In this driver, there is a "port" created for the connection to each of
the other ldoms; a netdev queue is mapped to each port, and they are
collected under a single netdev. The generic netdev statistics show
us all the traffic in and out of our network device, but don't show
individual queue/port stats. This patch breaks out the traffic counts
for the individual ports and gives us a little view into the state of
those connections.
Orabug: 25190537
Signed-off-by: Shannon Nelson <shannon.nelson@oracle.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
There have been several changes since the first version of this code, so
we bump the version number. While we're at it, we can simplify the
version printing a bit and drop a couple lines of code.
Signed-off-by: Shannon Nelson <shannon.nelson@oracle.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
cassini: min_mtu 60, max_mtu 9000
niu: min_mtu 68, max_mtu 9216
sungem: min_mtu 68, max_mtu 1500 (comments say jumbo mode is broken)
sunvnet: min_mtu 68, max_mtu 65535
- removed sunvnet_change_mut_common as it does nothing now
CC: netdev@vger.kernel.org
Signed-off-by: Jarod Wilson <jarod@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Checkpatch updates for sunvnet.c and sunvnet_common.c.
Signed-off-by: Aaron Young <aaron.young@oracle.com>
Signed-off-by: Rashmi Narasimhan <rashmi.narasimhan@oracle.com>
Reviewed-by: Sowmini Varadhan <sowmini.varadhan@oracle.com>
Reviewed-by: Alexandre Chartre <Alexandre.Chartre@oracle.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Modify sunvnet common code and data structures to be compatible
with both sunvnet and ldmvsw drivers.
Details:
Sunvnet operates on "vnet-port" nodes which appear in the Machine
Description (MD) in a guest domain. Ldmvsw operates on "vsw-port"
nodes which appear in the MD of a service domain.
A difference between the sunvnet driver and the ldmvsw driver is
the sunvnet driver creates a network interface (i.e. a struct net_device)
for every vnet-port *parent* "network" node. Several vnet-ports may appear
under this common parent network node - each corresponding to a common parent
network interface. Conversely, since bridge/vswitch software will need
to interface with every vsw-port in a system, the ldmvsw driver creates
a network interface (i.e. a struct net_device) for every vsw-port - not
every parent node as with sunvnet. This difference required some special
handling in the common code as explained below.
There are 2 key data structures used by the sunvnet and ldmvsw drivers
(which are now found in sunvnet_common.h):
1. struct vnet_port
This structure represents a vnet-port node in sunvnet and a vsw-port
in the ldmvsw driver.
2. struct vnet
This structure represents a parent "network" node in sunvnet and a parent
"virtual-network-switch" node in ldmvsw.
Since the sunvnet driver allocates a net_device for every parent "network"
node, a net_device member appears in the struct vnet. Since the ldmvsw
driver allocates a net_device for every port, a net_device member was
added to the vnet_port. The common code distinguishes which structure
net_device member to use by checking a 'vsw' bit that was added to the
vnet_port structure. See the VNET_PORT_TO_NET_DEVICE() marco in
sunvnet_common.h.
The netdev_priv() in sunvnet is allocated as a vnet. The netdev_priv()
in ldmvsw is a vnet_port. Therefore, any place in the common code
where a netdev_priv() call was made, a wrapper function was implemented
in each driver to first get the vnet and/or vnet_port (in a driver
specific way) and pass them as newly added parameters to the common
functions (see wrapper funcs: vnet_set_rx_mode() and vnet_poll_controller()).
Since these wrapper functions call __tx_port_find(), __tx_port_find() was
moved from the common code back into sunvnet.c. Note - ldmvsw.c does not
require this function.
These changes also required that port_is_up() be made
into a common function and thus it was given a _common suffix and
exported like the other common functions.
A wrapper function was also added for vnet_start_xmit_common() to pass a
driver-specific function arg to return the port associated with a given
struct sk_buff and struct net_device. This was required because
vnet_start_xmit_common() grabs a lock prior to getting the associated
port. Using a function pointer arg allowed the code to work unchanged
without risking changes to the non-trivial locking logic in
vnet_start_xmit_common().
