Home | History | Annotate | Line # | Download | only in pci
if_bge.c revision 1.133
      1 /*	$NetBSD: if_bge.c,v 1.133 2007/08/06 12:29:36 markd 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 1.133 2007/08/06 12:29:36 markd 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 static const 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 static int	bge_rxthresh_nodenum;
    183 
    184 static int	bge_probe(device_t, cfdata_t, void *);
    185 static void	bge_attach(device_t, device_t, void *);
    186 static void	bge_powerhook(int, void *);
    187 static void	bge_release_resources(struct bge_softc *);
    188 static void	bge_txeof(struct bge_softc *);
    189 static void	bge_rxeof(struct bge_softc *);
    190 
    191 static void	bge_tick(void *);
    192 static void	bge_stats_update(struct bge_softc *);
    193 static int	bge_encap(struct bge_softc *, struct mbuf *, u_int32_t *);
    194 
    195 static int	bge_intr(void *);
    196 static void	bge_start(struct ifnet *);
    197 static int	bge_ioctl(struct ifnet *, u_long, void *);
    198 static int	bge_init(struct ifnet *);
    199 static void	bge_stop(struct bge_softc *);
    200 static void	bge_watchdog(struct ifnet *);
    201 static void	bge_shutdown(void *);
    202 static int	bge_ifmedia_upd(struct ifnet *);
    203 static void	bge_ifmedia_sts(struct ifnet *, struct ifmediareq *);
    204 
    205 static void	bge_setmulti(struct bge_softc *);
    206 
    207 static void	bge_handle_events(struct bge_softc *);
    208 static int	bge_alloc_jumbo_mem(struct bge_softc *);
    209 #if 0 /* XXX */
    210 static void	bge_free_jumbo_mem(struct bge_softc *);
    211 #endif
    212 static void	*bge_jalloc(struct bge_softc *);
    213 static void	bge_jfree(struct mbuf *, void *, size_t, void *);
    214 static int	bge_newbuf_std(struct bge_softc *, int, struct mbuf *,
    215 			       bus_dmamap_t);
    216 static int	bge_newbuf_jumbo(struct bge_softc *, int, struct mbuf *);
    217 static int	bge_init_rx_ring_std(struct bge_softc *);
    218 static void	bge_free_rx_ring_std(struct bge_softc *);
    219 static int	bge_init_rx_ring_jumbo(struct bge_softc *);
    220 static void	bge_free_rx_ring_jumbo(struct bge_softc *);
    221 static void	bge_free_tx_ring(struct bge_softc *);
    222 static int	bge_init_tx_ring(struct bge_softc *);
    223 
    224 static int	bge_chipinit(struct bge_softc *);
    225 static int	bge_blockinit(struct bge_softc *);
    226 static int	bge_setpowerstate(struct bge_softc *, int);
    227 
    228 static void	bge_reset(struct bge_softc *);
    229 
    230 #define BGE_DEBUG
    231 #ifdef BGE_DEBUG
    232 #define DPRINTF(x)	if (bgedebug) printf x
    233 #define DPRINTFN(n,x)	if (bgedebug >= (n)) printf x
    234 #define BGE_TSO_PRINTF(x)  do { if (bge_tso_debug) printf x ;} while (0)
    235 int	bgedebug = 0;
    236 int	bge_tso_debug = 0;
    237 #else
    238 #define DPRINTF(x)
    239 #define DPRINTFN(n,x)
    240 #define BGE_TSO_PRINTF(x)
    241 #endif
    242 
    243 #ifdef BGE_EVENT_COUNTERS
    244 #define	BGE_EVCNT_INCR(ev)	(ev).ev_count++
    245 #define	BGE_EVCNT_ADD(ev, val)	(ev).ev_count += (val)
    246 #define	BGE_EVCNT_UPD(ev, val)	(ev).ev_count = (val)
    247 #else
    248 #define	BGE_EVCNT_INCR(ev)	/* nothing */
    249 #define	BGE_EVCNT_ADD(ev, val)	/* nothing */
    250 #define	BGE_EVCNT_UPD(ev, val)	/* nothing */
    251 #endif
    252 
    253 /* Various chip quirks. */
    254 #define	BGE_QUIRK_LINK_STATE_BROKEN	0x00000001
    255 #define	BGE_QUIRK_CSUM_BROKEN		0x00000002
    256 #define	BGE_QUIRK_ONLY_PHY_1		0x00000004
    257 #define	BGE_QUIRK_5700_SMALLDMA		0x00000008
    258 #define	BGE_QUIRK_5700_PCIX_REG_BUG	0x00000010
    259 #define	BGE_QUIRK_PRODUCER_BUG		0x00000020
    260 #define	BGE_QUIRK_PCIX_DMA_ALIGN_BUG	0x00000040
    261 #define	BGE_QUIRK_5705_CORE		0x00000080
    262 #define	BGE_QUIRK_FEWER_MBUFS		0x00000100
    263 
    264 /*
    265  * XXX: how to handle variants based on 5750 and derivatives:
    266  * 5750 5751, 5721, possibly 5714, 5752, and 5708?, which
    267  * in general behave like a 5705, except with additional quirks.
    268  * This driver's current handling of the 5721 is wrong;
    269  * how we map ASIC revision to "quirks" needs more thought.
    270  * (defined here until the thought is done).
    271  */
    272 #define BGE_IS_5714_FAMILY(sc) \
    273 	(BGE_ASICREV(sc->bge_chipid) == BGE_ASICREV_BCM5714_A0 || \
    274 	 BGE_ASICREV(sc->bge_chipid) == BGE_ASICREV_BCM5780 ||	\
    275 	 BGE_ASICREV(sc->bge_chipid) == BGE_ASICREV_BCM5714 )
    276 
    277 #define BGE_IS_5750_OR_BEYOND(sc)  \
    278 	(BGE_ASICREV(sc->bge_chipid) == BGE_ASICREV_BCM5750 || \
    279 	 BGE_ASICREV(sc->bge_chipid) == BGE_ASICREV_BCM5752 || \
    280 	 BGE_ASICREV(sc->bge_chipid) == BGE_ASICREV_BCM5755 || \
    281 	 BGE_ASICREV(sc->bge_chipid) == BGE_ASICREV_BCM5787 || \
    282 	 BGE_IS_5714_FAMILY(sc) )
    283 
    284 #define BGE_IS_5705_OR_BEYOND(sc)  \
    285 	( ((sc)->bge_quirks & BGE_QUIRK_5705_CORE) || \
    286 	  BGE_IS_5750_OR_BEYOND(sc) )
    287 
    288 
    289 /* following bugs are common to bcm5700 rev B, all flavours */
    290 #define BGE_QUIRK_5700_COMMON \
    291 	(BGE_QUIRK_5700_SMALLDMA|BGE_QUIRK_PRODUCER_BUG)
    292 
    293 CFATTACH_DECL(bge, sizeof(struct bge_softc),
    294     bge_probe, bge_attach, NULL, NULL);
    295 
    296 static u_int32_t
    297 bge_readmem_ind(struct bge_softc *sc, int off)
    298 {
    299 	struct pci_attach_args	*pa = &(sc->bge_pa);
    300 	pcireg_t val;
    301 
    302 	pci_conf_write(pa->pa_pc, pa->pa_tag, BGE_PCI_MEMWIN_BASEADDR, off);
    303 	val = pci_conf_read(pa->pa_pc, pa->pa_tag, BGE_PCI_MEMWIN_DATA);
    304 	return val;
    305 }
    306 
    307 static void
    308 bge_writemem_ind(struct bge_softc *sc, int off, int val)
    309 {
    310 	struct pci_attach_args	*pa = &(sc->bge_pa);
    311 
    312 	pci_conf_write(pa->pa_pc, pa->pa_tag, BGE_PCI_MEMWIN_BASEADDR, off);
    313 	pci_conf_write(pa->pa_pc, pa->pa_tag, BGE_PCI_MEMWIN_DATA, val);
    314 }
    315 
    316 #ifdef notdef
    317 static u_int32_t
    318 bge_readreg_ind(struct bge_softc *sc, int off)
    319 {
    320 	struct pci_attach_args	*pa = &(sc->bge_pa);
    321 
    322 	pci_conf_write(pa->pa_pc, pa->pa_tag, BGE_PCI_REG_BASEADDR, off);
    323 	return(pci_conf_read(pa->pa_pc, pa->pa_tag, BGE_PCI_REG_DATA));
    324 }
    325 #endif
    326 
    327 static void
    328 bge_writereg_ind(struct bge_softc *sc, int off, int val)
    329 {
    330 	struct pci_attach_args	*pa = &(sc->bge_pa);
    331 
    332 	pci_conf_write(pa->pa_pc, pa->pa_tag, BGE_PCI_REG_BASEADDR, off);
    333 	pci_conf_write(pa->pa_pc, pa->pa_tag, BGE_PCI_REG_DATA, val);
    334 }
    335 
    336 #ifdef notdef
    337 static u_int8_t
    338 bge_vpd_readbyte(struct bge_softc *sc, int addr)
    339 {
    340 	int i;
    341 	u_int32_t val;
    342 	struct pci_attach_args	*pa = &(sc->bge_pa);
    343 
    344 	pci_conf_write(pa->pa_pc, pa->pa_tag, BGE_PCI_VPD_ADDR, addr);
    345 	for (i = 0; i < BGE_TIMEOUT * 10; i++) {
    346 		DELAY(10);
    347 		if (pci_conf_read(pa->pa_pc, pa->pa_tag, BGE_PCI_VPD_ADDR) &
    348 		    BGE_VPD_FLAG)
    349 			break;
    350 	}
    351 
    352 	if (i == BGE_TIMEOUT) {
    353 		printf("%s: VPD read timed out\n", sc->bge_dev.dv_xname);
    354 		return(0);
    355 	}
    356 
    357 	val = pci_conf_read(pa->pa_pc, pa->pa_tag, BGE_PCI_VPD_DATA);
    358 
    359 	return((val >> ((addr % 4) * 8)) & 0xFF);
    360 }
    361 
    362 static void
    363 bge_vpd_read_res(struct bge_softc *sc, struct vpd_res *res, int addr)
    364 {
    365 	int i;
    366 	u_int8_t *ptr;
    367 
    368 	ptr = (u_int8_t *)res;
    369 	for (i = 0; i < sizeof(struct vpd_res); i++)
    370 		ptr[i] = bge_vpd_readbyte(sc, i + addr);
    371 }
    372 
    373 static void
    374 bge_vpd_read(struct bge_softc *sc)
    375 {
    376 	int pos = 0, i;
    377 	struct vpd_res res;
    378 
    379 	if (sc->bge_vpd_prodname != NULL)
    380 		free(sc->bge_vpd_prodname, M_DEVBUF);
    381 	if (sc->bge_vpd_readonly != NULL)
    382 		free(sc->bge_vpd_readonly, M_DEVBUF);
    383 	sc->bge_vpd_prodname = NULL;
    384 	sc->bge_vpd_readonly = NULL;
    385 
    386 	bge_vpd_read_res(sc, &res, pos);
    387 
    388 	if (res.vr_id != VPD_RES_ID) {
    389 		printf("%s: bad VPD resource id: expected %x got %x\n",
    390 			sc->bge_dev.dv_xname, VPD_RES_ID, res.vr_id);
    391 		return;
    392 	}
    393 
    394 	pos += sizeof(res);
    395 	sc->bge_vpd_prodname = malloc(res.vr_len + 1, M_DEVBUF, M_NOWAIT);
    396 	if (sc->bge_vpd_prodname == NULL)
    397 		panic("bge_vpd_read");
    398 	for (i = 0; i < res.vr_len; i++)
    399 		sc->bge_vpd_prodname[i] = bge_vpd_readbyte(sc, i + pos);
    400 	sc->bge_vpd_prodname[i] = '\0';
    401 	pos += i;
    402 
    403 	bge_vpd_read_res(sc, &res, pos);
    404 
    405 	if (res.vr_id != VPD_RES_READ) {
    406 		printf("%s: bad VPD resource id: expected %x got %x\n",
    407 		    sc->bge_dev.dv_xname, VPD_RES_READ, res.vr_id);
    408 		return;
    409 	}
    410 
    411 	pos += sizeof(res);
    412 	sc->bge_vpd_readonly = malloc(res.vr_len, M_DEVBUF, M_NOWAIT);
    413 	if (sc->bge_vpd_readonly == NULL)
    414 		panic("bge_vpd_read");
    415 	for (i = 0; i < res.vr_len + 1; i++)
    416 		sc->bge_vpd_readonly[i] = bge_vpd_readbyte(sc, i + pos);
    417 }
    418 #endif
    419 
    420 /*
    421  * Read a byte of data stored in the EEPROM at address 'addr.' The
    422  * BCM570x supports both the traditional bitbang interface and an
    423  * auto access interface for reading the EEPROM. We use the auto
    424  * access method.
    425  */
    426 static u_int8_t
    427 bge_eeprom_getbyte(struct bge_softc *sc, int addr, u_int8_t *dest)
    428 {
    429 	int i;
    430 	u_int32_t byte = 0;
    431 
    432 	/*
    433 	 * Enable use of auto EEPROM access so we can avoid
    434 	 * having to use the bitbang method.
    435 	 */
    436 	BGE_SETBIT(sc, BGE_MISC_LOCAL_CTL, BGE_MLC_AUTO_EEPROM);
    437 
    438 	/* Reset the EEPROM, load the clock period. */
    439 	CSR_WRITE_4(sc, BGE_EE_ADDR,
    440 	    BGE_EEADDR_RESET|BGE_EEHALFCLK(BGE_HALFCLK_384SCL));
    441 	DELAY(20);
    442 
    443 	/* Issue the read EEPROM command. */
    444 	CSR_WRITE_4(sc, BGE_EE_ADDR, BGE_EE_READCMD | addr);
    445 
    446 	/* Wait for completion */
    447 	for(i = 0; i < BGE_TIMEOUT * 10; i++) {
    448 		DELAY(10);
    449 		if (CSR_READ_4(sc, BGE_EE_ADDR) & BGE_EEADDR_DONE)
    450 			break;
    451 	}
    452 
    453 	if (i == BGE_TIMEOUT) {
    454 		printf("%s: eeprom read timed out\n", sc->bge_dev.dv_xname);
    455 		return(0);
    456 	}
    457 
    458 	/* Get result. */
    459 	byte = CSR_READ_4(sc, BGE_EE_DATA);
    460 
    461 	*dest = (byte >> ((addr % 4) * 8)) & 0xFF;
    462 
    463 	return(0);
    464 }
    465 
    466 /*
    467  * Read a sequence of bytes from the EEPROM.
    468  */
    469 static int
    470 bge_read_eeprom(struct bge_softc *sc, void *destv, int off, int cnt)
    471 {
    472 	int err = 0, i;
    473 	u_int8_t byte = 0;
    474 	char *dest = destv;
    475 
    476 	for (i = 0; i < cnt; i++) {
    477 		err = bge_eeprom_getbyte(sc, off + i, &byte);
    478 		if (err)
    479 			break;
    480 		*(dest + i) = byte;
    481 	}
    482 
    483 	return(err ? 1 : 0);
    484 }
    485 
    486 static int
    487 bge_miibus_readreg(device_t dev, int phy, int reg)
    488 {
    489 	struct bge_softc *sc = (struct bge_softc *)dev;
    490 	u_int32_t val;
    491 	u_int32_t saved_autopoll;
    492 	int i;
    493 
    494 	/*
    495 	 * Several chips with builtin PHYs will incorrectly answer to
    496 	 * other PHY instances than the builtin PHY at id 1.
    497 	 */
    498 	if (phy != 1 && (sc->bge_quirks & BGE_QUIRK_ONLY_PHY_1))
    499 		return(0);
    500 
    501 	/* Reading with autopolling on may trigger PCI errors */
    502 	saved_autopoll = CSR_READ_4(sc, BGE_MI_MODE);
    503 	if (saved_autopoll & BGE_MIMODE_AUTOPOLL) {
    504 		CSR_WRITE_4(sc, BGE_MI_MODE,
    505 		    saved_autopoll &~ BGE_MIMODE_AUTOPOLL);
    506 		DELAY(40);
    507 	}
    508 
    509 	CSR_WRITE_4(sc, BGE_MI_COMM, BGE_MICMD_READ|BGE_MICOMM_BUSY|
    510 	    BGE_MIPHY(phy)|BGE_MIREG(reg));
    511 
    512 	for (i = 0; i < BGE_TIMEOUT; i++) {
    513 		val = CSR_READ_4(sc, BGE_MI_COMM);
    514 		if (!(val & BGE_MICOMM_BUSY))
    515 			break;
    516 		delay(10);
    517 	}
    518 
    519 	if (i == BGE_TIMEOUT) {
    520 		printf("%s: PHY read timed out\n", sc->bge_dev.dv_xname);
    521 		val = 0;
    522 		goto done;
    523 	}
    524 
    525 	val = CSR_READ_4(sc, BGE_MI_COMM);
    526 
    527 done:
    528 	if (saved_autopoll & BGE_MIMODE_AUTOPOLL) {
    529 		CSR_WRITE_4(sc, BGE_MI_MODE, saved_autopoll);
    530 		DELAY(40);
    531 	}
    532 
    533 	if (val & BGE_MICOMM_READFAIL)
    534 		return(0);
    535 
    536 	return(val & 0xFFFF);
    537 }
    538 
    539 static void
    540 bge_miibus_writereg(device_t dev, int phy, int reg, int val)
    541 {
    542 	struct bge_softc *sc = (struct bge_softc *)dev;
    543 	u_int32_t saved_autopoll;
    544 	int i;
    545 
    546 	/* Touching the PHY while autopolling is on may trigger PCI errors */
    547 	saved_autopoll = CSR_READ_4(sc, BGE_MI_MODE);
    548 	if (saved_autopoll & BGE_MIMODE_AUTOPOLL) {
    549 		delay(40);
    550 		CSR_WRITE_4(sc, BGE_MI_MODE,
    551 		    saved_autopoll & (~BGE_MIMODE_AUTOPOLL));
    552 		delay(10); /* 40 usec is supposed to be adequate */
    553 	}
    554 
    555 	CSR_WRITE_4(sc, BGE_MI_COMM, BGE_MICMD_WRITE|BGE_MICOMM_BUSY|
    556 	    BGE_MIPHY(phy)|BGE_MIREG(reg)|val);
    557 
    558 	for (i = 0; i < BGE_TIMEOUT; i++) {
    559 		if (!(CSR_READ_4(sc, BGE_MI_COMM) & BGE_MICOMM_BUSY))
    560 			break;
    561 		delay(10);
    562 	}
    563 
    564 	if (saved_autopoll & BGE_MIMODE_AUTOPOLL) {
    565 		CSR_WRITE_4(sc, BGE_MI_MODE, saved_autopoll);
    566 		delay(40);
    567 	}
    568 
    569 	if (i == BGE_TIMEOUT) {
    570 		printf("%s: PHY read timed out\n", sc->bge_dev.dv_xname);
    571 	}
    572 }
    573 
    574 static void
    575 bge_miibus_statchg(device_t dev)
    576 {
    577 	struct bge_softc *sc = (struct bge_softc *)dev;
    578 	struct mii_data *mii = &sc->bge_mii;
    579 
    580 	/*
    581 	 * Get flow control negotiation result.
    582 	 */
    583 	if (IFM_SUBTYPE(mii->mii_media.ifm_cur->ifm_media) == IFM_AUTO &&
    584 	    (mii->mii_media_active & IFM_ETH_FMASK) != sc->bge_flowflags) {
    585 		sc->bge_flowflags = mii->mii_media_active & IFM_ETH_FMASK;
    586 		mii->mii_media_active &= ~IFM_ETH_FMASK;
    587 	}
    588 
    589 	BGE_CLRBIT(sc, BGE_MAC_MODE, BGE_MACMODE_PORTMODE);
    590 	if (IFM_SUBTYPE(mii->mii_media_active) == IFM_1000_T) {
    591 		BGE_SETBIT(sc, BGE_MAC_MODE, BGE_PORTMODE_GMII);
    592 	} else {
    593 		BGE_SETBIT(sc, BGE_MAC_MODE, BGE_PORTMODE_MII);
    594 	}
    595 
    596 	if ((mii->mii_media_active & IFM_GMASK) == IFM_FDX) {
    597 		BGE_CLRBIT(sc, BGE_MAC_MODE, BGE_MACMODE_HALF_DUPLEX);
    598 	} else {
    599 		BGE_SETBIT(sc, BGE_MAC_MODE, BGE_MACMODE_HALF_DUPLEX);
    600 	}
    601 
    602 	/*
    603 	 * 802.3x flow control
    604 	 */
    605 	if (sc->bge_flowflags & IFM_ETH_RXPAUSE) {
    606 		BGE_SETBIT(sc, BGE_RX_MODE, BGE_RXMODE_FLOWCTL_ENABLE);
    607 	} else {
    608 		BGE_CLRBIT(sc, BGE_RX_MODE, BGE_RXMODE_FLOWCTL_ENABLE);
    609 	}
    610 	if (sc->bge_flowflags & IFM_ETH_TXPAUSE) {
    611 		BGE_SETBIT(sc, BGE_TX_MODE, BGE_TXMODE_FLOWCTL_ENABLE);
    612 	} else {
    613 		BGE_CLRBIT(sc, BGE_TX_MODE, BGE_TXMODE_FLOWCTL_ENABLE);
    614 	}
    615 }
    616 
    617 /*
    618  * Update rx threshold levels to values in a particular slot
    619  * of the interrupt-mitigation table bge_rx_threshes.
    620  */
    621 static void
    622 bge_set_thresh(struct ifnet *ifp, int lvl)
    623 {
    624 	struct bge_softc *sc = ifp->if_softc;
    625 	int s;
    626 
    627 	/* For now, just save the new Rx-intr thresholds and record
    628 	 * that a threshold update is pending.  Updating the hardware
    629 	 * registers here (even at splhigh()) is observed to
    630 	 * occasionaly cause glitches where Rx-interrupts are not
    631 	 * honoured for up to 10 seconds. jonathan (at) NetBSD.org, 2003-04-05
    632 	 */
    633 	s = splnet();
    634 	sc->bge_rx_coal_ticks = bge_rx_threshes[lvl].rx_ticks;
    635 	sc->bge_rx_max_coal_bds = bge_rx_threshes[lvl].rx_max_bds;
    636 	sc->bge_pending_rxintr_change = 1;
    637 	splx(s);
    638 
    639 	 return;
    640 }
    641 
    642 
    643 /*
    644  * Update Rx thresholds of all bge devices
    645  */
    646 static void
    647 bge_update_all_threshes(int lvl)
    648 {
    649 	struct ifnet *ifp;
    650 	const char * const namebuf = "bge";
    651 	int namelen;
    652 
    653 	if (lvl < 0)
    654 		lvl = 0;
    655 	else if( lvl >= NBGE_RX_THRESH)
    656 		lvl = NBGE_RX_THRESH - 1;
    657 
    658 	namelen = strlen(namebuf);
    659 	/*
    660 	 * Now search all the interfaces for this name/number
    661 	 */
    662 	IFNET_FOREACH(ifp) {
    663 		if (strncmp(ifp->if_xname, namebuf, namelen) != 0)
    664 		      continue;
    665 		/* We got a match: update if doing auto-threshold-tuning */
    666 		if (bge_auto_thresh)
    667 			bge_set_thresh(ifp, lvl);
    668 	}
    669 }
    670 
    671 /*
    672  * Handle events that have triggered interrupts.
    673  */
    674 static void
    675 bge_handle_events(struct bge_softc *sc)
    676 {
    677 
    678 	return;
    679 }
    680 
    681 /*
    682  * Memory management for jumbo frames.
    683  */
    684 
    685 static int
    686 bge_alloc_jumbo_mem(struct bge_softc *sc)
    687 {
    688 	char *ptr, *kva;
    689 	bus_dma_segment_t	seg;
    690 	int		i, rseg, state, error;
    691 	struct bge_jpool_entry   *entry;
    692 
    693 	state = error = 0;
    694 
    695 	/* Grab a big chunk o' storage. */
    696 	if (bus_dmamem_alloc(sc->bge_dmatag, BGE_JMEM, PAGE_SIZE, 0,
    697 	     &seg, 1, &rseg, BUS_DMA_NOWAIT)) {
    698 		printf("%s: can't alloc rx buffers\n", sc->bge_dev.dv_xname);
    699 		return ENOBUFS;
    700 	}
    701 
    702 	state = 1;
    703 	if (bus_dmamem_map(sc->bge_dmatag, &seg, rseg, BGE_JMEM, (void **)&kva,
    704 	    BUS_DMA_NOWAIT)) {
    705 		printf("%s: can't map DMA buffers (%d bytes)\n",
    706 		    sc->bge_dev.dv_xname, (int)BGE_JMEM);
    707 		error = ENOBUFS;
    708 		goto out;
    709 	}
    710 
    711 	state = 2;
    712 	if (bus_dmamap_create(sc->bge_dmatag, BGE_JMEM, 1, BGE_JMEM, 0,
    713 	    BUS_DMA_NOWAIT, &sc->bge_cdata.bge_rx_jumbo_map)) {
    714 		printf("%s: can't create DMA map\n", sc->bge_dev.dv_xname);
    715 		error = ENOBUFS;
    716 		goto out;
    717 	}
    718 
    719 	state = 3;
    720 	if (bus_dmamap_load(sc->bge_dmatag, sc->bge_cdata.bge_rx_jumbo_map,
    721 	    kva, BGE_JMEM, NULL, BUS_DMA_NOWAIT)) {
    722 		printf("%s: can't load DMA map\n", sc->bge_dev.dv_xname);
    723 		error = ENOBUFS;
    724 		goto out;
    725 	}
    726 
    727 	state = 4;
    728 	sc->bge_cdata.bge_jumbo_buf = (void *)kva;
    729 	DPRINTFN(1,("bge_jumbo_buf = %p\n", sc->bge_cdata.bge_jumbo_buf));
    730 
    731 	SLIST_INIT(&sc->bge_jfree_listhead);
    732 	SLIST_INIT(&sc->bge_jinuse_listhead);
    733 
    734 	/*
    735 	 * Now divide it up into 9K pieces and save the addresses
    736 	 * in an array.
    737 	 */
    738 	ptr = sc->bge_cdata.bge_jumbo_buf;
    739 	for (i = 0; i < BGE_JSLOTS; i++) {
    740 		sc->bge_cdata.bge_jslots[i] = ptr;
    741 		ptr += BGE_JLEN;
    742 		entry = malloc(sizeof(struct bge_jpool_entry),
    743 		    M_DEVBUF, M_NOWAIT);
    744 		if (entry == NULL) {
    745 			printf("%s: no memory for jumbo buffer queue!\n",
    746 			    sc->bge_dev.dv_xname);
    747 			error = ENOBUFS;
    748 			goto out;
    749 		}
    750 		entry->slot = i;
    751 		SLIST_INSERT_HEAD(&sc->bge_jfree_listhead,
    752 				 entry, jpool_entries);
    753 	}
    754 out:
    755 	if (error != 0) {
    756 		switch (state) {
    757 		case 4:
    758 			bus_dmamap_unload(sc->bge_dmatag,
    759 			    sc->bge_cdata.bge_rx_jumbo_map);
    760 		case 3:
    761 			bus_dmamap_destroy(sc->bge_dmatag,
    762 			    sc->bge_cdata.bge_rx_jumbo_map);
    763 		case 2:
    764 			bus_dmamem_unmap(sc->bge_dmatag, kva, BGE_JMEM);
    765 		case 1:
    766 			bus_dmamem_free(sc->bge_dmatag, &seg, rseg);
    767 			break;
    768 		default:
    769 			break;
    770 		}
    771 	}
    772 
    773 	return error;
    774 }
    775 
    776 /*
    777  * Allocate a jumbo buffer.
