE_MEMWIN_START + BGE_STATS_BLOCK;
   3242 	bus_size_t rstats = BGE_RX_STATS;
   3243 
   3244 #define READ_RSTAT(sc, stats, stat) \
   3245 	  CSR_READ_4(sc, stats + offsetof(struct bge_mac_stats_regs, stat))
   3246 
   3247 	if (sc->bge_quirks & BGE_QUIRK_5705_CORE) {
   3248 		ifp->if_collisions +=
   3249 		    READ_RSTAT(sc, rstats, dot3StatsSingleCollisionFrames) +
   3250 		    READ_RSTAT(sc, rstats, dot3StatsMultipleCollisionFrames) +
   3251 		    READ_RSTAT(sc, rstats, dot3StatsExcessiveCollisions) +
   3252 		    READ_RSTAT(sc, rstats, dot3StatsLateCollisions);
   3253 
   3254 		BGE_EVCNT_ADD(sc->bge_ev_tx_xoff,
   3255 			      READ_RSTAT(sc, rstats, outXoffSent));
   3256 		BGE_EVCNT_ADD(sc->bge_ev_tx_xon,
   3257 			      READ_RSTAT(sc, rstats, outXonSent));
   3258 		BGE_EVCNT_ADD(sc->bge_ev_rx_xoff,
   3259 			      READ_RSTAT(sc, rstats, xoffPauseFramesReceived));
   3260 		BGE_EVCNT_ADD(sc->bge_ev_rx_xon,
   3261 			      READ_RSTAT(sc, rstats, xonPauseFramesReceived));
   3262 		BGE_EVCNT_ADD(sc->bge_ev_rx_macctl,
   3263 			      READ_RSTAT(sc, rstats, macControlFramesReceived));
   3264 		BGE_EVCNT_ADD(sc->bge_ev_xoffentered,
   3265 			      READ_RSTAT(sc, rstats, xoffStateEntered));
   3266 		return;
   3267 	}
   3268 
   3269 #undef READ_RSTAT
   3270 #define READ_STAT(sc, stats, stat) \
   3271 	  CSR_READ_4(sc, stats + offsetof(struct bge_stats, stat))
   3272 
   3273 	ifp->if_collisions +=
   3274 	  (READ_STAT(sc, stats, dot3StatsSingleCollisionFrames.bge_addr_lo) +
   3275 	   READ_STAT(sc, stats, dot3StatsMultipleCollisionFrames.bge_addr_lo) +
   3276 	   READ_STAT(sc, stats, dot3StatsExcessiveCollisions.bge_addr_lo) +
   3277 	   READ_STAT(sc, stats, dot3StatsLateCollisions.bge_addr_lo)) -
   3278 	  ifp->if_collisions;
   3279 
   3280 	BGE_EVCNT_UPD(sc->bge_ev_tx_xoff,
   3281 		      READ_STAT(sc, stats, outXoffSent.bge_addr_lo));
   3282 	BGE_EVCNT_UPD(sc->bge_ev_tx_xon,
   3283 		      READ_STAT(sc, stats, outXonSent.bge_addr_lo));
   3284 	BGE_EVCNT_UPD(sc->bge_ev_rx_xoff,
   3285 		      READ_STAT(sc, stats,
   3286 		      		xoffPauseFramesReceived.bge_addr_lo));
   3287 	BGE_EVCNT_UPD(sc->bge_ev_rx_xon,
   3288 		      READ_STAT(sc, stats, xonPauseFramesReceived.bge_addr_lo));
   3289 	BGE_EVCNT_UPD(sc->bge_ev_rx_macctl,
   3290 		      READ_STAT(sc, stats,
   3291 		      		macControlFramesReceived.bge_addr_lo));
   3292 	BGE_EVCNT_UPD(sc->bge_ev_xoffentered,
   3293 		      READ_STAT(sc, stats, xoffStateEntered.bge_addr_lo));
   3294 
   3295 #undef READ_STAT
   3296 
   3297 #ifdef notdef
   3298 	ifp->if_collisions +=
   3299 	   (sc->bge_rdata->bge_info.bge_stats.dot3StatsSingleCollisionFrames +
   3300 	   sc->bge_rdata->bge_info.bge_stats.dot3StatsMultipleCollisionFrames +
   3301 	   sc->bge_rdata->bge_info.bge_stats.dot3StatsExcessiveCollisions +
   3302 	   sc->bge_rdata->bge_info.bge_stats.dot3StatsLateCollisions) -
   3303 	   ifp->if_collisions;
   3304 #endif
   3305 }
   3306 
   3307 /*
   3308  * Pad outbound frame to ETHER_MIN_NOPAD for an unusual reason.
   3309  * The bge hardware will pad out Tx runts to ETHER_MIN_NOPAD,
   3310  * but when such padded frames employ the  bge IP/TCP checksum offload,
   3311  * the hardware checksum assist gives incorrect results (possibly
   3312  * from incorporating its own padding into the UDP/TCP checksum; who knows).
   3313  * If we pad such runts with zeros, the onboard checksum comes out correct.
   3314  */
   3315 static __inline int
   3316 bge_cksum_pad(struct mbuf *pkt)
   3317 {
   3318 	struct mbuf *last = NULL;
   3319 	int padlen;
   3320 
   3321 	padlen = ETHER_MIN_NOPAD - pkt->m_pkthdr.len;
   3322 
   3323 	/* if there's only the packet-header and we can pad there, use it. */
   3324 	if (pkt->m_pkthdr.len == pkt->m_len &&
   3325 	    !M_READONLY(pkt) && M_TRAILINGSPACE(pkt) >= padlen) {
   3326 		last = pkt;
   3327 	} else {
   3328 		/*
   3329 		 * Walk packet chain to find last mbuf. We will either
   3330 		 * pad there, or append a new mbuf and pad it
   3331 		 * (thus perhaps avoiding the bcm5700 dma-min bug).
   3332 		 */
   3333 		for (last = pkt; last->m_next != NULL; last = last->m_next) {
   3334 	      	       (void) 0; /* do nothing*/
   3335 		}
   3336 
   3337 		/* `last' now points to last in chain. */
   3338 		if (!M_READONLY(last) && M_TRAILINGSPACE(last) >= padlen) {
   3339 			(void) 0; /* we can pad here, in-place. */
   3340 		} else {
   3341 			/* Allocate new empty mbuf, pad it. Compact later. */
   3342 			struct mbuf *n;
   3343 			MGET(n, M_DONTWAIT, MT_DATA);
   3344 			n->m_len = 0;
   3345 			last->m_next = n;
   3346 			last = n;
   3347 		}
   3348 	}
   3349 
   3350 #ifdef DEBUG
   3351 	  /*KASSERT(M_WRITABLE(last), ("to-pad mbuf not writeable\n"));*/
   3352 	  KASSERT(M_TRAILINGSPACE(last) >= padlen /*, ("insufficient space to pad\n")*/ );
   3353 #endif
   3354 	/* Now zero the pad area, to avoid the bge cksum-assist bug */
   3355 	memset(mtod(last, caddr_t) + last->m_len, 0, padlen);
   3356 	last->m_len += padlen;
   3357 	pkt->m_pkthdr.len += padlen;
   3358 	return 0;
   3359 }
   3360 
   3361 /*
   3362  * Compact outbound packets to avoid bug with DMA segments less than 8 bytes.