Signed-off-by: Aaron Young <aaron.young@oracle.com>
Signed-off-by: Rashmi Narasimhan <rashmi.narasimhan@oracle.com>
Reviewed-by: Sowmini Varadhan <sowmini.varadhan@oracle.com>
Reviewed-by: Alexandre Chartre <Alexandre.Chartre@oracle.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Split sunvnet.c into sunvnet.c and sunvnet_common.c.
Details:
Since the sunvnet and ldmvsw drivers will both use common sunvnet code,
move the functions (and support functions) anticipated to be common code
from sunvnet.c to sunvnet_common.c. Similarly, sunvnet.h was renamed to
sunvnet_common.h. The sunvnet_common.c code will be compiled into the
kernel and act as a library of functions that are linked by either
(or both) drivers when loaded.
Function names for external functions in sunvnet_common.c (to be
called by both the sunvnet and ldmvsw drivers) were tagged with a "_common"
suffix to clearly designate them as common functions.
No functional changes as of yet... just moved code verbatim to the new
sunvnet_common.c/h files.
Makefile/Kconfig support added to build sunvnet_common.c file. The code
is included in the kernel if SUN_LDOMS is defined/selected.
NOTE - per the SubmittingPatches documentation, since the code was just
moved from one file another, the code was NOT checkpatch'd in this commit
to aid in review.
Signed-off-by: Aaron Young <aaron.young@oracle.com>
Signed-off-by: Rashmi Narasimhan <rashmi.narasimhan@oracle.com>
Reviewed-by: Sowmini Varadhan <sowmini.varadhan@oracle.com>
Reviewed-by: Alexandre Chartre <Alexandre.Chartre@oracle.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Use sunvnet perf trace macros to monitor LDC message exchange state.
Signed-off-by: Sowmini Varadhan <sowmini.varadhan@oracle.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
vnet_fullcsum() accesses ip_hdr() and transport header to compute
the checksum for IPv4 packets, so these need to be initialized in
skb created in vnet_rx_one().
Signed-off-by: Sowmini Varadhan <sowmini.varadhan@oracle.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
`ls /sys/devices/channel-devices/vnet-port-0-0/net' is missing without
this change, and applications like NetworkManager are looking in
sysfs for the information.
Signed-off-by: Sowmini Varadhan <sowmini.varadhan@oracle.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch goes through and replaces wmb/rmb with dma_wmb/dma_rmb in cases
where the barrier is being used to order writes or reads to just memory and
doesn't involve any programmed I/O.
Signed-off-by: Alexander Duyck <alexander.h.duyck@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
An error return from __vnet_tx_trigger() sets the TX descriptor to
VIO_DESC_FREE while leaving port->tx_bufs[txi].skb set. This leads
to a BUG_ON() the next time this descriptor is used.
This patch frees the pending skb when getting a trigger error to
match the VIO_DESC_FREE state.
Signed-off-by: David L Stevens <david.stevens@oracle.com>
Acked-by: Sowmini Varadhan <sowmini.varadhan@oracle.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch unclones an skb for the case where the sunvnet driver needs to
change the segmentation size so that it doesn't interfere with TCP SACK's
use of them.
Signed-off-by: David L Stevens <david.stevens@oracle.com>
Acked-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Conflicts:
drivers/net/vxlan.c
drivers/vhost/net.c
include/linux/if_vlan.h
net/core/dev.c
The net/core/dev.c conflict was the overlap of one commit marking an
existing function static whilst another was adding a new function.
In the include/linux/if_vlan.h case, the type used for a local
variable was changed in 'net', whereas the function got rewritten
to fix a stacked vlan bug in 'net-next'.
In drivers/vhost/net.c, Al Viro's iov_iter conversions in 'net-next'
overlapped with an endainness fix for VHOST 1.0 in 'net'.
In drivers/net/vxlan.c, vxlan_find_vni() added a 'flags' parameter
in 'net-next' whereas in 'net' there was a bug fix to pass in the
correct network namespace pointer in calls to this function.
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch fixes a bug where vnet_skb_shape() didn't set the already-selected
queue mapping when a packet copy was required. This results in using the
wrong queue index for stops/starts, hung tx queues and watchdog timeouts
under heavy load.
