Using the MCQI and MCQS registers, we query the running and pending
fw version of the HCA.
The MCQS is queried with sequentially increasing component index, until
a component of type BOOT_IMG is found. Querying this component's version
using the MCQI register yields the running and pending fw version of the
HCA.
Querying MCQI for the pending fw version should be done only after
validating that such fw version exists. This is done my checking
'component update state' field in MCQS output.
Signed-off-by: Shay Agroskin <shayag@mellanox.com>
Reviewed-by: Jakub Kicinski <jakub.kicinski@netronome.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
Misc updates from mlx5-next branch:
1) Add the required HW definitions and structures for upcoming TLS
support.
2) Add support for MCQI and MCQS hardware registers for fw version query.
3) Added hardware bits and structures definitions for sub-functions
4) Small code cleanup and improvement for PF pci driver.
5) Bluefield (ECPF) updates and refactoring for better E-Switch
management on ECPF embedded CPU NIC:
5.1) Consolidate querying eswitch number of VFs
5.2) Register event handler at the correct E-Switch init stage
5.3) Setup PF's inline mode and vlan pop when the ECPF is the
E-Swtich manager ( the host PF is basically a VF ).
5.4) Handle Vport UC address changes in switchdev mode.
6) Cleanup the rep and netdev reference when unloading IB rep.
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
i# All conflicts fixed but you are still merging.
Both MTK_TRGMII_MT7621_CLK and MTK_PATH_BIT are defined as bit 10.
This can causes issues on non-MT7621 devices which has the
MTK_PATH_BIT(MTK_ETH_PATH_GMAC1_RGMII) and MTK_TRGMII capability set.
The wrong TRGMII setup code can be executed. The current wrongly executed
code doesn’t do any harm on MT7623 and the TRGMII setup for the MT7623
SOC side is done in MT7530 driver So it wasn’t noticed in the test.
Move all capability bits in one enum so that they are all unique and easy
to expand in the future.
Because mtk_eth_path enum is merged in to mkt_eth_capabilities, the
variable path value is no longer between 0 to number of paths,
mtk_eth_path_name can’t be used anymore in this form. Convert the
mtk_eth_path_name array to a function to lookup the pathname.
The old code walked thru the mtk_eth_path enum, which is also merged
with mkt_eth_capabilities. Expand array mtk_eth_muxc so it can store the
name and capability bit of the mux. Convert the code so it can walk thru
the mtk_eth_muxc array.
Fixes: 8efaa653a8 ("net: ethernet: mediatek: Add MT7621 TRGMII mode support")
Signed-off-by: René van Dorst <opensource@vdorst.com>
v1->v2:
- Move all capability bits in one enum, suggested by Willem de Bruijn
- Convert the mtk_eth_path_name array to a function to lookup the pathname
- Expand array mtk_eth_muxc so it can also store the name and capability
bit of the mux
- Updated commit message
Signed-off-by: David S. Miller <davem@davemloft.net>
Enable GMAC v4.xx and beyond to support 16KiB buffer.
Signed-off-by: Weifeng Voon <weifeng.voon@intel.com>
Signed-off-by: Ong Boon Leong <boon.leong.ong@intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Currently, gratuitous ARP/ND packets are sent every `miimon'
milliseconds. This commit allows a user to specify a custom delay
through a new option, `peer_notif_delay'.
Like for `updelay' and `downdelay', this delay should be a multiple of
`miimon' to avoid managing an additional work queue. The configuration
logic is copied from `updelay' and `downdelay'. However, the default
value cannot be set using a module parameter: Netlink or sysfs should
be used to configure this feature.
When setting `miimon' to 100 and `peer_notif_delay' to 500, we can
observe the 500 ms delay is respected:
20:30:19.354693 ARP, Request who-has 203.0.113.10 tell 203.0.113.10, length 28
20:30:19.874892 ARP, Request who-has 203.0.113.10 tell 203.0.113.10, length 28
20:30:20.394919 ARP, Request who-has 203.0.113.10 tell 203.0.113.10, length 28
20:30:20.914963 ARP, Request who-has 203.0.113.10 tell 203.0.113.10, length 28
In bond_mii_monitor(), I have tried to keep the lock logic readable.