   3363  */
   3364 static __inline int
   3365 bge_compact_dma_runt(struct mbuf *pkt)
   3366 {
   3367 	struct mbuf	*m, *prev;
   3368 	int 		totlen, prevlen;
   3369 
   3370 	prev = NULL;
   3371 	totlen = 0;
   3372 	prevlen = -1;
   3373 
   3374 	for (m = pkt; m != NULL; prev = m,m = m->m_next) {
   3375 		int mlen = m->m_len;
   3376 		int shortfall = 8 - mlen ;
   3377 
   3378 		totlen += mlen;
   3379 		if (mlen == 0) {
   3380 			continue;
   3381 		}
   3382 		if (mlen >= 8)
   3383 			continue;
   3384 
   3385 		/* If we get here, mbuf data is too small for DMA engine.
   3386 		 * Try to fix by shuffling data to prev or next in chain.
   3387 		 * If that fails, do a compacting deep-copy of the whole chain.
   3388 		 */
   3389 
   3390 		/* Internal frag. If fits in prev, copy it there. */
   3391 		if (prev && !M_READONLY(prev) &&
   3392 		      M_TRAILINGSPACE(prev) >= m->m_len) {
   3393 		  	bcopy(m->m_data,
   3394 			      prev->m_data+prev->m_len,
   3395 			      mlen);
   3396 			prev->m_len += mlen;
   3397 			m->m_len = 0;
   3398 			/* XXX stitch chain */
   3399 			prev->m_next = m_free(m);
   3400 			m = prev;
   3401 			continue;
   3402 		}
   3403 		else if (m->m_next != NULL && !M_READONLY(m) &&
   3404 			     M_TRAILINGSPACE(m) >= shortfall &&
   3405 			     m->m_next->m_len >= (8 + shortfall)) {
   3406 		    /* m is writable and have enough data in next, pull up. */
   3407 
   3408 		  	bcopy(m->m_next->m_data,
   3409 			      m->m_data+m->m_len,
   3410 			      shortfall);
   3411 			m->m_len += shortfall;
   3412 			m->m_next->m_len -= shortfall;
   3413 			m->m_next->m_data += shortfall;
   3414 		}
   3415 		else if (m->m_next == NULL || 1) {
   3416 		  	/* Got a runt at the very end of the packet.
   3417 			 * borrow data from the tail of the preceding mbuf and
   3418 			 * update its length in-place. (The original data is still
   3419 			 * valid, so we can do this even if prev is not writable.)
   3420 			 */
   3421 
   3422 			/* if we'd make prev a runt, just move all of its data. */
   3423 #ifdef DEBUG
   3424 			KASSERT(prev != NULL /*, ("runt but null PREV")*/);
   3425 			KASSERT(prev->m_len >= 8 /*, ("runt prev")*/);
   3426 #endif
   3427 			if ((prev->m_len - shortfall) < 8)
   3428 				shortfall = prev->m_len;
   3429 
   3430 #ifdef notyet	/* just do the safe slow thing for now */
   3431 			if (!M_READONLY(m)) {
   3432 				if (M_LEADINGSPACE(m) < shorfall) {
   3433 					void *m_dat;
   3434 					m_dat = (m->m_flags & M_PKTHDR) ?
   3435 					  m->m_pktdat : m->dat;
   3436 					memmove(m_dat, mtod(m, void*), m->m_len);
   3437 					m->m_data = m_dat;
   3438 				    }
   3439 			} else
   3440 #endif	/* just do the safe slow thing */
   3441 			{
   3442 				struct mbuf * n = NULL;
   3443 				int newprevlen = prev->m_len - shortfall;
   3444 
   3445 				MGET(n, M_NOWAIT, MT_DATA);
   3446 				if (n == NULL)
   3447 				   return ENOBUFS;
   3448 				KASSERT(m->m_len + shortfall < MLEN
   3449 					/*,
   3450 					  ("runt %d +prev %d too big\n", m->m_len, shortfall)*/);
   3451 
   3452 				/* first copy the data we're stealing from prev */
   3453 				bcopy(prev->m_data + newprevlen, n->m_data, shortfall);
   3454 
   3455 				/* update prev->m_len accordingly */
   3456 				prev->m_len -= shortfall;
   3457 
   3458 				/* copy data from runt m */
   3459 				bcopy(m->m_data, n->m_data + shortfall, m->m_len);
   3460 
   3461 				/* n holds what we stole from prev, plus m */
   3462 				n->m_len = shortfall + m->m_len;
   3463 
   3464 				/* stitch n into chain and free m */
   3465 				n->m_next = m->m_next;
   3466 				prev->m_next = n;
   3467 				/* KASSERT(m->m_next == NULL); */
   3468 				m->m_next = NULL;
   3469 				m_free(m);
   3470 				m = n;	/* for continuing loop */
   3471 			}
   3472 		}
   3473 		prevlen = m->m_len;
   3474 	}
   3475 	return 0;
   3476 }
   3477 
   3478 /*
   3479  * Encapsulate an mbuf chain in the tx ring  by coupling the mbuf data
   3480  * pointers to descriptors.
   3481  */
   3482 int
   3483 bge_encap(sc, m_head, txidx)
   3484 	struct bge_softc *sc;
   3485 	struct mbuf *m_head;
   3486 	u_int32_t *txidx;
   3487 {
   3488 	struct bge_tx_bd	*f = NULL;
   3489 	u_int32_t		frag, cur, cnt = 0;
   3490 	u_int16_t		csum_flags = 0;
   3491 	u_int16_t		txbd_tso_flags = 0;
   3492 	struct txdmamap_pool_entry *dma;
   3493 	bus_dmamap_t dmamap;
   3494 	int			i = 0;
   3495 	struct m_tag		*mtag;
   3496 	int			use_tso, maxsegsize, error;
   3497 
   3498 	cur = frag = *txidx;
   3499 
   3500 	if (m_head->m_pkthdr.csum_flags) {
   3501 		if (m_head->m_pkthdr.csum_flags & M_CSUM_IPv4)
   3502 			csum_flags |= BGE_TXBDFLAG_IP_CSUM;
   3503 		if (m_head->m_pkthdr.csum_flags & (M_CSUM_TCPv4|M_CSUM_UDPv4))
   3504 			csum_flags |= BGE_TXBDFLAG_TCP_UDP_CSUM;
   3505 	}
   3506 
   3507 	/*
   3508 	 * If we were asked to do an outboard checksum, and the NIC
   3509 	 * has the bug where it sometimes adds in the Ethernet padding,
   3510 	 * explicitly pad with zeros so the cksum will be correct either way.
