- Only update the rx ring consumer pointer after running through the rx loop,
not with each iteration through the loop.
- If possible, use a fast interupt handler instead of an ithread handler. Use
the interrupt handler to check and squelch the interrupt, then schedule a
taskqueue to do the actual work. This has three benefits:
- Eliminates the 'interrupt aliasing' problem found in many chipsets by
allowing the driver to mask the interrupt in the NIC instead of the
OS masking the interrupt in the APIC.
- Allows the driver to control the amount of work done in the interrupt
handler. This results in what I call 'adaptive polling', where you get
the latency benefits of a quick response to interrupts with the
interrupt mitigation and work partitioning of polling. Polling is still
an option in the driver, but I consider it orthogonal to this work.
- Don't hold the driver lock in the RX handler. The handler and all data
associated is effectively serialized already. This eliminates the cost of
dropping and reaquiring the lock for every receieved packet. The result
is much lower contention for the driver lock, resulting in lower CPU usage
and lower latency for interactive workloads.
The amount of work done in the taskqueue is controlled by the sysctl
dev.em.N.rx_processing_limit
and tunable
hw.em.rx_process_limit
Setting these to -1 effectively removes the limit.
The fast interrupt and taskqueue can be disabled by defining NO_EM_FASTINTR.
This work has been shown to increase fast-forwarding from ~570 kpps to
~750 kpps (note that the same NIC hardware seems unable to transmit more than
800 kpps, so this increase appears to be limited almost solely by the
hardware). Gains have been shown in other workloads, ranging from better
performance to elimination of over-saturation livelocks.
Thanks to Andre Opperman for his time and resources from his network
performance project in performing much of the testing. Thanks to Gleb
Smirnoff and Danny Braniss for their help in testing also.
case if memory allocation failed.
- Remove fourth argument from VLAN_INPUT_TAG(), that was used
incorrectly in almost all drivers. Indicate failure with
mbuf value of NULL.
In collaboration with: yongari, ru, sam
- don't force busdma to pre-allocate bounce pages for parent tag.
- use system supplied roundup2 macro instead of rolling its own version.
- TX/RX decriptor length should be multiple of 128. There is no
no need to expand the size with the multiple of 4096.
- don't create/destroy DMA maps in TX/RX handlers. Use pre-allocated
DMA maps. Since creating DMA maps on sparc64 is time consuming
operations(resource mananger overhead), this change should boost
performance on sparc64. I could get > 2x speedup on Ultra60.
- TX/RX descriptors could be aligned on 128 boundary. Aligning them
on PAGE_SIZE is waste of resource.
- don't blindly create TX DMA tag with size of MCLBYTES * 8. The size
is only valid under jumbo frame environments. Instead of using the
hardcoded value, re-compute necessary size on the fly.
- RX side bus_dmamap_load_mbuf_sg(9) support.
- remove unused macro EM_ROUNDUP and constant EM_MMBA.
Reviewed by: scottl
Tested by: glebius
for a notebook with em(4) adapter.
- Introduce tunables em.hw.txd and em.hw.rxd, which allow administrator
to configure number of transmit and receive descriptors.
- Check em.hw.txd and em.hw.rxd against hardware limits [*] and require
them to be multiple of 128.
[*] According to comments in if_em.h the 82540EM/82541ER chips can handle
more than 256 descriptors. Since we don't have this hardware to test,
we decided to mimic NetBSD wm(4) driver, that limits these chips to
256 descriptors.
In collaboration with: yongari
rather than in ifindex_table[]; all (except one) accesses are
through ifp anyway. IF_LLADDR() works faster, and all (except
one) ifaddr_byindex() users were converted to use ifp->if_addr.
- Stop storing a (pointer to) Ethernet address in "struct arpcom",
and drop the IFP2ENADDR() macro; all users have been converted
to use IF_LLADDR() instead.
overruns and number of watchdog timeouts.
- Do not log(9) RX overrun events, since this pessimizes
things under load [1].
- Do not increase if->if_oerrors in em_watchdog(), since
this leads to counter slipping back, when if->if_oerrors
is recalculated in em_update_stats_counters(). Instead
increase watchdog counter in em_watchdog() and take it
into account in em_update_stats_counters().
Submitted by: ade [1]
- disable jumbo frame support on strict alignment architectures due
to the limitation of hardware. The driver needs a fix-up code for
RX side. The fix will show up in near future.
- fix endian issue for 82544 on PCI-X bus. I couldn't test this as
I don't have the NIC/hardware.
- prefer PCIR_BAR to hardcoded EM_MMBA.
- Properly checks for for 64bit BAR [1]
- replace inl/outl with bus_space(9) [1]
- fix endian issue on VLAN handling.
- reorder header files and remove unnecessary one.
Reviewed by: cognet
No response from: pdeuskar, tackerman
Obtained from: OpenBSD [1]
since the link takes a bit to negotiate, the information is pretty
much never available during the probe. As such, the boot output
pretty much always prints N/A for speed and duplex. Since we print
out the output of ifconfig during the user space boot, this early
boot information is also generally redundant, and added to the noise.
