control. Controller does not automatically generate pause frames
based on number of available RX buffers so it's very hard to
know when driver should generate XON frame in time. The only
mechanism driver can detect low number of RX buffer condition is
ET_INTR_RXRING0_LOW or ET_INTR_RXRING1_LOW interrupt. This
interrupt is generated whenever controller notices the number of
available RX buffers are lower than pre-programmed value(
ET_RX_RING0_MINCNT and ET_RX_RING1_MINCNT register). This scheme
does not provide a way to detect when controller sees enough number
of RX buffers again such that efficient generation of XON/XOFF
frame is not easy.
While here, add more flow control related register definition.
interrupt is ours. Note, interrupts are automatically ACKed when
the status register is read.
Add RX/TX DMA error to interrupt handler and do full controller
reset if driver happen to encounter these errors. There is no way
to recover from these DMA errors without controller reset.
Rename local variable name intrs with status to enhance
readability.
While I'm here, rename ET_INTR_TXEOF and ET_INTR_RXEOF to
ET_INTR_TXDMA and ET_INTR_RXDMA respectively. These interrupts
indicate that a frame is successfully DMAed to controller's
internal FIFO and they have nothing to do with EOF(end of frame).
Driver does not need to wait actual end of TX/RX of a frame(e.g.
no need to wait the end signal of TX which is generated when a
frame in TX FIFO is emptied by MAC). Previous names were somewhat
confusing.
of a devfs file descriptor in devfs_close_f(). The passed in td argument
may be NULL if the close was invoked by garbage collection of open
file descriptors in pending control messages in the socket buffer of a
UNIX domain socket after it was closed.
PR: kern/151758
Submitted by: Andrey Shidakov andrey shidakov ru
Submitted by: Ruben van Staveren ruben verweg com
Reviewed by: kib
MFC after: 2 weeks
thread_free(newtd). This to avoid a possible page fault in
cpu_thread_clean() as seen on amd64 with syscall fuzzing.
Reviewed by: kib
MFC after: 1 week
suspend state. This will save more power.
On resume, make sure to enable all clocks. While I'm here, if
controller is not fast ethernet, enable gigabit PHY.
Don't blindly re-initialize controller whenever MTU is changed.
Now, reinitializing is done only when driver is running.
While here, remove unnecessary assignment of error value since it
was already initialized to 0.
o Do not report link status if driver is not running.
o TX/RX MAC configuration should be done with resolved speed,
duplex and flow control after establishing a link so it can't
be done in driver initialization routine.
Move the configuration to miibus_statchg callback which will be
called whenever any link state change is detected.
At this moment, flow-control is not enabled yet mainly because
I was not able to set correct flow control parameters to
generate TX pause frames.
o Now TX/RX MAC is enabled only when a valid link is detected.
Rearragnge hardware initialization routine a bit to leave
enabling MAC to miibus_statchg callback. In order to that,
TX/RX DMA engine is enabled in et_init_locked().
o Introduce ET_FLAG_LINK flag to track current link state.
o Introduce ET_FLAG_FASTETHER flag to mark whether controller is
fast ethernet. This flag is checked in miibus_statchg callback
to know whether PHY established a valid link.
o In et_stop(), TX/RX MAC is explicitly disabled instead of
relying on et_reset(). And move et_reset() from et_stop() to
controller initialization. Controler reset is not required here
and it would also clear critial registers(i.e station address,
RX filter configuration, WOL etc) that are required to make WOL
work.
o Switching to current media is done in et_init_locked() after
setting IFF_DRV_RUNNING flag. This should ensure reliable
auto-negotiation/manual link establishment.
o In et_start_locked(), check whether driver got a valid link
before trying to send frames.
o Remove checking a link in et_tick() as this is done by
miibus_statchg callback.
identifier reserved for the implementation in C99 and earlier so there is
no sensible reason for introducing yet another reserved identifier when we
could just use the one C1x uses.