The change is due to the fact we cannot rely on a notification to
lower the value of `bond->send_peer_notif' as `NETDEV_NOTIFY_PEERS' is
only triggered once every N times, while we need to decrement the
counter each time.
iproute2 also needs to be updated to be able to specify this new
attribute through `ip link'.
Signed-off-by: Vincent Bernat <vincent@bernat.ch>
Signed-off-by: David S. Miller <davem@davemloft.net>
The variable err is being assigned with a value that is never
read and it is being updated in the next statement with a new value.
The assignment is redundant and can be removed.
Addresses-Coverity: ("Unused value")
Signed-off-by: Colin Ian King <colin.king@canonical.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Add rts2rts_qp_counters_set_id field in hca cap so that RTS2RTS
qp modification can be used to change the counter of a QP.
Signed-off-by: Mark Zhang <markz@mellanox.com>
Reviewed-by: Majd Dibbiny <majd@mellanox.com>
Acked-by: Saeed Mahameed <saeedm@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Currently the check to see if a page is allocated is incorrect
and is checking if the pointer page is null, not *page as
intended. Fix this.
Addresses-Coverity: ("Dereference before null check")
Fixes: f5cedc84a3 ("gve: Add transmit and receive support")
Signed-off-by: Colin Ian King <colin.king@canonical.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Paolo Abeni says:
====================
net: use ICW for sk_proto->{send,recv}msg
This series extends ICW usage to one of the few remaining spots in fast-path
still hitting per packet retpoline overhead, namely the sk_proto->{send,recv}msg
calls.
The first 3 patches in this series refactor the existing code so that applying
the ICW macros is straight-forward: we demux inet_{recv,send}msg in ipv4 and
ipv6 variants so that each of them can easily select the appropriate TCP or UDP
direct call. While at it, a new helper is created to avoid excessive code
duplication, and the current ICWs for inet_{recv,send}msg are adjusted
accordingly.
The last 2 patches really introduce the new ICW use-case, respectively for the
ipv6 and the ipv4 code path.
This gives up to 5% performance improvement under UDP flood, and smaller but
measurable gains for TCP RR workloads.
v1 -> v2:
- drop inet6_{recv,send}msg declaration from header file,
prefer ICW macro instead
- avoid unneeded reclaration for udp_sendmsg, as suggested by Willem
====================
Acked-by: Willem de Bruijn <willemb@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This avoids an indirect call per syscall for common ipv4 transports
v1 -> v2:
- avoid unneeded reclaration for udp_sendmsg, as suggested by Willem
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This avoids an indirect call per syscall for common ipv6 transports
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
After the previous patch we have ipv{6,4} variants for {recv,send}msg,
we should use the generic _INET ICW variant to call into the proper
build-in.
This also allows dropping the now unused and rather ugly _INET4 ICW macro
v1 -> v2:
- use ICW macro to declare inet6_{recv,send}msg
- fix a couple of checkpatch offender in the code context
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This will simplify indirect call wrapper invocation in the following
patch.
No functional change intended, any - out-of-tree - IPv6 user of
inet_{recv,send}msg can keep using the existing functions.
SCTP code still uses the existing version even for ipv6: as this series
will not add ICW for SCTP, moving to the new helper would not give
any benefit.
The only other in-kernel user of inet_{recv,send}msg is
pvcalls_conn_back_read(), but psvcalls explicitly creates only IPv4 socket,
so no need to update that code path, too.
v1 -> v2: drop inet6_{recv,send}msg declaration from header file,
prefer ICW macro instead
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The same code is replicated verbatim in multiple places, and the next
patches will introduce an additional user for it. Factor out a
helper and use it where appropriate. No functional change intended.
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
A generic WQE control field is used for different purposes
in different cases.
Use union to allow using the proper name in each case.