    778  */
    779 static void *
    780 bge_jalloc(struct bge_softc *sc)
    781 {
    782 	struct bge_jpool_entry   *entry;
    783 
    784 	entry = SLIST_FIRST(&sc->bge_jfree_listhead);
    785 
    786 	if (entry == NULL) {
    787 		printf("%s: no free jumbo buffers\n", sc->bge_dev.dv_xname);
    788 		return(NULL);
    789 	}
    790 
    791 	SLIST_REMOVE_HEAD(&sc->bge_jfree_listhead, jpool_entries);
    792 	SLIST_INSERT_HEAD(&sc->bge_jinuse_listhead, entry, jpool_entries);
    793 	return(sc->bge_cdata.bge_jslots[entry->slot]);
    794 }
    795 
    796 /*
    797  * Release a jumbo buffer.
    798  */
    799 static void
    800 bge_jfree(struct mbuf *m, void *buf, size_t size, void *arg)
    801 {
    802 	struct bge_jpool_entry *entry;
    803 	struct bge_softc *sc;
    804 	int i, s;
    805 
    806 	/* Extract the softc struct pointer. */
    807 	sc = (struct bge_softc *)arg;
    808 
    809 	if (sc == NULL)
    810 		panic("bge_jfree: can't find softc pointer!");
    811 
    812 	/* calculate the slot this buffer belongs to */
    813 
    814 	i = ((char *)buf
    815 	     - (char *)sc->bge_cdata.bge_jumbo_buf) / BGE_JLEN;
    816 
    817 	if ((i < 0) || (i >= BGE_JSLOTS))
    818 		panic("bge_jfree: asked to free buffer that we don't manage!");
    819 
    820 	s = splvm();
    821 	entry = SLIST_FIRST(&sc->bge_jinuse_listhead);
    822 	if (entry == NULL)
    823 		panic("bge_jfree: buffer not in use!");
    824 	entry->slot = i;
    825 	SLIST_REMOVE_HEAD(&sc->bge_jinuse_listhead, jpool_entries);
    826 	SLIST_INSERT_HEAD(&sc->bge_jfree_listhead, entry, jpool_entries);
    827 
    828 	if (__predict_true(m != NULL))
    829   		pool_cache_put(&mbpool_cache, m);
    830 	splx(s);
    831 }
    832 
    833 
    834 /*
    835  * Intialize a standard receive ring descriptor.
    836  */
    837 static int
    838 bge_newbuf_std(struct bge_softc *sc, int i, struct mbuf *m, bus_dmamap_t dmamap)
    839 {
    840 	struct mbuf		*m_new = NULL;
    841 	struct bge_rx_bd	*r;
    842 	int			error;
    843 
    844 	if (dmamap == NULL) {
    845 		error = bus_dmamap_create(sc->bge_dmatag, MCLBYTES, 1,
    846 		    MCLBYTES, 0, BUS_DMA_NOWAIT, &dmamap);
    847 		if (error != 0)
    848 			return error;
    849 	}
    850 
    851 	sc->bge_cdata.bge_rx_std_map[i] = dmamap;
    852 
    853 	if (m == NULL) {
    854 		MGETHDR(m_new, M_DONTWAIT, MT_DATA);
    855 		if (m_new == NULL) {
    856 			return(ENOBUFS);
    857 		}
    858 
    859 		MCLGET(m_new, M_DONTWAIT);
    860 		if (!(m_new->m_flags & M_EXT)) {
    861 			m_freem(m_new);
    862 			return(ENOBUFS);
    863 		}
    864 		m_new->m_len = m_new->m_pkthdr.len = MCLBYTES;
    865 
    866 	} else {
    867 		m_new = m;
    868 		m_new->m_len = m_new->m_pkthdr.len = MCLBYTES;
    869 		m_new->m_data = m_new->m_ext.ext_buf;
    870 	}
    871 	if (!sc->bge_rx_alignment_bug)
    872 	    m_adj(m_new, ETHER_ALIGN);
    873 	if (bus_dmamap_load_mbuf(sc->bge_dmatag, dmamap, m_new,
    874 	    BUS_DMA_READ|BUS_DMA_NOWAIT))
    875 		return(ENOBUFS);
    876 	bus_dmamap_sync(sc->bge_dmatag, dmamap, 0, dmamap->dm_mapsize,
    877 	    BUS_DMASYNC_PREREAD);
    878 
    879 	sc->bge_cdata.bge_rx_std_chain[i] = m_new;
    880 	r = &sc->bge_rdata->bge_rx_std_ring[i];
    881 	bge_set_hostaddr(&r->bge_addr,
    882 	    dmamap->dm_segs[0].ds_addr);
    883 	r->bge_flags = BGE_RXBDFLAG_END;
    884 	r->bge_len = m_new->m_len;
    885 	r->bge_idx = i;
    886 
    887 	bus_dmamap_sync(sc->bge_dmatag, sc->bge_ring_map,
    888 	    offsetof(struct bge_ring_data, bge_rx_std_ring) +
    889 		i * sizeof (struct bge_rx_bd),
    890 	    sizeof (struct bge_rx_bd),
    891 	    BUS_DMASYNC_PREWRITE|BUS_DMASYNC_PREREAD);
    892 
    893 	return(0);
    894 }
    895 
    896 /*
    897  * Initialize a jumbo receive ring descriptor. This allocates
    898  * a jumbo buffer from the pool managed internally by the driver.
    899  */
    900 static int
    901 bge_newbuf_jumbo(struct bge_softc *sc, int i, struct mbuf *m)
    902 {
    903 	struct mbuf *m_new = NULL;
    904 	struct bge_rx_bd *r;
    905 	void *buf = NULL;
    906 
    907 	if (m == NULL) {
    908 
    909 		/* Allocate the mbuf. */
    910 		MGETHDR(m_new, M_DONTWAIT, MT_DATA);
    911 		if (m_new == NULL) {
    912 			return(ENOBUFS);
    913 		}
    914 
    915 		/* Allocate the jumbo buffer */
    916 		buf = bge_jalloc(sc);
    917 		if (buf == NULL) {
    918 			m_freem(m_new);
    919 			printf("%s: jumbo allocation failed "
    920 			    "-- packet dropped!\n", sc->bge_dev.dv_xname);
    921 			return(ENOBUFS);
    922 		}
    923 
    924 		/* Attach the buffer to the mbuf. */
    925 		m_new->m_len = m_new->m_pkthdr.len = BGE_JUMBO_FRAMELEN;
    926 		MEXTADD(m_new, buf, BGE_JUMBO_FRAMELEN, M_DEVBUF,
    927 		    bge_jfree, sc);
    928 		m_new->m_flags |= M_EXT_RW;
    929 	} else {
    930 		m_new = m;
    931 		buf = m_new->m_data = m_new->m_ext.ext_buf;
    932 		m_new->m_ext.ext_size = BGE_JUMBO_FRAMELEN;
    933 	}
    934 	if (!sc->bge_rx_alignment_bug)
    935 	    m_adj(m_new, ETHER_ALIGN);
    936 	bus_dmamap_sync(sc->bge_dmatag, sc->bge_cdata.bge_rx_jumbo_map,
    937 	    mtod(m_new, char *) - (char *)sc->bge_cdata.bge_jumbo_buf, BGE_JLEN,
    938 	    BUS_DMASYNC_PREREAD);
    939 	/* Set up the descriptor. */
    940 	r = &sc->bge_rdata->bge_rx_jumbo_ring[i];
    941 	sc->bge_cdata.bge_rx_jumbo_chain[i] = m_new;
    942 	bge_set_hostaddr(&r->bge_addr, BGE_JUMBO_DMA_ADDR(sc, m_new));
    943 	r->bge_flags = BGE_RXBDFLAG_END|BGE_RXBDFLAG_JUMBO_RING;
    944 	r->bge_len = m_new->m_len;
    945 	r->bge_idx = i;
    946 
    947 	bus_dmamap_sync(sc->bge_dmatag, sc->bge_ring_map,
    948 	    offsetof(struct bge_ring_data, bge_rx_jumbo_ring) +
    949 		i * sizeof (struct bge_rx_bd),
    950 	    sizeof (struct bge_rx_bd),
    951 	    BUS_DMASYNC_PREWRITE|BUS_DMASYNC_PREREAD);
    952 
    953 	return(0);
    954 }
    955 
    956 /*
    957  * The standard receive ring has 512 entries in it. At 2K per mbuf cluster,
    958  * that's 1MB or memory, which is a lot. For now, we fill only the first
    959  * 256 ring entries and hope that our CPU is fast enough to keep up with
    960  * the NIC.
    961  */
    962 static int
    963 bge_init_rx_ring_std(struct bge_softc *sc)
    964 {
    965 	int i;
    966 
    967 	if (sc->bge_flags & BGE_RXRING_VALID)
    968 		return 0;
    969 
    970 	for (i = 0; i < BGE_SSLOTS; i++) {
    971 		if (bge_newbuf_std(sc, i, NULL, 0) == ENOBUFS)
    972 			return(ENOBUFS);
    973 	}
    974 
    975 	sc->bge_std = i - 1;
    976 	CSR_WRITE_4(sc, BGE_MBX_RX_STD_PROD_LO, sc->bge_std);
    977 
    978 	sc->bge_flags |= BGE_RXRING_VALID;
    979 
    980 	return(0);
    981 }
    982 
    983 static void
    984 bge_free_rx_ring_std(struct bge_softc *sc)
    985 {
    986 	int i;
    987 
    988 	if (!(sc->bge_flags & BGE_RXRING_VALID))
    989 		return;
    990 
    991 	for (i = 0; i < BGE_STD_RX_RING_CNT; i++) {
    992 		if (sc->bge_cdata.bge_rx_std_chain[i] != NULL) {
    993 			m_freem(sc->bge_cdata.bge_rx_std_chain[i]);
    994 			sc->bge_cdata.bge_rx_std_chain[i] = NULL;
    995 			bus_dmamap_destroy(sc->bge_dmatag,
    996 			    sc->bge_cdata.bge_rx_std_map[i]);
    997 		}
    998 		memset((char *)&sc->bge_rdata->bge_rx_std_ring[i], 0,
    999 		    sizeof(struct bge_rx_bd));
   1000 	}
   1001 
   1002 	sc->bge_flags &= ~BGE_RXRING_VALID;
   1003 }
   1004 
   1005 static int
   1006 bge_init_rx_ring_jumbo(struct bge_softc *sc)
   1007 {
   1008 	int i;
   1009 	volatile struct bge_rcb *rcb;
   1010 
   1011 	if (sc->bge_flags & BGE_JUMBO_RXRING_VALID)
   1012 		return 0;
   1013 
   1014 	for (i = 0; i < BGE_JUMBO_RX_RING_CNT; i++) {
   1015 		if (bge_newbuf_jumbo(sc, i, NULL) == ENOBUFS)
   1016 			return(ENOBUFS);
   1017 	};
   1018 
   1019 	sc->bge_jumbo = i - 1;
   1020 	sc->bge_flags |= BGE_JUMBO_RXRING_VALID;
   1021 
   1022 	rcb = &sc->bge_rdata->bge_info.bge_jumbo_rx_rcb;
   1023 	rcb->bge_maxlen_flags = 0;
   1024 	CSR_WRITE_4(sc, BGE_RX_JUMBO_RCB_MAXLEN_FLAGS, rcb->bge_maxlen_flags);
   1025 
   1026 	CSR_WRITE_4(sc, BGE_MBX_RX_JUMBO_PROD_LO, sc->bge_jumbo);
   1027 
   1028 	return(0);
   1029 }
   1030 
   1031 static void
   1032 bge_free_rx_ring_jumbo(struct bge_softc *sc)
   1033 {
   1034 	int i;
   1035 
   1036 	if (!(sc->bge_flags & BGE_JUMBO_RXRING_VALID))
   1037 		return;
   1038 
   1039 	for (i = 0; i < BGE_JUMBO_RX_RING_CNT; i++) {
   1040 		if (sc->bge_cdata.bge_rx_jumbo_chain[i] != NULL) {
   1041 			m_freem(sc->bge_cdata.bge_rx_jumbo_chain[i]);
   1042 			sc->bge_cdata.bge_rx_jumbo_chain[i] = NULL;
   1043 		}
   1044 		memset((char *)&sc->bge_rdata->bge_rx_jumbo_ring[i], 0,
   1045 		    sizeof(struct bge_rx_bd));
   1046 	}
   1047 
   1048 	sc->bge_flags &= ~BGE_JUMBO_RXRING_VALID;
   1049 }
   1050 
   1051 static void
   1052 bge_free_tx_ring(struct bge_softc *sc)
   1053 {
   1054 	int i, freed;
   1055 	struct txdmamap_pool_entry *dma;
   1056 
   1057 	if (!(sc->bge_flags & BGE_TXRING_VALID))
   1058 		return;
   1059 
   1060 	freed = 0;
   1061 
   1062 	for (i = 0; i < BGE_TX_RING_CNT; i++) {
   1063 		if (sc->bge_cdata.bge_tx_chain[i] != NULL) {
   1064 			freed++;
   1065 			m_freem(sc->bge_cdata.bge_tx_chain[i]);
   1066 			sc->bge_cdata.bge_tx_chain[i] = NULL;
   1067 			SLIST_INSERT_HEAD(&sc->txdma_list, sc->txdma[i],
   1068 					    link);
   1069 			sc->txdma[i] = 0;
   1070 		}
   1071 		memset((char *)&sc->bge_rdata->bge_tx_ring[i], 0,
   1072 		    sizeof(struct bge_tx_bd));
   1073 	}
   1074 
   1075 	while ((dma = SLIST_FIRST(&sc->txdma_list))) {
   1076 		SLIST_REMOVE_HEAD(&sc->txdma_list, link);
   1077 		bus_dmamap_destroy(sc->bge_dmatag, dma->dmamap);
   1078 		free(dma, M_DEVBUF);
   1079 	}
   1080 
   1081 	sc->bge_flags &= ~BGE_TXRING_VALID;
   1082 }
   1083 
   1084 static int
   1085 bge_init_tx_ring(struct bge_softc *sc)
   1086 {
   1087 	int i;
   1088 	bus_dmamap_t dmamap;
   1089 	struct txdmamap_pool_entry *dma;
   1090 
   1091 	if (sc->bge_flags & BGE_TXRING_VALID)
   1092 		return 0;
   1093 
   1094 	sc->bge_txcnt = 0;
   1095 	sc->bge_tx_saved_considx = 0;
   1096 
   1097 	/* Initialize transmit producer index for host-memory send ring. */
   1098 	sc->bge_tx_prodidx = 0;
   1099 	CSR_WRITE_4(sc, BGE_MBX_TX_HOST_PROD0_LO, sc->bge_tx_prodidx);
   1100 	if (sc->bge_quirks & BGE_QUIRK_PRODUCER_BUG)	/* 5700 b2 errata */
   1101 		CSR_WRITE_4(sc, BGE_MBX_TX_HOST_PROD0_LO, sc->bge_tx_prodidx);
   1102 
   1103 	/* NIC-memory send ring  not used; initialize to zero. */
   1104 	CSR_WRITE_4(sc, BGE_MBX_TX_NIC_PROD0_LO, 0);
   1105 	if (sc->bge_quirks & BGE_QUIRK_PRODUCER_BUG)	/* 5700 b2 errata */
   1106 		CSR_WRITE_4(sc, BGE_MBX_TX_HOST_PROD0_LO, 0);
   1107 
   1108 	SLIST_INIT(&sc->txdma_list);
   1109 	for (i = 0; i < BGE_RSLOTS; i++) {
   1110 		if (bus_dmamap_create(sc->bge_dmatag, BGE_TXDMA_MAX,
   1111 		    BGE_NTXSEG, ETHER_MAX_LEN_JUMBO, 0, BUS_DMA_NOWAIT,
   1112 		    &dmamap))
   1113 			return(ENOBUFS);
   1114 		if (dmamap == NULL)
   1115 			panic("dmamap NULL in bge_init_tx_ring");
   1116 		dma = malloc(sizeof(*dma), M_DEVBUF, M_NOWAIT);
   1117 		if (dma == NULL) {
   1118 			printf("%s: can't alloc txdmamap_pool_entry\n",
   1119 			    sc->bge_dev.dv_xname);
   1120 			bus_dmamap_destroy(sc->bge_dmatag, dmamap);
   1121 			return (ENOMEM);
   1122 		}
   1123 		dma->dmamap = dmamap;
   1124 		SLIST_INSERT_HEAD(&sc->txdma_list, dma, link);
   1125 	}
   1126 
   1127 	sc->bge_flags |= BGE_TXRING_VALID;
   1128 
   1129 	return(0);
   1130 }
   1131 
   1132 static void
   1133 bge_setmulti(struct bge_softc *sc)
   1134 {
   1135 	struct ethercom		*ac = &sc->ethercom;
   1136 	struct ifnet		*ifp = &ac->ec_if;
   1137 	struct ether_multi	*enm;
   1138 	struct ether_multistep  step;
   1139 	u_int32_t		hashes[4] = { 0, 0, 0, 0 };
   1140 	u_int32_t		h;
   1141 	int			i;
   1142 
   1143 	if (ifp->if_flags & IFF_PROMISC)
   1144 		goto allmulti;
   1145 
   1146 	/* Now program new ones. */
   1147 	ETHER_FIRST_MULTI(step, ac, enm);
   1148 	while (enm != NULL) {
   1149 		if (memcmp(enm->enm_addrlo, enm->enm_addrhi, ETHER_ADDR_LEN)) {
   1150 			/*
   1151 			 * We must listen to a range of multicast addresses.
   1152 			 * For now, just accept all multicasts, rather than
   1153 			 * trying to set only those filter bits needed to match
   1154 			 * the range.  (At this time, the only use of address
   1155 			 * ranges is for IP multicast routing, for which the
   1156 			 * range is big enough to require all bits set.)
   1157 			 */
   1158 			goto allmulti;
   1159 		}
   1160 
   1161 		h = ether_crc32_le(enm->enm_addrlo, ETHER_ADDR_LEN);
   1162 
   1163 		/* Just want the 7 least-significant bits. */
   1164 		h &= 0x7f;
   1165 
   1166 		hashes[(h & 0x60) >> 5] |= 1 << (h & 0x1F);
   1167 		ETHER_NEXT_MULTI(step, enm);
   1168 	}
   1169 
   1170 	ifp->if_flags &= ~IFF_ALLMULTI;
   1171 	goto setit;
   1172 
   1173  allmulti:
   1174 	ifp->if_flags |= IFF_ALLMULTI;
   1175 	hashes[0] = hashes[1] = hashes[2] = hashes[3] = 0xffffffff;
   1176 
   1177  setit:
   1178 	for (i = 0; i < 4; i++)
   1179 		CSR_WRITE_4(sc, BGE_MAR0 + (i * 4), hashes[i]);
   1180 }
   1181 
   1182 const int bge_swapbits[] = {
   1183 	0,
   1184 	BGE_MODECTL_BYTESWAP_DATA,
   1185 	BGE_MODECTL_WORDSWAP_DATA,
   1186 	BGE_MODECTL_BYTESWAP_NONFRAME,
   1187 	BGE_MODECTL_WORDSWAP_NONFRAME,
   1188 
   1189 	BGE_MODECTL_BYTESWAP_DATA|BGE_MODECTL_WORDSWAP_DATA,
   1190 	BGE_MODECTL_BYTESWAP_DATA|BGE_MODECTL_BYTESWAP_NONFRAME,
   1191 	BGE_MODECTL_BYTESWAP_DATA|BGE_MODECTL_WORDSWAP_NONFRAME,
   1192 
   1193 	BGE_MODECTL_WORDSWAP_DATA|BGE_MODECTL_BYTESWAP_NONFRAME,
   1194 	BGE_MODECTL_WORDSWAP_DATA|BGE_MODECTL_WORDSWAP_NONFRAME,
   1195 
   1196 	BGE_MODECTL_BYTESWAP_NONFRAME|BGE_MODECTL_WORDSWAP_NONFRAME,
   1197 
   1198 	BGE_MODECTL_BYTESWAP_DATA|BGE_MODECTL_WORDSWAP_DATA|
   1199 	    BGE_MODECTL_BYTESWAP_NONFRAME,
   1200 	BGE_MODECTL_BYTESWAP_DATA|BGE_MODECTL_WORDSWAP_DATA|
   1201 	    BGE_MODECTL_WORDSWAP_NONFRAME,
   1202 	BGE_MODECTL_BYTESWAP_DATA|BGE_MODECTL_BYTESWAP_NONFRAME|
   1203 	    BGE_MODECTL_WORDSWAP_NONFRAME,
   1204 	BGE_MODECTL_WORDSWAP_DATA|BGE_MODECTL_BYTESWAP_NONFRAME|
   1205 	    BGE_MODECTL_WORDSWAP_NONFRAME,
   1206 
   1207 	BGE_MODECTL_BYTESWAP_DATA|BGE_MODECTL_WORDSWAP_DATA|
   1208 	    BGE_MODECTL_BYTESWAP_NONFRAME|BGE_MODECTL_WORDSWAP_NONFRAME,
   1209 };
   1210 
   1211 int bge_swapindex = 0;
   1212 
   1213 /*
   1214  * Do endian, PCI and DMA initialization. Also check the on-board ROM
   1215  * self-test results.
   1216  */
   1217 static int
   1218 bge_chipinit(struct bge_softc *sc)
   1219 {
   1220 	u_int32_t		cachesize;
   1221 	int			i;
   1222 	u_int32_t		dma_rw_ctl;
   1223 	struct pci_attach_args	*pa = &(sc->bge_pa);
   1224 
   1225 
   1226 	/* Set endianness before we access any non-PCI registers. */
   1227 	pci_conf_write(pa->pa_pc, pa->pa_tag, BGE_PCI_MISC_CTL,
   1228 	    BGE_INIT);
   1229 
   1230 	/* Set power state to D0. */
   1231 	bge_setpowerstate(sc, 0);
   1232 
   1233 	/*
   1234 	 * Check the 'ROM failed' bit on the RX CPU to see if
   1235 	 * self-tests passed.
   1236 	 */
   1237 	if (CSR_READ_4(sc, BGE_RXCPU_MODE) & BGE_RXCPUMODE_ROMFAIL) {
   1238 		printf("%s: RX CPU self-diagnostics failed!\n",
   1239 		    sc->bge_dev.dv_xname);
   1240 		return(ENODEV);
   1241 	}
   1242 
   1243 	/* Clear the MAC control register */
   1244 	CSR_WRITE_4(sc, BGE_MAC_MODE, 0);
   1245 
   1246 	/*
   1247 	 * Clear the MAC statistics block in the NIC's
   1248 	 * internal memory.
   1249 	 */
   1250 	for (i = BGE_STATS_BLOCK;
   1251 	    i < BGE_STATS_BLOCK_END + 1; i += sizeof(u_int32_t))
   1252 		BGE_MEMWIN_WRITE(pa->pa_pc, pa->pa_tag, i, 0);
   1253 
   1254 	for (i = BGE_STATUS_BLOCK;
   1255 	    i < BGE_STATUS_BLOCK_END + 1; i += sizeof(u_int32_t))
   1256 		BGE_MEMWIN_WRITE(pa->pa_pc, pa->pa_tag, i, 0);
   1257 
   1258 	/* Set up the PCI DMA control register. */
   1259 	if (sc->bge_pcie) {
   1260 	  u_int32_t device_ctl;
   1261 
   1262 		/* From FreeBSD */
   1263 		DPRINTFN(4, ("(%s: PCI-Express DMA setting)\n",
   1264 		    sc->bge_dev.dv_xname));
   1265 		dma_rw_ctl = (BGE_PCI_READ_CMD | BGE_PCI_WRITE_CMD |
   1266 		    (0xf << BGE_PCIDMARWCTL_RD_WAT_SHIFT) |
   1267 		    (0x2 << BGE_PCIDMARWCTL_WR_WAT_SHIFT));
   1268 
   1269 		/* jonathan: alternative from Linux driver */
   1270 #define DMA_CTRL_WRITE_PCIE_H20MARK_128         0x00180000
   1271 #define DMA_CTRL_WRITE_PCIE_H20MARK_256         0x00380000
   1272 
   1273 		dma_rw_ctl =   0x76000000; /* XXX XXX XXX */;
   1274 		device_ctl = pci_conf_read(pa->pa_pc, pa->pa_tag,
   1275 					   BGE_PCI_CONF_DEV_CTRL);
   1276 		aprint_debug("%s: pcie mode=0x%x\n", sc->bge_dev.dv_xname,
   1277 		    device_ctl);
   1278 
   1279 		if ((device_ctl & 0x00e0) && 0) {
   1280 			/*
   1281 			 * XXX jonathan (at) NetBSD.org:
   1282 			 * This clause is exactly what the Broadcom-supplied
   1283 			 * Linux does; but given overall register programming
   1284 			 * by if_bge(4), this larger DMA-write watermark
   1285 			 * value causes bcm5721 chips to totally wedge.
   1286 			 */
   1287 			dma_rw_ctl |= BGE_PCIDMA_RWCTL_PCIE_WRITE_WATRMARK_256;
   1288 		} else {
   1289 			dma_rw_ctl |= BGE_PCIDMA_RWCTL_PCIE_WRITE_WATRMARK_128;
   1290 		}
   1291 	} else if (pci_conf_read(pa->pa_pc, pa->pa_tag,BGE_PCI_PCISTATE) &
   1292 	    BGE_PCISTATE_PCI_BUSMODE) {
   1293 		/* Conventional PCI bus */
   1294 	  	DPRINTFN(4, ("(%s: PCI 2.2 DMA setting)\n", sc->bge_dev.dv_xname));
   1295 		dma_rw_ctl = (BGE_PCI_READ_CMD | BGE_PCI_WRITE_CMD |
   1296 		   (0x7 << BGE_PCIDMARWCTL_RD_WAT_SHIFT) |
   1297 		   (0x7 << BGE_PCIDMARWCTL_WR_WAT_SHIFT));
   1298 		if ((sc->bge_quirks & BGE_QUIRK_5705_CORE) == 0) {
   1299 			dma_rw_ctl |= 0x0F;
   1300 		}
   1301 	} else {
   1302 	  	DPRINTFN(4, ("(:%s: PCI-X DMA setting)\n", sc->bge_dev.dv_xname));
   1303 		/* PCI-X bus */
   1304 		dma_rw_ctl = BGE_PCI_READ_CMD|BGE_PCI_WRITE_CMD |
   1305 		    (0x3 << BGE_PCIDMARWCTL_RD_WAT_SHIFT) |
   1306 		    (0x3 << BGE_PCIDMARWCTL_WR_WAT_SHIFT) |
   1307 		    (0x0F);
   1308 		/*
   1309 		 * 5703 and 5704 need ONEDMA_AT_ONCE as a workaround
   1310 		 * for hardware bugs, which means we should also clear
   1311 		 * the low-order MINDMA bits.  In addition, the 5704
   1312 		 * uses a different encoding of read/write watermarks.
