call gem_setladrf() if the NIC is already running and only the
PROMISC or ALLMULTI flags are being adjusted or if adjusting
the multicast addresses instead of doing a full re-init of the
chip.

ok ke

call gem_setladrf() if the NIC is already running and only the
PROMISC or ALLMULTI flags are being adjusted or if adjusting
the multicast addresses instead of doing a full re-init of the
chip.

ok kettenis@

show more ...

add IFQ_SET_MAXLEN().

splimp -> splnet

sync mbuf before unloading in gem_rxdrain().

From NetBSD

handle TX underrun and packet too long errors by resetting the chip.

From NetBSD

ok krw@

- some DEBUG should be GEM_DEBUG
- introduce gem_bitwait() to factor out some of the register wait code

From NetBSD

- remove some statics
- remove parameter names from prototypes

ok kettenis@

typo

Start using the unused sc_variant to identify the GMAC vendor and chipset
model to the MI part of the driver. On Apple systems only attach the PHY to
location 0, unless using a K2 GMAC, then use loca

Start using the unused sc_variant to identify the GMAC vendor and chipset
model to the MI part of the driver. On Apple systems only attach the PHY to
location 0, unless using a K2 GMAC, then use location 1.

Tested by drahn@ and kettenis@

show more ...

have gem strip off the Ethernet FCS before passing it to bpf.

ok krw@, also tested by Peter Hessler on macppc

copy the MAC address from sc_arpcom.ac_enaddr to sc_enaddr in
gem_init() just before calling gem_init_regs() so the proper
MAC address gets programmed into the card (if it has been
changed by say ifc

copy the MAC address from sc_arpcom.ac_enaddr to sc_enaddr in
gem_init() just before calling gem_init_regs() so the proper
MAC address gets programmed into the card (if it has been
changed by say ifconfig lladdr for example).

Noticed and diff tested by: Troex Nevelin <troex at fury dot scancode dot ru>

show more ...

remove whitespace

clear IFF_RUNNING & IFF_OACTIVE in foo_stop() before de-allocating resources.

remove netns crud.
some drivers actually had hooks for SIOCSIFADDR, most just useless includes
"looks good" deraadt miod brad

make sure interface is in RUNNING state before touching the multicast filters

From NetBSD

NetBSD PR 27678 for details

ok mcbride@

allow for VLAN-sized frames.

take 2 times less room in dmesg, prettier too
allocate interrupts early on (before PHY attachment, etc) just like all
other drivers do to allow the above; ok mcbride

Don't return random garbage in the default case; return the smallest ring
descriptor instead.

From NetBSD

ok deraadt@

remove prototypes for non-existant functions.

ok deraadt@

no vtophys(), don't need uvm_extern.h anymore.

tested on alpha, i386, macppc and sparc64.

ok millert@ mickey@

correct some debugging printfs

Eliminate creation/destruction of dmamap's for every xmit by reusing
statically allocated maps.

Significant performance boost on both sparc64 and macppc.

Tested on macppc by drahn@.

ok jason@ drah

Eliminate creation/destruction of dmamap's for every xmit by reusing
statically allocated maps.

Significant performance boost on both sparc64 and macppc.

Tested on macppc by drahn@.

ok jason@ drahn@.

show more ...

strcpy/sprintf cleanup of sys/dev. miod@, deraadt@ says to commit.

The gem driver will sometimes unmap bus_dma mappings that are still
in use by the hardware. Since the sparc64 has IOMMUs, an access by
the hardware after the unmap operation will cause a fault.

Thi

The gem driver will sometimes unmap bus_dma mappings that are still
in use by the hardware. Since the sparc64 has IOMMUs, an access by
the hardware after the unmap operation will cause a fault.

This replaces "disable" calls by "reset" calls. This increases
the time between the command to disable DMA and when the maps
are actually unloaded.

show more ...

- Either memset() or loop and set to zero, not both
- reading RX_COMPLETION is a waste of time (and a -slow- PCI read vs. an,
albeit uncached, memory access to determine the same thing).
- on RX_OVER

- Either memset() or loop and set to zero, not both
- reading RX_COMPLETION is a waste of time (and a -slow- PCI read vs. an,
albeit uncached, memory access to determine the same thing).
- on RX_OVERFLOW, reset the board; the RX unit has probably wedged

show more ...

1000baseTX -> 1000baseT
- More technically correct
- Matches FreeBSD and NetBSD
- Preserved #define for 1000baseTX for backwards compatibility
ok jason@