MFC after: 2 weeks
- Destroy mutex in case of attach failure. [1]
- Lock properly em_watchdog(). [1]
- Lock properly em_sysctl_int_delay(). [1]
- Remove unused global adapter linked list.
- Remove unused dma_size field from struct em_dma_alloc.
- Do not touch interface statistics, that must be edited
only by upper layers. [1]
Submitted by: yongari [1]
o Do not mask the RX overrun interrupt.
o Rewrite em_intr():
- Axe EM_MAX_INTR.
- Cycle acknowledging interrupts and processing
packets until zero interrupt cause register is
read.
- If RX overrun comes in log this fact. [ NetBSD also
resets adapter in this case, but my tests showed that
this is not needed and only pessimizes behavior under
heavy load. ]
- Since almost all functions is rewritten, style the
remaining lines.
This fixes em(4) interfaces wedging under high load.
In collaboration with: wpaul, cognet
Obtained from: NetBSD
flag. This fixes panic, when 'ifconfig em0 down' was called and it calls
em_stop() while the em_process_receive_interrupts() has temporarily
dropped the lock.
the descriptors set.
- In em_process_receive_interrupts(), call bus_dmamap_sync() for the
descriptors set each time we modify one descriptor, instead of doing it only
at the function exit, to make sure the adapters know he can re-use the
descriptor.
This helps on arm with write-back data cache (and possibly on other arches
with bounce pages, I don't know) under heavy network load. Without this,
if we attempt to process more than num_rx_desc descriptors, the adapter
would just stop processing rx interrupts.
The receive function em_process_receive_interrupts() unlocks the
adapter while ether_input() processes the packet, and then locks
it back. In the meantime, em_init() may be called, either from
em_watchdog() from softclock interrupt or from the ifconfig(8)
program. The em_init() resets the card, in particular it sets
adapter->next_rx_desc_to_check to 0 and resets hardware RX Head
and Tail descriptor pointers. The loop in
em_process_receive_interrupts() does not expect these things to
change, and a mess may result.
This fixes long wedges of em(4) interfaces receive part under high
load and IP fastforwarding enabled.
PR: kern/87418
Submitted by: Dmitrij Tejblum <tejblum yandex-team.ru>
opt_device_polling.h
- Include opt_device_polling.h into appropriate files.
- Embrace with HAVE_KERNEL_OPTION_HEADERS the include in the files that
can be compiled as loadable modules.
Reviewed by: bde
o Axe poll in trap.
o Axe IFF_POLLING flag from if_flags.
o Rework revision 1.21 (Giant removal), in such a way that
poll_mtx is not dropped during call to polling handler.
This fixes problem with idle polling.
o Make registration and deregistration from polling in a
functional way, insted of next tick/interrupt.
o Obsolete kern.polling.enable. Polling is turned on/off
with ifconfig.
Detailed kern_poll.c changes:
- Remove polling handler flags, introduced in 1.21. The are not
needed now.
- Forget and do not check if_flags, if_capenable and if_drv_flags.
- Call all registered polling handlers unconditionally.
- Do not drop poll_mtx, when entering polling handlers.
- In ether_poll() NET_LOCK_GIANT prior to locking poll_mtx.
- In netisr_poll() axe the block, where polling code asks drivers
to unregister.
- In netisr_poll() and ether_poll() do polling always, if any
handlers are present.
- In ether_poll_[de]register() remove a lot of error hiding code. Assert
that arguments are correct, instead.
- In ether_poll_[de]register() use standard return values in case of
error or success.
- Introduce poll_switch() that is a sysctl handler for kern.polling.enable.
poll_switch() goes through interface list and enabled/disables polling.
A message that kern.polling.enable is deprecated is printed.
Detailed driver changes:
- On attach driver announces IFCAP_POLLING in if_capabilities, but
not in if_capenable.
- On detach driver calls ether_poll_deregister() if polling is enabled.
- In polling handler driver obtains its lock and checks IFF_DRV_RUNNING
flag. If there is no, then unlocks and returns.
- In ioctl handler driver checks for IFCAP_POLLING flag requested to
be set or cleared. Driver first calls ether_poll_[de]register(), then
obtains driver lock and [dis/en]ables interrupts.
- In interrupt handler driver checks IFCAP_POLLING flag in if_capenable.
If present, then returns.This is important to protect from spurious
interrupts.
Reviewed by: ru, sam, jhb
make function reenterable. In the runtime the race is masked by serializing
of em_process_receive_interrupts() either by interrupt thread, or by
polling. The race can be triggered when polling is switched on or off.
could get an interrupt after we free the ifp, and the interrupt
handler depended on the ifp being still alive, this could, in theory,
cause a crash. Eliminate this possibility by moving the if_free to
after the bus_teardown_intr() call.
IFF_DRV_RUNNING, as well as the move from ifnet.if_flags to
ifnet.if_drv_flags. Device drivers are now responsible for
synchronizing access to these flags, as they are in if_drv_flags. This
helps prevent races between the network stack and device driver in
maintaining the interface flags field.