Signed-off-by: David L Stevens <david.stevens@oracle.com>
Acked-by: Sowmini Varadhan <sowmini.varadhan@oracle.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
If a remote machine crashes while there are pending transmit buffers, the
sunvnet driver reallocates the ring descriptors giving us enries that have
state VIO_DESC_FREE but also an allocated skb. This results in a BUG_ON()
call when the remote reboots and we reach that point in the ring.
This patch:
1) clears pending tx packets in the ring on port reset
2) changes a BUG_ON() to a pr_warn() when a remote host has given us an invalid
descriptor state
3) collapses multiple active buffer frees in a ring to a single message per
ring and adds the device name and remote MAC address
This fixes the particular problem of not cleaning up pending buffers on a
reset, but also prevents us from crashing if the remote handles descriptors
out of order or sets an unexpected state for a descriptor.
Signed-off-by: David L Stevens <david.stevens@oracle.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch moves the clearing of ring data in vnet_port_free_tx_bufs to after
the freeing of pending buffers in the ring. Otherwise, this can result in
dereferencing a NULL pointer.
Reported-by: Sowmini Varadhan <sowmini.varadhan@oracle.com>
Signed-off-by: David L Stevens <david.stevens@oracle.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch fixes the rx packet length check in the sunvnet driver to allow
for a TSO max packet length greater than the LDC channel negotiated MTU.
These are negotiated separately and there is no requirement that
port->tsolen be less than port->rmtu, but if it isn't, it'll drop packets
with rx length errors.
Signed-off-by: David L Stevens <david.stevens@oracle.com>
Acked-by: Sowmini Varadhan <sowmini.varadhan@oracle.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
when skb_gso_segment returns error, the original skb should be freed
Signed-off-by: Li RongQing <roy.qing.li@gmail.com>
Acked-by: David L Stevens <david.stevens@oracle.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Both sunvdc and sunvnet implemented distinct functionality for incrementing
and decrementing dring indexes. Create common functions for use by both
from the sunvnet versions, which were chosen since they will still work
correctly in case a non power of two ring size is used.
Signed-off-by: Dwight Engen <dwight.engen@oracle.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch removes an extra rcu_read_unlock() on an allocation failure
in vnet_skb_shape(). The needed rcu_read_unlock() is already done in
the out_dropped label.
Reported-by: Rashmi Narasimhan <rashmi.narasimhan@oracle.com>
Signed-off-by: David L Stevens <david.stevens@oracle.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch adds TSO support for the sunvnet driver.
Signed-off-by: David L Stevens <david.stevens@oracle.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch adds GSO support to the sunvnet driver.
Signed-off-by: David L Stevens <david.stevens@oracle.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch adds support for sender-side checksum offloading.
Signed-off-by: David L Stevens <david.stevens@oracle.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch adds scatter/gather support to the sunvnet driver.
Signed-off-by: David L Stevens <david.stevens@oracle.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch adds support for VIO v1.7 (extended descriptor format)
and v1.8 (receive-side checksumming) to the sunvnet driver.
Signed-off-by: David L Stevens <david.stevens@oracle.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch changes the name of vnet_port_alloc_tx_bufs to
vnet_port_alloc_tx_ring, since there are no buffer allocations after
transmit zero copy support was added. This patch also moves the ring
allocation to after VIO version negotiation to allow for
different-sized descriptors in later VIO versions.
Signed-off-by: David L Stevens <david.stevens@oracle.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch fixes a NULL pointer dereference when __tx_port_find() doesn't
find a matching port.
Signed-off-by: David L Stevens <david.stevens@oracle.com>
Acked-by: Sowmini Varadhan <sowmini.varadhan@oracle.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The out_dropped label will only do rcu_read_unlock for non-null port.
So add the missing rcu_read_unlock() when bailing due to non-null port.