   1313 		 */
   1314 		if (BGE_ASICREV(sc->bge_chipid) == BGE_ASICREV_BCM5704) {
   1315 			dma_rw_ctl = BGE_PCI_READ_CMD|BGE_PCI_WRITE_CMD |
   1316 			  /* should be 0x1f0000 */
   1317 			  (0x7 << BGE_PCIDMARWCTL_RD_WAT_SHIFT) |
   1318 			  (0x3 << BGE_PCIDMARWCTL_WR_WAT_SHIFT);
   1319 			dma_rw_ctl |= BGE_PCIDMARWCTL_ONEDMA_ATONCE;
   1320 		}
   1321 		else if (BGE_ASICREV(sc->bge_chipid) == BGE_ASICREV_BCM5703) {
   1322 			dma_rw_ctl &=  0xfffffff0;
   1323 			dma_rw_ctl |= BGE_PCIDMARWCTL_ONEDMA_ATONCE;
   1324 		}
   1325 		else if (BGE_IS_5714_FAMILY(sc)) {
   1326 			dma_rw_ctl = BGE_PCI_READ_CMD|BGE_PCI_WRITE_CMD;
   1327 			dma_rw_ctl &= ~BGE_PCIDMARWCTL_ONEDMA_ATONCE; /* XXX */
   1328 			/* XXX magic values, Broadcom-supplied Linux driver */
   1329 			if (BGE_ASICREV(sc->bge_chipid) == BGE_ASICREV_BCM5780)
   1330 				dma_rw_ctl |= (1 << 20) | (1 << 18) |
   1331 				  BGE_PCIDMARWCTL_ONEDMA_ATONCE;
   1332 			else
   1333 				dma_rw_ctl |= (1<<20) | (1<<18) | (1 << 15);
   1334 		}
   1335 	}
   1336 
   1337 	pci_conf_write(pa->pa_pc, pa->pa_tag, BGE_PCI_DMA_RW_CTL, dma_rw_ctl);
   1338 
   1339 	/*
   1340 	 * Set up general mode register.
   1341 	 */
   1342 	CSR_WRITE_4(sc, BGE_MODE_CTL, BGE_DMA_SWAP_OPTIONS|
   1343 		    BGE_MODECTL_MAC_ATTN_INTR|BGE_MODECTL_HOST_SEND_BDS|
   1344 		    BGE_MODECTL_TX_NO_PHDR_CSUM|BGE_MODECTL_RX_NO_PHDR_CSUM);
   1345 
   1346 	/* Get cache line size. */
   1347 	cachesize = pci_conf_read(pa->pa_pc, pa->pa_tag, BGE_PCI_CACHESZ);
   1348 
   1349 	/*
   1350 	 * Avoid violating PCI spec on certain chip revs.
   1351 	 */
   1352 	if (pci_conf_read(pa->pa_pc, pa->pa_tag, BGE_PCI_CMD) &
   1353 	    PCIM_CMD_MWIEN) {
   1354 		switch(cachesize) {
   1355 		case 1:
   1356 			PCI_SETBIT(pa->pa_pc, pa->pa_tag, BGE_PCI_DMA_RW_CTL,
   1357 				   BGE_PCI_WRITE_BNDRY_16BYTES);
   1358 			break;
   1359 		case 2:
   1360 			PCI_SETBIT(pa->pa_pc, pa->pa_tag, BGE_PCI_DMA_RW_CTL,
   1361 				   BGE_PCI_WRITE_BNDRY_32BYTES);
   1362 			break;
   1363 		case 4:
   1364 			PCI_SETBIT(pa->pa_pc, pa->pa_tag, BGE_PCI_DMA_RW_CTL,
   1365 				   BGE_PCI_WRITE_BNDRY_64BYTES);
   1366 			break;
   1367 		case 8:
   1368 			PCI_SETBIT(pa->pa_pc, pa->pa_tag, BGE_PCI_DMA_RW_CTL,
   1369 				   BGE_PCI_WRITE_BNDRY_128BYTES);
   1370 			break;
   1371 		case 16:
   1372 			PCI_SETBIT(pa->pa_pc, pa->pa_tag, BGE_PCI_DMA_RW_CTL,
   1373 				   BGE_PCI_WRITE_BNDRY_256BYTES);
   1374 			break;
   1375 		case 32:
   1376 			PCI_SETBIT(pa->pa_pc, pa->pa_tag, BGE_PCI_DMA_RW_CTL,
   1377 				   BGE_PCI_WRITE_BNDRY_512BYTES);
   1378 			break;
   1379 		case 64:
   1380 			PCI_SETBIT(pa->pa_pc, pa->pa_tag, BGE_PCI_DMA_RW_CTL,
   1381 				   BGE_PCI_WRITE_BNDRY_1024BYTES);
   1382 			break;
   1383 		default:
   1384 		/* Disable PCI memory write and invalidate. */
   1385 #if 0
   1386 			if (bootverbose)
   1387 				printf("%s: cache line size %d not "
   1388 				    "supported; disabling PCI MWI\n",
   1389 				    sc->bge_dev.dv_xname, cachesize);
   1390 #endif
   1391 			PCI_CLRBIT(pa->pa_pc, pa->pa_tag, BGE_PCI_CMD,
   1392 			    PCIM_CMD_MWIEN);
   1393 			break;
   1394 		}
   1395 	}
   1396 
   1397 	/*
   1398 	 * Disable memory write invalidate.  Apparently it is not supported
   1399 	 * properly by these devices.
   1400 	 */
   1401 	PCI_CLRBIT(pa->pa_pc, pa->pa_tag, BGE_PCI_CMD, PCIM_CMD_MWIEN);
   1402 
   1403 
   1404 #ifdef __brokenalpha__
   1405 	/*
   1406 	 * Must insure that we do not cross an 8K (bytes) boundary
   1407 	 * for DMA reads.  Our highest limit is 1K bytes.  This is a
   1408 	 * restriction on some ALPHA platforms with early revision
   1409 	 * 21174 PCI chipsets, such as the AlphaPC 164lx
   1410 	 */
   1411 	PCI_SETBIT(sc, BGE_PCI_DMA_RW_CTL, BGE_PCI_READ_BNDRY_1024, 4);
   1412 #endif
   1413 
   1414 	/* Set the timer prescaler (always 66MHz) */
   1415 	CSR_WRITE_4(sc, BGE_MISC_CFG, 65 << 1/*BGE_32BITTIME_66MHZ*/);
   1416 
   1417 	return(0);
   1418 }
   1419 
   1420 static int
   1421 bge_blockinit(struct bge_softc *sc)
   1422 {
   1423 	volatile struct bge_rcb		*rcb;
   1424 	bus_size_t		rcb_addr;
   1425 	int			i;
   1426 	struct ifnet		*ifp = &sc->ethercom.ec_if;
   1427 	bge_hostaddr		taddr;
   1428 
   1429 	/*
   1430 	 * Initialize the memory window pointer register so that
   1431 	 * we can access the first 32K of internal NIC RAM. This will
   1432 	 * allow us to set up the TX send ring RCBs and the RX return
   1433 	 * ring RCBs, plus other things which live in NIC memory.
   1434 	 */
   1435 
   1436 	pci_conf_write(sc->bge_pa.pa_pc, sc->bge_pa.pa_tag,
   1437 	    BGE_PCI_MEMWIN_BASEADDR, 0);
   1438 
   1439 	/* Configure mbuf memory pool */
   1440 	if ((sc->bge_quirks & BGE_QUIRK_5705_CORE) == 0) {
   1441 		if (sc->bge_extram) {
   1442 			CSR_WRITE_4(sc, BGE_BMAN_MBUFPOOL_BASEADDR,
   1443 			    BGE_EXT_SSRAM);
   1444 			if ((sc->bge_quirks & BGE_QUIRK_FEWER_MBUFS) != 0)
   1445 				CSR_WRITE_4(sc, BGE_BMAN_MBUFPOOL_LEN, 0x10000);
   1446 			else
   1447 				CSR_WRITE_4(sc, BGE_BMAN_MBUFPOOL_LEN, 0x18000);
   1448 		} else {
   1449 			CSR_WRITE_4(sc, BGE_BMAN_MBUFPOOL_BASEADDR,
   1450 			    BGE_BUFFPOOL_1);
   1451 			if ((sc->bge_quirks & BGE_QUIRK_FEWER_MBUFS) != 0)
   1452 				CSR_WRITE_4(sc, BGE_BMAN_MBUFPOOL_LEN, 0x10000);
   1453 			else
   1454 				CSR_WRITE_4(sc, BGE_BMAN_MBUFPOOL_LEN, 0x18000);
   1455 		}
   1456 
   1457 		/* Configure DMA resource pool */
   1458 		CSR_WRITE_4(sc, BGE_BMAN_DMA_DESCPOOL_BASEADDR,
   1459 		    BGE_DMA_DESCRIPTORS);
   1460 		CSR_WRITE_4(sc, BGE_BMAN_DMA_DESCPOOL_LEN, 0x2000);
   1461 	}
   1462 
   1463 	/* Configure mbuf pool watermarks */
   1464 #ifdef ORIG_WPAUL_VALUES
   1465 	CSR_WRITE_4(sc, BGE_BMAN_MBUFPOOL_READDMA_LOWAT, 24);
   1466 	CSR_WRITE_4(sc, BGE_BMAN_MBUFPOOL_MACRX_LOWAT, 24);
   1467 	CSR_WRITE_4(sc, BGE_BMAN_MBUFPOOL_HIWAT, 48);
   1468 #else
   1469 	/* new broadcom docs strongly recommend these: */
   1470 	if ((sc->bge_quirks & BGE_QUIRK_5705_CORE) == 0) {
   1471 		if (ifp->if_mtu > ETHER_MAX_LEN) {
   1472 			CSR_WRITE_4(sc, BGE_BMAN_MBUFPOOL_READDMA_LOWAT, 0x50);
   1473 			CSR_WRITE_4(sc, BGE_BMAN_MBUFPOOL_MACRX_LOWAT, 0x20);
   1474 			CSR_WRITE_4(sc, BGE_BMAN_MBUFPOOL_HIWAT, 0x60);
   1475 		} else {
   1476 			/* Values from Linux driver... */
   1477 			CSR_WRITE_4(sc, BGE_BMAN_MBUFPOOL_READDMA_LOWAT, 304);
   1478 			CSR_WRITE_4(sc, BGE_BMAN_MBUFPOOL_MACRX_LOWAT, 152);
   1479 			CSR_WRITE_4(sc, BGE_BMAN_MBUFPOOL_HIWAT, 380);
   1480 		}
   1481 	} else {
   1482 		CSR_WRITE_4(sc, BGE_BMAN_MBUFPOOL_READDMA_LOWAT, 0x0);
   1483 		CSR_WRITE_4(sc, BGE_BMAN_MBUFPOOL_MACRX_LOWAT, 0x10);
   1484 		CSR_WRITE_4(sc, BGE_BMAN_MBUFPOOL_HIWAT, 0x60);
   1485 	}
   1486 #endif
   1487 
   1488 	/* Configure DMA resource watermarks */
   1489 	CSR_WRITE_4(sc, BGE_BMAN_DMA_DESCPOOL_LOWAT, 5);
   1490 	CSR_WRITE_4(sc, BGE_BMAN_DMA_DESCPOOL_HIWAT, 10);
   1491 
   1492 	/* Enable buffer manager */
   1493 	CSR_WRITE_4(sc, BGE_BMAN_MODE,
   1494 	    BGE_BMANMODE_ENABLE|BGE_BMANMODE_LOMBUF_ATTN);
   1495 
   1496 	/* Poll for buffer manager start indication */
   1497 	for (i = 0; i < BGE_TIMEOUT; i++) {
   1498 		if (CSR_READ_4(sc, BGE_BMAN_MODE) & BGE_BMANMODE_ENABLE)
   1499 			break;
   1500 		DELAY(10);
   1501 	}
   1502 
   1503 	if (i == BGE_TIMEOUT) {
   1504 		printf("%s: buffer manager failed to start\n",
   1505 		    sc->bge_dev.dv_xname);
   1506 		return(ENXIO);
   1507 	}
   1508 
   1509 	/* Enable flow-through queues */
   1510 	CSR_WRITE_4(sc, BGE_FTQ_RESET, 0xFFFFFFFF);
   1511 	CSR_WRITE_4(sc, BGE_FTQ_RESET, 0);
   1512 
   1513 	/* Wait until queue initialization is complete */
   1514 	for (i = 0; i < BGE_TIMEOUT; i++) {
   1515 		if (CSR_READ_4(sc, BGE_FTQ_RESET) == 0)
   1516 			break;
   1517 		DELAY(10);
   1518 	}
   1519 
   1520 	if (i == BGE_TIMEOUT) {
   1521 		printf("%s: flow-through queue init failed\n",
   1522 		    sc->bge_dev.dv_xname);
   1523 		return(ENXIO);
   1524 	}
   1525 
   1526 	/* Initialize the standard RX ring control block */
   1527 	rcb = &sc->bge_rdata->bge_info.bge_std_rx_rcb;
   1528 	bge_set_hostaddr(&rcb->bge_hostaddr,
   1529 	    BGE_RING_DMA_ADDR(sc, bge_rx_std_ring));
   1530 	if ((sc->bge_quirks & BGE_QUIRK_5705_CORE) == 0) {
   1531 		rcb->bge_maxlen_flags =
   1532 		    BGE_RCB_MAXLEN_FLAGS(BGE_MAX_FRAMELEN, 0);
   1533 	} else {
   1534 		rcb->bge_maxlen_flags = BGE_RCB_MAXLEN_FLAGS(512, 0);
   1535 	}
   1536 	if (sc->bge_extram)
   1537 		rcb->bge_nicaddr = BGE_EXT_STD_RX_RINGS;
   1538 	else
   1539 		rcb->bge_nicaddr = BGE_STD_RX_RINGS;
   1540 	CSR_WRITE_4(sc, BGE_RX_STD_RCB_HADDR_HI, rcb->bge_hostaddr.bge_addr_hi);
   1541 	CSR_WRITE_4(sc, BGE_RX_STD_RCB_HADDR_LO, rcb->bge_hostaddr.bge_addr_lo);
   1542 	CSR_WRITE_4(sc, BGE_RX_STD_RCB_MAXLEN_FLAGS, rcb->bge_maxlen_flags);
   1543 	CSR_WRITE_4(sc, BGE_RX_STD_RCB_NICADDR, rcb->bge_nicaddr);
   1544 
   1545 	if ((sc->bge_quirks & BGE_QUIRK_5705_CORE) == 0) {
   1546 		sc->bge_return_ring_cnt = BGE_RETURN_RING_CNT;
   1547 	} else {
   1548 		sc->bge_return_ring_cnt = BGE_RETURN_RING_CNT_5705;
   1549 	}
   1550 
   1551 	/*
   1552 	 * Initialize the jumbo RX ring control block
   1553 	 * We set the 'ring disabled' bit in the flags
   1554 	 * field until we're actually ready to start
   1555 	 * using this ring (i.e. once we set the MTU
   1556 	 * high enough to require it).
   1557 	 */
   1558 	if ((sc->bge_quirks & BGE_QUIRK_5705_CORE) == 0) {
   1559 		rcb = &sc->bge_rdata->bge_info.bge_jumbo_rx_rcb;
   1560 		bge_set_hostaddr(&rcb->bge_hostaddr,
   1561 		    BGE_RING_DMA_ADDR(sc, bge_rx_jumbo_ring));
   1562 		rcb->bge_maxlen_flags =
   1563 		    BGE_RCB_MAXLEN_FLAGS(BGE_MAX_FRAMELEN,
   1564 			BGE_RCB_FLAG_RING_DISABLED);
   1565 		if (sc->bge_extram)
   1566 			rcb->bge_nicaddr = BGE_EXT_JUMBO_RX_RINGS;
   1567 		else
   1568 			rcb->bge_nicaddr = BGE_JUMBO_RX_RINGS;
   1569 
   1570 		CSR_WRITE_4(sc, BGE_RX_JUMBO_RCB_HADDR_HI,
   1571 		    rcb->bge_hostaddr.bge_addr_hi);
   1572 		CSR_WRITE_4(sc, BGE_RX_JUMBO_RCB_HADDR_LO,
   1573 		    rcb->bge_hostaddr.bge_addr_lo);
   1574 		CSR_WRITE_4(sc, BGE_RX_JUMBO_RCB_MAXLEN_FLAGS,
   1575 		    rcb->bge_maxlen_flags);
   1576 		CSR_WRITE_4(sc, BGE_RX_JUMBO_RCB_NICADDR, rcb->bge_nicaddr);
   1577 
   1578 		/* Set up dummy disabled mini ring RCB */
   1579 		rcb = &sc->bge_rdata->bge_info.bge_mini_rx_rcb;
   1580 		rcb->bge_maxlen_flags = BGE_RCB_MAXLEN_FLAGS(0,
   1581 		    BGE_RCB_FLAG_RING_DISABLED);
   1582 		CSR_WRITE_4(sc, BGE_RX_MINI_RCB_MAXLEN_FLAGS,
   1583 		    rcb->bge_maxlen_flags);
   1584 
   1585 		bus_dmamap_sync(sc->bge_dmatag, sc->bge_ring_map,
   1586 		    offsetof(struct bge_ring_data, bge_info),
   1587 		    sizeof (struct bge_gib),
   1588 		    BUS_DMASYNC_PREREAD|BUS_DMASYNC_PREWRITE);
   1589 	}
   1590 
   1591 	/*
   1592 	 * Set the BD ring replenish thresholds. The recommended
   1593 	 * values are 1/8th the number of descriptors allocated to
   1594 	 * each ring.
   1595 	 */
   1596 	i = BGE_STD_RX_RING_CNT / 8;
   1597 
   1598 	/*
   1599  	 * Use a value of 8 for the following chips to workaround HW errata.
   1600 	 * Some of these chips have been added based on empirical
   1601 	 * evidence (they don't work unless this is done).
   1602 	 */
   1603 	if (BGE_ASICREV(sc->bge_chipid) == BGE_ASICREV_BCM5750 ||
   1604 	    BGE_ASICREV(sc->bge_chipid) == BGE_ASICREV_BCM5752 ||
   1605 	    BGE_ASICREV(sc->bge_chipid) == BGE_ASICREV_BCM5755 ||
   1606 	    BGE_ASICREV(sc->bge_chipid) == BGE_ASICREV_BCM5787)
   1607 		i = 8;
   1608 
   1609 	CSR_WRITE_4(sc, BGE_RBDI_STD_REPL_THRESH, i);
   1610 	CSR_WRITE_4(sc, BGE_RBDI_JUMBO_REPL_THRESH, BGE_JUMBO_RX_RING_CNT/8);
   1611 
   1612 	/*
   1613 	 * Disable all unused send rings by setting the 'ring disabled'
   1614 	 * bit in the flags field of all the TX send ring control blocks.
   1615 	 * These are located in NIC memory.
   1616 	 */
   1617 	rcb_addr = BGE_MEMWIN_START + BGE_SEND_RING_RCB;
   1618 	for (i = 0; i < BGE_TX_RINGS_EXTSSRAM_MAX; i++) {
   1619 		RCB_WRITE_4(sc, rcb_addr, bge_maxlen_flags,
   1620 		    BGE_RCB_MAXLEN_FLAGS(0,BGE_RCB_FLAG_RING_DISABLED));
   1621 		RCB_WRITE_4(sc, rcb_addr, bge_nicaddr, 0);
   1622 		rcb_addr += sizeof(struct bge_rcb);
   1623 	}
   1624 
   1625 	/* Configure TX RCB 0 (we use only the first ring) */
   1626 	rcb_addr = BGE_MEMWIN_START + BGE_SEND_RING_RCB;
   1627 	bge_set_hostaddr(&taddr, BGE_RING_DMA_ADDR(sc, bge_tx_ring));
   1628 	RCB_WRITE_4(sc, rcb_addr, bge_hostaddr.bge_addr_hi, taddr.bge_addr_hi);
   1629 	RCB_WRITE_4(sc, rcb_addr, bge_hostaddr.bge_addr_lo, taddr.bge_addr_lo);
   1630 	RCB_WRITE_4(sc, rcb_addr, bge_nicaddr,
   1631 		    BGE_NIC_TXRING_ADDR(0, BGE_TX_RING_CNT));
   1632 	if ((sc->bge_quirks & BGE_QUIRK_5705_CORE) == 0) {
   1633 		RCB_WRITE_4(sc, rcb_addr, bge_maxlen_flags,
   1634 		    BGE_RCB_MAXLEN_FLAGS(BGE_TX_RING_CNT, 0));
   1635 	}
   1636 
   1637 	/* Disable all unused RX return rings */
   1638 	rcb_addr = BGE_MEMWIN_START + BGE_RX_RETURN_RING_RCB;
   1639 	for (i = 0; i < BGE_RX_RINGS_MAX; i++) {
   1640 		RCB_WRITE_4(sc, rcb_addr, bge_hostaddr.bge_addr_hi, 0);
   1641 		RCB_WRITE_4(sc, rcb_addr, bge_hostaddr.bge_addr_lo, 0);
   1642 		RCB_WRITE_4(sc, rcb_addr, bge_maxlen_flags,
   1643 			    BGE_RCB_MAXLEN_FLAGS(sc->bge_return_ring_cnt,
   1644                                      BGE_RCB_FLAG_RING_DISABLED));
   1645 		RCB_WRITE_4(sc, rcb_addr, bge_nicaddr, 0);
   1646 		CSR_WRITE_4(sc, BGE_MBX_RX_CONS0_LO +
   1647 		    (i * (sizeof(u_int64_t))), 0);
   1648 		rcb_addr += sizeof(struct bge_rcb);
   1649 	}
   1650 
   1651 	/* Initialize RX ring indexes */
   1652 	CSR_WRITE_4(sc, BGE_MBX_RX_STD_PROD_LO, 0);
   1653 	CSR_WRITE_4(sc, BGE_MBX_RX_JUMBO_PROD_LO, 0);
   1654 	CSR_WRITE_4(sc, BGE_MBX_RX_MINI_PROD_LO, 0);
   1655 
   1656 	/*
   1657 	 * Set up RX return ring 0
   1658 	 * Note that the NIC address for RX return rings is 0x00000000.
   1659 	 * The return rings live entirely within the host, so the
   1660 	 * nicaddr field in the RCB isn't used.
   1661 	 */
   1662 	rcb_addr = BGE_MEMWIN_START + BGE_RX_RETURN_RING_RCB;
   1663 	bge_set_hostaddr(&taddr, BGE_RING_DMA_ADDR(sc, bge_rx_return_ring));
   1664 	RCB_WRITE_4(sc, rcb_addr, bge_hostaddr.bge_addr_hi, taddr.bge_addr_hi);
   1665 	RCB_WRITE_4(sc, rcb_addr, bge_hostaddr.bge_addr_lo, taddr.bge_addr_lo);
   1666 	RCB_WRITE_4(sc, rcb_addr, bge_nicaddr, 0x00000000);
   1667 	RCB_WRITE_4(sc, rcb_addr, bge_maxlen_flags,
   1668 	    BGE_RCB_MAXLEN_FLAGS(sc->bge_return_ring_cnt, 0));
   1669 
   1670 	/* Set random backoff seed for TX */
   1671 	CSR_WRITE_4(sc, BGE_TX_RANDOM_BACKOFF,
   1672 	    LLADDR(ifp->if_sadl)[0] + LLADDR(ifp->if_sadl)[1] +
   1673 	    LLADDR(ifp->if_sadl)[2] + LLADDR(ifp->if_sadl)[3] +
   1674 	    LLADDR(ifp->if_sadl)[4] + LLADDR(ifp->if_sadl)[5] +
   1675 	    BGE_TX_BACKOFF_SEED_MASK);
   1676 
   1677 	/* Set inter-packet gap */
   1678 	CSR_WRITE_4(sc, BGE_TX_LENGTHS, 0x2620);
   1679 
   1680 	/*
   1681 	 * Specify which ring to use for packets that don't match
   1682 	 * any RX rules.
   1683 	 */
   1684 	CSR_WRITE_4(sc, BGE_RX_RULES_CFG, 0x08);
   1685 
   1686 	/*
   1687 	 * Configure number of RX lists. One interrupt distribution
   1688 	 * list, sixteen active lists, one bad frames class.