Approved by: brooks (mentor)
manipulation of interrupt register access is done through
CSR_WRITE_4 macro. Also add disabling interrupt into et_reset()
because we want interrupt disabled state after controller reset.
While I'm here slightly change interrupt handler to be more
readable one.
send a single TX command after setting up all TX frames. This
removes unnecessary register accesses and bus_dmamap_sync(9) calls.
et(4) uses TX interrupt moderation so it's possible to have TX
buffers that were already transmitted but waiting for TX completion
interrupt. If the number of available TX descriptor is less then
1/3 of total TX descriptor, try reclaiming first to get enough free
TX descriptors before setting up TX descriptors.
After r228325, et_txeof() no longer tries to send frames after
reclaiming TX buffers. That change was made to give more chance
to transmit frames in main interrupt handler since we can still
send frames in interrupt handler with RX interrupt. So right
before exiting interrupt hander, after enabling interrupt, try to
send more frames. This gives slightly better performance numbers.
While I'm here reduce number of spare TX descriptors from 8 to 4.
Controller does not require reserved TX descriptors, it was just to
reduce TX overhead. After r228325, driver has much lower TX
overhead so it does not make sense to reserve 8 TX descriptors.
change should make et(4) work on any architectures.
o Remove m_getl inline function and replace it with stanard mbuf
interfaces. Previous code tried to minimize code duplication
but this came from incorrect use of common DMA tag.
Driver may be still use a common RX allocation handler with
additional structure changes but I don't see much point to do
that it would make it hard to understand the code.
o Remove DragonflyBSD specific constant EVL_ENCAPLEN, use
ETHER_VLAN_ENCAP_LEN instead.
o Add bunch of new RX status definition. It seems controller
supports RX checksum offloading but I was not able to make the
feature work yet. Currently driver checks whether recevied
frame is good one or not.
o Avoid a typedef ending in '_t' as style(9) says.
o Controller has no restriction on DMA address space, so there
is no reason to limit the DMA address to 32bit. Descriptor
rings, status blocks and TX/RX buffers now use full 64bit DMA
addressing.
o Allocate DMA memory shared between host and controller as
coherent.
o Create 3 separate DMA tags to be used as TX, mini RX ring and
stanard RX ring. Previously it created a single DMA tag and it
was used to all three rings.
o et(4) does not support jumbo frame at this moment and I still
don't quite understand how jumbo frame works on this controller
so use two RX rings to handle small sized frame and normal sized
frame respectively. The mini RX ring will be used to receive
frames that are less than or equal to 127 bytes. The second RX
ring is used to receive frames that are not handled by the first
RX ring.
If jumbo frame support is implemented, driver may have to choose
better RX scheme by letting the second RX ring handle jumbo
frames. This scheme will mimic Broadcom's efficient jumbo frame
handling feature. However RAM buffer size(16KB) of the
controller is too small to hold 2 jumbo frames, if 9KB
jumbo frame is used, I'm not sure how good performance would it
have.
o In et_rxeof(), make sure to check whether controller received
good frame or not. Passing corrupted frame to upper layer is
bad idea.
o If driver receives a bad frame or driver fails to allocate RX
buffer due to resource shortage condition, reuse previously
loaded DMA map for RX buffer instead of unloading/loading RX
buffer again.
o et_init_tx_ring() never fails so change return type to void.
o In watchdog handler, show TX DMA write back status of errored
frame which could be used as a clue to debug watchdog timeout.
o Add missing bus_dmamap_sync() in various places such that et(4)
should work with bounce buffers(e.g. PAE).
o TX side bus_dmamap_load_mbuf_sg(9) support.
o RX side bus_dmamap_load_mbuf_sg(9) support.
o Controller has no DMA alignment limit in RX buffer so use
m_adj(9) in RX buffer allocation to make IP header align on 2
bytes boundary. Otherwise it would trigger unaligned access
error in upper layer on strict alignment architectures.