   3511 	 * (For now, do this for all chip versions, until newer
   3512 	 * are confirmed to not require the workaround.)
   3513 	 */
   3514 	if ((csum_flags & BGE_TXBDFLAG_TCP_UDP_CSUM) == 0 ||
   3515 #ifdef notyet
   3516 	    (sc->bge_quirks & BGE_QUIRK_SHORT_CKSUM_BUG) == 0 ||
   3517 #endif
   3518 	    m_head->m_pkthdr.len >= ETHER_MIN_NOPAD)
   3519 		goto check_dma_bug;
   3520 
   3521 	if (bge_cksum_pad(m_head) != 0) {
   3522 	    return ENOBUFS;
   3523 	}
   3524 
   3525 check_dma_bug:
   3526 	if (!(sc->bge_quirks & BGE_QUIRK_5700_SMALLDMA))
   3527 		goto doit;
   3528 	/*
   3529 	 * bcm5700 Revision B silicon cannot handle DMA descriptors with
   3530 	 * less than eight bytes.  If we encounter a teeny mbuf
   3531 	 * at the end of a chain, we can pad.  Otherwise, copy.
   3532 	 */
   3533 	if (bge_compact_dma_runt(m_head) != 0)
   3534 		return ENOBUFS;
   3535 
   3536 doit:
   3537 	dma = SLIST_FIRST(&sc->txdma_list);
   3538 	if (dma == NULL)
   3539 		return ENOBUFS;
   3540 	dmamap = dma->dmamap;
   3541 
   3542 	/*
   3543 	 * Set up any necessary TSO state before we start packing...
   3544 	 */
   3545 	use_tso = (m_head->m_pkthdr.csum_flags & M_CSUM_TSOv4) != 0;
   3546 	if (!use_tso) {
   3547 		maxsegsize = 0;
   3548 	} else {	/* TSO setup */
   3549 		unsigned  mss;
   3550 		struct ether_header *eh;
   3551 		unsigned ip_tcp_hlen, iptcp_opt_words, tcp_seg_flags, offset;
   3552 		struct mbuf * m0 = m_head;
   3553 		struct ip *ip;
   3554 		struct tcphdr *th;
   3555 		int iphl, hlen;
   3556 
   3557 		/*
   3558 		 * XXX It would be nice if the mbuf pkthdr had offset
   3559 		 * fields for the protocol headers.
   3560 		 */
   3561 
   3562 		eh = mtod(m0, struct ether_header *);
   3563 		switch (htons(eh->ether_type)) {
   3564 		case ETHERTYPE_IP:
   3565 			offset = ETHER_HDR_LEN;
   3566 			break;
   3567 
   3568 		case ETHERTYPE_VLAN:
   3569 			offset = ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN;
   3570 			break;
   3571 
   3572 		default:
   3573 			/*
   3574 			 * Don't support this protocol or encapsulation.
   3575 			 */
   3576 			return (ENOBUFS);
   3577 		}
   3578 
   3579 		/*
   3580 		 * TCP/IP headers are in the first mbuf; we can do
   3581 		 * this the easy way.
   3582 		 */
   3583 		iphl = M_CSUM_DATA_IPv4_IPHL(m0->m_pkthdr.csum_data);
   3584 		hlen = iphl + offset;
   3585 		if (__predict_false(m0->m_len <
   3586 				    (hlen + sizeof(struct tcphdr)))) {
   3587 
   3588 			  printf("TSO: hard case m0->m_len == %d <"
   3589 				 " ip/tcp hlen %zd, not handled yet\n",
   3590 				 m0->m_len, hlen+ sizeof(struct tcphdr));
   3591 #ifdef NOTYET
   3592 			/*
   3593 			 * XXX jonathan (at) NetBSD.org: untested.
   3594 			 * how to force  this branch to be taken?
   3595 			 */
   3596 			BGE_EVCNT_INCR(&sc->sc_ev_txtsopain);
   3597 
   3598 			m_copydata(m0, offset, sizeof(ip), &ip);
   3599 			m_copydata(m0, hlen, sizeof(th), &th);
   3600 
   3601 			ip.ip_len = 0;
   3602 
   3603 			m_copyback(m0, hlen + offsetof(struct ip, ip_len),
   3604 			    sizeof(ip.ip_len), &ip.ip_len);
   3605 
   3606 			th.th_sum = in_cksum_phdr(ip.ip_src.s_addr,
   3607 			    ip.ip_dst.s_addr, htons(IPPROTO_TCP));
   3608 
   3609 			m_copyback(m0, hlen + offsetof(struct tcphdr, th_sum),
   3610 			    sizeof(th.th_sum), &th.th_sum);
   3611 
   3612 			hlen += th.th_off << 2;
   3613 			iptcp_opt_words	= hlen;
   3614 #else
   3615 			/*
   3616 			 * if_wm "hard" case not yet supported, can we not
   3617 			 * mandate it out of existence?
   3618 			 */
   3619 			(void) ip; (void)th; (void) ip_tcp_hlen;
   3620 
   3621 			return ENOBUFS;
   3622 #endif
   3623 		} else {
   3624 			ip = (struct ip *) (mtod(m0, caddr_t) + offset);
   3625 			th = (struct tcphdr *) (mtod(m0, caddr_t) + hlen);
   3626 			ip_tcp_hlen = iphl +  (th->th_off << 2);
   3627 
   3628 			/* Total IP/TCP options, in 32-bit words */
   3629 			iptcp_opt_words = (ip_tcp_hlen
   3630 					   - sizeof(struct tcphdr)
   3631 					   - sizeof(struct ip)) >> 2;
   3632 		}
   3633 		if (BGE_IS_5750_OR_BEYOND(sc)) {
   3634 			th->th_sum = 0;
   3635 			csum_flags &= ~(BGE_TXBDFLAG_TCP_UDP_CSUM);
   3636 		} else {
   3637 			/*
   3638 			 * XXX jonathan (at) NetBSD.org: 5705 untested.