Signed-off-by: Tariq Toukan <tariqt@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
Functions change event output data size changes when functions other
than VFs will be enabled in HCA CAP.
With current API, multiple callers needs to align, calculate accurate
size of the output data depending on number on non VF functions enabled
in the device.
Instead of duplicating such math at multiple places, refactor
mlx5_esw_query_functions() to return raw output allocated by itself.
Caller must free the allocated memory using kvfree() as described in the
function comment section.
This hides calcuation within mlx5_esw_query_functions() and provides
simpler API.
Signed-off-by: Parav Pandit <parav@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
Eswitch function change handler will service multiple type of events for
VFs and non VF functions update.
Hence, introduce and use the helper function
esw_vfs_changed_event_handler() for handling change in num VFs to improve
the code readability.
Signed-off-by: Parav Pandit <parav@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
Instead MLX5_TOTAL_VPORTS, use mlx5_eswitch_get_total_vports().
mlx5_eswitch_get_total_vports() in subsequent patch accounts for SF
vports as well.
Expanding MLX5_TOTAL_VPORTS macro would require exposing SF internals to
more generic vport.h header file. Such exposure is not desired.
Hence a mlx5_eswitch_get_total_vports() is introduced.
Given that mlx5_eswitch_get_total_vports() API wants to work on const
mlx5_core_dev*, change its helper functions also to accept const *dev.
Signed-off-by: Parav Pandit <parav@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
The variable err is being initialized with a value that is never
read and it is being updated later with a new value. The
initialization is redundant and can be removed.
Addresses-Coverity: ("Unused value")
Signed-off-by: Colin Ian King <colin.king@canonical.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The variable tpd_req is being initialized with a value that is never
read and it is being updated later with a new value. The
initialization is redundant and can be removed.
Addresses-Coverity: ("Unused value")
Signed-off-by: Colin Ian King <colin.king@canonical.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch adds driver changes to detect/timestamp the unicast PTP packets.
Changes from previous version:
-------------------------------
v2: Defined a macro for unicast ptp param mask.
Please consider applying this to "net-next".
Signed-off-by: Sudarsana Reddy Kalluru <skalluru@marvell.com>
Signed-off-by: Ariel Elior <aelior@marvell.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
u64_stats_fetch_begin needs to initialize start.
Signed-off-by: Catherine Sullivan <csully@google.com>
Reported-by: kbuild test robot <lkp@intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Expose an extra device definitions for objects events.
It includes: object_type values for legacy objects and generic data
header for any other object.
Signed-off-by: Yishai Hadas <yishaih@mellanox.com>
Acked-by: Saeed Mahameed <saeedm@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Report EQE data upon CQ completion to let upper layers use this data.
Signed-off-by: Yishai Hadas <yishaih@mellanox.com>
Acked-by: Saeed Mahameed <saeedm@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Report a CQ error event only when a handler was set.
This enables mlx5_ib to not set a handler upon CQ creation and use some
other mechanism to get this event as of other events by the
mlx5_eq_notifier_register API.
Signed-off-by: Yishai Hadas <yishaih@mellanox.com>
Acked-by: Saeed Mahameed <saeedm@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Enhance mlx5_core_create_cq() to get the command out buffer from the
callers to let them use the output.
Signed-off-by: Yishai Hadas <yishaih@mellanox.com>
Acked-by: Saeed Mahameed <saeedm@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Expose the API to register for ANY event, mlx5_ib will be able to use
this functionality for its needs.
Signed-off-by: Yishai Hadas <yishaih@mellanox.com>
Acked-by: Saeed Mahameed <saeedm@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Use the reported device capabilities for the supported user events (i.e.
affiliated and un-affiliated) to set the EQ mask.
As the event mask can be up to 256 defined by 4 entries of u64 change
the applicable code to work accordingly.
Signed-off-by: Yishai Hadas <yishaih@mellanox.com>
Acked-by: Saeed Mahameed <saeedm@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
The firmware command to destroy a CQ might fail when the object is
referenced by other object and the ref count is managed by the firmware.