Many __FreeBSD__ and __FreeBSD_version checks maintained and continued;
some less so.
Reviewed by: pjd, bz
MFC after: 7 days
over iteration of their multicast address lists when synchronizing the
hardware address filter with the network stack-maintained list.
Problem reported by: Ed Maste (emaste at phaedrus dot sandvine dot ca>
MFC after: 1 week
struct ifnet or the layer 2 common structure it was embedded in have
been replaced with a struct ifnet pointer to be filled by a call to the
new function, if_alloc(). The layer 2 common structure is also allocated
via if_alloc() based on the interface type. It is hung off the new
struct ifnet member, if_l2com.
This change removes the size of these structures from the kernel ABI and
will allow us to better manage them as interfaces come and go.
Other changes of note:
- Struct arpcom is no longer referenced in normal interface code.
Instead the Ethernet address is accessed via the IFP2ENADDR() macro.
To enforce this ac_enaddr has been renamed to _ac_enaddr.
- The second argument to ether_ifattach is now always the mac address
from driver private storage rather than sometimes being ac_enaddr.
Reviewed by: sobomax, sam
- Changed from using explicit devices id to using descriptive labels.
- Added support for 82573 and 82546 Quad adapters.
- Corrected support for 82547EI and 82541ER (mac_type was not assigned)
- Removed #ifdef DBG_STATS and extraneous code.
if_em_hw.c/if_em_hw.h
- Added support for 82573 and 82546 Quad adapters.
- Brought forward Intel's most current mac and phy changes.
the interface when going to toggle VLAN support for
internal reasons. If the IFCAP_VLAN_HWTAGGING bit is
cleared, we should rely on the (re)init routine to turn
VLAN support off and never touch the relevant hardware bits.
This applies to other capability bits, too. The user
obviously has a reason for clearing a capability bit,
e.g., if his particular NIC is buggy and hangs if a
certain hardware capability is turned on even for a
fraction of a second.
The flag adapter->em_insert_vlan_header still is set or
reset irrespective of the IFCAP_VLAN_HWTAGGING setting,
as before, in order to handle the case when a user sets
promiscuous mode on an interface first and later turns
its IFCAP_VLAN_HWTAGGING bit on.
This change might look orthogonal to rev#1.85, but in fact
it is not. It introduces bugfixes that hopefully will make
implementing the general scheme mentioned in the commit
message of rev#1.85 easier.
configuration: it appears to work properly in the non-promiscuous case, but
we've not yet implemented a more general solution that maintains full
functionality with promiscuous mode enabled. While my hope is that we can
get one implemented soon, this will improve functionality substantially in
the mean time.
MFC after: 3 days
- Because em_encap() can now fail in a way that leaves us without an
mbuf chain, potentially set *m_headp to NULL if that happens, so that
the caller can do the right thing. This case can occur when we try
to prepend the vlan header mbuf but can't allocate additional memory.
- Modify the caller of em_encap() to detect a NULL m_head and not try
to queue the mbuf if that happens.
- When em_encap() fails, make sure to call bus_dmamap_destroy() to
clean up.
promiscuous mode introduced in 1.45, which programs the em card not
to strip or prepend tags when in promiscuous mode without also
modifying behavior to manually prepend a vlan header in the event
that the card isn't doing it on transmit. Due to a feature of card
operation, if the global VLAN prepend/strip register isn't set,
setting the VLAN tag flag on individual packet descriptors will
cause the packet to be transmitted using ISL encapsulation rather
than 802.1Q VLAN encapsulation.
This fix causes em_encap() to prepend the header by tracking whether
the card is configured to temporarily disable prepending/stripping
due to promiscuous mode. As a result, entering promiscuous mode on
the parent em interface no longer causes vlans to appear to "wedge"
or transmit ISL-encapsulated frames, which typically will not be
configured/spoken by the other endpoints on the VLAN trunk. This
bug may also exist in other drivers, and the additional vlan
encapsulation logic should be abstracted and centralized in
if_vlan.c if so.
RELENG_5_3 candidate.
MFC after: 1 week
Tested by: pjd, rwatson
Reported by: astesin at ukrtelecom dot net
Reported by: Mike Tancsa <mike at sentex dot net>
Reported by: Iasen Kostov <tbyte at OTEL dot net>
with a weak memory model or x86 + PAE (or more specifically, your
driver is using bounce pages) and you have had problems with em(4),
this may fix it. At least this is needed to have em(4) work properly
on FreeBSD/arm.
Original version by: cognet
Reviewed by: tackerman
Tested by: cognet
Since the e1000 DMA engines hava no constraints on the alignment of buffer
transfers, there is no reason to tell busdma that there is. This save a
minimum of 1 malloc call per packet, which translates to eliminating 4 locks.
It also means that buffers are not needlessly bounced when transfered. The
end result is a 38% improvement in pps in a 4 way bridging environment.
Obtained from: Sandvine, Inc.