   3639 			 * Requires TSO firmware patch for 5701/5703/5704.
   3640 			 */
   3641 			th->th_sum = in_cksum_phdr(ip->ip_src.s_addr,
   3642 			    ip->ip_dst.s_addr, htons(IPPROTO_TCP));
   3643 		}
   3644 
   3645 		mss = m_head->m_pkthdr.segsz;
   3646 		txbd_tso_flags |=
   3647 		    BGE_TXBDFLAG_CPU_PRE_DMA |
   3648 		    BGE_TXBDFLAG_CPU_POST_DMA;
   3649 
   3650 		/*
   3651 		 * Our NIC TSO-assist assumes TSO has standard, optionless
   3652 		 * IPv4 and TCP headers, which total 40 bytes. By default,
   3653 		 * the NIC copies 40 bytes of IP/TCP header from the
   3654 		 * supplied header into the IP/TCP header portion of
   3655 		 * each post-TSO-segment. If the supplied packet has IP or
   3656 		 * TCP options, we need to tell the NIC to copy those extra
   3657 		 * bytes into each  post-TSO header, in addition to the normal
   3658 		 * 40-byte IP/TCP header (and to leave space accordingly).
   3659 		 * Unfortunately, the driver encoding of option length
   3660 		 * varies across different ASIC families.
   3661 		 */
   3662 		tcp_seg_flags = 0;
   3663 		if (iptcp_opt_words) {
   3664 			if ( BGE_IS_5705_OR_BEYOND(sc)) {
   3665 				tcp_seg_flags =
   3666 					iptcp_opt_words << 11;
   3667 			} else {
   3668 				txbd_tso_flags |=
   3669 					iptcp_opt_words << 12;
   3670 			}
   3671 		}
   3672 		maxsegsize = mss | tcp_seg_flags;
   3673 		ip->ip_len = htons(mss + ip_tcp_hlen);
   3674 
   3675 	}	/* TSO setup */
   3676 
   3677 	/*
   3678 	 * Start packing the mbufs in this chain into
   3679 	 * the fragment pointers. Stop when we run out
   3680 	 * of fragments or hit the end of the mbuf chain.
   3681 	 */
   3682 	error = bus_dmamap_load_mbuf(sc->bge_dmatag, dmamap, m_head,
   3683 	    BUS_DMA_NOWAIT);
   3684 	if (error) {
   3685 		return(ENOBUFS);
   3686 	}
   3687 
   3688 	mtag = sc->ethercom.ec_nvlans ?
   3689 	    m_tag_find(m_head, PACKET_TAG_VLAN, NULL) : NULL;
   3690 
   3691 
   3692 	/* Iterate over dmap-map fragments. */
   3693 	for (i = 0; i < dmamap->dm_nsegs; i++) {
   3694 		f = &sc->bge_rdata->bge_tx_ring[frag];
   3695 		if (sc->bge_cdata.bge_tx_chain[frag] != NULL)
   3696 			break;
   3697 
   3698 		bge_set_hostaddr(&f->bge_addr, dmamap->dm_segs[i].ds_addr);
   3699 		f->bge_len = dmamap->dm_segs[i].ds_len;
   3700 
   3701 		/*
   3702 		 * For 5751 and follow-ons, for TSO we must turn
   3703 		 * off checksum-assist flag in the tx-descr, and
   3704 		 * supply the ASIC-revision-specific encoding
   3705 		 * of TSO flags and segsize.
   3706 		 */
   3707 		if (use_tso) {
   3708 			if (BGE_IS_5750_OR_BEYOND(sc) || i == 0) {
   3709 				f->bge_rsvd = maxsegsize;
   3710 				f->bge_flags = csum_flags | txbd_tso_flags;
   3711 			} else {
   3712 				f->bge_rsvd = 0;
   3713 				f->bge_flags =
   3714 				  (csum_flags | txbd_tso_flags) & 0x0fff;
   3715 			}
   3716 		} else {
   3717 			f->bge_rsvd = 0;
   3718 			f->bge_flags = csum_flags;
   3719 		}
   3720 
   3721 		if (mtag != NULL) {
   3722 			f->bge_flags |= BGE_TXBDFLAG_VLAN_TAG;
   3723 			f->bge_vlan_tag = VLAN_TAG_VALUE(mtag);
   3724 		} else {
   3725 			f->bge_vlan_tag = 0;
   3726 		}
   3727 		/*
   3728 		 * Sanity check: avoid coming within 16 descriptors
   3729 		 * of the end of the ring.
   3730 		 */
   3731 		if ((BGE_TX_RING_CNT - (sc->bge_txcnt + cnt)) < 16) {
   3732 			BGE_TSO_PRINTF(("%s: "
   3733 			    " dmamap_load_mbuf too close to ring wrap\n",
   3734 			    sc->bge_dev.dv_xname));
   3735 			return(ENOBUFS);
   3736 		}
   3737 		cur = frag;
   3738 		BGE_INC(frag, BGE_TX_RING_CNT);
   3739 		cnt++;
   3740 	}
   3741 
   3742 	if (i < dmamap->dm_nsegs) {
   3743 		BGE_TSO_PRINTF(("%s: reached %d < dm_nsegs %d\n",
   3744 		    sc->bge_dev.dv_xname, i, dmamap->dm_nsegs));
   3745 		return ENOBUFS;
   3746 	}
   3747 
   3748 	bus_dmamap_sync(sc->bge_dmatag, dmamap, 0, dmamap->dm_mapsize,
   3749 	    BUS_DMASYNC_PREWRITE);
   3750 
   3751 	if (frag == sc->bge_tx_saved_considx) {
   3752 		BGE_TSO_PRINTF(("%s: frag %d = wrapped id %d?\n",
   3753 		    sc->bge_dev.dv_xname, frag, sc->bge_tx_saved_considx));
   3754 
   3755 		return(ENOBUFS);
   3756 	}
   3757 
   3758 	sc->bge_rdata->bge_tx_ring[cur].bge_flags |= BGE_TXBDFLAG_END;
   3759 	sc->bge_cdata.bge_tx_chain[cur] = m_head;
   3760 	SLIST_REMOVE_HEAD(&sc->txdma_list, link);
   3761 	sc->txdma[cur] = dma;
   3762 	sc->bge_txcnt += cnt;
   3763 
   3764 	*txidx = frag;
   3765 
   3766 	return(0);
   3767 }
   3768 
   3769 /*
   3770  * Main transmit routine. To avoid having to do mbuf copies, we put pointers
   3771  * to the mbuf data regions directly in the transmit descriptors.