   1689 	 */
   1690 	CSR_WRITE_4(sc, BGE_RXLP_CFG, 0x181);
   1691 
   1692 	/* Inialize RX list placement stats mask. */
   1693 	CSR_WRITE_4(sc, BGE_RXLP_STATS_ENABLE_MASK, 0x007FFFFF);
   1694 	CSR_WRITE_4(sc, BGE_RXLP_STATS_CTL, 0x1);
   1695 
   1696 	/* Disable host coalescing until we get it set up */
   1697 	CSR_WRITE_4(sc, BGE_HCC_MODE, 0x00000000);
   1698 
   1699 	/* Poll to make sure it's shut down. */
   1700 	for (i = 0; i < BGE_TIMEOUT; i++) {
   1701 		if (!(CSR_READ_4(sc, BGE_HCC_MODE) & BGE_HCCMODE_ENABLE))
   1702 			break;
   1703 		DELAY(10);
   1704 	}
   1705 
   1706 	if (i == BGE_TIMEOUT) {
   1707 		printf("%s: host coalescing engine failed to idle\n",
   1708 		    sc->bge_dev.dv_xname);
   1709 		return(ENXIO);
   1710 	}
   1711 
   1712 	/* Set up host coalescing defaults */
   1713 	CSR_WRITE_4(sc, BGE_HCC_RX_COAL_TICKS, sc->bge_rx_coal_ticks);
   1714 	CSR_WRITE_4(sc, BGE_HCC_TX_COAL_TICKS, sc->bge_tx_coal_ticks);
   1715 	CSR_WRITE_4(sc, BGE_HCC_RX_MAX_COAL_BDS, sc->bge_rx_max_coal_bds);
   1716 	CSR_WRITE_4(sc, BGE_HCC_TX_MAX_COAL_BDS, sc->bge_tx_max_coal_bds);
   1717 	if ((sc->bge_quirks & BGE_QUIRK_5705_CORE) == 0) {
   1718 		CSR_WRITE_4(sc, BGE_HCC_RX_COAL_TICKS_INT, 0);
   1719 		CSR_WRITE_4(sc, BGE_HCC_TX_COAL_TICKS_INT, 0);
   1720 	}
   1721 	CSR_WRITE_4(sc, BGE_HCC_RX_MAX_COAL_BDS_INT, 0);
   1722 	CSR_WRITE_4(sc, BGE_HCC_TX_MAX_COAL_BDS_INT, 0);
   1723 
   1724 	/* Set up address of statistics block */
   1725 	if ((sc->bge_quirks & BGE_QUIRK_5705_CORE) == 0) {
   1726 		bge_set_hostaddr(&taddr,
   1727 		    BGE_RING_DMA_ADDR(sc, bge_info.bge_stats));
   1728 		CSR_WRITE_4(sc, BGE_HCC_STATS_TICKS, sc->bge_stat_ticks);
   1729 		CSR_WRITE_4(sc, BGE_HCC_STATS_BASEADDR, BGE_STATS_BLOCK);
   1730 		CSR_WRITE_4(sc, BGE_HCC_STATS_ADDR_HI, taddr.bge_addr_hi);
   1731 		CSR_WRITE_4(sc, BGE_HCC_STATS_ADDR_LO, taddr.bge_addr_lo);
   1732 	}
   1733 
   1734 	/* Set up address of status block */
   1735 	bge_set_hostaddr(&taddr, BGE_RING_DMA_ADDR(sc, bge_status_block));
   1736 	CSR_WRITE_4(sc, BGE_HCC_STATUSBLK_BASEADDR, BGE_STATUS_BLOCK);
   1737 	CSR_WRITE_4(sc, BGE_HCC_STATUSBLK_ADDR_HI, taddr.bge_addr_hi);
   1738 	CSR_WRITE_4(sc, BGE_HCC_STATUSBLK_ADDR_LO, taddr.bge_addr_lo);
   1739 	sc->bge_rdata->bge_status_block.bge_idx[0].bge_rx_prod_idx = 0;
   1740 	sc->bge_rdata->bge_status_block.bge_idx[0].bge_tx_cons_idx = 0;
   1741 
   1742 	/* Turn on host coalescing state machine */
   1743 	CSR_WRITE_4(sc, BGE_HCC_MODE, BGE_HCCMODE_ENABLE);
   1744 
   1745 	/* Turn on RX BD completion state machine and enable attentions */
   1746 	CSR_WRITE_4(sc, BGE_RBDC_MODE,
   1747 	    BGE_RBDCMODE_ENABLE|BGE_RBDCMODE_ATTN);
   1748 
   1749 	/* Turn on RX list placement state machine */
   1750 	CSR_WRITE_4(sc, BGE_RXLP_MODE, BGE_RXLPMODE_ENABLE);
   1751 
   1752 	/* Turn on RX list selector state machine. */
   1753 	if ((sc->bge_quirks & BGE_QUIRK_5705_CORE) == 0) {
   1754 		CSR_WRITE_4(sc, BGE_RXLS_MODE, BGE_RXLSMODE_ENABLE);
   1755 	}
   1756 
   1757 	/* Turn on DMA, clear stats */
   1758 	CSR_WRITE_4(sc, BGE_MAC_MODE, BGE_MACMODE_TXDMA_ENB|
   1759 	    BGE_MACMODE_RXDMA_ENB|BGE_MACMODE_RX_STATS_CLEAR|
   1760 	    BGE_MACMODE_TX_STATS_CLEAR|BGE_MACMODE_RX_STATS_ENB|
   1761 	    BGE_MACMODE_TX_STATS_ENB|BGE_MACMODE_FRMHDR_DMA_ENB|
   1762 	    (sc->bge_tbi ? BGE_PORTMODE_TBI : BGE_PORTMODE_MII));
   1763 
   1764 	/* Set misc. local control, enable interrupts on attentions */
   1765 	sc->bge_local_ctrl_reg = BGE_MLC_INTR_ONATTN | BGE_MLC_AUTO_EEPROM;
   1766 
   1767 #ifdef notdef
   1768 	/* Assert GPIO pins for PHY reset */
   1769 	BGE_SETBIT(sc, BGE_MISC_LOCAL_CTL, BGE_MLC_MISCIO_OUT0|
   1770 	    BGE_MLC_MISCIO_OUT1|BGE_MLC_MISCIO_OUT2);
   1771 	BGE_SETBIT(sc, BGE_MISC_LOCAL_CTL, BGE_MLC_MISCIO_OUTEN0|
   1772 	    BGE_MLC_MISCIO_OUTEN1|BGE_MLC_MISCIO_OUTEN2);
   1773 #endif
   1774 
   1775 #if defined(not_quite_yet)
   1776 	/* Linux driver enables enable gpio pin #1 on 5700s */
   1777 	if (sc->bge_chipid == BGE_CHIPID_BCM5700) {
   1778 		sc->bge_local_ctrl_reg |=
   1779 		  (BGE_MLC_MISCIO_OUT1|BGE_MLC_MISCIO_OUTEN1);
   1780 	}
   1781 #endif
   1782 	CSR_WRITE_4(sc, BGE_MISC_LOCAL_CTL, sc->bge_local_ctrl_reg);
   1783 
   1784 	/* Turn on DMA completion state machine */
   1785 	if ((sc->bge_quirks & BGE_QUIRK_5705_CORE) == 0) {
   1786 		CSR_WRITE_4(sc, BGE_DMAC_MODE, BGE_DMACMODE_ENABLE);
   1787 	}
   1788 
   1789 	/* Turn on write DMA state machine */
   1790 	{
   1791 		uint32_t bge_wdma_mode =
   1792 			BGE_WDMAMODE_ENABLE|BGE_WDMAMODE_ALL_ATTNS;
   1793 
   1794 		if (BGE_ASICREV(sc->bge_chipid) == BGE_ASICREV_BCM5755 ||
   1795 		    BGE_ASICREV(sc->bge_chipid) == BGE_ASICREV_BCM5787)
   1796 		  /* Enable host coalescing bug fix; see Linux tg3.c */
   1797 		  bge_wdma_mode |= (1 << 29);
   1798 
   1799 		CSR_WRITE_4(sc, BGE_WDMA_MODE, bge_wdma_mode);
   1800         }
   1801 
   1802 	/* Turn on read DMA state machine */
   1803 	{
   1804 		uint32_t dma_read_modebits;
   1805 
   1806 		dma_read_modebits =
   1807 		  BGE_RDMAMODE_ENABLE | BGE_RDMAMODE_ALL_ATTNS;
   1808 
   1809 		if (sc->bge_pcie && 0) {
   1810 			dma_read_modebits |= BGE_RDMA_MODE_FIFO_LONG_BURST;
   1811 		} else if ((sc->bge_quirks & BGE_QUIRK_5705_CORE)) {
   1812 			dma_read_modebits |= BGE_RDMA_MODE_FIFO_SIZE_128;
   1813 		}
   1814 
   1815 		/* XXX broadcom-supplied linux driver; undocumented */
   1816 		if (BGE_IS_5750_OR_BEYOND(sc)) {
   1817  			/*
   1818 			 * XXX: magic values.
   1819 			 * From Broadcom-supplied Linux driver;  apparently
   1820 			 * required to workaround a DMA bug affecting TSO
   1821 			 * on bcm575x/bcm5721?
   1822 			 */
   1823 			dma_read_modebits |= (1 << 27);
   1824 		}
   1825 		CSR_WRITE_4(sc, BGE_RDMA_MODE, dma_read_modebits);
   1826 	}
   1827 
   1828 	/* Turn on RX data completion state machine */
   1829 	CSR_WRITE_4(sc, BGE_RDC_MODE, BGE_RDCMODE_ENABLE);
   1830 
   1831 	/* Turn on RX BD initiator state machine */
   1832 	CSR_WRITE_4(sc, BGE_RBDI_MODE, BGE_RBDIMODE_ENABLE);
   1833 
   1834 	/* Turn on RX data and RX BD initiator state machine */
   1835 	CSR_WRITE_4(sc, BGE_RDBDI_MODE, BGE_RDBDIMODE_ENABLE);
   1836 
   1837 	/* Turn on Mbuf cluster free state machine */
   1838 	if ((sc->bge_quirks & BGE_QUIRK_5705_CORE) == 0) {
   1839 		CSR_WRITE_4(sc, BGE_MBCF_MODE, BGE_MBCFMODE_ENABLE);
   1840 	}
   1841 
   1842 	/* Turn on send BD completion state machine */
   1843 	CSR_WRITE_4(sc, BGE_SBDC_MODE, BGE_SBDCMODE_ENABLE);
   1844 
   1845 	/* Turn on send data completion state machine */
   1846 	CSR_WRITE_4(sc, BGE_SDC_MODE, BGE_SDCMODE_ENABLE);
   1847 
   1848 	/* Turn on send data initiator state machine */
   1849 	if (BGE_IS_5750_OR_BEYOND(sc)) {
   1850 		/* XXX: magic value from Linux driver */
   1851 		CSR_WRITE_4(sc, BGE_SDI_MODE, BGE_SDIMODE_ENABLE | 0x08);
   1852 	} else {
   1853 		CSR_WRITE_4(sc, BGE_SDI_MODE, BGE_SDIMODE_ENABLE);
   1854 	}
   1855 
   1856 	/* Turn on send BD initiator state machine */
   1857 	CSR_WRITE_4(sc, BGE_SBDI_MODE, BGE_SBDIMODE_ENABLE);
   1858 
   1859 	/* Turn on send BD selector state machine */
   1860 	CSR_WRITE_4(sc, BGE_SRS_MODE, BGE_SRSMODE_ENABLE);
   1861 
   1862 	CSR_WRITE_4(sc, BGE_SDI_STATS_ENABLE_MASK, 0x007FFFFF);
   1863 	CSR_WRITE_4(sc, BGE_SDI_STATS_CTL,
   1864 	    BGE_SDISTATSCTL_ENABLE|BGE_SDISTATSCTL_FASTER);
   1865 
   1866 	/* ack/clear link change events */
   1867 	CSR_WRITE_4(sc, BGE_MAC_STS, BGE_MACSTAT_SYNC_CHANGED|
   1868 	    BGE_MACSTAT_CFG_CHANGED);
   1869 	CSR_WRITE_4(sc, BGE_MI_STS, 0);
   1870 
   1871 	/* Enable PHY auto polling (for MII/GMII only) */
   1872 	if (sc->bge_tbi) {
   1873 		CSR_WRITE_4(sc, BGE_MI_STS, BGE_MISTS_LINK);
   1874  	} else {
   1875 		BGE_SETBIT(sc, BGE_MI_MODE, BGE_MIMODE_AUTOPOLL|10<<16);
   1876 		if (sc->bge_quirks & BGE_QUIRK_LINK_STATE_BROKEN)
   1877 			CSR_WRITE_4(sc, BGE_MAC_EVT_ENB,
   1878 			    BGE_EVTENB_MI_INTERRUPT);
   1879 	}
   1880 
   1881 	/* Enable link state change attentions. */
   1882 	BGE_SETBIT(sc, BGE_MAC_EVT_ENB, BGE_EVTENB_LINK_CHANGED);
   1883 
   1884 	return(0);
   1885 }
   1886 
   1887 static const struct bge_revision {
   1888 	uint32_t		br_chipid;
   1889 	uint32_t		br_quirks;
   1890 	const char		*br_name;
   1891 } bge_revisions[] = {
   1892 	{ BGE_CHIPID_BCM5700_A0,
   1893 	  BGE_QUIRK_LINK_STATE_BROKEN,
   1894 	  "BCM5700 A0" },
   1895 
   1896 	{ BGE_CHIPID_BCM5700_A1,
   1897 	  BGE_QUIRK_LINK_STATE_BROKEN,
   1898 	  "BCM5700 A1" },
   1899 
   1900 	{ BGE_CHIPID_BCM5700_B0,
   1901 	  BGE_QUIRK_LINK_STATE_BROKEN|BGE_QUIRK_CSUM_BROKEN|BGE_QUIRK_5700_COMMON,
   1902 	  "BCM5700 B0" },
   1903 
   1904 	{ BGE_CHIPID_BCM5700_B1,
   1905 	  BGE_QUIRK_LINK_STATE_BROKEN|BGE_QUIRK_5700_COMMON,
   1906 	  "BCM5700 B1" },
   1907 
   1908 	{ BGE_CHIPID_BCM5700_B2,
   1909 	  BGE_QUIRK_LINK_STATE_BROKEN|BGE_QUIRK_5700_COMMON,
   1910 	  "BCM5700 B2" },
   1911 
   1912 	{ BGE_CHIPID_BCM5700_B3,
   1913 	  BGE_QUIRK_LINK_STATE_BROKEN|BGE_QUIRK_5700_COMMON,
   1914 	  "BCM5700 B3" },
   1915 
   1916 	/* This is treated like a BCM5700 Bx */
   1917 	{ BGE_CHIPID_BCM5700_ALTIMA,
   1918 	  BGE_QUIRK_LINK_STATE_BROKEN|BGE_QUIRK_5700_COMMON,
   1919 	  "BCM5700 Altima" },
   1920 
   1921 	{ BGE_CHIPID_BCM5700_C0,
   1922 	  0,
   1923 	  "BCM5700 C0" },
   1924 
   1925 	{ BGE_CHIPID_BCM5701_A0,
   1926 	  0, /*XXX really, just not known */
   1927 	  "BCM5701 A0" },
   1928 
   1929 	{ BGE_CHIPID_BCM5701_B0,
   1930 	  BGE_QUIRK_PCIX_DMA_ALIGN_BUG,
   1931 	  "BCM5701 B0" },
   1932 
   1933 	{ BGE_CHIPID_BCM5701_B2,
   1934 	  BGE_QUIRK_ONLY_PHY_1|BGE_QUIRK_PCIX_DMA_ALIGN_BUG,
   1935 	  "BCM5701 B2" },
   1936 
   1937 	{ BGE_CHIPID_BCM5701_B5,
   1938 	  BGE_QUIRK_ONLY_PHY_1|BGE_QUIRK_PCIX_DMA_ALIGN_BUG,
   1939 	  "BCM5701 B5" },
   1940 
   1941 	{ BGE_CHIPID_BCM5703_A0,
   1942 	  0,
   1943 	  "BCM5703 A0" },
   1944 
   1945 	{ BGE_CHIPID_BCM5703_A1,
   1946 	  0,
   1947 	  "BCM5703 A1" },
   1948 
   1949 	{ BGE_CHIPID_BCM5703_A2,
   1950 	  BGE_QUIRK_ONLY_PHY_1,
   1951 	  "BCM5703 A2" },
   1952 
   1953 	{ BGE_CHIPID_BCM5703_A3,
   1954 	  BGE_QUIRK_ONLY_PHY_1,
   1955 	  "BCM5703 A3" },
   1956 
   1957 	{ BGE_CHIPID_BCM5703_B0,
   1958 	  BGE_QUIRK_ONLY_PHY_1,
   1959 	  "BCM5703 B0" },
   1960 
   1961 	{ BGE_CHIPID_BCM5704_A0,
   1962   	  BGE_QUIRK_ONLY_PHY_1|BGE_QUIRK_FEWER_MBUFS,
   1963 	  "BCM5704 A0" },
   1964 
   1965 	{ BGE_CHIPID_BCM5704_A1,
   1966   	  BGE_QUIRK_ONLY_PHY_1|BGE_QUIRK_FEWER_MBUFS,
   1967 	  "BCM5704 A1" },
   1968 
   1969 	{ BGE_CHIPID_BCM5704_A2,
   1970   	  BGE_QUIRK_ONLY_PHY_1|BGE_QUIRK_FEWER_MBUFS,
   1971 	  "BCM5704 A2" },
   1972 
   1973 	{ BGE_CHIPID_BCM5704_A3,
   1974   	  BGE_QUIRK_ONLY_PHY_1|BGE_QUIRK_FEWER_MBUFS,
   1975 	  "BCM5704 A3" },
   1976 
   1977 	{ BGE_CHIPID_BCM5705_A0,
   1978 	  BGE_QUIRK_ONLY_PHY_1|BGE_QUIRK_5705_CORE,
   1979 	  "BCM5705 A0" },
   1980 
   1981 	{ BGE_CHIPID_BCM5705_A1,
   1982 	  BGE_QUIRK_ONLY_PHY_1|BGE_QUIRK_5705_CORE,
   1983 	  "BCM5705 A1" },
   1984 
   1985 	{ BGE_CHIPID_BCM5705_A2,
   1986 	  BGE_QUIRK_ONLY_PHY_1|BGE_QUIRK_5705_CORE,
   1987 	  "BCM5705 A2" },
   1988 
   1989 	{ BGE_CHIPID_BCM5705_A3,
   1990 	  BGE_QUIRK_ONLY_PHY_1|BGE_QUIRK_5705_CORE,
   1991 	  "BCM5705 A3" },
   1992 
   1993 	{ BGE_CHIPID_BCM5750_A0,
   1994 	  BGE_QUIRK_ONLY_PHY_1|BGE_QUIRK_5705_CORE,
   1995 	  "BCM5750 A0" },
   1996 
   1997 	{ BGE_CHIPID_BCM5750_A1,
   1998 	  BGE_QUIRK_ONLY_PHY_1|BGE_QUIRK_5705_CORE,
   1999 	  "BCM5750 A1" },
   2000 
   2001 	{ BGE_CHIPID_BCM5751_A1,
   2002 	  BGE_QUIRK_ONLY_PHY_1|BGE_QUIRK_5705_CORE,
   2003 	  "BCM5751 A1" },
   2004 
   2005 	{ BGE_CHIPID_BCM5752_A0,
   2006 	  BGE_QUIRK_ONLY_PHY_1|BGE_QUIRK_5705_CORE,
   2007 	  "BCM5752 A0" },
   2008 
   2009 	{ BGE_CHIPID_BCM5752_A1,
   2010 	  BGE_QUIRK_ONLY_PHY_1|BGE_QUIRK_5705_CORE,
   2011 	  "BCM5752 A1" },
   2012 
   2013 	{ BGE_CHIPID_BCM5752_A2,
   2014 	  BGE_QUIRK_ONLY_PHY_1|BGE_QUIRK_5705_CORE,
   2015 	  "BCM5752 A2" },
   2016 
   2017 	{ BGE_CHIPID_BCM5787_A0,
   2018 	  BGE_QUIRK_ONLY_PHY_1|BGE_QUIRK_5705_CORE,
   2019 	  "BCM5754/5787 A0" },
   2020 
   2021 	{ BGE_CHIPID_BCM5787_A1,
   2022 	  BGE_QUIRK_ONLY_PHY_1|BGE_QUIRK_5705_CORE,
   2023 	  "BCM5754/5787 A1" },
   2024 
   2025 	{ BGE_CHIPID_BCM5787_A2,
   2026 	  BGE_QUIRK_ONLY_PHY_1|BGE_QUIRK_5705_CORE,
   2027 	  "BCM5754/5787 A2" },
   2028 
   2029 	{ 0, 0, NULL }
   2030 };
   2031 
   2032 /*
   2033  * Some defaults for major revisions, so that newer steppings
   2034  * that we don't know about have a shot at working.
   2035  */
   2036 static const struct bge_revision bge_majorrevs[] = {
   2037 	{ BGE_ASICREV_BCM5700,
   2038 	  BGE_QUIRK_LINK_STATE_BROKEN,
   2039 	  "unknown BCM5700" },
   2040 
   2041 	{ BGE_ASICREV_BCM5701,
   2042 	  BGE_QUIRK_PCIX_DMA_ALIGN_BUG,
   2043 	  "unknown BCM5701" },
   2044 
   2045 	{ BGE_ASICREV_BCM5703,
   2046 	  0,
   2047 	  "unknown BCM5703" },
   2048 
   2049 	{ BGE_ASICREV_BCM5704,
   2050 	  BGE_QUIRK_ONLY_PHY_1,
   2051 	  "unknown BCM5704" },
   2052 
   2053 	{ BGE_ASICREV_BCM5705,
   2054 	  BGE_QUIRK_ONLY_PHY_1|BGE_QUIRK_5705_CORE,
   2055 	  "unknown BCM5705" },
   2056 
   2057 	{ BGE_ASICREV_BCM5750,
   2058 	  BGE_QUIRK_ONLY_PHY_1|BGE_QUIRK_5705_CORE,
   2059 	  "unknown BCM575x family" },
   2060 
   2061 	{ BGE_ASICREV_BCM5714_A0,
   2062 	  BGE_QUIRK_ONLY_PHY_1|BGE_QUIRK_5705_CORE,
   2063 	  "unknown BCM5714" },
   2064 
   2065 	{ BGE_ASICREV_BCM5714,
   2066 	  BGE_QUIRK_ONLY_PHY_1|BGE_QUIRK_5705_CORE,
   2067 	  "unknown BCM5714" },
   2068 
   2069 	{ BGE_ASICREV_BCM5752,
   2070 	  BGE_QUIRK_ONLY_PHY_1|BGE_QUIRK_5705_CORE,
   2071 	  "unknown BCM5752 family" },
   2072 
   2073 	{ BGE_ASICREV_BCM5755,
   2074 	  BGE_QUIRK_ONLY_PHY_1|BGE_QUIRK_5705_CORE,
   2075 	  "unknown BCM5755" },
   2076 
   2077 	{ BGE_ASICREV_BCM5780,
   2078 	  BGE_QUIRK_ONLY_PHY_1|BGE_QUIRK_5705_CORE,
   2079 	  "unknown BCM5780" },
   2080 
   2081 	{ BGE_ASICREV_BCM5787,
   2082 	  BGE_QUIRK_ONLY_PHY_1|BGE_QUIRK_5705_CORE,
   2083 	  "unknown BCM5787" },
   2084 
   2085 	{ 0,
   2086 	  0,
   2087 	  NULL }
   2088 };
   2089 
   2090 
   2091 static const struct bge_revision *
   2092 bge_lookup_rev(uint32_t chipid)
   2093 {
   2094 	const struct bge_revision *br;
   2095 
   2096 	for (br = bge_revisions; br->br_name != NULL; br++) {
   2097 		if (br->br_chipid == chipid)
   2098 			return (br);
   2099 	}
   2100 
   2101 	for (br = bge_majorrevs; br->br_name != NULL; br++) {
   2102 		if (br->br_chipid == BGE_ASICREV(chipid))
   2103 			return (br);
   2104 	}
   2105 
   2106 	return (NULL);
   2107 }
   2108 
   2109 static const struct bge_product {
   2110 	pci_vendor_id_t		bp_vendor;
   2111 	pci_product_id_t	bp_product;
   2112 	const char		*bp_name;
   2113 } bge_products[] = {
   2114 	/*
   2115 	 * The BCM5700 documentation seems to indicate that the hardware
   2116 	 * still has the Alteon vendor ID burned into it, though it
   2117 	 * should always be overridden by the value in the EEPROM.  We'll
   2118 	 * check for it anyway.