To enable a second successful destruction post the first failure need to
change mlx5_eq_del_cq() to be a void function.
As an error in mlx5_eq_del_cq() is quite fatal from the option to
recover, a debug message inside it should be good enougth and it was
changed to be void.
Signed-off-by: Yishai Hadas <yishaih@mellanox.com>
Acked-by: Saeed Mahameed <saeedm@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Before mlxsw_sp1_ptp_packet_finish() sends the packet back, it validates
whether the corresponding port is still valid. However the condition is
incorrect: when mlxsw_sp_port == NULL, the code dereferences the port to
compare it to skb->dev.
The condition needs to check whether the port is present and skb->dev still
refers to that port (or else is NULL). If that does not hold, bail out.
Add a pair of parentheses to fix the condition.
Fixes: d92e4e6e33 ("mlxsw: spectrum: PTP: Support timestamping on Spectrum-1")
Reported-by: Colin Ian King <colin.king@canonical.com>
Signed-off-by: Petr Machata <petrm@mellanox.com>
Signed-off-by: Colin Ian King <colin.king@canonical.com>
Reviewed-by: Ido Schimmel <idosch@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
It was reported that the GPD MicroPC is broken in a way that no valid
MAC address can be read from the network chip. The vendor driver deals
with this by assigning a random MAC address as fallback. So let's do
the same.
Signed-off-by: Heiner Kallweit <hkallweit1@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This reverts commit 759d095741.
The patch was based on a misunderstanding. As Al Viro pointed out [0]
it's simply wrong on big endian. So let's revert it.
[0] https://marc.info/?t=156200975600004&r=1&w=2
Reported-by: Al Viro <viro@zeniv.linux.org.uk>
Signed-off-by: Heiner Kallweit <hkallweit1@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Commit 760f1dc295 ("net: stmmac: add sanity check to
device_property_read_u32_array call") introduced error checking of the
device_property_read_u32_array() call in stmmac_mdio_reset().
This results in the following error when the "snps,reset-delays-us"
property is not defined in devicetree:
invalid property snps,reset-delays-us
This sanity check made sense until commit 84ce4d0f9f ("net: stmmac:
initialize the reset delay array") ensured that there are fallback
values for the reset delay if the "snps,reset-delays-us" property is
absent. That was at the cost of making that property mandatory though.
Drop the sanity check for device_property_read_u32_array() and thus make
the "snps,reset-delays-us" property optional again (avoiding the error
message while loading the stmmac driver with a .dtb where the property
is absent).
Fixes: 760f1dc295 ("net: stmmac: add sanity check to device_property_read_u32_array call")
Signed-off-by: Martin Blumenstingl <martin.blumenstingl@googlemail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Both tipc_udp_enable and tipc_udp_disable are called under rtnl_lock,
ub->ubsock could never be NULL in tipc_udp_disable and cleanup_bearer,
so remove the check.
Also remove the one in tipc_udp_enable by adding "free" label.
Signed-off-by: Xin Long <lucien.xin@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
In commit ee28906fd7 ("ipv4: Dump route exceptions if requested") I
added a counter of per-node dumped routes (including actual routes and
exceptions), analogous to the existing counter for dumped nodes. Dumping
exceptions means we need to also keep track of how many routes are dumped
for each node: this would be just one route per node, without exceptions.
When netlink strict checking is not enabled, we dump both routes and
exceptions at the same time: the RTM_F_CLONED flag is not used as a
filter. In this case, the per-node counter 'i_fa' is incremented by one
to track the single dumped route, then also incremented by one for each
exception dumped, and then stored as netlink callback argument as skip
counter, 's_fa', to be used when a partial dump operation restarts.
The per-node counter needs to be increased by one also when we skip a
route (exception) due to a previous non-zero skip counter, because it
needs to match the existing skip counter, if we are dumping both routes
and exceptions. I missed this, and only incremented the counter, for
regular routes, if the previous skip counter was zero. This means that,
in case of a mixed dump, partial dump operations after the first one
will start with a mismatching skip counter value, one less than expected.