   3772  */
   3773 void
   3774 bge_start(ifp)
   3775 	struct ifnet *ifp;
   3776 {
   3777 	struct bge_softc *sc;
   3778 	struct mbuf *m_head = NULL;
   3779 	u_int32_t prodidx;
   3780 	int pkts = 0;
   3781 
   3782 	sc = ifp->if_softc;
   3783 
   3784 	if (!sc->bge_link && ifp->if_snd.ifq_len < 10)
   3785 		return;
   3786 
   3787 	prodidx = sc->bge_tx_prodidx;
   3788 
   3789 	while(sc->bge_cdata.bge_tx_chain[prodidx] == NULL) {
   3790 		IFQ_POLL(&ifp->if_snd, m_head);
   3791 		if (m_head == NULL)
   3792 			break;
   3793 
   3794 #if 0
   3795 		/*
   3796 		 * XXX
   3797 		 * safety overkill.  If this is a fragmented packet chain
   3798 		 * with delayed TCP/UDP checksums, then only encapsulate
   3799 		 * it if we have enough descriptors to handle the entire
   3800 		 * chain at once.
   3801 		 * (paranoia -- may not actually be needed)
   3802 		 */
   3803 		if (m_head->m_flags & M_FIRSTFRAG &&
   3804 		    m_head->m_pkthdr.csum_flags & (CSUM_DELAY_DATA)) {
   3805 			if ((BGE_TX_RING_CNT - sc->bge_txcnt) <
   3806 			    M_CSUM_DATA_IPv4_OFFSET(m_head->m_pkthdr.csum_data) + 16) {
   3807 				ifp->if_flags |= IFF_OACTIVE;
   3808 				break;
   3809 			}
   3810 		}
   3811 #endif
   3812 
   3813 		/*
   3814 		 * Pack the data into the transmit ring. If we
   3815 		 * don't have room, set the OACTIVE flag and wait
   3816 		 * for the NIC to drain the ring.
   3817 		 */
   3818 		if (bge_encap(sc, m_head, &prodidx)) {
   3819 			printf("bge: failed on len %d?\n", m_head->m_pkthdr.len);
   3820 			ifp->if_flags |= IFF_OACTIVE;
   3821 			break;
   3822 		}
   3823 
   3824 		/* now we are committed to transmit the packet */
   3825 		IFQ_DEQUEUE(&ifp->if_snd, m_head);
   3826 		pkts++;
   3827 
   3828 #if NBPFILTER > 0
   3829 		/*
   3830 		 * If there's a BPF listener, bounce a copy of this frame
   3831 		 * to him.
   3832 		 */
   3833 		if (ifp->if_bpf)
   3834 			bpf_mtap(ifp->if_bpf, m_head);
   3835 #endif
   3836 	}
   3837 	if (pkts == 0)
   3838 		return;
   3839 
   3840 	/* Transmit */
   3841 	CSR_WRITE_4(sc, BGE_MBX_TX_HOST_PROD0_LO, prodidx);
   3842 	if (sc->bge_quirks & BGE_QUIRK_PRODUCER_BUG)	/* 5700 b2 errata */
   3843 		CSR_WRITE_4(sc, BGE_MBX_TX_HOST_PROD0_LO, prodidx);
   3844 
   3845 	sc->bge_tx_prodidx = prodidx;
   3846 
   3847 	/*
   3848 	 * Set a timeout in case the chip goes out to lunch.
   3849 	 */
   3850 	ifp->if_timer = 5;
   3851 }
   3852 
   3853 int
   3854 bge_init(ifp)
   3855 	struct ifnet *ifp;
   3856 {
   3857 	struct bge_softc *sc = ifp->if_softc;
   3858 	u_int16_t *m;
   3859 	int s, error;
   3860 
   3861 	s = splnet();
   3862 
   3863 	ifp = &sc->ethercom.ec_if;
   3864 
   3865 	/* Cancel pending I/O and flush buffers. */
   3866 	bge_stop(sc);
   3867 	bge_reset(sc);
   3868 	bge_chipinit(sc);
   3869 
   3870 	/*
   3871 	 * Init the various state machines, ring
   3872 	 * control blocks and firmware.
   3873 	 */
   3874 	error = bge_blockinit(sc);
   3875 	if (error != 0) {
   3876 		printf("%s: initialization error %d\n", sc->bge_dev.dv_xname,
   3877 		    error);
   3878 		splx(s);
   3879 		return error;
   3880 	}
   3881 
   3882 	ifp = &sc->ethercom.ec_if;
   3883 
   3884 	/* Specify MTU. */
   3885 	CSR_WRITE_4(sc, BGE_RX_MTU, ifp->if_mtu +
   3886 	    ETHER_HDR_LEN + ETHER_CRC_LEN + ETHER_VLAN_ENCAP_LEN);
   3887 
   3888 	/* Load our MAC address. */
   3889 	m = (u_int16_t *)&(LLADDR(ifp->if_sadl)[0]);
   3890 	CSR_WRITE_4(sc, BGE_MAC_ADDR1_LO, htons(m[0]));
   3891 	CSR_WRITE_4(sc, BGE_MAC_ADDR1_HI, (htons(m[1]) << 16) | htons(m[2]));
   3892 
   3893 	/* Enable or disable promiscuous mode as needed. */
   3894 	if (ifp->if_flags & IFF_PROMISC) {
   3895 		BGE_SETBIT(sc, BGE_RX_MODE, BGE_RXMODE_RX_PROMISC);
   3896 	} else {
   3897 		BGE_CLRBIT(sc, BGE_RX_MODE, BGE_RXMODE_RX_PROMISC);
   3898 	}
   3899 
   3900 	/* Program multicast filter. */
   3901 	bge_setmulti(sc);
   3902 
   3903 	/* Init RX ring. */
   3904 	bge_init_rx_ring_std(sc);
   3905 
   3906 	/* Init jumbo RX ring. */
   3907 	if (ifp->if_mtu > (ETHERMTU + ETHER_HDR_LEN + ETHER_CRC_LEN))
   3908 		bge_init_rx_ring_jumbo(sc);
   3909 
   3910 	/* Init our RX return ring index */
   3911 	sc->bge_rx_saved_considx = 0;
   3912 
   3913 	/* Init TX ring. */
   3914 	bge_init_tx_ring(sc);
   3915 
   3916 	/* Turn on transmitter */
   3917 	BGE_SETBIT(sc, BGE_TX_MODE, BGE_TXMODE_ENABLE);
   3918 
   3919 	/* Turn on receiver */
   3920 	BGE_SETBIT(sc, BGE_RX_MODE, BGE_RXMODE_ENABLE);
   3921 
   3922 	CSR_WRITE_4(sc, BGE_MAX_RX_FRAME_LOWAT, 2);
   3923 
   3924 	/* Tell firmware we're alive. */
   3925 	BGE_SETBIT(sc, BGE_MODE_CTL, BGE_MODECTL_STACKUP);
   3926 
   3927 	/* Enable host interrupts. */
   3928 	BGE_SETBIT(sc, BGE_PCI_MISC_CTL, BGE_PCIMISCCTL_CLEAR_INTA);
   3929 	BGE_CLRBIT(sc, BGE_PCI_MISC_CTL, BGE_PCIMISCCTL_MASK_PCI_INTR);
   3930 	CSR_WRITE_4(sc, BGE_MBX_IRQ0_LO, 0);
   3931 
   3932 	bge_ifmedia_upd(ifp);
   3933 
   3934 	ifp->if_flags |= IFF_RUNNING;
   3935 	ifp->if_flags &= ~IFF_OACTIVE;
   3936 
   3937 	splx(s);
   3938 
   3939 	callout_reset(&sc->bge_timeout, hz, bge_tick, sc);
   3940 
   3941 	return 0;
   3942 }
   3943 
   3944 /*
   3945  * Set media options.