   2119 	 */
   2120 	{ PCI_VENDOR_ALTEON,
   2121 	  PCI_PRODUCT_ALTEON_BCM5700,
   2122 	  "Broadcom BCM5700 Gigabit Ethernet",
   2123 	  },
   2124 	{ PCI_VENDOR_ALTEON,
   2125 	  PCI_PRODUCT_ALTEON_BCM5701,
   2126 	  "Broadcom BCM5701 Gigabit Ethernet",
   2127 	  },
   2128 
   2129 	{ PCI_VENDOR_ALTIMA,
   2130 	  PCI_PRODUCT_ALTIMA_AC1000,
   2131 	  "Altima AC1000 Gigabit Ethernet",
   2132 	  },
   2133 	{ PCI_VENDOR_ALTIMA,
   2134 	  PCI_PRODUCT_ALTIMA_AC1001,
   2135 	  "Altima AC1001 Gigabit Ethernet",
   2136 	   },
   2137 	{ PCI_VENDOR_ALTIMA,
   2138 	  PCI_PRODUCT_ALTIMA_AC9100,
   2139 	  "Altima AC9100 Gigabit Ethernet",
   2140 	  },
   2141 
   2142 	{ PCI_VENDOR_BROADCOM,
   2143 	  PCI_PRODUCT_BROADCOM_BCM5700,
   2144 	  "Broadcom BCM5700 Gigabit Ethernet",
   2145 	  },
   2146 	{ PCI_VENDOR_BROADCOM,
   2147 	  PCI_PRODUCT_BROADCOM_BCM5701,
   2148 	  "Broadcom BCM5701 Gigabit Ethernet",
   2149 	  },
   2150 	{ PCI_VENDOR_BROADCOM,
   2151 	  PCI_PRODUCT_BROADCOM_BCM5702,
   2152 	  "Broadcom BCM5702 Gigabit Ethernet",
   2153 	  },
   2154 	{ PCI_VENDOR_BROADCOM,
   2155 	  PCI_PRODUCT_BROADCOM_BCM5702X,
   2156 	  "Broadcom BCM5702X Gigabit Ethernet" },
   2157 
   2158 	{ PCI_VENDOR_BROADCOM,
   2159 	  PCI_PRODUCT_BROADCOM_BCM5703,
   2160 	  "Broadcom BCM5703 Gigabit Ethernet",
   2161 	  },
   2162 	{ PCI_VENDOR_BROADCOM,
   2163 	  PCI_PRODUCT_BROADCOM_BCM5703X,
   2164 	  "Broadcom BCM5703X Gigabit Ethernet",
   2165 	  },
   2166 	{ PCI_VENDOR_BROADCOM,
   2167 	  PCI_PRODUCT_BROADCOM_BCM5703_ALT,
   2168 	  "Broadcom BCM5703 Gigabit Ethernet",
   2169 	  },
   2170 
   2171    	{ PCI_VENDOR_BROADCOM,
   2172 	  PCI_PRODUCT_BROADCOM_BCM5704C,
   2173 	  "Broadcom BCM5704C Dual Gigabit Ethernet",
   2174 	  },
   2175    	{ PCI_VENDOR_BROADCOM,
   2176 	  PCI_PRODUCT_BROADCOM_BCM5704S,
   2177 	  "Broadcom BCM5704S Dual Gigabit Ethernet",
   2178 	  },
   2179 
   2180    	{ PCI_VENDOR_BROADCOM,
   2181 	  PCI_PRODUCT_BROADCOM_BCM5705,
   2182 	  "Broadcom BCM5705 Gigabit Ethernet",
   2183 	  },
   2184    	{ PCI_VENDOR_BROADCOM,
   2185 	  PCI_PRODUCT_BROADCOM_BCM5705K,
   2186 	  "Broadcom BCM5705K Gigabit Ethernet",
   2187 	  },
   2188    	{ PCI_VENDOR_BROADCOM,
   2189 	  PCI_PRODUCT_BROADCOM_BCM5705M,
   2190 	  "Broadcom BCM5705M Gigabit Ethernet",
   2191 	  },
   2192    	{ PCI_VENDOR_BROADCOM,
   2193 	  PCI_PRODUCT_BROADCOM_BCM5705M_ALT,
   2194 	  "Broadcom BCM5705M Gigabit Ethernet",
   2195 	  },
   2196 
   2197 	{ PCI_VENDOR_BROADCOM,
   2198 	  PCI_PRODUCT_BROADCOM_BCM5714,
   2199 	  "Broadcom BCM5714/5715 Gigabit Ethernet",
   2200 	  },
   2201 	{ PCI_VENDOR_BROADCOM,
   2202 	  PCI_PRODUCT_BROADCOM_BCM5715,
   2203 	  "Broadcom BCM5714/5715 Gigabit Ethernet",
   2204 	  },
   2205 	{ PCI_VENDOR_BROADCOM,
   2206 	  PCI_PRODUCT_BROADCOM_BCM5789,
   2207 	  "Broadcom BCM5789 Gigabit Ethernet",
   2208 	  },
   2209 
   2210 	{ PCI_VENDOR_BROADCOM,
   2211 	  PCI_PRODUCT_BROADCOM_BCM5721,
   2212 	  "Broadcom BCM5721 Gigabit Ethernet",
   2213 	  },
   2214 
   2215 	{ PCI_VENDOR_BROADCOM,
   2216 	  PCI_PRODUCT_BROADCOM_BCM5750,
   2217 	  "Broadcom BCM5750 Gigabit Ethernet",
   2218 	  },
   2219 
   2220 	{ PCI_VENDOR_BROADCOM,
   2221 	  PCI_PRODUCT_BROADCOM_BCM5750M,
   2222 	  "Broadcom BCM5750M Gigabit Ethernet",
   2223 	  },
   2224 
   2225 	{ PCI_VENDOR_BROADCOM,
   2226 	  PCI_PRODUCT_BROADCOM_BCM5751,
   2227 	  "Broadcom BCM5751 Gigabit Ethernet",
   2228 	  },
   2229 
   2230 	{ PCI_VENDOR_BROADCOM,
   2231 	  PCI_PRODUCT_BROADCOM_BCM5751M,
   2232 	  "Broadcom BCM5751M Gigabit Ethernet",
   2233 	  },
   2234 
   2235 	{ PCI_VENDOR_BROADCOM,
   2236 	  PCI_PRODUCT_BROADCOM_BCM5752,
   2237 	  "Broadcom BCM5752 Gigabit Ethernet",
   2238 	  },
   2239 
   2240 	{ PCI_VENDOR_BROADCOM,
   2241 	  PCI_PRODUCT_BROADCOM_BCM5752M,
   2242 	  "Broadcom BCM5752M Gigabit Ethernet",
   2243 	  },
   2244 
   2245 	{ PCI_VENDOR_BROADCOM,
   2246 	  PCI_PRODUCT_BROADCOM_BCM5753,
   2247 	  "Broadcom BCM5753 Gigabit Ethernet",
   2248 	  },
   2249 
   2250 	{ PCI_VENDOR_BROADCOM,
   2251 	  PCI_PRODUCT_BROADCOM_BCM5753M,
   2252 	  "Broadcom BCM5753M Gigabit Ethernet",
   2253 	  },
   2254 
   2255 	{ PCI_VENDOR_BROADCOM,
   2256 	  PCI_PRODUCT_BROADCOM_BCM5754,
   2257 	  "Broadcom BCM5754 Gigabit Ethernet",
   2258 	},
   2259 
   2260 	{ PCI_VENDOR_BROADCOM,
   2261 	  PCI_PRODUCT_BROADCOM_BCM5754M,
   2262 	  "Broadcom BCM5754M Gigabit Ethernet",
   2263 	},
   2264 
   2265 	{ PCI_VENDOR_BROADCOM,
   2266 	  PCI_PRODUCT_BROADCOM_BCM5755,
   2267 	  "Broadcom BCM5755 Gigabit Ethernet",
   2268 	},
   2269 
   2270 	{ PCI_VENDOR_BROADCOM,
   2271 	  PCI_PRODUCT_BROADCOM_BCM5755M,
   2272 	  "Broadcom BCM5755M Gigabit Ethernet",
   2273 	},
   2274 
   2275    	{ PCI_VENDOR_BROADCOM,
   2276 	  PCI_PRODUCT_BROADCOM_BCM5780,
   2277 	  "Broadcom BCM5780 Gigabit Ethernet",
   2278 	  },
   2279 
   2280    	{ PCI_VENDOR_BROADCOM,
   2281 	  PCI_PRODUCT_BROADCOM_BCM5780S,
   2282 	  "Broadcom BCM5780S Gigabit Ethernet",
   2283 	  },
   2284 
   2285    	{ PCI_VENDOR_BROADCOM,
   2286 	  PCI_PRODUCT_BROADCOM_BCM5782,
   2287 	  "Broadcom BCM5782 Gigabit Ethernet",
   2288 	},
   2289 
   2290 	{ PCI_VENDOR_BROADCOM,
   2291 	  PCI_PRODUCT_BROADCOM_BCM5787,
   2292 	  "Broadcom BCM5787 Gigabit Ethernet",
   2293 	},
   2294 
   2295 	{ PCI_VENDOR_BROADCOM,
   2296 	  PCI_PRODUCT_BROADCOM_BCM5787M,
   2297 	  "Broadcom BCM5787M Gigabit Ethernet",
   2298 	},
   2299 
   2300    	{ PCI_VENDOR_BROADCOM,
   2301 	  PCI_PRODUCT_BROADCOM_BCM5788,
   2302 	  "Broadcom BCM5788 Gigabit Ethernet",
   2303 	  },
   2304    	{ PCI_VENDOR_BROADCOM,
   2305 	  PCI_PRODUCT_BROADCOM_BCM5789,
   2306 	  "Broadcom BCM5789 Gigabit Ethernet",
   2307 	  },
   2308 
   2309    	{ PCI_VENDOR_BROADCOM,
   2310 	  PCI_PRODUCT_BROADCOM_BCM5901,
   2311 	  "Broadcom BCM5901 Fast Ethernet",
   2312 	  },
   2313    	{ PCI_VENDOR_BROADCOM,
   2314 	  PCI_PRODUCT_BROADCOM_BCM5901A2,
   2315 	  "Broadcom BCM5901A2 Fast Ethernet",
   2316 	  },
   2317 
   2318 	{ PCI_VENDOR_SCHNEIDERKOCH,
   2319 	  PCI_PRODUCT_SCHNEIDERKOCH_SK_9DX1,
   2320 	  "SysKonnect SK-9Dx1 Gigabit Ethernet",
   2321 	  },
   2322 
   2323 	{ PCI_VENDOR_3COM,
   2324 	  PCI_PRODUCT_3COM_3C996,
   2325 	  "3Com 3c996 Gigabit Ethernet",
   2326 	  },
   2327 
   2328 	{ 0,
   2329 	  0,
   2330 	  NULL },
   2331 };
   2332 
   2333 static const struct bge_product *
   2334 bge_lookup(const struct pci_attach_args *pa)
   2335 {
   2336 	const struct bge_product *bp;
   2337 
   2338 	for (bp = bge_products; bp->bp_name != NULL; bp++) {
   2339 		if (PCI_VENDOR(pa->pa_id) == bp->bp_vendor &&
   2340 		    PCI_PRODUCT(pa->pa_id) == bp->bp_product)
   2341 			return (bp);
   2342 	}
   2343 
   2344 	return (NULL);
   2345 }
   2346 
   2347 static int
   2348 bge_setpowerstate(struct bge_softc *sc, int powerlevel)
   2349 {
   2350 #ifdef NOTYET
   2351 	u_int32_t pm_ctl = 0;
   2352 
   2353 	/* XXX FIXME: make sure indirect accesses enabled? */
   2354 	pm_ctl = pci_conf_read(sc->bge_dev, BGE_PCI_MISC_CTL, 4);
   2355 	pm_ctl |= BGE_PCIMISCCTL_INDIRECT_ACCESS;
   2356 	pci_write_config(sc->bge_dev, BGE_PCI_MISC_CTL, pm_ctl, 4);
   2357 
   2358 	/* clear the PME_assert bit and power state bits, enable PME */
   2359 	pm_ctl = pci_conf_read(sc->bge_dev, BGE_PCI_PWRMGMT_CMD, 2);
   2360 	pm_ctl &= ~PCIM_PSTAT_DMASK;
   2361 	pm_ctl |= (1 << 8);
   2362 
   2363 	if (powerlevel == 0) {
   2364 		pm_ctl |= PCIM_PSTAT_D0;
   2365 		pci_write_config(sc->bge_dev, BGE_PCI_PWRMGMT_CMD,
   2366 		    pm_ctl, 2);
   2367 		DELAY(10000);
   2368 		CSR_WRITE_4(sc, BGE_MISC_LOCAL_CTL, sc->bge_local_ctrl_reg);
   2369 		DELAY(10000);
   2370 
   2371 #ifdef NOTYET
   2372 		/* XXX FIXME: write 0x02 to phy aux_Ctrl reg */
   2373 		bge_miibus_writereg(sc->bge_dev, 1, 0x18, 0x02);
   2374 #endif
   2375 		DELAY(40); DELAY(40); DELAY(40);
   2376 		DELAY(10000);	/* above not quite adequate on 5700 */
   2377 		return 0;
   2378 	}
   2379 
   2380 
   2381 	/*
   2382 	 * Entering ACPI power states D1-D3 is achieved by wiggling
   2383 	 * GMII gpio pins. Example code assumes all hardware vendors
   2384 	 * followed Broadom's sample pcb layout. Until we verify that
   2385 	 * for all supported OEM cards, states D1-D3 are  unsupported.
   2386 	 */
   2387 	printf("%s: power state %d unimplemented; check GPIO pins\n",
   2388 	       sc->bge_dev.dv_xname, powerlevel);
   2389 #endif
   2390 	return EOPNOTSUPP;
   2391 }
   2392 
   2393 
   2394 /*
   2395  * Probe for a Broadcom chip. Check the PCI vendor and device IDs
   2396  * against our list and return its name if we find a match. Note
   2397  * that since the Broadcom controller contains VPD support, we
   2398  * can get the device name string from the controller itself instead
   2399  * of the compiled-in string. This is a little slow, but it guarantees
   2400  * we'll always announce the right product name.
   2401  */
   2402 static int
   2403 bge_probe(device_t parent, cfdata_t match, void *aux)
   2404 {
   2405 	struct pci_attach_args *pa = (struct pci_attach_args *)aux;
   2406 
   2407 	if (bge_lookup(pa) != NULL)
   2408 		return (1);
   2409 
   2410 	return (0);
   2411 }
   2412 
   2413 static void
   2414 bge_attach(device_t parent, device_t self, void *aux)
   2415 {
   2416 	struct bge_softc	*sc = (struct bge_softc *)self;
   2417 	struct pci_attach_args	*pa = aux;
   2418 	const struct bge_product *bp;
   2419 	const struct bge_revision *br;
   2420 	pci_chipset_tag_t	pc = pa->pa_pc;
   2421 	pci_intr_handle_t	ih;
   2422 	const char		*intrstr = NULL;
   2423 	bus_dma_segment_t	seg;
   2424 	int			rseg;
   2425 	u_int32_t		hwcfg = 0;
   2426 	u_int32_t		mac_addr = 0;
   2427 	u_int32_t		command;
   2428 	struct ifnet		*ifp;
   2429 	void *			kva;
   2430 	u_char			eaddr[ETHER_ADDR_LEN];
   2431 	pcireg_t		memtype;
   2432 	bus_addr_t		memaddr;
   2433 	bus_size_t		memsize;
   2434 	u_int32_t		pm_ctl;
   2435 
   2436 	bp = bge_lookup(pa);
   2437 	KASSERT(bp != NULL);
   2438 
   2439 	sc->bge_pa = *pa;
   2440 
   2441 	aprint_naive(": Ethernet controller\n");
   2442 	aprint_normal(": %s\n", bp->bp_name);
   2443 
   2444 	/*
   2445 	 * Map control/status registers.
   2446 	 */
   2447 	DPRINTFN(5, ("Map control/status regs\n"));
   2448 	command = pci_conf_read(pc, pa->pa_tag, PCI_COMMAND_STATUS_REG);
   2449 	command |= PCI_COMMAND_MEM_ENABLE | PCI_COMMAND_MASTER_ENABLE;
   2450 	pci_conf_write(pc, pa->pa_tag, PCI_COMMAND_STATUS_REG, command);
   2451 	command = pci_conf_read(pc, pa->pa_tag, PCI_COMMAND_STATUS_REG);
   2452 
   2453 	if (!(command & PCI_COMMAND_MEM_ENABLE)) {
   2454 		aprint_error("%s: failed to enable memory mapping!\n",
   2455 		    sc->bge_dev.dv_xname);
   2456 		return;
   2457 	}
   2458 
   2459 	DPRINTFN(5, ("pci_mem_find\n"));
   2460 	memtype = pci_mapreg_type(pa->pa_pc, pa->pa_tag, BGE_PCI_BAR0);
   2461  	switch (memtype) {
   2462 	case PCI_MAPREG_TYPE_MEM | PCI_MAPREG_MEM_TYPE_32BIT:
   2463 	case PCI_MAPREG_TYPE_MEM | PCI_MAPREG_MEM_TYPE_64BIT:
   2464 		if (pci_mapreg_map(pa, BGE_PCI_BAR0,
   2465 		    memtype, 0, &sc->bge_btag, &sc->bge_bhandle,
   2466 		    &memaddr, &memsize) == 0)
   2467 			break;
   2468 	default:
   2469 		aprint_error("%s: can't find mem space\n",
   2470 		    sc->bge_dev.dv_xname);
   2471 		return;
   2472 	}
   2473 
   2474 	DPRINTFN(5, ("pci_intr_map\n"));
   2475 	if (pci_intr_map(pa, &ih)) {
   2476 		aprint_error("%s: couldn't map interrupt\n",
   2477 		    sc->bge_dev.dv_xname);
   2478 		return;
   2479 	}
   2480 
   2481 	DPRINTFN(5, ("pci_intr_string\n"));
   2482 	intrstr = pci_intr_string(pc, ih);
   2483 
   2484 	DPRINTFN(5, ("pci_intr_establish\n"));
   2485 	sc->bge_intrhand = pci_intr_establish(pc, ih, IPL_NET, bge_intr, sc);
   2486 
   2487 	if (sc->bge_intrhand == NULL) {
   2488 		aprint_error("%s: couldn't establish interrupt",
   2489 		    sc->bge_dev.dv_xname);
   2490 		if (intrstr != NULL)
   2491 			aprint_normal(" at %s", intrstr);
   2492 		aprint_normal("\n");
   2493 		return;
   2494 	}
   2495 	aprint_normal("%s: interrupting at %s\n",
   2496 	    sc->bge_dev.dv_xname, intrstr);
   2497 
   2498 	/*
   2499 	 * Kludge for 5700 Bx bug: a hardware bug (PCIX byte enable?)
   2500 	 * can clobber the chip's PCI config-space power control registers,
   2501 	 * leaving the card in D3 powersave state.
   2502 	 * We do not have memory-mapped registers in this state,
   2503 	 * so force device into D0 state before starting initialization.
   2504 	 */
   2505 	pm_ctl = pci_conf_read(pc, pa->pa_tag, BGE_PCI_PWRMGMT_CMD);
   2506 	pm_ctl &= ~(PCI_PWR_D0|PCI_PWR_D1|PCI_PWR_D2|PCI_PWR_D3);
   2507 	pm_ctl |= (1 << 8) | PCI_PWR_D0 ; /* D0 state */
   2508 	pci_conf_write(pc, pa->pa_tag, BGE_PCI_PWRMGMT_CMD, pm_ctl);
   2509 	DELAY(1000);	/* 27 usec is allegedly sufficent */
   2510 
   2511 	/*
   2512 	 * Save ASIC rev.  Look up any quirks associated with this
   2513 	 * ASIC.
   2514 	 */
   2515 	sc->bge_chipid =
   2516 	    pci_conf_read(pa->pa_pc, pa->pa_tag, BGE_PCI_MISC_CTL) &
   2517 	    BGE_PCIMISCCTL_ASICREV;
   2518 
   2519 	/*
   2520 	 * Detect PCI-Express devices
   2521 	 * XXX: guessed from Linux/FreeBSD; no documentation
   2522 	 */
   2523 	if (pci_get_capability(pa->pa_pc, pa->pa_tag, PCI_CAP_PCIEXPRESS,
   2524 	        NULL, NULL) != 0)
   2525 		sc->bge_pcie = 1;
   2526 	else
   2527 		sc->bge_pcie = 0;
   2528 
   2529 	/* Try to reset the chip. */
   2530 	DPRINTFN(5, ("bge_reset\n"));
   2531 	bge_reset(sc);
   2532 
   2533 	if (bge_chipinit(sc)) {
   2534 		aprint_error("%s: chip initialization failed\n",
   2535 		    sc->bge_dev.dv_xname);
   2536 		bge_release_resources(sc);
   2537 		return;
   2538 	}
   2539 
   2540 	/*
   2541 	 * Get station address from the EEPROM.
   2542 	 */
   2543 	mac_addr = bge_readmem_ind(sc, 0x0c14);
   2544 	if ((mac_addr >> 16) == 0x484b) {
   2545 		eaddr[0] = (u_char)(mac_addr >> 8);
   2546 		eaddr[1] = (u_char)(mac_addr >> 0);
   2547 		mac_addr = bge_readmem_ind(sc, 0x0c18);
   2548 		eaddr[2] = (u_char)(mac_addr >> 24);
   2549 		eaddr[3] = (u_char)(mac_addr >> 16);
   2550 		eaddr[4] = (u_char)(mac_addr >> 8);
   2551 		eaddr[5] = (u_char)(mac_addr >> 0);
   2552 	} else if (bge_read_eeprom(sc, (void *)eaddr,
   2553 	    BGE_EE_MAC_OFFSET + 2, ETHER_ADDR_LEN)) {
   2554 		aprint_error("%s: failed to read station address\n",
   2555 		    sc->bge_dev.dv_xname);
   2556 		bge_release_resources(sc);
   2557 		return;
   2558 	}
   2559 
   2560 	br = bge_lookup_rev(sc->bge_chipid);
   2561 	aprint_normal("%s: ", sc->bge_dev.dv_xname);
   2562 
   2563 	if (br == NULL) {
   2564 		aprint_normal("unknown ASIC (0x%04x)", sc->bge_chipid >> 16);
   2565 		sc->bge_quirks = 0;
   2566 	} else {
   2567 		aprint_normal("ASIC %s (0x%04x)",
   2568 		    br->br_name, sc->bge_chipid >> 16);
   2569 		sc->bge_quirks |= br->br_quirks;
   2570 	}
   2571 	aprint_normal(", Ethernet address %s\n", ether_sprintf(eaddr));
   2572 
   2573 	/* Allocate the general information block and ring buffers. */
   2574 	if (pci_dma64_available(pa))
   2575 		sc->bge_dmatag = pa->pa_dmat64;
   2576 	else
   2577 		sc->bge_dmatag = pa->pa_dmat;
   2578 	DPRINTFN(5, ("bus_dmamem_alloc\n"));
   2579 	if (bus_dmamem_alloc(sc->bge_dmatag, sizeof(struct bge_ring_data),
   2580 			     PAGE_SIZE, 0, &seg, 1, &rseg, BUS_DMA_NOWAIT)) {
   2581 		aprint_error("%s: can't alloc rx buffers\n",
   2582 		    sc->bge_dev.dv_xname);
   2583 		return;
   2584 	}
   2585 	DPRINTFN(5, ("bus_dmamem_map\n"));
   2586 	if (bus_dmamem_map(sc->bge_dmatag, &seg, rseg,
   2587 			   sizeof(struct bge_ring_data), &kva,
   2588 			   BUS_DMA_NOWAIT)) {
   2589 		aprint_error("%s: can't map DMA buffers (%d bytes)\n",
   2590 		    sc->bge_dev.dv_xname, (int)sizeof(struct bge_ring_data));
   2591 		bus_dmamem_free(sc->bge_dmatag, &seg, rseg);
   2592 		return;
   2593 	}
   2594 	DPRINTFN(5, ("bus_dmamem_create\n"));
   2595 	if (bus_dmamap_create(sc->bge_dmatag, sizeof(struct bge_ring_data), 1,
   2596 	    sizeof(struct bge_ring_data), 0,
   2597 	    BUS_DMA_NOWAIT, &sc->bge_ring_map)) {
   2598 		aprint_error("%s: can't create DMA map\n",
   2599 		    sc->bge_dev.dv_xname);
   2600 		bus_dmamem_unmap(sc->bge_dmatag, kva,
   2601 				 sizeof(struct bge_ring_data));
   2602 		bus_dmamem_free(sc->bge_dmatag, &seg, rseg);
   2603 		return;
   2604 	}
   2605 	DPRINTFN(5, ("bus_dmamem_load\n"));
   2606 	if (bus_dmamap_load(sc->bge_dmatag, sc->bge_ring_map, kva,
   2607 			    sizeof(struct bge_ring_data), NULL,
   2608 			    BUS_DMA_NOWAIT)) {
   2609 		bus_dmamap_destroy(sc->bge_dmatag, sc->bge_ring_map);
   2610 		bus_dmamem_unmap(sc->bge_dmatag, kva,
   2611 				 sizeof(struct bge_ring_data));
   2612 		bus_dmamem_free(sc->bge_dmatag, &seg, rseg);
   2613 		return;
   2614 	}
   2615 
   2616 	DPRINTFN(5, ("bzero\n"));
   2617 	sc->bge_rdata = (struct bge_ring_data *)kva;
   2618 
   2619 	memset(sc->bge_rdata, 0, sizeof(struct bge_ring_data));
   2620 
   2621 	/* Try to allocate memory for jumbo buffers. */
   2622 	if ((sc->bge_quirks & BGE_QUIRK_5705_CORE) == 0) {
   2623 		if (bge_alloc_jumbo_mem(sc)) {
   2624 			aprint_error("%s: jumbo buffer allocation failed\n",
   2625 			    sc->bge_dev.dv_xname);
   2626 		} else
   2627 			sc->ethercom.ec_capabilities |= ETHERCAP_JUMBO_MTU;
   2628 	}
   2629 
   2630 	/* Set default tuneable values. */
   2631 	sc->bge_stat_ticks = BGE_TICKS_PER_SEC;
   2632 	sc->bge_rx_coal_ticks = 150;
   2633 	sc->bge_rx_max_coal_bds = 64;
   2634 #ifdef ORIG_WPAUL_VALUES
   2635 	sc->bge_tx_coal_ticks = 150;
   2636 	sc->bge_tx_max_coal_bds = 128;
   2637 #else
   2638 	sc->bge_tx_coal_ticks = 300;
   2639 	sc->bge_tx_max_coal_bds = 400;
   2640 #endif
   2641 	if (sc->bge_quirks & BGE_QUIRK_5705_CORE) {
   2642 		sc->bge_tx_coal_ticks = (12 * 5);
   2643 		sc->bge_rx_max_coal_bds = (12 * 5);
   2644 			aprint_verbose("%s: setting short Tx thresholds\n",
   2645 			    sc->bge_dev.dv_xname);
   2646 	}
   2647 
   2648 	/* Set up ifnet structure */
   2649 	ifp = &sc->ethercom.ec_if;
   2650 	ifp->if_softc = sc;
   2651 	ifp->if_flags = IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST;
   2652 	ifp->if_ioctl = bge_ioctl;
   2653 	ifp->if_start = bge_start;
   2654 	ifp->if_init = bge_init;
   2655 	ifp->if_watchdog = bge_watchdog;
   2656 	IFQ_SET_MAXLEN(&ifp->if_snd, max(BGE_TX_RING_CNT - 1, IFQ_MAXLEN));
   2657 	IFQ_SET_READY(&ifp->if_snd);
   2658 	DPRINTFN(5, ("strcpy if_xname\n"));
   2659 	strcpy(ifp->if_xname, sc->bge_dev.dv_xname);
   2660 
   2661 	if ((sc->bge_quirks & BGE_QUIRK_CSUM_BROKEN) == 0)
   2662 		sc->ethercom.ec_if.if_capabilities |=
   2663 		    IFCAP_CSUM_IPv4_Tx | IFCAP_CSUM_IPv4_Rx |
   2664 		    IFCAP_CSUM_TCPv4_Tx | IFCAP_CSUM_TCPv4_Rx |
   2665 		    IFCAP_CSUM_UDPv4_Tx | IFCAP_CSUM_UDPv4_Rx;
   2666 	sc->ethercom.ec_capabilities |=
   2667 	    ETHERCAP_VLAN_HWTAGGING | ETHERCAP_VLAN_MTU;
   2668 
   2669 	if (sc->bge_pcie)
   2670 		sc->ethercom.ec_if.if_capabilities |= IFCAP_TSOv4;
   2671 
   2672 	/*
   2673 	 * Do MII setup.
   2674 	 */
   2675 	DPRINTFN(5, ("mii setup\n"));
   2676 	sc->bge_mii.mii_ifp = ifp;
   2677 	sc->bge_mii.mii_readreg = bge_miibus_readreg;
   2678 	sc->bge_mii.mii_writereg = bge_miibus_writereg;
   2679 	sc->bge_mii.mii_statchg = bge_miibus_statchg;
   2680 
   2681 	/*
   2682 	 * Figure out what sort of media we have by checking the
   2683 	 * hardware config word in the first 32k of NIC internal memory,
   2684 	 * or fall back to the config word in the EEPROM. Note: on some BCM5700
   2685 	 * cards, this value appears to be unset. If that's the
   2686 	 * case, we have to rely on identifying the NIC by its PCI
   2687 	 * subsystem ID, as we do below for the SysKonnect SK-9D41.
   2688 	 */
   2689 	if (bge_readmem_ind(sc, BGE_SOFTWARE_GENCOMM_SIG) == BGE_MAGIC_NUMBER) {
   2690 		hwcfg = bge_readmem_ind(sc, BGE_SOFTWARE_GENCOMM_NICCFG);
   2691 	} else {
   2692 		bge_read_eeprom(sc, (void *)&hwcfg,
   2693 		    BGE_EE_HWCFG_OFFSET, sizeof(hwcfg));
   2694 		hwcfg = be32toh(hwcfg);
   2695 	}
   2696 	if ((hwcfg & BGE_HWCFG_MEDIA) == BGE_MEDIA_FIBER)
   2697 		sc->bge_tbi = 1;
   2698 
   2699 	/* The SysKonnect SK-9D41 is a 1000baseSX card. */
   2700 	if ((pci_conf_read(pa->pa_pc, pa->pa_tag, BGE_PCI_SUBSYS) >> 16) ==
   2701 	    SK_SUBSYSID_9D41)
   2702 		sc->bge_tbi = 1;
   2703 
   2704 	if (sc->bge_tbi) {
   2705 		ifmedia_init(&sc->bge_ifmedia, IFM_IMASK, bge_ifmedia_upd,
   2706 		    bge_ifmedia_sts);
   2707 		ifmedia_add(&sc->bge_ifmedia, IFM_ETHER|IFM_1000_SX, 0, NULL);
   2708 		ifmedia_add(&sc->bge_ifmedia, IFM_ETHER|IFM_1000_SX|IFM_FDX,
   2709 			    0, NULL);
   2710 		ifmedia_add(&sc->bge_ifmedia, IFM_ETHER|IFM_AUTO, 0, NULL);
   2711 		ifmedia_set(&sc->bge_ifmedia, IFM_ETHER|IFM_AUTO);
   2712 	} else {
   2713 		/*
   2714 		 * Do transceiver setup.
   2715 		 */
   2716 		ifmedia_init(&sc->bge_mii.mii_media, 0, bge_ifmedia_upd,
   2717 			     bge_ifmedia_sts);
   2718 		mii_attach(&sc->bge_dev, &sc->bge_mii, 0xffffffff,
   2719 			   MII_PHY_ANY, MII_OFFSET_ANY,
   2720 			   MIIF_FORCEANEG|MIIF_DOPAUSE);
   2721 
   2722 		if (LIST_FIRST(&sc->bge_mii.mii_phys) == NULL) {
   2723 			printf("%s: no PHY found!\n", sc->bge_dev.dv_xname);
   2724 			ifmedia_add(&sc->bge_mii.mii_media,
   2725 				    IFM_ETHER|IFM_MANUAL, 0, NULL);
   2726 			ifmedia_set(&sc->bge_mii.mii_media,
   2727 				    IFM_ETHER|IFM_MANUAL);
   2728 		} else
   2729 			ifmedia_set(&sc->bge_mii.mii_media,
   2730 				    IFM_ETHER|IFM_AUTO);
   2731 	}
   2732 
   2733 	/*
   2734 	 * When using the BCM5701 in PCI-X mode, data corruption has
   2735 	 * been observed in the first few bytes of some received packets.
   2736 	 * Aligning the packet buffer in memory eliminates the corruption.