This means in turn that the first exception for a given node is skipped
every time a partial dump operation restarts, if netlink strict checking
is not enabled (iproute < 5.0).
It turns out I didn't repeat the test in its final version, commit
de755a8513 ("selftests: pmtu: Introduce list_flush_ipv4_exception test
case"), which also counts the number of route exceptions returned, with
iproute2 versions < 5.0 -- I was instead using the equivalent of the IPv6
test as it was before commit b964641e99 ("selftests: pmtu: Make
list_flush_ipv6_exception test more demanding").
Always increment the per-node counter by one if we previously dumped
a regular route, so that it matches the current skip counter.
Fixes: ee28906fd7 ("ipv4: Dump route exceptions if requested")
Signed-off-by: Stefano Brivio <sbrivio@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
No reason to error out on a MT7621 device with DDR2 memory when non
TRGMII mode is selected.
Only MT7621 DDR2 clock setup is not supported for TRGMII mode.
But non TRGMII mode doesn't need any special clock setup.
Signed-off-by: René van Dorst <opensource@vdorst.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
With gcc 4.1:
net/rxrpc/output.c: In function ‘rxrpc_send_data_packet’:
net/rxrpc/output.c:338: warning: ‘ret’ may be used uninitialized in this function
Indeed, if the first jump to the send_fragmentable label is made, and
the address family is not handled in the switch() statement, ret will be
used uninitialized.
Fix this by BUG()'ing as is done in other places in rxrpc where internal
support for future address families will need adding. It should not be
possible to reach this normally as the address families are checked
up-front.
Fixes: 5a924b8951 ("rxrpc: Don't store the rxrpc header in the Tx queue sk_buffs")
Reported-by: Geert Uytterhoeven <geert@linux-m68k.org>
Signed-off-by: David Howells <dhowells@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The variable r is being initialized with a value that is never
read and it is being updated later with a new value. The
initialization is redundant and can be removed.
Addresses-Coverity: ("Unused value")
Signed-off-by: Colin Ian King <colin.king@canonical.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch removes standard netdev stats in ethtool -S.
Suggested-by: Jakub Kicinski <jakub.kicinski@netronome.com>
Signed-off-by: Xue Chaojing <xuechaojing@huawei.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
We support many speeds and it doesn't make much sense to list them all
in the Kconfig. Let's just call it Multi-Gigabit.
Suggested-by: David S. Miller <davem@davemloft.net>
Signed-off-by: Jose Abreu <joabreu@synopsys.com>
Cc: Joao Pinto <jpinto@synopsys.com>
Cc: David S. Miller <davem@davemloft.net>
Cc: Giuseppe Cavallaro <peppe.cavallaro@st.com>
Cc: Alexandre Torgue <alexandre.torgue@st.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Catherine Sullivan says:
====================
Add gve driver
This patch series adds the gve driver which will support the
Compute Engine Virtual NIC that will be available in the future.
v2:
- Patch 1:
- Remove gve_size_assert.h and use static_assert instead.