   3946  */
   3947 int
   3948 bge_ifmedia_upd(ifp)
   3949 	struct ifnet *ifp;
   3950 {
   3951 	struct bge_softc *sc = ifp->if_softc;
   3952 	struct mii_data *mii = &sc->bge_mii;
   3953 	struct ifmedia *ifm = &sc->bge_ifmedia;
   3954 
   3955 	/* If this is a 1000baseX NIC, enable the TBI port. */
   3956 	if (sc->bge_tbi) {
   3957 		if (IFM_TYPE(ifm->ifm_media) != IFM_ETHER)
   3958 			return(EINVAL);
   3959 		switch(IFM_SUBTYPE(ifm->ifm_media)) {
   3960 		case IFM_AUTO:
   3961 			break;
   3962 		case IFM_1000_SX:
   3963 			if ((ifm->ifm_media & IFM_GMASK) == IFM_FDX) {
   3964 				BGE_CLRBIT(sc, BGE_MAC_MODE,
   3965 				    BGE_MACMODE_HALF_DUPLEX);
   3966 			} else {
   3967 				BGE_SETBIT(sc, BGE_MAC_MODE,
   3968 				    BGE_MACMODE_HALF_DUPLEX);
   3969 			}
   3970 			break;
   3971 		default:
   3972 			return(EINVAL);
   3973 		}
   3974 		/* XXX 802.3x flow control for 1000BASE-SX */
   3975 		return(0);
   3976 	}
   3977 
   3978 	sc->bge_link = 0;
   3979 	mii_mediachg(mii);
   3980 
   3981 	return(0);
   3982 }
   3983 
   3984 /*
   3985  * Report current media status.
   3986  */
   3987 void
   3988 bge_ifmedia_sts(ifp, ifmr)
   3989 	struct ifnet *ifp;
   3990 	struct ifmediareq *ifmr;
   3991 {
   3992 	struct bge_softc *sc = ifp->if_softc;
   3993 	struct mii_data *mii = &sc->bge_mii;
   3994 
   3995 	if (sc->bge_tbi) {
   3996 		ifmr->ifm_status = IFM_AVALID;
   3997 		ifmr->ifm_active = IFM_ETHER;
   3998 		if (CSR_READ_4(sc, BGE_MAC_STS) &
   3999 		    BGE_MACSTAT_TBI_PCS_SYNCHED)
   4000 			ifmr->ifm_status |= IFM_ACTIVE;
   4001 		ifmr->ifm_active |= IFM_1000_SX;
   4002 		if (CSR_READ_4(sc, BGE_MAC_MODE) & BGE_MACMODE_HALF_DUPLEX)
   4003 			ifmr->ifm_active |= IFM_HDX;
   4004 		else
   4005 			ifmr->ifm_active |= IFM_FDX;
   4006 		return;
   4007 	}
   4008 
   4009 	mii_pollstat(mii);
   4010 	ifmr->ifm_status = mii->mii_media_status;
   4011 	ifmr->ifm_active = (mii->mii_media_active & ~IFM_ETH_FMASK) |
   4012 	    sc->bge_flowflags;
   4013 }
   4014 
   4015 int
   4016 bge_ioctl(ifp, command, data)
   4017 	struct ifnet *ifp;
   4018 	u_long command;
   4019 	caddr_t data;
   4020 {
   4021 	struct bge_softc *sc = ifp->if_softc;
   4022 	struct ifreq *ifr = (struct ifreq *) data;
   4023 	int s, error = 0;
   4024 	struct mii_data *mii;
   4025 
   4026 	s = splnet();
   4027 
   4028 	switch(command) {
   4029 	case SIOCSIFFLAGS:
   4030 		if (ifp->if_flags & IFF_UP) {
   4031 			/*
   4032 			 * If only the state of the PROMISC flag changed,
   4033 			 * then just use the 'set promisc mode' command
   4034 			 * instead of reinitializing the entire NIC. Doing
   4035 			 * a full re-init means reloading the firmware and
   4036 			 * waiting for it to start up, which may take a
   4037 			 * second or two.