   2737 	 * Unfortunately, this misaligns the packet payloads.  On platforms
   2738 	 * which do not support unaligned accesses, we will realign the
   2739 	 * payloads by copying the received packets.
   2740 	 */
   2741 	if (sc->bge_quirks & BGE_QUIRK_PCIX_DMA_ALIGN_BUG) {
   2742 		/* If in PCI-X mode, work around the alignment bug. */
   2743 		if ((pci_conf_read(pa->pa_pc, pa->pa_tag, BGE_PCI_PCISTATE) &
   2744                     (BGE_PCISTATE_PCI_BUSMODE | BGE_PCISTATE_PCI_BUSSPEED)) ==
   2745                          BGE_PCISTATE_PCI_BUSSPEED)
   2746 		sc->bge_rx_alignment_bug = 1;
   2747         }
   2748 
   2749 	/*
   2750 	 * Call MI attach routine.
   2751 	 */
   2752 	DPRINTFN(5, ("if_attach\n"));
   2753 	if_attach(ifp);
   2754 	DPRINTFN(5, ("ether_ifattach\n"));
   2755 	ether_ifattach(ifp, eaddr);
   2756 #ifdef BGE_EVENT_COUNTERS
   2757 	/*
   2758 	 * Attach event counters.
   2759 	 */
   2760 	evcnt_attach_dynamic(&sc->bge_ev_intr, EVCNT_TYPE_INTR,
   2761 	    NULL, sc->bge_dev.dv_xname, "intr");
   2762 	evcnt_attach_dynamic(&sc->bge_ev_tx_xoff, EVCNT_TYPE_MISC,
   2763 	    NULL, sc->bge_dev.dv_xname, "tx_xoff");
   2764 	evcnt_attach_dynamic(&sc->bge_ev_tx_xon, EVCNT_TYPE_MISC,
   2765 	    NULL, sc->bge_dev.dv_xname, "tx_xon");
   2766 	evcnt_attach_dynamic(&sc->bge_ev_rx_xoff, EVCNT_TYPE_MISC,
   2767 	    NULL, sc->bge_dev.dv_xname, "rx_xoff");
   2768 	evcnt_attach_dynamic(&sc->bge_ev_rx_xon, EVCNT_TYPE_MISC,
   2769 	    NULL, sc->bge_dev.dv_xname, "rx_xon");
   2770 	evcnt_attach_dynamic(&sc->bge_ev_rx_macctl, EVCNT_TYPE_MISC,
   2771 	    NULL, sc->bge_dev.dv_xname, "rx_macctl");
   2772 	evcnt_attach_dynamic(&sc->bge_ev_xoffentered, EVCNT_TYPE_MISC,
   2773 	    NULL, sc->bge_dev.dv_xname, "xoffentered");
   2774 #endif /* BGE_EVENT_COUNTERS */
   2775 	DPRINTFN(5, ("callout_init\n"));
   2776 	callout_init(&sc->bge_timeout, 0);
   2777 
   2778 	sc->bge_powerhook = powerhook_establish(sc->bge_dev.dv_xname,
   2779 	    bge_powerhook, sc);
   2780 	if (sc->bge_powerhook == NULL)
   2781 		printf("%s: WARNING: unable to establish PCI power hook\n",
   2782 		    sc->bge_dev.dv_xname);
   2783 }
   2784 
   2785 static void
   2786 bge_release_resources(struct bge_softc *sc)
   2787 {
   2788 	if (sc->bge_vpd_prodname != NULL)
   2789 		free(sc->bge_vpd_prodname, M_DEVBUF);
   2790 
   2791 	if (sc->bge_vpd_readonly != NULL)
   2792 		free(sc->bge_vpd_readonly, M_DEVBUF);
   2793 }
   2794 
   2795 static void
   2796 bge_reset(struct bge_softc *sc)
   2797 {
   2798 	struct pci_attach_args *pa = &sc->bge_pa;
   2799 	u_int32_t cachesize, command, pcistate, new_pcistate;
   2800 	int i, val;
   2801 
   2802 	/* Save some important PCI state. */
   2803 	cachesize = pci_conf_read(pa->pa_pc, pa->pa_tag, BGE_PCI_CACHESZ);
   2804 	command = pci_conf_read(pa->pa_pc, pa->pa_tag, BGE_PCI_CMD);
   2805 	pcistate = pci_conf_read(pa->pa_pc, pa->pa_tag, BGE_PCI_PCISTATE);
   2806 
   2807 	pci_conf_write(pa->pa_pc, pa->pa_tag, BGE_PCI_MISC_CTL,
   2808 	    BGE_PCIMISCCTL_INDIRECT_ACCESS|BGE_PCIMISCCTL_MASK_PCI_INTR|
   2809 	    BGE_HIF_SWAP_OPTIONS|BGE_PCIMISCCTL_PCISTATE_RW);
   2810 
   2811 	/*
   2812 	 * Disable the firmware fastboot feature on 5752 ASIC
   2813 	 * to avoid firmware timeout.
   2814 	 */
   2815 	if (BGE_ASICREV(sc->bge_chipid) == BGE_ASICREV_BCM5752)
   2816 		CSR_WRITE_4(sc, BGE_FASTBOOT_PC, 0);
   2817 
   2818 	val = BGE_MISCCFG_RESET_CORE_CLOCKS | (65<<1);
   2819 	/*
   2820 	 * XXX: from FreeBSD/Linux; no documentation
   2821 	 */
   2822 	if (sc->bge_pcie) {
   2823 		if (CSR_READ_4(sc, BGE_PCIE_CTL1) == 0x60)
   2824 			CSR_WRITE_4(sc, BGE_PCIE_CTL1, 0x20);
   2825 		if (sc->bge_chipid != BGE_CHIPID_BCM5750_A0) {
   2826 			/* No idea what that actually means */
   2827 			CSR_WRITE_4(sc, BGE_MISC_CFG, 1 << 29);
   2828 			val |= (1<<29);
   2829 		}
   2830 	}
   2831 	/*
   2832 	 * Write the magic number to the firmware mailbox at 0xb50
   2833 	 * so that the driver can synchronize with the firmware.
   2834 	 */
   2835 	bge_writemem_ind(sc, BGE_SOFTWARE_GENCOMM, BGE_MAGIC_NUMBER);
   2836 
   2837 	/* Issue global reset */
   2838 	bge_writereg_ind(sc, BGE_MISC_CFG, val);
   2839 
   2840 	DELAY(1000);
   2841 
   2842 	/*
   2843 	 * XXX: from FreeBSD/Linux; no documentation
   2844 	 */
   2845 	if (sc->bge_pcie) {
   2846 		if (sc->bge_chipid == BGE_CHIPID_BCM5750_A0) {
   2847 			pcireg_t reg;
   2848 
   2849 			DELAY(500000);
   2850 			/* XXX: Magic Numbers */
   2851 			reg = pci_conf_read(pa->pa_pc, pa->pa_tag, BGE_PCI_UNKNOWN0);
   2852 			pci_conf_write(pa->pa_pc, pa->pa_tag, BGE_PCI_UNKNOWN0,
   2853 			    reg | (1 << 15));
   2854 		}
   2855 		/*
   2856 		 * XXX: Magic Numbers.
   2857 		 * Sets maximal PCI-e payload and clears any PCI-e errors.
   2858 		 * Should be replaced with references to PCI config-space
   2859 		 * capability block for PCI-Express.
   2860 		 */
   2861 		pci_conf_write(pa->pa_pc, pa->pa_tag,
   2862 		    BGE_PCI_CONF_DEV_CTRL, 0xf5000);
   2863 
   2864 	}
   2865 
   2866 	/* Reset some of the PCI state that got zapped by reset */
   2867 	pci_conf_write(pa->pa_pc, pa->pa_tag, BGE_PCI_MISC_CTL,
   2868 	    BGE_PCIMISCCTL_INDIRECT_ACCESS|BGE_PCIMISCCTL_MASK_PCI_INTR|
   2869 	    BGE_HIF_SWAP_OPTIONS|BGE_PCIMISCCTL_PCISTATE_RW);
   2870 	pci_conf_write(pa->pa_pc, pa->pa_tag, BGE_PCI_CMD, command);
   2871 	pci_conf_write(pa->pa_pc, pa->pa_tag, BGE_PCI_CACHESZ, cachesize);
   2872 	bge_writereg_ind(sc, BGE_MISC_CFG, (65 << 1));
   2873 
   2874 	/* Enable memory arbiter. */
   2875 	{
   2876 		uint32_t marbmode = 0;
   2877 		if (BGE_IS_5714_FAMILY(sc)) {
   2878 			marbmode = CSR_READ_4(sc, BGE_MARB_MODE);
   2879 		}
   2880  		CSR_WRITE_4(sc, BGE_MARB_MODE, BGE_MARBMODE_ENABLE | marbmode);
   2881 	}
   2882 
   2883 	/*
   2884 	 * Poll the value location we just wrote until
   2885 	 * we see the 1's complement of the magic number.
   2886 	 * This indicates that the firmware initialization
   2887 	 * is complete.
   2888 	 */
   2889 	for (i = 0; i < BGE_TIMEOUT; i++) {
   2890 		val = bge_readmem_ind(sc, BGE_SOFTWARE_GENCOMM);
   2891 		if (val == ~BGE_MAGIC_NUMBER)
   2892 			break;
   2893 		DELAY(1000);
   2894 	}
   2895 
   2896 	if (i >= BGE_TIMEOUT) {
   2897 		printf("%s: firmware handshake timed out, val = %x\n",
   2898 		    sc->bge_dev.dv_xname, val);
   2899 		/*
   2900 		 * XXX: occasionally fired on bcm5721, but without
   2901 		 * apparent harm.  For now, keep going if we timeout
   2902 		 * against PCI-E devices.
   2903 		 */
   2904 		 if (!sc->bge_pcie)
   2905 		  return;
   2906 	}
   2907 
   2908 	/*
   2909 	 * XXX Wait for the value of the PCISTATE register to
   2910 	 * return to its original pre-reset state. This is a
   2911 	 * fairly good indicator of reset completion. If we don't
   2912 	 * wait for the reset to fully complete, trying to read
   2913 	 * from the device's non-PCI registers may yield garbage
   2914 	 * results.
   2915 	 */
   2916 	for (i = 0; i < BGE_TIMEOUT; i++) {
   2917 		new_pcistate = pci_conf_read(pa->pa_pc, pa->pa_tag,
   2918 		    BGE_PCI_PCISTATE);
   2919 		if ((new_pcistate & ~BGE_PCISTATE_RESERVED) ==
   2920 		    (pcistate & ~BGE_PCISTATE_RESERVED))
   2921 			break;
   2922 		DELAY(10);
   2923 	}
   2924 	if ((new_pcistate & ~BGE_PCISTATE_RESERVED) !=
   2925 	    (pcistate & ~BGE_PCISTATE_RESERVED)) {
   2926 		printf("%s: pcistate failed to revert\n",
   2927 		    sc->bge_dev.dv_xname);
   2928 	}
   2929 
   2930 	/* XXX: from FreeBSD/Linux; no documentation */
   2931 	if (sc->bge_pcie && sc->bge_chipid != BGE_CHIPID_BCM5750_A0)
   2932 		CSR_WRITE_4(sc, BGE_PCIE_CTL0, CSR_READ_4(sc, BGE_PCIE_CTL0) | (1<<25));
   2933 
   2934 	/* Enable memory arbiter. */
   2935 	/* XXX why do this twice? */
   2936 	{
   2937 		uint32_t marbmode = 0;
   2938 		if (BGE_IS_5714_FAMILY(sc)) {
   2939 			marbmode = CSR_READ_4(sc, BGE_MARB_MODE);
   2940 		}
   2941  		CSR_WRITE_4(sc, BGE_MARB_MODE, BGE_MARBMODE_ENABLE | marbmode);
   2942 	}
   2943 
   2944 	/* Fix up byte swapping */
   2945 	CSR_WRITE_4(sc, BGE_MODE_CTL, BGE_DMA_SWAP_OPTIONS);
   2946 
   2947 	CSR_WRITE_4(sc, BGE_MAC_MODE, 0);
   2948 
   2949 	DELAY(10000);
   2950 }
   2951 
   2952 /*
   2953  * Frame reception handling. This is called if there's a frame
   2954  * on the receive return list.
   2955  *
   2956  * Note: we have to be able to handle two possibilities here:
   2957  * 1) the frame is from the jumbo recieve ring
   2958  * 2) the frame is from the standard receive ring
   2959  */
   2960 
   2961 static void
   2962 bge_rxeof(struct bge_softc *sc)
   2963 {
   2964 	struct ifnet *ifp;
   2965 	int stdcnt = 0, jumbocnt = 0;
   2966 	bus_dmamap_t dmamap;
   2967 	bus_addr_t offset, toff;
   2968 	bus_size_t tlen;
   2969 	int tosync;
   2970 
   2971 	ifp = &sc->ethercom.ec_if;
   2972 
   2973 	bus_dmamap_sync(sc->bge_dmatag, sc->bge_ring_map,
   2974 	    offsetof(struct bge_ring_data, bge_status_block),
   2975 	    sizeof (struct bge_status_block),
   2976 	    BUS_DMASYNC_POSTREAD);
   2977 
   2978 	offset = offsetof(struct bge_ring_data, bge_rx_return_ring);
   2979 	tosync = sc->bge_rdata->bge_status_block.bge_idx[0].bge_rx_prod_idx -
   2980 	    sc->bge_rx_saved_considx;
   2981 
   2982 	toff = offset + (sc->bge_rx_saved_considx * sizeof (struct bge_rx_bd));
   2983 
   2984 	if (tosync < 0) {
   2985 		tlen = (sc->bge_return_ring_cnt - sc->bge_rx_saved_considx) *
   2986 		    sizeof (struct bge_rx_bd);
   2987 		bus_dmamap_sync(sc->bge_dmatag, sc->bge_ring_map,
   2988 		    toff, tlen, BUS_DMASYNC_POSTREAD);
   2989 		tosync = -tosync;
   2990 	}
   2991 
   2992 	bus_dmamap_sync(sc->bge_dmatag, sc->bge_ring_map,
   2993 	    offset, tosync * sizeof (struct bge_rx_bd),
   2994 	    BUS_DMASYNC_POSTREAD);
   2995 
   2996 	while(sc->bge_rx_saved_considx !=
   2997 	    sc->bge_rdata->bge_status_block.bge_idx[0].bge_rx_prod_idx) {
   2998 		struct bge_rx_bd	*cur_rx;
   2999 		u_int32_t		rxidx;
   3000 		struct mbuf		*m = NULL;
   3001 
   3002 		cur_rx = &sc->bge_rdata->
   3003 			bge_rx_return_ring[sc->bge_rx_saved_considx];
   3004 
   3005 		rxidx = cur_rx->bge_idx;
   3006 		BGE_INC(sc->bge_rx_saved_considx, sc->bge_return_ring_cnt);
   3007 
   3008 		if (cur_rx->bge_flags & BGE_RXBDFLAG_JUMBO_RING) {
   3009 			BGE_INC(sc->bge_jumbo, BGE_JUMBO_RX_RING_CNT);
   3010 			m = sc->bge_cdata.bge_rx_jumbo_chain[rxidx];
   3011 			sc->bge_cdata.bge_rx_jumbo_chain[rxidx] = NULL;
   3012 			jumbocnt++;
   3013 			bus_dmamap_sync(sc->bge_dmatag,
   3014 			    sc->bge_cdata.bge_rx_jumbo_map,
   3015 			    mtod(m, char *) - (char *)sc->bge_cdata.bge_jumbo_buf,
   3016 			    BGE_JLEN, BUS_DMASYNC_POSTREAD);
   3017 			if (cur_rx->bge_flags & BGE_RXBDFLAG_ERROR) {
   3018 				ifp->if_ierrors++;
   3019 				bge_newbuf_jumbo(sc, sc->bge_jumbo, m);
   3020 				continue;
   3021 			}
   3022 			if (bge_newbuf_jumbo(sc, sc->bge_jumbo,
   3023 					     NULL)== ENOBUFS) {
   3024 				ifp->if_ierrors++;
   3025 				bge_newbuf_jumbo(sc, sc->bge_jumbo, m);
   3026 				continue;
   3027 			}
   3028 		} else {
   3029 			BGE_INC(sc->bge_std, BGE_STD_RX_RING_CNT);
   3030 			m = sc->bge_cdata.bge_rx_std_chain[rxidx];
   3031 
   3032 			sc->bge_cdata.bge_rx_std_chain[rxidx] = NULL;
   3033 			stdcnt++;
   3034 			dmamap = sc->bge_cdata.bge_rx_std_map[rxidx];
   3035 			sc->bge_cdata.bge_rx_std_map[rxidx] = 0;
   3036 			bus_dmamap_sync(sc->bge_dmatag, dmamap, 0,
   3037 			    dmamap->dm_mapsize, BUS_DMASYNC_POSTREAD);
   3038 			bus_dmamap_unload(sc->bge_dmatag, dmamap);
   3039 			if (cur_rx->bge_flags & BGE_RXBDFLAG_ERROR) {
   3040 				ifp->if_ierrors++;
   3041 				bge_newbuf_std(sc, sc->bge_std, m, dmamap);
   3042 				continue;
   3043 			}
   3044 			if (bge_newbuf_std(sc, sc->bge_std,
   3045 			    NULL, dmamap) == ENOBUFS) {
   3046 				ifp->if_ierrors++;
   3047 				bge_newbuf_std(sc, sc->bge_std, m, dmamap);
   3048 				continue;
   3049 			}
   3050 		}
   3051 
   3052 		ifp->if_ipackets++;
   3053 #ifndef __NO_STRICT_ALIGNMENT
   3054                 /*
   3055                  * XXX: if the 5701 PCIX-Rx-DMA workaround is in effect,
   3056                  * the Rx buffer has the layer-2 header unaligned.
   3057                  * If our CPU requires alignment, re-align by copying.
   3058                  */
   3059 		if (sc->bge_rx_alignment_bug) {
   3060 			memmove(mtod(m, char *) + ETHER_ALIGN, m->m_data,
   3061                                 cur_rx->bge_len);
   3062 			m->m_data += ETHER_ALIGN;
   3063 		}
   3064 #endif
   3065 
   3066 		m->m_pkthdr.len = m->m_len = cur_rx->bge_len - ETHER_CRC_LEN;
   3067 		m->m_pkthdr.rcvif = ifp;
   3068 
   3069 #if NBPFILTER > 0
   3070 		/*
   3071 		 * Handle BPF listeners. Let the BPF user see the packet.
   3072 		 */
   3073 		if (ifp->if_bpf)
   3074 			bpf_mtap(ifp->if_bpf, m);
   3075 #endif
   3076 
   3077 		m->m_pkthdr.csum_flags = M_CSUM_IPv4;
   3078 
   3079 		if ((cur_rx->bge_ip_csum ^ 0xffff) != 0)
   3080 			m->m_pkthdr.csum_flags |= M_CSUM_IPv4_BAD;
   3081 		/*
   3082 		 * Rx transport checksum-offload may also
   3083 		 * have bugs with packets which, when transmitted,
   3084 		 * were `runts' requiring padding.
   3085 		 */
   3086 		if (cur_rx->bge_flags & BGE_RXBDFLAG_TCP_UDP_CSUM &&
   3087 		    (/* (sc->_bge_quirks & BGE_QUIRK_SHORT_CKSUM_BUG) == 0 ||*/
   3088 		     m->m_pkthdr.len >= ETHER_MIN_NOPAD)) {
   3089 			m->m_pkthdr.csum_data =
   3090 			    cur_rx->bge_tcp_udp_csum;
   3091 			m->m_pkthdr.csum_flags |=
   3092 			    (M_CSUM_TCPv4|M_CSUM_UDPv4|
   3093 			     M_CSUM_DATA|M_CSUM_NO_PSEUDOHDR);
   3094 		}
   3095 
   3096 		/*
   3097 		 * If we received a packet with a vlan tag, pass it
   3098 		 * to vlan_input() instead of ether_input().
   3099 		 */
   3100 		if (cur_rx->bge_flags & BGE_RXBDFLAG_VLAN_TAG)
   3101 			VLAN_INPUT_TAG(ifp, m, cur_rx->bge_vlan_tag, continue);
   3102 
   3103 		(*ifp->if_input)(ifp, m);
   3104 	}
   3105 
   3106 	CSR_WRITE_4(sc, BGE_MBX_RX_CONS0_LO, sc->bge_rx_saved_considx);
   3107 	if (stdcnt)
   3108 		CSR_WRITE_4(sc, BGE_MBX_RX_STD_PROD_LO, sc->bge_std);
   3109 	if (jumbocnt)
   3110 		CSR_WRITE_4(sc, BGE_MBX_RX_JUMBO_PROD_LO, sc->bge_jumbo);
   3111 }
   3112 
   3113 static void
   3114 bge_txeof(struct bge_softc *sc)
   3115 {
   3116 	struct bge_tx_bd *cur_tx = NULL;
   3117 	struct ifnet *ifp;
   3118 	struct txdmamap_pool_entry *dma;
   3119 	bus_addr_t offset, toff;
   3120 	bus_size_t tlen;
   3121 	int tosync;
   3122 	struct mbuf *m;
   3123 
   3124 	ifp = &sc->ethercom.ec_if;
   3125 
   3126 	bus_dmamap_sync(sc->bge_dmatag, sc->bge_ring_map,
   3127 	    offsetof(struct bge_ring_data, bge_status_block),
   3128 	    sizeof (struct bge_status_block),
   3129 	    BUS_DMASYNC_POSTREAD);
   3130 
   3131 	offset = offsetof(struct bge_ring_data, bge_tx_ring);
   3132 	tosync = sc->bge_rdata->bge_status_block.bge_idx[0].bge_tx_cons_idx -
   3133 	    sc->bge_tx_saved_considx;
   3134 
   3135 	toff = offset + (sc->bge_tx_saved_considx * sizeof (struct bge_tx_bd));
   3136 
   3137 	if (tosync < 0) {
   3138 		tlen = (BGE_TX_RING_CNT - sc->bge_tx_saved_considx) *
   3139 		    sizeof (struct bge_tx_bd);
   3140 		bus_dmamap_sync(sc->bge_dmatag, sc->bge_ring_map,
   3141 		    toff, tlen, BUS_DMASYNC_POSTREAD|BUS_DMASYNC_POSTWRITE);
   3142 		tosync = -tosync;
   3143 	}
   3144 
   3145 	bus_dmamap_sync(sc->bge_dmatag, sc->bge_ring_map,
   3146 	    offset, tosync * sizeof (struct bge_tx_bd),
   3147 	    BUS_DMASYNC_POSTREAD|BUS_DMASYNC_POSTWRITE);
   3148 
   3149 	/*
   3150 	 * Go through our tx ring and free mbufs for those
   3151 	 * frames that have been sent.
   3152 	 */
   3153 	while (sc->bge_tx_saved_considx !=
   3154 	    sc->bge_rdata->bge_status_block.bge_idx[0].bge_tx_cons_idx) {
   3155 		u_int32_t		idx = 0;
   3156 
   3157 		idx = sc->bge_tx_saved_considx;
   3158 		cur_tx = &sc->bge_rdata->bge_tx_ring[idx];
   3159 		if (cur_tx->bge_flags & BGE_TXBDFLAG_END)
   3160 			ifp->if_opackets++;
   3161 		m = sc->bge_cdata.bge_tx_chain[idx];
   3162 		if (m != NULL) {
   3163 			sc->bge_cdata.bge_tx_chain[idx] = NULL;
   3164 			dma = sc->txdma[idx];
   3165 			bus_dmamap_sync(sc->bge_dmatag, dma->dmamap, 0,
   3166 			    dma->dmamap->dm_mapsize, BUS_DMASYNC_POSTWRITE);
   3167 			bus_dmamap_unload(sc->bge_dmatag, dma->dmamap);
   3168 			SLIST_INSERT_HEAD(&sc->txdma_list, dma, link);
   3169 			sc->txdma[idx] = NULL;
   3170 
   3171 			m_freem(m);
   3172 		}
   3173 		sc->bge_txcnt--;
   3174 		BGE_INC(sc->bge_tx_saved_considx, BGE_TX_RING_CNT);
   3175 		ifp->if_timer = 0;
   3176 	}
   3177 
   3178 	if (cur_tx != NULL)
   3179 		ifp->if_flags &= ~IFF_OACTIVE;
   3180 }
   3181 
   3182 static int
   3183 bge_intr(void *xsc)
   3184 {
   3185 	struct bge_softc *sc;
   3186 	struct ifnet *ifp;
   3187 
   3188 	sc = xsc;
   3189 	ifp = &sc->ethercom.ec_if;
   3190 
   3191 #ifdef notdef
   3192 	/* Avoid this for now -- checking this register is expensive. */
   3193 	/* Make sure this is really our interrupt. */
   3194 	if (!(CSR_READ_4(sc, BGE_MISC_LOCAL_CTL) & BGE_MLC_INTR_STATE))
   3195 		return (0);
   3196 #endif
   3197 	/* Ack interrupt and stop others from occuring. */
   3198 	CSR_WRITE_4(sc, BGE_MBX_IRQ0_LO, 1);
   3199 
   3200 	BGE_EVCNT_INCR(sc->bge_ev_intr);
   3201 
   3202 	/*
   3203 	 * Process link state changes.
   3204 	 * Grrr. The link status word in the status block does
   3205 	 * not work correctly on the BCM5700 rev AX and BX chips,
   3206 	 * according to all available information. Hence, we have
   3207 	 * to enable MII interrupts in order to properly obtain
   3208 	 * async link changes. Unfortunately, this also means that
   3209 	 * we have to read the MAC status register to detect link
   3210 	 * changes, thereby adding an additional register access to
   3211 	 * the interrupt handler.