Signed-off-by: Sowmini Varadhan <sowmini.varadhan@oracle.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
As per comments in vnet_start_xmit, for the edge case
when outgoing vnet_start_xmit() data and an incoming STOPPED
ACK cross each other in flight, we may need to send the missed
START trigger from maybe_tx_wakeup() after checking for a
false value of start_cons
Signed-off-by: Sowmini Varadhan <sowmini.varadhan@oracle.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
When vnet_start_xmit() is concurrent with vnet_ack(), we may
have a race that looks like:
thread 1 thread 2
vnet_start_xmit vnet_event_napi -> vnet_rx
__vnet_tx_trigger for some desc X
at this point dr->prod == X
peer sends back a stopped ack for X
we process X, but X == dr->prod
so we bail out in vnet_ack with
!idx_is_pending
update dr->prod
As a result of the fact that we never processed the stopped ack for X,
the Tx path is led to incorrectly believe that the peer is still
"started" and reading, but the peer has stopped reading, which will
ultimately end in flow-control assertions.
The fix is to synchronize the above 2 paths on the netif_tx_lock.
Signed-off-by: Sowmini Varadhan <sowmini.varadhan@oracle.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
vnet_event_napi() may be called as part of the NAPI ->poll,
to resume reading descriptor rings. When no data is available,
descriptor ring state (e.g., rcv_nxt) needs to be reset
carefully to stay in lock-step with ldc_read(). In the interest
of simplicity, the best way to do this is to return from
vnet_event_napi() when there are no more packets to read.
The next trip through ldc_rx will correctly set up the dring state.
Signed-off-by: Sowmini Varadhan <sowmini.varadhan@oracle.com>
Tested-by: David Stevens <david.stevens@oracle.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
remove redundant tab.
Signed-off-by: Sowmini Varadhan <sowmini.varadhan@oracle.com>
Reported-by: Ben Hutchings <ben@decadent.org.uk>
Signed-off-by: David S. Miller <davem@davemloft.net>
Use multple Tx netdev queues for sunvnet by supporting a one-to-one
mapping between vnet_port and Tx queue. Provide a ndo_select_queue
indirection (vnet_select_queue()) which selects the queue based
on the peer that would be selected in vnet_start_xmit()
Signed-off-by: Sowmini Varadhan <sowmini.varadhan@oracle.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
When vnet_event_napi re-enables interrupts, it should
reset LDC_EVENT_DATA_READY as an optimization.
Signed-off-by: Sowmini Varadhan <sowmini.varadhan@oracle.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
After the NAPIfication of sunvnet, we no longer need to
synchronize by doing irqsave/restore on vio.lock in the
I/O fastpath.
NAPI ->poll() is non-reentrant, so all RX processing occurs
strictly in a serialized environment. TX reclaim is done in NAPI
context, so the netif_tx_lock can be used to serialize
critical sections between Tx and Rx paths.
Signed-off-by: Sowmini Varadhan <sowmini.varadhan@oracle.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
A vnet_port_remove could be triggered as a result of an ldm-unbind
operation by the peer, module unload, or other changes to the
inter-vnet-link configuration. When this is concurrent with
vnet_start_xmit(), there are several race sequences possible,
such as
thread 1 thread 2
vnet_start_xmit
-> tx_port_find
spin_lock_irqsave(&vp->lock..)
ret = __tx_port_find(..)
spin_lock_irqrestore(&vp->lock..)
vio_remove -> ..
->vnet_port_remove
spin_lock_irqsave(&vp->lock..)
cleanup
spin_lock_irqrestore(&vp->lock..)
kfree(port)
/* attempt to use ret will bomb */
This patch adds RCU locking for port access so that vnet_port_remove
will correctly clean up port-related state.
Signed-off-by: Sowmini Varadhan <sowmini.varadhan@oracle.com>
Acked-by: Dwight Engen <dwight.engen@oracle.com>
Acked-by: Bob Picco <bob.picco@oracle.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Move Rx packet procssing to the NAPI poll callback.
Disable VIO interrupt and unconditioanlly go into NAPI
context from vnet_event.
Note that we want to minimize the number of LDC
STOP/START messages sent. Specifically, do not send a STOP
message if vnet_walk_rx does not read all the available descriptors
because of the NAPI budget limitation. Instead, note the end index
as part of port state, and resume from this index when the
next poll callback is triggered.