   4038 			 */
   4039 			if (ifp->if_flags & IFF_RUNNING &&
   4040 			    ifp->if_flags & IFF_PROMISC &&
   4041 			    !(sc->bge_if_flags & IFF_PROMISC)) {
   4042 				BGE_SETBIT(sc, BGE_RX_MODE,
   4043 				    BGE_RXMODE_RX_PROMISC);
   4044 			} else if (ifp->if_flags & IFF_RUNNING &&
   4045 			    !(ifp->if_flags & IFF_PROMISC) &&
   4046 			    sc->bge_if_flags & IFF_PROMISC) {
   4047 				BGE_CLRBIT(sc, BGE_RX_MODE,
   4048 				    BGE_RXMODE_RX_PROMISC);
   4049 			} else
   4050 				bge_init(ifp);
   4051 		} else {
   4052 			if (ifp->if_flags & IFF_RUNNING) {
   4053 				bge_stop(sc);
   4054 			}
   4055 		}
   4056 		sc->bge_if_flags = ifp->if_flags;
   4057 		error = 0;
   4058 		break;
   4059 	case SIOCSIFMEDIA:
   4060 		/* XXX Flow control is not supported for 1000BASE-SX */
   4061 		if (sc->bge_tbi) {
   4062 			ifr->ifr_media &= ~IFM_ETH_FMASK;
   4063 			sc->bge_flowflags = 0;
   4064 		}
   4065 
   4066 		/* Flow control requires full-duplex mode. */
   4067 		if (IFM_SUBTYPE(ifr->ifr_media) == IFM_AUTO ||
   4068 		    (ifr->ifr_media & IFM_FDX) == 0) {
   4069 		    	ifr->ifr_media &= ~IFM_ETH_FMASK;
   4070 		}
   4071 		if (IFM_SUBTYPE(ifr->ifr_media) != IFM_AUTO) {
   4072 			if ((ifr->ifr_media & IFM_ETH_FMASK) == IFM_FLOW) {
   4073 				/* We an do both TXPAUSE and RXPAUSE. */
   4074 				ifr->ifr_media |=
   4075 				    IFM_ETH_TXPAUSE | IFM_ETH_RXPAUSE;
   4076 			}
   4077 			sc->bge_flowflags = ifr->ifr_media & IFM_ETH_FMASK;
   4078 		}
   4079 		/* FALLTHROUGH */
   4080 	case SIOCGIFMEDIA:
   4081 		if (sc->bge_tbi) {
   4082 			error = ifmedia_ioctl(ifp, ifr, &sc->bge_ifmedia,
   4083 			    command);
   4084 		} else {
   4085 			mii = &sc->bge_mii;
   4086 			error = ifmedia_ioctl(ifp, ifr, &mii->mii_media,
   4087 			    command);
   4088 		}
   4089 		break;
   4090 	default:
   4091 		error = ether_ioctl(ifp, command, data);
   4092 		if (error == ENETRESET) {
   4093 			if (ifp->if_flags & IFF_RUNNING)
   4094 				bge_setmulti(sc);
   4095 			error = 0;
   4096 		}
   4097 		break;
   4098 	}
   4099 
   4100 	splx(s);
   4101 
   4102 	return(error);
   4103 }
   4104 
   4105 void
   4106 bge_watchdog(ifp)
   4107 	struct ifnet *ifp;
   4108 {
   4109 	struct bge_softc *sc;
   4110 
   4111 	sc = ifp->if_softc;
   4112 
   4113 	printf("%s: watchdog timeout -- resetting\n", sc->bge_dev.dv_xname);
   4114 
   4115 	ifp->if_flags &= ~IFF_RUNNING;
   4116 	bge_init(ifp);
   4117 
   4118 	ifp->if_oerrors++;
   4119 }
   4120 
   4121 static void
   4122 bge_stop_block(struct bge_softc *sc, bus_addr_t reg, uint32_t bit)
   4123 {
   4124 	int i;
   4125 
   4126 	BGE_CLRBIT(sc, reg, bit);
   4127 
   4128 	for (i = 0; i < BGE_TIMEOUT; i++) {
   4129 		if ((CSR_READ_4(sc, reg) & bit) == 0)
   4130 			return;
   4131 		delay(100);
   4132 		if (sc->bge_pcie)
   4133 		  DELAY(1000);
   4134 	}
   4135 
   4136 	printf("%s: block failed to stop: reg 0x%lx, bit 0x%08x\n",
   4137 	    sc->bge_dev.dv_xname, (u_long) reg, bit);
   4138 }
   4139 
   4140 /*
   4141  * Stop the adapter and free any mbufs allocated to the
   4142  * RX and TX lists.
   4143  */
   4144 void
   4145 bge_stop(sc)
   4146 	struct bge_softc *sc;
   4147 {
   4148 	struct ifnet *ifp = &sc->ethercom.ec_if;
   4149 
   4150 	callout_stop(&sc->bge_timeout);
   4151 
   4152 	/*
   4153 	 * Disable all of the receiver blocks
   4154 	 */
   4155 	bge_stop_block(sc, BGE_RX_MODE, BGE_RXMODE_ENABLE);
   4156 	bge_stop_block(sc, BGE_RBDI_MODE, BGE_RBDIMODE_ENABLE);
   4157 	bge_stop_block(sc, BGE_RXLP_MODE, BGE_RXLPMODE_ENABLE);
   4158 	if ((sc->bge_quirks & BGE_QUIRK_5705_CORE) == 0) {
   4159 		bge_stop_block(sc, BGE_RXLS_MODE, BGE_RXLSMODE_ENABLE);
   4160 	}
   4161 	bge_stop_block(sc, BGE_RDBDI_MODE, BGE_RBDIMODE_ENABLE);
   4162 	bge_stop_block(sc, BGE_RDC_MODE, BGE_RDCMODE_ENABLE);
   4163 	bge_stop_block(sc, BGE_RBDC_MODE, BGE_RBDCMODE_ENABLE);
   4164 
   4165 	/*
   4166 	 * Disable all of the transmit blocks
   4167 	 */
   4168 	bge_stop_block(sc, BGE_SRS_MODE, BGE_SRSMODE_ENABLE);
   4169 	bge_stop_block(sc, BGE_SBDI_MODE, BGE_SBDIMODE_ENABLE);
   4170 	bge_stop_block(sc, BGE_SDI_MODE, BGE_SDIMODE_ENABLE);
   4171 	bge_stop_block(sc, BGE_RDMA_MODE, BGE_RDMAMODE_ENABLE);
   4172 	bge_stop_block(sc, BGE_SDC_MODE, BGE_SDCMODE_ENABLE);
   4173 	if ((sc->bge_quirks & BGE_QUIRK_5705_CORE) == 0) {
   4174 		bge_stop_block(sc, BGE_DMAC_MODE, BGE_DMACMODE_ENABLE);
   4175 	}
   4176 	bge_stop_block(sc, BGE_SBDC_MODE, BGE_SBDCMODE_ENABLE);
   4177 
   4178 	/*
   4179 	 * Shut down all of the memory managers and related
   4180 	 * state machines.
   4181 	 */
   4182 	bge_stop_block(sc, BGE_HCC_MODE, BGE_HCCMODE_ENABLE);
   4183 	bge_stop_block(sc, BGE_WDMA_MODE, BGE_WDMAMODE_ENABLE);
   4184 	if ((sc->bge_quirks & BGE_QUIRK_5705_CORE) == 0) {
   4185 		bge_stop_block(sc, BGE_MBCF_MODE, BGE_MBCFMODE_ENABLE);
   4186 	}
   4187 
   4188 	CSR_WRITE_4(sc, BGE_FTQ_RESET, 0xFFFFFFFF);
   4189 	CSR_WRITE_4(sc, BGE_FTQ_RESET, 0);
   4190 
   4191 	if ((sc->bge_quirks & BGE_QUIRK_5705_CORE) == 0) {
   4192 		bge_stop_block(sc, BGE_BMAN_MODE, BGE_BMANMODE_ENABLE);
   4193 		bge_stop_block(sc, BGE_MARB_MODE, BGE_MARBMODE_ENABLE);
   4194 	}
   4195 
   4196 	/* Disable host interrupts. */
   4197 	BGE_SETBIT(sc, BGE_PCI_MISC_CTL, BGE_PCIMISCCTL_MASK_PCI_INTR);
   4198 	CSR_WRITE_4(sc, BGE_MBX_IRQ0_LO, 1);
   4199 
   4200 	/*
   4201 	 * Tell firmware we're shutting down.