One of down side of controller is it provides limited set of RX
buffer length like most Intel controllers. This is not problem
at this moment because driver does not support jumbo frame yet
but it may require alignment fixup code to support jumbo frame
on strict alignment architectures.
o In et_txeof(), don't zero TX descriptors for transmitted frames.
TX descriptors don't need write access after transmission.
Driver sets IFF_DRV_OACTIVE when the number of available TX
descriptors are less than or equal to ET_NSEG_SPARE. Make sure
to clear IFF_DRV_OACTIVE only when the number of available TX
descriptor is greater than ET_NSEG_SPARE.
__noreturn macro and modify the other exiting functions to use it.
The __noreturn macro, unlike __dead2, must be used BEFORE the function.
This is in line with the C and C++ specifications that place _Noreturn (c1x)
and [[noreturn]] (C++11) in front of the functions. As with __dead2, this
macro falls back to using the GCC attribute.
Unfortunately, clang currently sets the same value for the C version macro
in C99 and C1x modes, so these functions are hidden by default. At some
point before 10.0, I need to go through the headers and clean up the C1x /
C++11 visibility.
Reviewed by: brooks (mentor)
of vm_kmem_size that may occur if the system administrator has specified a
vm.vm_kmem_size tunable value that exceeds the hard cap.
PR: 162741
Submitted by: Adam McDougall
Reviewed by: bde@
MFC after: 3 weeks
Optimize for the case, by lazily allocating the pipe inode number at the
fstat(2) time. If alloc_unr(9) returns failure, do not fail fstat(2), since
uses of inode numbers are even rare then fstat(2), but provide zero inode
forever. Note that alloc_unr() failure is unlikely due to total number
of pipes in the system limited by the number of file descriptors.
Based on the submission by: gianni
MFC after: 2 weeks
c162516
Remove vtblk_sector_size
c162515
Wrap long license lines
c162514
Remove vtblk_unit
c162513
Wrap long lines in the license.
c162512
Remove verbose messages when link goes up/down.
A similar message is printed elsewhere as a result of
if_link_state_change().
c162511
Explicity compare pointer to NULL
c162510
Allocate the mac filter table at attach time.
c162509
Add real BSD licenses to the header files copied from Linux.
The chases upstream changes made in Linux awhile ago.
c162508
Only notify if we actually dequeued something.
c162507
Change a couple of if () { KASSERT(...) } to just KASSERTs.
In non-debug kernels, the if() { } probably get optomized
away, but I guess this is clearer.
c162506
Remove VIRTIO_BLK_F_TOPOLOGY fields in the config.
TOPOLOGY has since been removed from the spec, and the FreeBSD
didn't really do anything with the fields anyways.
c162505
Move vtblk_enqueue_request() outside the locks when getting the ident.
c162504
Remove soon to be uneeded trylock during dump [1].
http://lists.freebsd.org/pipermail/freebsd-current/2011-November/029226.html
c162503
Remove emtpy line
c162502
Drop frame if cannot allocate a vtnet_tx_header.
If we don't, we set OACTIVE, but if there are no
other frames in flight, vtnet_txeof() will never
be called to unset OACTIVE. The interface would
have to be down/up'ed in order to become usable.
We could be cuter here and only do this if the
virtqueue is emtpy, but its probably not worth
the complication.
c162501
Start mbuf replacement loop at 1 for clarity
Obtained from: Bryan Venteicher bryanv at daemoninthecloset dot org
driver that has high precedence for the controller override et(4).
Add missing callout_drain(9) in device detach and rework detach
routine. While I'm here use rman_get_rid(9) instead of using
cached resource id because bus methods are free to change the
id.
While I'm here remove initializing if_mtu, it is set by
ether_ifattach(9). Also move callout_init_mtx(9) to the right below
driver lock initialization.
descriptors before trying to send frames. If we're not able to
send a frame, make sure to prepend it to if_snd queue such that
alt(4) should work.