Signed-off-by: Sowmini Varadhan <sowmini.varadhan@oracle.com>
Acked-by: Raghuram Kothakota <raghuram.kothakota@oracle.com>
Acked-by: Dwight Engen <dwight.engen@oracle.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Pull sparc updates from David Miller:
1) Move to 4-level page tables on sparc64 and support up to 53-bits of
physical addressing. Kernel static image BSS size reduced by
several megabytes.
2) M6/M7 cpu support, from Allan Pais.
3) Move to sparse IRQs, handle hypervisor TLB call errors more
gracefully, and add T5 perf_event support. From Bob Picco.
4) Recognize cdroms and compute geometry from capacity in virtual disk
driver, also from Allan Pais.
5) Fix memset() return value on sparc32, from Andreas Larsson.
6) Respect gfp flags in dma_alloc_coherent on sparc32, from Daniel
Hellstrom.
7) Fix handling of compound pages in virtual disk driver, from Dwight
Engen.
8) Fix lockdep warnings in LDC layer by moving IRQ requesting to
ldc_alloc() from ldc_bind().
9) Increase boot string length to 1024 bytes, from Dave Kleikamp.
* git://git.kernel.org/pub/scm/linux/kernel/git/davem/sparc: (31 commits)
sparc64: Fix lockdep warnings on reboot on Ultra-5
sparc64: Increase size of boot string to 1024 bytes
sparc64: Kill unnecessary tables and increase MAX_BANKS.
sparc64: sparse irq
sparc64: Adjust vmalloc region size based upon available virtual address bits.
sparc64: Increase MAX_PHYS_ADDRESS_BITS to 53.
sparc64: Use kernel page tables for vmemmap.
sparc64: Fix physical memory management regressions with large max_phys_bits.
sparc64: Adjust KTSB assembler to support larger physical addresses.
sparc64: Define VA hole at run time, rather than at compile time.
sparc64: Switch to 4-level page tables.
sparc64: Fix reversed start/end in flush_tlb_kernel_range()
sparc64: Add vio_set_intr() to enable/disable Rx interrupts
vio: fix reuse of vio_dring slot
sunvdc: limit each sg segment to a page
sunvdc: compute vdisk geometry from capacity
sunvdc: add cdrom and v1.1 protocol support
sparc: VIO protocol version 1.6
sparc64: Fix hibernation code refrence to PAGE_OFFSET.
sparc64: Move request_irq() from ldc_bind() to ldc_alloc()
...
One of the error cases for vnet_start_xmit()'s "out_dropped" label
is port == NULL, so only mess with port->clean_timer when port is not NULL.
Signed-off-by: David L Stevens <david.stevens@oracle.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
vio_dring_avail() will allow use of every dring entry, but when the last
entry is allocated then dr->prod == dr->cons which is indistinguishable from
the ring empty condition. This causes the next allocation to reuse an entry.
When this happens in sunvdc, the server side vds driver begins nack'ing the
messages and ends up resetting the ldc channel. This problem does not effect
sunvnet since it checks for < 2.
The fix here is to just never allocate the very last dring slot so that full
and empty are not the same condition. The request start path was changed to
check for the ring being full a bit earlier, and to stop the blk_queue if
there is no space left. The blk_queue will be restarted once the ring is
only half full again. The number of ring entries was increased to 512 which
matches the sunvnet and Solaris vdc drivers, and greatly reduces the
frequency of hitting the ring full condition and the associated blk_queue
stop/starting. The checks in sunvent were adjusted to account for
vio_dring_avail() returning 1 less.
Orabug: 19441666
OraBZ: 14983
Signed-off-by: Dwight Engen <dwight.engen@oracle.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch sends ICMP and ICMPv6 messages for Path MTU Discovery when a remote
port MTU is smaller than the device MTU. This allows mixing newer VIO protocol
devices that support MTU negotiation with older devices that do not on the
same vswitch. It also allows Linux-Linux LDOMs to use 64K-1 data packets even
though Solaris vswitch is limited to <16K MTU.
Signed-off-by: David L Stevens <david.stevens@oracle.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch allows an admin to set the MTU on a sunvnet device to arbitrary
values between the minimum (68) and maximum (65535) IPv4 packet sizes.
Signed-off-by: David L Stevens <david.stevens@oracle.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch removes pre-allocated transmit buffers and instead directly maps
pending packets on demand. This saves O(n^2) maximum-sized transmit buffers,
for n hosts on a vswitch, as well as a copy to those buffers.