   3212 	 */
   3213 
   3214 	if (sc->bge_quirks & BGE_QUIRK_LINK_STATE_BROKEN) {
   3215 		u_int32_t		status;
   3216 
   3217 		status = CSR_READ_4(sc, BGE_MAC_STS);
   3218 		if (status & BGE_MACSTAT_MI_INTERRUPT) {
   3219 			sc->bge_link = 0;
   3220 			callout_stop(&sc->bge_timeout);
   3221 			bge_tick(sc);
   3222 			/* Clear the interrupt */
   3223 			CSR_WRITE_4(sc, BGE_MAC_EVT_ENB,
   3224 			    BGE_EVTENB_MI_INTERRUPT);
   3225 			bge_miibus_readreg(&sc->bge_dev, 1, BRGPHY_MII_ISR);
   3226 			bge_miibus_writereg(&sc->bge_dev, 1, BRGPHY_MII_IMR,
   3227 			    BRGPHY_INTRS);
   3228 		}
   3229 	} else {
   3230 		if (sc->bge_rdata->bge_status_block.bge_status &
   3231 		    BGE_STATFLAG_LINKSTATE_CHANGED) {
   3232 			sc->bge_link = 0;
   3233 			callout_stop(&sc->bge_timeout);
   3234 			bge_tick(sc);
   3235 			/* Clear the interrupt */
   3236 			CSR_WRITE_4(sc, BGE_MAC_STS, BGE_MACSTAT_SYNC_CHANGED|
   3237 			    BGE_MACSTAT_CFG_CHANGED|BGE_MACSTAT_MI_COMPLETE|
   3238 			    BGE_MACSTAT_LINK_CHANGED);
   3239 		}
   3240 	}
   3241 
   3242 	if (ifp->if_flags & IFF_RUNNING) {
   3243 		/* Check RX return ring producer/consumer */
   3244 		bge_rxeof(sc);
   3245 
   3246 		/* Check TX ring producer/consumer */
   3247 		bge_txeof(sc);
   3248 	}
   3249 
   3250 	if (sc->bge_pending_rxintr_change) {
   3251 		uint32_t rx_ticks = sc->bge_rx_coal_ticks;
   3252 		uint32_t rx_bds = sc->bge_rx_max_coal_bds;
   3253 		uint32_t junk;
   3254 
   3255 		CSR_WRITE_4(sc, BGE_HCC_RX_COAL_TICKS, rx_ticks);
   3256 		DELAY(10);
   3257 		junk = CSR_READ_4(sc, BGE_HCC_RX_COAL_TICKS);
   3258 
   3259 		CSR_WRITE_4(sc, BGE_HCC_RX_MAX_COAL_BDS, rx_bds);
   3260 		DELAY(10);
   3261 		junk = CSR_READ_4(sc, BGE_HCC_RX_MAX_COAL_BDS);
   3262 
   3263 		sc->bge_pending_rxintr_change = 0;
   3264 	}
   3265 	bge_handle_events(sc);
   3266 
   3267 	/* Re-enable interrupts. */
   3268 	CSR_WRITE_4(sc, BGE_MBX_IRQ0_LO, 0);
   3269 
   3270 	if (ifp->if_flags & IFF_RUNNING && !IFQ_IS_EMPTY(&ifp->if_snd))
   3271 		bge_start(ifp);
   3272 
   3273 	return (1);
   3274 }
   3275 
   3276 static void
   3277 bge_tick(void *xsc)
   3278 {
   3279 	struct bge_softc *sc = xsc;
   3280 	struct mii_data *mii = &sc->bge_mii;
   3281 	struct ifmedia *ifm = NULL;
   3282 	struct ifnet *ifp = &sc->ethercom.ec_if;
   3283 	int s;
   3284 
   3285 	s = splnet();
   3286 
   3287 	bge_stats_update(sc);
   3288 	callout_reset(&sc->bge_timeout, hz, bge_tick, sc);
   3289 	if (sc->bge_link) {
   3290 		splx(s);
   3291 		return;
   3292 	}
   3293 
   3294 	if (sc->bge_tbi) {
   3295 		ifm = &sc->bge_ifmedia;
   3296 		if (CSR_READ_4(sc, BGE_MAC_STS) &
   3297 		    BGE_MACSTAT_TBI_PCS_SYNCHED) {
   3298 			sc->bge_link++;
   3299 			CSR_WRITE_4(sc, BGE_MAC_STS, 0xFFFFFFFF);
   3300 			if (!IFQ_IS_EMPTY(&ifp->if_snd))
   3301 				bge_start(ifp);
   3302 		}
   3303 		splx(s);
   3304 		return;
   3305 	}
   3306 
   3307 	mii_tick(mii);
   3308 
   3309 	if (!sc->bge_link && mii->mii_media_status & IFM_ACTIVE &&
   3310 	    IFM_SUBTYPE(mii->mii_media_active) != IFM_NONE) {
   3311 		sc->bge_link++;
   3312 		if (!IFQ_IS_EMPTY(&ifp->if_snd))
   3313 			bge_start(ifp);
   3314 	}
   3315 
   3316 	splx(s);
   3317 }
   3318 
   3319 static void
   3320 bge_stats_update(struct bge_softc *sc)
   3321 {
   3322 	struct ifnet *ifp = &sc->ethercom.ec_if;
   3323 	bus_size_t stats = BGE_MEMWIN_START + BGE_STATS_BLOCK;
   3324 	bus_size_t rstats = BGE_RX_STATS;
   3325 
   3326 #define READ_RSTAT(sc, stats, stat) \
   3327 	  CSR_READ_4(sc, stats + offsetof(struct bge_mac_stats_regs, stat))
   3328 
   3329 	if (sc->bge_quirks & BGE_QUIRK_5705_CORE) {
   3330 		ifp->if_collisions +=
   3331 		    READ_RSTAT(sc, rstats, dot3StatsSingleCollisionFrames) +
   3332 		    READ_RSTAT(sc, rstats, dot3StatsMultipleCollisionFrames) +
   3333 		    READ_RSTAT(sc, rstats, dot3StatsExcessiveCollisions) +
   3334 		    READ_RSTAT(sc, rstats, dot3StatsLateCollisions);
   3335 
   3336 		BGE_EVCNT_ADD(sc->bge_ev_tx_xoff,
   3337 			      READ_RSTAT(sc, rstats, outXoffSent));
   3338 		BGE_EVCNT_ADD(sc->bge_ev_tx_xon,
   3339 			      READ_RSTAT(sc, rstats, outXonSent));
   3340 		BGE_EVCNT_ADD(sc->bge_ev_rx_xoff,
   3341 			      READ_RSTAT(sc, rstats, xoffPauseFramesReceived));
   3342 		BGE_EVCNT_ADD(sc->bge_ev_rx_xon,
   3343 			      READ_RSTAT(sc, rstats, xonPauseFramesReceived));
   3344 		BGE_EVCNT_ADD(sc->bge_ev_rx_macctl,
   3345 			      READ_RSTAT(sc, rstats, macControlFramesReceived));
   3346 		BGE_EVCNT_ADD(sc->bge_ev_xoffentered,
   3347 			      READ_RSTAT(sc, rstats, xoffStateEntered));
   3348 		return;
   3349 	}
   3350 
   3351 #undef READ_RSTAT
   3352 #define READ_STAT(sc, stats, stat) \
   3353 	  CSR_READ_4(sc, stats + offsetof(struct bge_stats, stat))
   3354 
   3355 	ifp->if_collisions +=
   3356 	  (READ_STAT(sc, stats, dot3StatsSingleCollisionFrames.bge_addr_lo) +
   3357 	   READ_STAT(sc, stats, dot3StatsMultipleCollisionFrames.bge_addr_lo) +
   3358 	   READ_STAT(sc, stats, dot3StatsExcessiveCollisions.bge_addr_lo) +
   3359 	   READ_STAT(sc, stats, dot3StatsLateCollisions.bge_addr_lo)) -
   3360 	  ifp->if_collisions;
   3361 
   3362 	BGE_EVCNT_UPD(sc->bge_ev_tx_xoff,
   3363 		      READ_STAT(sc, stats, outXoffSent.bge_addr_lo));
   3364 	BGE_EVCNT_UPD(sc->bge_ev_tx_xon,
   3365 		      READ_STAT(sc, stats, outXonSent.bge_addr_lo));
   3366 	BGE_EVCNT_UPD(sc->bge_ev_rx_xoff,
   3367 		      READ_STAT(sc, stats,
   3368 		      		xoffPauseFramesReceived.bge_addr_lo));
   3369 	BGE_EVCNT_UPD(sc->bge_ev_rx_xon,
   3370 		      READ_STAT(sc, stats, xonPauseFramesReceived.bge_addr_lo));
   3371 	BGE_EVCNT_UPD(sc->bge_ev_rx_macctl,
   3372 		      READ_STAT(sc, stats,
   3373 		      		macControlFramesReceived.bge_addr_lo));
   3374 	BGE_EVCNT_UPD(sc->bge_ev_xoffentered,
   3375 		      READ_STAT(sc, stats, xoffStateEntered.bge_addr_lo));
   3376 
   3377 #undef READ_STAT
   3378 
   3379 #ifdef notdef
   3380 	ifp->if_collisions +=
   3381 	   (sc->bge_rdata->bge_info.bge_stats.dot3StatsSingleCollisionFrames +
   3382 	   sc->bge_rdata->bge_info.bge_stats.dot3StatsMultipleCollisionFrames +
   3383 	   sc->bge_rdata->bge_info.bge_stats.dot3StatsExcessiveCollisions +
   3384 	   sc->bge_rdata->bge_info.bge_stats.dot3StatsLateCollisions) -
   3385 	   ifp->if_collisions;
   3386 #endif
   3387 }
   3388 
   3389 /*
   3390  * Pad outbound frame to ETHER_MIN_NOPAD for an unusual reason.
   3391  * The bge hardware will pad out Tx runts to ETHER_MIN_NOPAD,
   3392  * but when such padded frames employ the  bge IP/TCP checksum offload,
   3393  * the hardware checksum assist gives incorrect results (possibly
   3394  * from incorporating its own padding into the UDP/TCP checksum; who knows).
   3395  * If we pad such runts with zeros, the onboard checksum comes out correct.
   3396  */
   3397 static inline int
   3398 bge_cksum_pad(struct mbuf *pkt)
   3399 {
   3400 	struct mbuf *last = NULL;
   3401 	int padlen;
   3402 
   3403 	padlen = ETHER_MIN_NOPAD - pkt->m_pkthdr.len;
   3404 
   3405 	/* if there's only the packet-header and we can pad there, use it. */
   3406 	if (pkt->m_pkthdr.len == pkt->m_len &&
   3407 	    M_TRAILINGSPACE(pkt) >= padlen) {
   3408 		last = pkt;
   3409 	} else {
   3410 		/*
   3411 		 * Walk packet chain to find last mbuf. We will either
   3412 		 * pad there, or append a new mbuf and pad it
   3413 		 * (thus perhaps avoiding the bcm5700 dma-min bug).
   3414 		 */
   3415 		for (last = pkt; last->m_next != NULL; last = last->m_next) {
   3416 	      	       continue; /* do nothing */
   3417 		}
   3418 
   3419 		/* `last' now points to last in chain. */
   3420 		if (M_TRAILINGSPACE(last) < padlen) {
   3421 			/* Allocate new empty mbuf, pad it. Compact later. */
   3422 			struct mbuf *n;
   3423 			MGET(n, M_DONTWAIT, MT_DATA);
   3424 			if (n == NULL)
   3425 				return ENOBUFS;
   3426 			n->m_len = 0;
   3427 			last->m_next = n;
   3428 			last = n;
   3429 		}
   3430 	}
   3431 
   3432 	KDASSERT(!M_READONLY(last));
   3433 	KDASSERT(M_TRAILINGSPACE(last) >= padlen);
   3434 
   3435 	/* Now zero the pad area, to avoid the bge cksum-assist bug */
   3436 	memset(mtod(last, char *) + last->m_len, 0, padlen);
   3437 	last->m_len += padlen;
   3438 	pkt->m_pkthdr.len += padlen;
   3439 	return 0;
   3440 }
   3441 
   3442 /*
   3443  * Compact outbound packets to avoid bug with DMA segments less than 8 bytes.
   3444  */
   3445 static inline int
   3446 bge_compact_dma_runt(struct mbuf *pkt)
   3447 {
   3448 	struct mbuf	*m, *prev;
   3449 	int 		totlen, prevlen;
   3450 
   3451 	prev = NULL;
   3452 	totlen = 0;
   3453 	prevlen = -1;
   3454 
   3455 	for (m = pkt; m != NULL; prev = m,m = m->m_next) {
   3456 		int mlen = m->m_len;
   3457 		int shortfall = 8 - mlen ;
   3458 
   3459 		totlen += mlen;
   3460 		if (mlen == 0) {
   3461 			continue;
   3462 		}
   3463 		if (mlen >= 8)
   3464 			continue;
   3465 
   3466 		/* If we get here, mbuf data is too small for DMA engine.
   3467 		 * Try to fix by shuffling data to prev or next in chain.
   3468 		 * If that fails, do a compacting deep-copy of the whole chain.
   3469 		 */
   3470 
   3471 		/* Internal frag. If fits in prev, copy it there. */
   3472 		if (prev && M_TRAILINGSPACE(prev) >= m->m_len) {
   3473 		  	memcpy(prev->m_data + prev->m_len, m->m_data, mlen);
   3474 			prev->m_len += mlen;
   3475 			m->m_len = 0;
   3476 			/* XXX stitch chain */
   3477 			prev->m_next = m_free(m);
   3478 			m = prev;
   3479 			continue;
   3480 		}
   3481 		else if (m->m_next != NULL &&
   3482 			     M_TRAILINGSPACE(m) >= shortfall &&
   3483 			     m->m_next->m_len >= (8 + shortfall)) {
   3484 		    /* m is writable and have enough data in next, pull up. */
   3485 
   3486 		  	memcpy(m->m_data + m->m_len, m->m_next->m_data,
   3487 			    shortfall);
   3488 			m->m_len += shortfall;
   3489 			m->m_next->m_len -= shortfall;
   3490 			m->m_next->m_data += shortfall;
   3491 		}
   3492 		else if (m->m_next == NULL || 1) {
   3493 		  	/* Got a runt at the very end of the packet.
   3494 			 * borrow data from the tail of the preceding mbuf and
   3495 			 * update its length in-place. (The original data is still
   3496 			 * valid, so we can do this even if prev is not writable.)
   3497 			 */
   3498 
   3499 			/* if we'd make prev a runt, just move all of its data. */
   3500 			KASSERT(prev != NULL /*, ("runt but null PREV")*/);
   3501 			KASSERT(prev->m_len >= 8 /*, ("runt prev")*/);
   3502 
   3503 			if ((prev->m_len - shortfall) < 8)
   3504 				shortfall = prev->m_len;
   3505 
   3506 #ifdef notyet	/* just do the safe slow thing for now */
   3507 			if (!M_READONLY(m)) {
   3508 				if (M_LEADINGSPACE(m) < shorfall) {
   3509 					void *m_dat;
   3510 					m_dat = (m->m_flags & M_PKTHDR) ?
   3511 					  m->m_pktdat : m->dat;
   3512 					memmove(m_dat, mtod(m, void*), m->m_len);
   3513 					m->m_data = m_dat;
   3514 				    }
   3515 			} else
   3516 #endif	/* just do the safe slow thing */
   3517 			{
   3518 				struct mbuf * n = NULL;
   3519 				int newprevlen = prev->m_len - shortfall;
   3520 
   3521 				MGET(n, M_NOWAIT, MT_DATA);
   3522 				if (n == NULL)
   3523 				   return ENOBUFS;
   3524 				KASSERT(m->m_len + shortfall < MLEN
   3525 					/*,
   3526 					  ("runt %d +prev %d too big\n", m->m_len, shortfall)*/);
   3527 
   3528 				/* first copy the data we're stealing from prev */
   3529 				memcpy(n->m_data, prev->m_data + newprevlen,
   3530 				    shortfall);
   3531 
   3532 				/* update prev->m_len accordingly */
   3533 				prev->m_len -= shortfall;
   3534 
   3535 				/* copy data from runt m */
   3536 				memcpy(n->m_data + shortfall, m->m_data,
   3537 				    m->m_len);
   3538 
   3539 				/* n holds what we stole from prev, plus m */
   3540 				n->m_len = shortfall + m->m_len;
   3541 
   3542 				/* stitch n into chain and free m */
   3543 				n->m_next = m->m_next;
   3544 				prev->m_next = n;
   3545 				/* KASSERT(m->m_next == NULL); */
   3546 				m->m_next = NULL;
   3547 				m_free(m);
   3548 				m = n;	/* for continuing loop */
   3549 			}
   3550 		}
   3551 		prevlen = m->m_len;
   3552 	}
   3553 	return 0;
   3554 }
   3555 
   3556 /*
   3557  * Encapsulate an mbuf chain in the tx ring  by coupling the mbuf data
   3558  * pointers to descriptors.
   3559  */
   3560 static int
   3561 bge_encap(struct bge_softc *sc, struct mbuf *m_head, u_int32_t *txidx)
   3562 {
   3563 	struct bge_tx_bd	*f = NULL;
   3564 	u_int32_t		frag, cur;
   3565 	u_int16_t		csum_flags = 0;
   3566 	u_int16_t		txbd_tso_flags = 0;
   3567 	struct txdmamap_pool_entry *dma;
   3568 	bus_dmamap_t dmamap;
   3569 	int			i = 0;
   3570 	struct m_tag		*mtag;
   3571 	int			use_tso, maxsegsize, error;
   3572 
   3573 	cur = frag = *txidx;
   3574 
   3575 	if (m_head->m_pkthdr.csum_flags) {
   3576 		if (m_head->m_pkthdr.csum_flags & M_CSUM_IPv4)
   3577 			csum_flags |= BGE_TXBDFLAG_IP_CSUM;
   3578 		if (m_head->m_pkthdr.csum_flags & (M_CSUM_TCPv4|M_CSUM_UDPv4))
   3579 			csum_flags |= BGE_TXBDFLAG_TCP_UDP_CSUM;
   3580 	}
   3581 
   3582 	/*
   3583 	 * If we were asked to do an outboard checksum, and the NIC
   3584 	 * has the bug where it sometimes adds in the Ethernet padding,
   3585 	 * explicitly pad with zeros so the cksum will be correct either way.
   3586 	 * (For now, do this for all chip versions, until newer
   3587 	 * are confirmed to not require the workaround.)
   3588 	 */
   3589 	if ((csum_flags & BGE_TXBDFLAG_TCP_UDP_CSUM) == 0 ||
   3590 #ifdef notyet
   3591 	    (sc->bge_quirks & BGE_QUIRK_SHORT_CKSUM_BUG) == 0 ||
   3592 #endif
   3593 	    m_head->m_pkthdr.len >= ETHER_MIN_NOPAD)
   3594 		goto check_dma_bug;
   3595 
   3596 	if (bge_cksum_pad(m_head) != 0) {
   3597 	    return ENOBUFS;
   3598 	}
   3599 
   3600 check_dma_bug:
   3601 	if (!(sc->bge_quirks & BGE_QUIRK_5700_SMALLDMA))
   3602 		goto doit;
   3603 	/*
   3604 	 * bcm5700 Revision B silicon cannot handle DMA descriptors with
   3605 	 * less than eight bytes.  If we encounter a teeny mbuf
   3606 	 * at the end of a chain, we can pad.  Otherwise, copy.
   3607 	 */
   3608 	if (bge_compact_dma_runt(m_head) != 0)
   3609 		return ENOBUFS;
   3610 
   3611 doit:
   3612 	dma = SLIST_FIRST(&sc->txdma_list);
   3613 	if (dma == NULL)
   3614 		return ENOBUFS;
   3615 	dmamap = dma->dmamap;
   3616 
   3617 	/*
   3618 	 * Set up any necessary TSO state before we start packing...
   3619 	 */
   3620 	use_tso = (m_head->m_pkthdr.csum_flags & M_CSUM_TSOv4) != 0;
   3621 	if (!use_tso) {
   3622 		maxsegsize = 0;
   3623 	} else {	/* TSO setup */
   3624 		unsigned  mss;
   3625 		struct ether_header *eh;
   3626 		unsigned ip_tcp_hlen, iptcp_opt_words, tcp_seg_flags, offset;
   3627 		struct mbuf * m0 = m_head;
   3628 		struct ip *ip;
   3629 		struct tcphdr *th;
   3630 		int iphl, hlen;
   3631 
   3632 		/*
   3633 		 * XXX It would be nice if the mbuf pkthdr had offset
   3634 		 * fields for the protocol headers.
   3635 		 */
   3636 
   3637 		eh = mtod(m0, struct ether_header *);
   3638 		switch (htons(eh->ether_type)) {
   3639 		case ETHERTYPE_IP:
   3640 			offset = ETHER_HDR_LEN;
   3641 			break;
   3642 
   3643 		case ETHERTYPE_VLAN:
   3644 			offset = ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN;
   3645 			break;
   3646 
   3647 		default:
   3648 			/*
   3649 			 * Don't support this protocol or encapsulation.
   3650 			 */
   3651 			return (ENOBUFS);
   3652 		}
   3653 
   3654 		/*
   3655 		 * TCP/IP headers are in the first mbuf; we can do
   3656 		 * this the easy way.
   3657 		 */
   3658 		iphl = M_CSUM_DATA_IPv4_IPHL(m0->m_pkthdr.csum_data);
   3659 		hlen = iphl + offset;
   3660 		if (__predict_false(m0->m_len <
   3661 				    (hlen + sizeof(struct tcphdr)))) {
   3662 
   3663 			  printf("TSO: hard case m0->m_len == %d <"
   3664 				 " ip/tcp hlen %zd, not handled yet\n",
   3665 				 m0->m_len, hlen+ sizeof(struct tcphdr));
   3666 #ifdef NOTYET
   3667 			/*
   3668 			 * XXX jonathan (at) NetBSD.org: untested.
   3669 			 * how to force  this branch to be taken?
   3670 			 */
   3671 			BGE_EVCNT_INCR(&sc->sc_ev_txtsopain);
   3672 
   3673 			m_copydata(m0, offset, sizeof(ip), &ip);
   3674 			m_copydata(m0, hlen, sizeof(th), &th);
   3675 
   3676 			ip.ip_len = 0;
   3677 
   3678 			m_copyback(m0, hlen + offsetof(struct ip, ip_len),
   3679 			    sizeof(ip.ip_len), &ip.ip_len);
   3680 
   3681 			th.th_sum = in_cksum_phdr(ip.ip_src.s_addr,
   3682 			    ip.ip_dst.s_addr, htons(IPPROTO_TCP));
   3683 
   3684 			m_copyback(m0, hlen + offsetof(struct tcphdr, th_sum),
   3685 			    sizeof(th.th_sum), &th.th_sum);
   3686 
   3687 			hlen += th.th_off << 2;
   3688 			iptcp_opt_words	= hlen;
   3689 #else
   3690 			/*
   3691 			 * if_wm "hard" case not yet supported, can we not
   3692 			 * mandate it out of existence?
   3693 			 */
   3694 			(void) ip; (void)th; (void) ip_tcp_hlen;
   3695 
   3696 			return ENOBUFS;
   3697 #endif
   3698 		} else {
   3699 			ip = (struct ip *) (mtod(m0, char *) + offset);
   3700 			th = (struct tcphdr *) (mtod(m0, char *) + hlen);
   3701 			ip_tcp_hlen = iphl +  (th->th_off << 2);
   3702 
   3703 			/* Total IP/TCP options, in 32-bit words */
   3704 			iptcp_opt_words = (ip_tcp_hlen
   3705 					   - sizeof(struct tcphdr)
   3706 					   - sizeof(struct ip)) >> 2;
   3707 		}
   3708 		if (BGE_IS_5750_OR_BEYOND(sc)) {
   3709 			th->th_sum = 0;
   3710 			csum_flags &= ~(BGE_TXBDFLAG_TCP_UDP_CSUM);
   3711 		} else {
   3712 			/*
   3713 			 * XXX jonathan (at) NetBSD.org: 5705 untested.
   3714 			 * Requires TSO firmware patch for 5701/5703/5704.
   3715 			 */
   3716 			th->th_sum = in_cksum_phdr(ip->ip_src.s_addr,
   3717 			    ip->ip_dst.s_addr, htons(IPPROTO_TCP));
   3718 		}
   3719 
   3720 		mss = m_head->m_pkthdr.segsz;
   3721 		txbd_tso_flags |=
   3722 		    BGE_TXBDFLAG_CPU_PRE_DMA |
   3723 		    BGE_TXBDFLAG_CPU_POST_DMA;
   3724 
   3725 		/*
   3726 		 * Our NIC TSO-assist assumes TSO has standard, optionless
   3727 		 * IPv4 and TCP headers, which total 40 bytes. By default,
   3728 		 * the NIC copies 40 bytes of IP/TCP header from the
   3729 		 * supplied header into the IP/TCP header portion of
   3730 		 * each post-TSO-segment. If the supplied packet has IP or
   3731 		 * TCP options, we need to tell the NIC to copy those extra
   3732 		 * bytes into each  post-TSO header, in addition to the normal
   3733 		 * 40-byte IP/TCP header (and to leave space accordingly).
   3734 		 * Unfortunately, the driver encoding of option length
   3735 		 * varies across different ASIC families.
   3736 		 */
   3737 		tcp_seg_flags = 0;
   3738 		if (iptcp_opt_words) {
   3739 			if ( BGE_IS_5705_OR_BEYOND(sc)) {
   3740 				tcp_seg_flags =
   3741 					iptcp_opt_words << 11;
   3742 			} else {
   3743 				txbd_tso_flags |=
   3744 					iptcp_opt_words << 12;
   3745 			}
   3746 		}
   3747 		maxsegsize = mss | tcp_seg_flags;
   3748 		ip->ip_len = htons(mss + ip_tcp_hlen);
   3749 
   3750 	}	/* TSO setup */
   3751 
   3752 	/*
   3753 	 * Start packing the mbufs in this chain into
   3754 	 * the fragment pointers. Stop when we run out
   3755 	 * of fragments or hit the end of the mbuf chain.
   3756 	 */
   3757 	error = bus_dmamap_load_mbuf(sc->bge_dmatag, dmamap, m_head,
   3758 	    BUS_DMA_NOWAIT);
   3759 	if (error) {
   3760 		return(ENOBUFS);
   3761 	}
   3762 	/*
   3763 	 * Sanity check: avoid coming within 16 descriptors
   3764 	 * of the end of the ring.
   3765 	 */
   3766 	if (dmamap->dm_nsegs > (BGE_TX_RING_CNT - sc->bge_txcnt - 16)) {
   3767 		BGE_TSO_PRINTF(("%s: "
   3768 		    " dmamap_load_mbuf too close to ring wrap\n",
   3769 		    sc->bge_dev.dv_xname));
   3770 		goto fail_unload;
   3771 	}
   3772 
   3773 	mtag = sc->ethercom.ec_nvlans ?
   3774 	    m_tag_find(m_head, PACKET_TAG_VLAN, NULL) : NULL;
   3775 
   3776 
   3777 	/* Iterate over dmap-map fragments. */
   3778 	for (i = 0; i < dmamap->dm_nsegs; i++) {
   3779 		f = &sc->bge_rdata->bge_tx_ring[frag];
   3780 		if (sc->bge_cdata.bge_tx_chain[frag] != NULL)
   3781 			break;
   3782 
   3783 		bge_set_hostaddr(&f->bge_addr, dmamap->dm_segs[i].ds_addr);
   3784 		f->bge_len = dmamap->dm_segs[i].ds_len;
   3785 
   3786 		/*
   3787 		 * For 5751 and follow-ons, for TSO we must turn
   3788 		 * off checksum-assist flag in the tx-descr, and
   3789 		 * supply the ASIC-revision-specific encoding
   3790 		 * of TSO flags and segsize.