While I'm here prefer ETHER_BPF_MTAP to BPF_MTAP. ETHER_BPF_MTAP
should be used for controllers that support VLAN hardware tag
insertion. The controller supports VLAN tag insertion but lacks
VLAN tag stripping in RX path though.
before calling bus_enumerate_hinted_children(9) (which is the minimum for
this to work) instead of fully probing it so later on we can just call
bus_generic_attach(9) on the parent of the miibus(4) instance. The latter
is necessary in order to work around what seems to be a bzzarre race in
newbus affecting a few machines since r227687, causing no driver being
probed for the newly added miibus(4) instance. Presumably this is the
same race that was the motivation for the work around done in r215348.
Reported and tested by: yongari
- Revert the removal of a static in r221913 in order to help compilers to
produce more optimal code.
Citing jilles:
If we are ever going to do ASLR, the AUXV information tells an attacker
where the stack, executable and RTLD are located, which defeats much of
the point of randomizing the addresses in the first place.
Given that the AUXV information seems to be used by debuggers only anyway,
I think it would be good to move it to p_candebug() now.
The full virtual memory maps (KERN_PROC_VMMAP, procstat -v) are already
under p_candebug().
Suggested by: jilles
Discussed with: rwatson
MFC after: 1 week
use superpage reservations. So, for the first time, kernel virtual memory
that is allocated by contigmalloc(), kmem_alloc_attr(), and
kmem_alloc_contig() can be promoted to superpages. In fact, even a series
of small contigmalloc() allocations may collectively result in a promoted
superpage.
Eliminate some duplication of code in vm_reserv_alloc_page().
Change the type of vm_reserv_reclaim_contig()'s first parameter in order
that it be consistent with other vm_*_contig() functions.
Tested by: marius (sparc64)
Where i386/bios/apm.c requires no per-descriptor state, the ACPI version
of these device do. Instead of using hackish clone lists that leave
stale device nodes lying around, use the cdevpriv API.
On my hardware, "em" in netmap mode does about 1.388 Mpps
on one card (on an Asus motherboard), and 1.1 Mpps on another
card (PCIe bus). Both seem to be NIC-limited, because
i have the same rate even with the CPU running at 150 MHz.
On the "re" driver the tx throughput is around 420-450 Kpps
on various (8111C and the like) chipsets. On the Rx side
performance seems much better, and i can receive the full
load generated by the "em" cards.
"igb" is untested as i don't have the hardware.
A link reset now is completely transparent for the netmap client:
even if the NIC resets its own ring (e.g. restarting from 0),
the client will not see any change in the current rx/tx positions,
because the driver will keep track of the offset between the two.
2. make the device-specific code more uniform across different drivers
There were some inconsistencies in the implementation of the netmap
support routines, now drivers have been aligned to a common
code structure.
3. import netmap support for ixgbe . This is implemented as a very
small patch for ixgbe.c (233 lines, 11 chunks, mostly comments:
in total the patch has only 54 lines of new code) , as most of
the code is in an external file sys/dev/netmap/ixgbe_netmap.h ,
following some initial comments from Jack Vogel about making
changes less intrusive.
(Note, i have emailed Jack multiple times asking if he had
comments on this structure of the code; i got no reply so
i assume he is fine with it).
Support for other drivers (em, lem, re, igb) will come later.
"ixgbe" is now the reference driver for netmap support. Both the
external file (sys/dev/netmap/ixgbe_netmap.h) and the device-specific
patches (in sys/dev/ixgbe/ixgbe.c) are heavily commented and should
serve as a reference for other device drivers.
Tested on i386 and amd64 with the pkt-gen program in tools/tools/netmap,
the sender does 14.88 Mpps at 1050 Mhz and 14.2 Mpps at 900 MHz
on an i7-860 with 4 cores and 82599 card. Haven't tried yet more
aggressive optimizations such as adding 'prefetch' instructions
in the time-critical parts of the code.