Single-stream TCP throughput linux-solaris dropped ~5% for 1500-byte MTU,
but linux-linux at 1500-bytes increased ~20%.
Signed-off-by: David L Stevens <david.stevens@oracle.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch upgrades the sunvnet driver to support VIO protocol version 1.6.
In particular, it adds per-port MTU negotiation, allowing MTUs other than
ETH_FRAMELEN with ports using newer VIO protocol versions.
Signed-off-by: David L Stevens <david.stevens@oracle.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Conflicts:
arch/mips/net/bpf_jit.c
drivers/net/can/flexcan.c
Both the flexcan and MIPS bpf_jit conflicts were cases of simple
overlapping changes.
Signed-off-by: David S. Miller <davem@davemloft.net>
When sending out a burst of packets across multiple descriptors,
it is sufficient to send one LDC "start" trigger for
the first descriptor, so do not send an LDC "start" for every
pass through vnet_start_xmit. Similarly, it is sufficient to send
one "DRING_STOPPED" trigger for the last dring (and if that
fails, hold off and send the trigger later).
Optimizations to the number of LDC messages helps avoid
filling up the LDC channel with superfluous LDC messages
that risk triggering flow-control on the channel,
and also boosts performance.
Signed-off-by: Sowmini Varadhan <sowmini.varadhan@oracle.com>
Acked-by: Raghuram Kothakota <raghuram.kothakota@oracle.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Use the much more common pr_warn instead of pr_warning.
Other miscellanea:
o Typo fixes submiting/submitting
o Coalesce formats
o Realign arguments
o Add missing terminating '\n' to formats
Signed-off-by: Joe Perches <joe@perches.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The sunvnet driver does not have an rmb() in the ring consumer corresponding
to the wmb() in the producer. According to Documentation/memory-barriers.txt:
"When dealing with CPU-CPU interactions, certain types of memory barrier should
always be paired. A lack of appropriate pairing is almost certainly an error."
In cases where an rmb() is not a no-op and a consumer is removing data from
the ring while a producer is adding new entries, a load reorder would allow
CPU1 CPU2
---- ----
LOAD desc.size [e.g]
STORE desc.size
<wmb>
set desc.hdr.state = VIO_DESC_READY
LOAD desc.hdr.state
[because VIO_DESC_READY, use
old desc.size, already loaded
out of order]
[CPU2 has reordered apparently unrelated LOADs]
To ensure other desc fields are not loaded before checking VIO_DESC_READY, we
need an rmb() between the check and desc data accesses.
I've also moved the viodbg() call to after the rmb() so that it, too, has
current descriptor data even with reordering, which has the side effect that
it won't print anything for descriptors that are not VIO_DESC_READY as before.
That's a) probably a good thing, since the fields are not necessarily set and,
b) better than adding another rmb() just for viodbg().
This would not be possible if strict-ordering is enforced, but then the
memory barriers should be no-ops in that case.
Signed-off-by: David L Stevens <david.stevens@oracle.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
At the tail of vnet_event(), if we hit the maybe_tx_wakeup()
condition, we try to take the netif_tx_lock() in the
recv-interrupt-context and can deadlock with dev_watchdog().
vnet_event() should schedule maybe_tx_wakeup() as a tasklet
to avoid this deadlock
Signed-off-by: Sowmini Varadhan <sowmini.varadhan@oracle.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
ldc_rx -> vnet_rx -> .. -> vnet_walk_rx->vnet_send_ack should not
spin into an infinite loop waiting EAGAIN to lift.
The sender could have sent us a burst, and gone to lunch without
doing any more ldc_read()'s. That should not cause the receiver to
loop infinitely till soft-lockup kicks in.
Similarly __vnet_tx_trigger should only loop on EAGAIN a finite
number of times. The caller (vnet_start_xmit()) already has code
to reset the dring state and bail on errors from __vnet_tx_trigger
Signed-off-by: Sowmini Varadhan <sowmini.varadhan@oracle.com>
Acked-by: Raghuram Kothakota <raghuram.kothakota@oracle.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