   4202 	 */
   4203 	BGE_CLRBIT(sc, BGE_MODE_CTL, BGE_MODECTL_STACKUP);
   4204 
   4205 	/* Free the RX lists. */
   4206 	bge_free_rx_ring_std(sc);
   4207 
   4208 	/* Free jumbo RX list. */
   4209 	bge_free_rx_ring_jumbo(sc);
   4210 
   4211 	/* Free TX buffers. */
   4212 	bge_free_tx_ring(sc);
   4213 
   4214 	/*
   4215 	 * Isolate/power down the PHY.
   4216 	 */
   4217 	if (!sc->bge_tbi)
   4218 		mii_down(&sc->bge_mii);
   4219 
   4220 	sc->bge_link = 0;
   4221 
   4222 	sc->bge_tx_saved_considx = BGE_TXCONS_UNSET;
   4223 
   4224 	ifp->if_flags &= ~(IFF_RUNNING | IFF_OACTIVE);
   4225 }
   4226 
   4227 /*
   4228  * Stop all chip I/O so that the kernel's probe routines don't
   4229  * get confused by errant DMAs when rebooting.
   4230  */
   4231 void
   4232 bge_shutdown(xsc)
   4233 	void *xsc;
   4234 {
   4235 	struct bge_softc *sc = (struct bge_softc *)xsc;
   4236 
   4237 	bge_stop(sc);
   4238 	bge_reset(sc);
   4239 }
   4240 
   4241 
   4242 static int
   4243 sysctl_bge_verify(SYSCTLFN_ARGS)
   4244 {
   4245 	int error, t;
   4246 	struct sysctlnode node;
   4247 
   4248 	node = *rnode;
   4249 	t = *(int*)rnode->sysctl_data;
   4250 	node.sysctl_data = &t;
   4251 	error = sysctl_lookup(SYSCTLFN_CALL(&node));
   4252 	if (error || newp == NULL)
   4253 		return (error);
   4254 
   4255 #if 0
   4256 	DPRINTF2(("%s: t = %d, nodenum = %d, rnodenum = %d\n", __func__, t,
   4257 	    node.sysctl_num, rnode->sysctl_num));
   4258 #endif
   4259 
   4260 	if (node.sysctl_num == bge_rxthresh_nodenum) {
   4261 		if (t < 0 || t >= NBGE_RX_THRESH)
   4262 			return (EINVAL);
   4263 		bge_update_all_threshes(t);
   4264 	} else
   4265 		return (EINVAL);
   4266 
   4267 	*(int*)rnode->sysctl_data = t;
   4268 
   4269 	return (0);
   4270 }
   4271 
   4272 /*
   4273  * Set up sysctl(3) MIB, hw.bge.*.
   4274  *
   4275  * TBD condition SYSCTL_PERMANENT on being an LKM or not
   4276  */
   4277 SYSCTL_SETUP(sysctl_bge, "sysctl bge subtree setup")
   4278 {
   4279 	int rc, bge_root_num;
   4280 	const struct sysctlnode *node;
   4281 
   4282 	if ((rc = sysctl_createv(clog, 0, NULL, NULL,
   4283 	    CTLFLAG_PERMANENT, CTLTYPE_NODE, "hw", NULL,
   4284 	    NULL, 0, NULL, 0, CTL_HW, CTL_EOL)) != 0) {
   4285 		goto err;
   4286 	}
   4287 
   4288 	if ((rc = sysctl_createv(clog, 0, NULL, &node,
   4289 	    CTLFLAG_PERMANENT, CTLTYPE_NODE, "bge",
   4290 	    SYSCTL_DESCR("BGE interface controls"),
   4291 	    NULL, 0, NULL, 0, CTL_HW, CTL_CREATE, CTL_EOL)) != 0) {
   4292 		goto err;
   4293 	}
   4294 
   4295 	bge_root_num = node->sysctl_num;
   4296 
   4297 	/* BGE Rx interrupt mitigation level */
   4298 	if ((rc = sysctl_createv(clog, 0, NULL, &node,
   4299 	    CTLFLAG_PERMANENT|CTLFLAG_READWRITE,
   4300 	    CTLTYPE_INT, "rx_lvl",
   4301 	    SYSCTL_DESCR("BGE receive interrupt mitigation level"),
   4302 	    sysctl_bge_verify, 0,
   4303 	    &bge_rx_thresh_lvl,
   4304 	    0, CTL_HW, bge_root_num, CTL_CREATE,
   4305 	    CTL_EOL)) != 0) {
   4306 		goto err;
   4307 	}
   4308 
   4309 	bge_rxthresh_nodenum = node->sysctl_num;
   4310 
   4311 	return;
   4312 
   4313 err:
   4314 	printf("%s: sysctl_createv failed (rc = %d)\n", __func__, rc);
   4315 }
   4316 
   4317 void
   4318 bge_powerhook(int why, void *hdl)
   4319 {
   4320 	struct bge_softc *sc = (struct bge_softc *)hdl;
   4321 	struct ifnet *ifp = &sc->ethercom.ec_if;
   4322 	struct pci_attach_args *pa = &(sc->bge_pa);
   4323 	pci_chipset_tag_t pc = pa->pa_pc;
   4324 	pcitag_t tag = pa->pa_tag;
   4325 
   4326 	switch (why) {
   4327 	case PWR_SOFTSUSPEND:
   4328 	case PWR_SOFTSTANDBY:
   4329 		bge_shutdown(sc);
   4330 		break;
   4331 	case PWR_SOFTRESUME:
   4332 		if (ifp->if_flags & IFF_UP) {
   4333 			ifp->if_flags &= ~IFF_RUNNING;
   4334 			bge_init(ifp);
   4335 		}
   4336 		break;
   4337 	case PWR_SUSPEND:
   4338 	case PWR_STANDBY:
   4339 		pci_conf_capture(pc, tag, &sc->bge_pciconf);
   4340 		break;
   4341 	case PWR_RESUME:
   4342 		pci_conf_restore(pc, tag, &sc->bge_pciconf);
   4343 		break;
   4344 	}
   4345 
   4346 	return;
   4347 }
   4348