   3791 		 */
   3792 		if (use_tso) {
   3793 			if (BGE_IS_5750_OR_BEYOND(sc) || i == 0) {
   3794 				f->bge_rsvd = maxsegsize;
   3795 				f->bge_flags = csum_flags | txbd_tso_flags;
   3796 			} else {
   3797 				f->bge_rsvd = 0;
   3798 				f->bge_flags =
   3799 				  (csum_flags | txbd_tso_flags) & 0x0fff;
   3800 			}
   3801 		} else {
   3802 			f->bge_rsvd = 0;
   3803 			f->bge_flags = csum_flags;
   3804 		}
   3805 
   3806 		if (mtag != NULL) {
   3807 			f->bge_flags |= BGE_TXBDFLAG_VLAN_TAG;
   3808 			f->bge_vlan_tag = VLAN_TAG_VALUE(mtag);
   3809 		} else {
   3810 			f->bge_vlan_tag = 0;
   3811 		}
   3812 		cur = frag;
   3813 		BGE_INC(frag, BGE_TX_RING_CNT);
   3814 	}
   3815 
   3816 	if (i < dmamap->dm_nsegs) {
   3817 		BGE_TSO_PRINTF(("%s: reached %d < dm_nsegs %d\n",
   3818 		    sc->bge_dev.dv_xname, i, dmamap->dm_nsegs));
   3819 		goto fail_unload;
   3820 	}
   3821 
   3822 	bus_dmamap_sync(sc->bge_dmatag, dmamap, 0, dmamap->dm_mapsize,
   3823 	    BUS_DMASYNC_PREWRITE);
   3824 
   3825 	if (frag == sc->bge_tx_saved_considx) {
   3826 		BGE_TSO_PRINTF(("%s: frag %d = wrapped id %d?\n",
   3827 		    sc->bge_dev.dv_xname, frag, sc->bge_tx_saved_considx));
   3828 
   3829 		goto fail_unload;
   3830 	}
   3831 
   3832 	sc->bge_rdata->bge_tx_ring[cur].bge_flags |= BGE_TXBDFLAG_END;
   3833 	sc->bge_cdata.bge_tx_chain[cur] = m_head;
   3834 	SLIST_REMOVE_HEAD(&sc->txdma_list, link);
   3835 	sc->txdma[cur] = dma;
   3836 	sc->bge_txcnt += dmamap->dm_nsegs;
   3837 
   3838 	*txidx = frag;
   3839 
   3840 	return(0);
   3841 
   3842  fail_unload:
   3843 	bus_dmamap_unload(sc->bge_dmatag, dmamap);
   3844 
   3845 	return ENOBUFS;
   3846 }
   3847 
   3848 /*
   3849  * Main transmit routine. To avoid having to do mbuf copies, we put pointers
   3850  * to the mbuf data regions directly in the transmit descriptors.
   3851  */
   3852 static void
   3853 bge_start(struct ifnet *ifp)
   3854 {
   3855 	struct bge_softc *sc;
   3856 	struct mbuf *m_head = NULL;
   3857 	u_int32_t prodidx;
   3858 	int pkts = 0;
   3859 
   3860 	sc = ifp->if_softc;
   3861 
   3862 	if ((ifp->if_flags & (IFF_RUNNING|IFF_OACTIVE)) != IFF_RUNNING)
   3863 		return;
   3864 
   3865 	prodidx = sc->bge_tx_prodidx;
   3866 
   3867 	while(sc->bge_cdata.bge_tx_chain[prodidx] == NULL) {
   3868 		IFQ_POLL(&ifp->if_snd, m_head);
   3869 		if (m_head == NULL)
   3870 			break;
   3871 
   3872 #if 0
   3873 		/*
   3874 		 * XXX
   3875 		 * safety overkill.  If this is a fragmented packet chain
   3876 		 * with delayed TCP/UDP checksums, then only encapsulate
   3877 		 * it if we have enough descriptors to handle the entire
   3878 		 * chain at once.
   3879 		 * (paranoia -- may not actually be needed)
   3880 		 */
   3881 		if (m_head->m_flags & M_FIRSTFRAG &&
   3882 		    m_head->m_pkthdr.csum_flags & (CSUM_DELAY_DATA)) {
   3883 			if ((BGE_TX_RING_CNT - sc->bge_txcnt) <
   3884 			    M_CSUM_DATA_IPv4_OFFSET(m_head->m_pkthdr.csum_data) + 16) {
   3885 				ifp->if_flags |= IFF_OACTIVE;
   3886 				break;
   3887 			}
   3888 		}
   3889 #endif
   3890 
   3891 		/*
   3892 		 * Pack the data into the transmit ring. If we
   3893 		 * don't have room, set the OACTIVE flag and wait
   3894 		 * for the NIC to drain the ring.
   3895 		 */
   3896 		if (bge_encap(sc, m_head, &prodidx)) {
   3897 			ifp->if_flags |= IFF_OACTIVE;
   3898 			break;
   3899 		}
   3900 
   3901 		/* now we are committed to transmit the packet */
   3902 		IFQ_DEQUEUE(&ifp->if_snd, m_head);
   3903 		pkts++;
   3904 
   3905 #if NBPFILTER > 0
   3906 		/*
   3907 		 * If there's a BPF listener, bounce a copy of this frame
   3908 		 * to him.
   3909 		 */
   3910 		if (ifp->if_bpf)
   3911 			bpf_mtap(ifp->if_bpf, m_head);
   3912 #endif
   3913 	}
   3914 	if (pkts == 0)
   3915 		return;
   3916 
   3917 	/* Transmit */
   3918 	CSR_WRITE_4(sc, BGE_MBX_TX_HOST_PROD0_LO, prodidx);
   3919 	if (sc->bge_quirks & BGE_QUIRK_PRODUCER_BUG)	/* 5700 b2 errata */
   3920 		CSR_WRITE_4(sc, BGE_MBX_TX_HOST_PROD0_LO, prodidx);
   3921 
   3922 	sc->bge_tx_prodidx = prodidx;
   3923 
   3924 	/*
   3925 	 * Set a timeout in case the chip goes out to lunch.
   3926 	 */
   3927 	ifp->if_timer = 5;
   3928 }
   3929 
   3930 static int
   3931 bge_init(struct ifnet *ifp)
   3932 {
   3933 	struct bge_softc *sc = ifp->if_softc;
   3934 	u_int16_t *m;
   3935 	int s, error;
   3936 
   3937 	s = splnet();
   3938 
   3939 	ifp = &sc->ethercom.ec_if;
   3940 
   3941 	/* Cancel pending I/O and flush buffers. */
   3942 	bge_stop(sc);
   3943 	bge_reset(sc);
   3944 	bge_chipinit(sc);
   3945 
   3946 	/*
   3947 	 * Init the various state machines, ring
   3948 	 * control blocks and firmware.
   3949 	 */
   3950 	error = bge_blockinit(sc);
   3951 	if (error != 0) {
   3952 		printf("%s: initialization error %d\n", sc->bge_dev.dv_xname,
   3953 		    error);
   3954 		splx(s);
   3955 		return error;
   3956 	}
   3957 
   3958 	ifp = &sc->ethercom.ec_if;
   3959 
   3960 	/* Specify MTU. */
   3961 	CSR_WRITE_4(sc, BGE_RX_MTU, ifp->if_mtu +
   3962 	    ETHER_HDR_LEN + ETHER_CRC_LEN + ETHER_VLAN_ENCAP_LEN);
   3963 
   3964 	/* Load our MAC address. */
   3965 	m = (u_int16_t *)&(LLADDR(ifp->if_sadl)[0]);
   3966 	CSR_WRITE_4(sc, BGE_MAC_ADDR1_LO, htons(m[0]));
   3967 	CSR_WRITE_4(sc, BGE_MAC_ADDR1_HI, (htons(m[1]) << 16) | htons(m[2]));
   3968 
   3969 	/* Enable or disable promiscuous mode as needed. */
   3970 	if (ifp->if_flags & IFF_PROMISC) {
   3971 		BGE_SETBIT(sc, BGE_RX_MODE, BGE_RXMODE_RX_PROMISC);
   3972 	} else {
   3973 		BGE_CLRBIT(sc, BGE_RX_MODE, BGE_RXMODE_RX_PROMISC);
   3974 	}
   3975 
   3976 	/* Program multicast filter. */
   3977 	bge_setmulti(sc);
   3978 
   3979 	/* Init RX ring. */
   3980 	bge_init_rx_ring_std(sc);
   3981 
   3982 	/* Init jumbo RX ring. */
   3983 	if (ifp->if_mtu > (ETHERMTU + ETHER_HDR_LEN + ETHER_CRC_LEN))
   3984 		bge_init_rx_ring_jumbo(sc);
   3985 
   3986 	/* Init our RX return ring index */
   3987 	sc->bge_rx_saved_considx = 0;
   3988 
   3989 	/* Init TX ring. */
   3990 	bge_init_tx_ring(sc);
   3991 
   3992 	/* Turn on transmitter */
   3993 	BGE_SETBIT(sc, BGE_TX_MODE, BGE_TXMODE_ENABLE);
   3994 
   3995 	/* Turn on receiver */
   3996 	BGE_SETBIT(sc, BGE_RX_MODE, BGE_RXMODE_ENABLE);
   3997 
   3998 	CSR_WRITE_4(sc, BGE_MAX_RX_FRAME_LOWAT, 2);
   3999 
   4000 	/* Tell firmware we're alive. */
   4001 	BGE_SETBIT(sc, BGE_MODE_CTL, BGE_MODECTL_STACKUP);
   4002 
   4003 	/* Enable host interrupts. */
   4004 	BGE_SETBIT(sc, BGE_PCI_MISC_CTL, BGE_PCIMISCCTL_CLEAR_INTA);
   4005 	BGE_CLRBIT(sc, BGE_PCI_MISC_CTL, BGE_PCIMISCCTL_MASK_PCI_INTR);
   4006 	CSR_WRITE_4(sc, BGE_MBX_IRQ0_LO, 0);
   4007 
   4008 	bge_ifmedia_upd(ifp);
   4009 
   4010 	ifp->if_flags |= IFF_RUNNING;
   4011 	ifp->if_flags &= ~IFF_OACTIVE;
   4012 
   4013 	splx(s);
   4014 
   4015 	callout_reset(&sc->bge_timeout, hz, bge_tick, sc);
   4016 
   4017 	return 0;
   4018 }
   4019 
   4020 /*
   4021  * Set media options.
   4022  */
   4023 static int
   4024 bge_ifmedia_upd(struct ifnet *ifp)
   4025 {
   4026 	struct bge_softc *sc = ifp->if_softc;
   4027 	struct mii_data *mii = &sc->bge_mii;
   4028 	struct ifmedia *ifm = &sc->bge_ifmedia;
   4029 
   4030 	/* If this is a 1000baseX NIC, enable the TBI port. */
   4031 	if (sc->bge_tbi) {
   4032 		if (IFM_TYPE(ifm->ifm_media) != IFM_ETHER)
   4033 			return(EINVAL);
   4034 		switch(IFM_SUBTYPE(ifm->ifm_media)) {
   4035 		case IFM_AUTO:
   4036 			break;
   4037 		case IFM_1000_SX:
   4038 			if ((ifm->ifm_media & IFM_GMASK) == IFM_FDX) {
   4039 				BGE_CLRBIT(sc, BGE_MAC_MODE,
   4040 				    BGE_MACMODE_HALF_DUPLEX);
   4041 			} else {
   4042 				BGE_SETBIT(sc, BGE_MAC_MODE,
   4043 				    BGE_MACMODE_HALF_DUPLEX);
   4044 			}
   4045 			break;
   4046 		default:
   4047 			return(EINVAL);
   4048 		}
   4049 		/* XXX 802.3x flow control for 1000BASE-SX */
   4050 		return(0);
   4051 	}
   4052 
   4053 	sc->bge_link = 0;
   4054 	mii_mediachg(mii);
   4055 
   4056 	return(0);
   4057 }
   4058 
   4059 /*
   4060  * Report current media status.
   4061  */
   4062 static void
   4063 bge_ifmedia_sts(struct ifnet *ifp, struct ifmediareq *ifmr)
   4064 {
   4065 	struct bge_softc *sc = ifp->if_softc;
   4066 	struct mii_data *mii = &sc->bge_mii;
   4067 
   4068 	if (sc->bge_tbi) {
   4069 		ifmr->ifm_status = IFM_AVALID;
   4070 		ifmr->ifm_active = IFM_ETHER;
   4071 		if (CSR_READ_4(sc, BGE_MAC_STS) &
   4072 		    BGE_MACSTAT_TBI_PCS_SYNCHED)
   4073 			ifmr->ifm_status |= IFM_ACTIVE;
   4074 		ifmr->ifm_active |= IFM_1000_SX;
   4075 		if (CSR_READ_4(sc, BGE_MAC_MODE) & BGE_MACMODE_HALF_DUPLEX)
   4076 			ifmr->ifm_active |= IFM_HDX;
   4077 		else
   4078 			ifmr->ifm_active |= IFM_FDX;
   4079 		return;
   4080 	}
   4081 
   4082 	mii_pollstat(mii);
   4083 	ifmr->ifm_status = mii->mii_media_status;
   4084 	ifmr->ifm_active = (mii->mii_media_active & ~IFM_ETH_FMASK) |
   4085 	    sc->bge_flowflags;
   4086 }
   4087 
   4088 static int
   4089 bge_ioctl(struct ifnet *ifp, u_long command, void *data)
   4090 {
   4091 	struct bge_softc *sc = ifp->if_softc;
   4092 	struct ifreq *ifr = (struct ifreq *) data;
   4093 	int s, error = 0;
   4094 	struct mii_data *mii;
   4095 
   4096 	s = splnet();
   4097 
   4098 	switch(command) {
   4099 	case SIOCSIFFLAGS:
   4100 		if (ifp->if_flags & IFF_UP) {
   4101 			/*
   4102 			 * If only the state of the PROMISC flag changed,
   4103 			 * then just use the 'set promisc mode' command
   4104 			 * instead of reinitializing the entire NIC. Doing
   4105 			 * a full re-init means reloading the firmware and
   4106 			 * waiting for it to start up, which may take a
   4107 			 * second or two.
   4108 			 */
   4109 			if (ifp->if_flags & IFF_RUNNING &&
   4110 			    ifp->if_flags & IFF_PROMISC &&
   4111 			    !(sc->bge_if_flags & IFF_PROMISC)) {
   4112 				BGE_SETBIT(sc, BGE_RX_MODE,
   4113 				    BGE_RXMODE_RX_PROMISC);
   4114 			} else if (ifp->if_flags & IFF_RUNNING &&
   4115 			    !(ifp->if_flags & IFF_PROMISC) &&
   4116 			    sc->bge_if_flags & IFF_PROMISC) {
   4117 				BGE_CLRBIT(sc, BGE_RX_MODE,
   4118 				    BGE_RXMODE_RX_PROMISC);
   4119 			} else if (!(sc->bge_if_flags & IFF_UP))
   4120 				bge_init(ifp);
   4121 		} else {
   4122 			if (ifp->if_flags & IFF_RUNNING) {
   4123 				bge_stop(sc);
   4124 			}
   4125 		}
   4126 		sc->bge_if_flags = ifp->if_flags;
   4127 		error = 0;
   4128 		break;
   4129 	case SIOCSIFMEDIA:
   4130 		/* XXX Flow control is not supported for 1000BASE-SX */
   4131 		if (sc->bge_tbi) {
   4132 			ifr->ifr_media &= ~IFM_ETH_FMASK;
   4133 			sc->bge_flowflags = 0;
   4134 		}
   4135 
   4136 		/* Flow control requires full-duplex mode. */
   4137 		if (IFM_SUBTYPE(ifr->ifr_media) == IFM_AUTO ||
   4138 		    (ifr->ifr_media & IFM_FDX) == 0) {
   4139 		    	ifr->ifr_media &= ~IFM_ETH_FMASK;
   4140 		}
   4141 		if (IFM_SUBTYPE(ifr->ifr_media) != IFM_AUTO) {
   4142 			if ((ifr->ifr_media & IFM_ETH_FMASK) == IFM_FLOW) {
   4143 				/* We an do both TXPAUSE and RXPAUSE. */
   4144 				ifr->ifr_media |=
   4145 				    IFM_ETH_TXPAUSE | IFM_ETH_RXPAUSE;
   4146 			}
   4147 			sc->bge_flowflags = ifr->ifr_media & IFM_ETH_FMASK;
   4148 		}
   4149 		/* FALLTHROUGH */
   4150 	case SIOCGIFMEDIA:
   4151 		if (sc->bge_tbi) {
   4152 			error = ifmedia_ioctl(ifp, ifr, &sc->bge_ifmedia,
   4153 			    command);
   4154 		} else {
   4155 			mii = &sc->bge_mii;
   4156 			error = ifmedia_ioctl(ifp, ifr, &mii->mii_media,
   4157 			    command);
   4158 		}
   4159 		break;
   4160 	default:
   4161 		error = ether_ioctl(ifp, command, data);
   4162 		if (error == ENETRESET) {
   4163 			if (ifp->if_flags & IFF_RUNNING)
   4164 				bge_setmulti(sc);
   4165 			error = 0;
   4166 		}
   4167 		break;
   4168 	}
   4169 
   4170 	splx(s);
   4171 
   4172 	return(error);
   4173 }
   4174 
   4175 static void
   4176 bge_watchdog(struct ifnet *ifp)
   4177 {
   4178 	struct bge_softc *sc;
   4179 
   4180 	sc = ifp->if_softc;
   4181 
   4182 	printf("%s: watchdog timeout -- resetting\n", sc->bge_dev.dv_xname);
   4183 
   4184 	ifp->if_flags &= ~IFF_RUNNING;
   4185 	bge_init(ifp);
   4186 
   4187 	ifp->if_oerrors++;
   4188 }
   4189 
   4190 static void
   4191 bge_stop_block(struct bge_softc *sc, bus_addr_t reg, uint32_t bit)
   4192 {
   4193 	int i;
   4194 
   4195 	BGE_CLRBIT(sc, reg, bit);
   4196 
   4197 	for (i = 0; i < BGE_TIMEOUT; i++) {
   4198 		if ((CSR_READ_4(sc, reg) & bit) == 0)
   4199 			return;
   4200 		delay(100);
   4201 		if (sc->bge_pcie)
   4202 		  DELAY(1000);
   4203 	}
   4204 
   4205 	printf("%s: block failed to stop: reg 0x%lx, bit 0x%08x\n",
   4206 	    sc->bge_dev.dv_xname, (u_long) reg, bit);
   4207 }
   4208 
   4209 /*
   4210  * Stop the adapter and free any mbufs allocated to the
   4211  * RX and TX lists.
   4212  */
   4213 static void
   4214 bge_stop(struct bge_softc *sc)
   4215 {
   4216 	struct ifnet *ifp = &sc->ethercom.ec_if;
   4217 
   4218 	callout_stop(&sc->bge_timeout);
   4219 
   4220 	/*
   4221 	 * Disable all of the receiver blocks
   4222 	 */
   4223 	bge_stop_block(sc, BGE_RX_MODE, BGE_RXMODE_ENABLE);
   4224 	bge_stop_block(sc, BGE_RBDI_MODE, BGE_RBDIMODE_ENABLE);
   4225 	bge_stop_block(sc, BGE_RXLP_MODE, BGE_RXLPMODE_ENABLE);
   4226 	if ((sc->bge_quirks & BGE_QUIRK_5705_CORE) == 0) {
   4227 		bge_stop_block(sc, BGE_RXLS_MODE, BGE_RXLSMODE_ENABLE);
   4228 	}
   4229 	bge_stop_block(sc, BGE_RDBDI_MODE, BGE_RBDIMODE_ENABLE);
   4230 	bge_stop_block(sc, BGE_RDC_MODE, BGE_RDCMODE_ENABLE);
   4231 	bge_stop_block(sc, BGE_RBDC_MODE, BGE_RBDCMODE_ENABLE);
   4232 
   4233 	/*
   4234 	 * Disable all of the transmit blocks
   4235 	 */
   4236 	bge_stop_block(sc, BGE_SRS_MODE, BGE_SRSMODE_ENABLE);
   4237 	bge_stop_block(sc, BGE_SBDI_MODE, BGE_SBDIMODE_ENABLE);
   4238 	bge_stop_block(sc, BGE_SDI_MODE, BGE_SDIMODE_ENABLE);
   4239 	bge_stop_block(sc, BGE_RDMA_MODE, BGE_RDMAMODE_ENABLE);
   4240 	bge_stop_block(sc, BGE_SDC_MODE, BGE_SDCMODE_ENABLE);
   4241 	if ((sc->bge_quirks & BGE_QUIRK_5705_CORE) == 0) {
   4242 		bge_stop_block(sc, BGE_DMAC_MODE, BGE_DMACMODE_ENABLE);
   4243 	}
   4244 	bge_stop_block(sc, BGE_SBDC_MODE, BGE_SBDCMODE_ENABLE);
   4245 
   4246 	/*
   4247 	 * Shut down all of the memory managers and related
   4248 	 * state machines.
   4249 	 */
   4250 	bge_stop_block(sc, BGE_HCC_MODE, BGE_HCCMODE_ENABLE);
   4251 	bge_stop_block(sc, BGE_WDMA_MODE, BGE_WDMAMODE_ENABLE);
   4252 	if ((sc->bge_quirks & BGE_QUIRK_5705_CORE) == 0) {
   4253 		bge_stop_block(sc, BGE_MBCF_MODE, BGE_MBCFMODE_ENABLE);
   4254 	}
   4255 
   4256 	CSR_WRITE_4(sc, BGE_FTQ_RESET, 0xFFFFFFFF);
   4257 	CSR_WRITE_4(sc, BGE_FTQ_RESET, 0);
   4258 
   4259 	if ((sc->bge_quirks & BGE_QUIRK_5705_CORE) == 0) {
   4260 		bge_stop_block(sc, BGE_BMAN_MODE, BGE_BMANMODE_ENABLE);
   4261 		bge_stop_block(sc, BGE_MARB_MODE, BGE_MARBMODE_ENABLE);
   4262 	}
   4263 
   4264 	/* Disable host interrupts. */
   4265 	BGE_SETBIT(sc, BGE_PCI_MISC_CTL, BGE_PCIMISCCTL_MASK_PCI_INTR);
   4266 	CSR_WRITE_4(sc, BGE_MBX_IRQ0_LO, 1);
   4267 
   4268 	/*
   4269 	 * Tell firmware we're shutting down.
   4270 	 */
   4271 	BGE_CLRBIT(sc, BGE_MODE_CTL, BGE_MODECTL_STACKUP);
   4272 
   4273 	/* Free the RX lists. */
   4274 	bge_free_rx_ring_std(sc);
   4275 
   4276 	/* Free jumbo RX list. */
   4277 	bge_free_rx_ring_jumbo(sc);
   4278 
   4279 	/* Free TX buffers. */
   4280 	bge_free_tx_ring(sc);
   4281 
   4282 	/*
   4283 	 * Isolate/power down the PHY.
   4284 	 */
   4285 	if (!sc->bge_tbi)
   4286 		mii_down(&sc->bge_mii);
   4287 
   4288 	sc->bge_link = 0;
   4289 
   4290 	sc->bge_tx_saved_considx = BGE_TXCONS_UNSET;
   4291 
   4292 	ifp->if_flags &= ~(IFF_RUNNING | IFF_OACTIVE);
   4293 }
   4294 
   4295 /*
   4296  * Stop all chip I/O so that the kernel's probe routines don't
   4297  * get confused by errant DMAs when rebooting.
   4298  */
   4299 static void
   4300 bge_shutdown(void *xsc)
   4301 {
   4302 	struct bge_softc *sc = (struct bge_softc *)xsc;
   4303 
   4304 	bge_stop(sc);
   4305 	bge_reset(sc);
   4306 }
   4307 
   4308 
   4309 static int
   4310 sysctl_bge_verify(SYSCTLFN_ARGS)
   4311 {
   4312 	int error, t;
   4313 	struct sysctlnode node;
   4314 
   4315 	node = *rnode;
   4316 	t = *(int*)rnode->sysctl_data;
   4317 	node.sysctl_data = &t;
   4318 	error = sysctl_lookup(SYSCTLFN_CALL(&node));
   4319 	if (error || newp == NULL)
   4320 		return (error);
   4321 
   4322 #if 0
   4323 	DPRINTF2(("%s: t = %d, nodenum = %d, rnodenum = %d\n", __func__, t,
   4324 	    node.sysctl_num, rnode->sysctl_num));
   4325 #endif
   4326 
   4327 	if (node.sysctl_num == bge_rxthresh_nodenum) {
   4328 		if (t < 0 || t >= NBGE_RX_THRESH)
   4329 			return (EINVAL);
   4330 		bge_update_all_threshes(t);
   4331 	} else
   4332 		return (EINVAL);
   4333 
   4334 	*(int*)rnode->sysctl_data = t;
   4335 
   4336 	return (0);
   4337 }
   4338 
   4339 /*
   4340  * Set up sysctl(3) MIB, hw.bge.*.
   4341  *
   4342  * TBD condition SYSCTL_PERMANENT on being an LKM or not
   4343  */
   4344 SYSCTL_SETUP(sysctl_bge, "sysctl bge subtree setup")
   4345 {
   4346 	int rc, bge_root_num;
   4347 	const struct sysctlnode *node;
   4348 
   4349 	if ((rc = sysctl_createv(clog, 0, NULL, NULL,
   4350 	    CTLFLAG_PERMANENT, CTLTYPE_NODE, "hw", NULL,
   4351 	    NULL, 0, NULL, 0, CTL_HW, CTL_EOL)) != 0) {
   4352 		goto err;
   4353 	}
   4354 
   4355 	if ((rc = sysctl_createv(clog, 0, NULL, &node,
   4356 	    CTLFLAG_PERMANENT, CTLTYPE_NODE, "bge",
   4357 	    SYSCTL_DESCR("BGE interface controls"),
   4358 	    NULL, 0, NULL, 0, CTL_HW, CTL_CREATE, CTL_EOL)) != 0) {
   4359 		goto err;
   4360 	}
   4361 
   4362 	bge_root_num = node->sysctl_num;
   4363 
   4364 	/* BGE Rx interrupt mitigation level */
   4365 	if ((rc = sysctl_createv(clog, 0, NULL, &node,
   4366 	    CTLFLAG_PERMANENT|CTLFLAG_READWRITE,
   4367 	    CTLTYPE_INT, "rx_lvl",
   4368 	    SYSCTL_DESCR("BGE receive interrupt mitigation level"),
   4369 	    sysctl_bge_verify, 0,
   4370 	    &bge_rx_thresh_lvl,
   4371 	    0, CTL_HW, bge_root_num, CTL_CREATE,
   4372 	    CTL_EOL)) != 0) {
   4373 		goto err;
   4374 	}
   4375 
   4376 	bge_rxthresh_nodenum = node->sysctl_num;
   4377 
   4378 	return;
   4379 
   4380 err:
   4381 	printf("%s: sysctl_createv failed (rc = %d)\n", __func__, rc);
   4382 }
   4383 
   4384 static void
   4385 bge_powerhook(int why, void *hdl)
   4386 {
   4387 	struct bge_softc *sc = (struct bge_softc *)hdl;
   4388 	struct ifnet *ifp = &sc->ethercom.ec_if;
   4389 	struct pci_attach_args *pa = &(sc->bge_pa);
   4390 	pci_chipset_tag_t pc = pa->pa_pc;
   4391 	pcitag_t tag = pa->pa_tag;
   4392 
   4393 	switch (why) {
   4394 	case PWR_SOFTSUSPEND:
   4395 	case PWR_SOFTSTANDBY:
   4396 		bge_shutdown(sc);
   4397 		break;
   4398 	case PWR_SOFTRESUME:
   4399 		if (ifp->if_flags & IFF_UP) {
   4400 			ifp->if_flags &= ~IFF_RUNNING;
   4401 			bge_init(ifp);
   4402 		}
   4403 		break;
   4404 	case PWR_SUSPEND:
   4405 	case PWR_STANDBY:
   4406 		pci_conf_capture(pc, tag, &sc->bge_pciconf);
   4407 		break;
   4408 	case PWR_RESUME:
   4409 		pci_conf_restore(pc, tag, &sc->bge_pciconf);
   4410 		break;
   4411 	}
   4412 
   4413 	return;
   4414 }
   4415