- Loop forever instead of bugging if the device won't reset
- Use module_pci_driver
- Patch 2:
- Use be16_to_cpu in the RX Seq No define
- Remove unneeded ndo_change_mtu
- Patch 3:
- No Changes
- Patch 4:
- Instead of checking netif_carrier_ok in ethtool stats, just make sure
v3:
- Patch 1:
- Remove X86 dep
- Patch 2:
- No changes
- Patch 3:
- No changes
- Patch 4:
- Remove unneeded memsets in ethtool stats
v4:
- Patch 1:
- Use io[read|write]32be instead of [read|write]l(cpu_to_be32())
- Explicitly add padding to gve_adminq_set_driver_parameter
- Use static where appropriate
- Patch 2:
- Use u64_stats_sync
- Explicity add padding to gve_adminq_create_rx_queue
- Fix some enianness typing issues found by kbuild
- Use static where appropriate
- Remove unused variables
- Patch 3:
- Use io[read|write]32be instead of [read|write]l(cpu_to_be32())
- Patch 4:
- Use u64_stats_sync
- Use static where appropriate
Warnings reported by:
Reported-by: kbuild test robot <lkp@intel.com>
Reported-by: Julia Lawall <julia.lawall@lip6.fr>
====================
Signed-off-by: David S. Miller <davem@davemloft.net>
Add support for the following ethtool commands:
ethtool -s|--change devname [msglvl N] [msglevel type on|off]
ethtool -S|--statistics devname
ethtool -i|--driver devname
ethtool -l|--show-channels devname
ethtool -L|--set-channels devname
ethtool -g|--show-ring devname
ethtool --reset devname
Signed-off-by: Catherine Sullivan <csully@google.com>
Signed-off-by: Sagi Shahar <sagis@google.com>
Signed-off-by: Jon Olson <jonolson@google.com>
Acked-by: Willem de Bruijn <willemb@google.com>
Reviewed-by: Luigi Rizzo <lrizzo@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Add support for the workqueue to handle management interrupts and
support for resets.
Signed-off-by: Catherine Sullivan <csully@google.com>
Signed-off-by: Sagi Shahar <sagis@google.com>
Signed-off-by: Jon Olson <jonolson@google.com>
Acked-by: Willem de Bruijn <willemb@google.com>
Reviewed-by: Luigi Rizzo <lrizzo@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Add support for passing traffic.
Signed-off-by: Catherine Sullivan <csully@google.com>
Signed-off-by: Sagi Shahar <sagis@google.com>
Signed-off-by: Jon Olson <jonolson@google.com>
Acked-by: Willem de Bruijn <willemb@google.com>
Reviewed-by: Luigi Rizzo <lrizzo@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Add a driver framework for the Compute Engine Virtual NIC that will be
available in the future.
At this point the only functionality is loading the driver.
Signed-off-by: Catherine Sullivan <csully@google.com>
Signed-off-by: Sagi Shahar <sagis@google.com>
Signed-off-by: Jon Olson <jonolson@google.com>
Acked-by: Willem de Bruijn <willemb@google.com>
Reviewed-by: Luigi Rizzo <lrizzo@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Mahesh Bandewar says:
====================
blackhole device to invalidate dst
When we invalidate dst or mark it "dead", we assign 'lo' to
dst->dev. First of all this assignment is racy and more over,
it has MTU implications.
The standard dev MTU is 1500 while the Loopback MTU is 64k. TCP
code when dereferencing the dst don't check if the dst is valid
or not. TCP when dereferencing a dead-dst while negotiating a
new connection, may use dst device which is 'lo' instead of
using the correct device. Consider the following scenario:
A SYN arrives on an interface and tcp-layer while processing
SYNACK finds a dst and associates it with SYNACK skb. Now before
skb gets passed to L3 for processing, if that dst gets "dead"
(because of the virtual device getting disappeared & then reappeared),
the 'lo' gets assigned to that dst (lo MTU = 64k). Let's assume
the SYN has ADV_MSS set as 9k while the output device through
which this SYNACK is going to go out has standard MTU of 1500.
The MTU check during the route check passes since MIN(9K, 64K)
is 9k and TCP successfully negotiates 9k MSS. The subsequent
data packet; bigger in size gets passed to the device and it
won't be marked as GSO since the assumed MTU of the device is
9k.
This either crashes the NIC and we have seen fixes that went
into drivers to handle this scenario. 8914a59511 ('bnx2x:
disable GSO where gso_size is too big for hardware') and
2b16f04872 ('net: create skb_gso_validate_mac_len()') and
with those fixes TCP eventually recovers but not before
few dropped segments.
Well, I'm not a TCP expert and though we have experienced
these corner cases in our environment, I could not reproduce
this case reliably in my test setup to try this fix myself.
However, Michael Chan <michael.chan@broadcom.com> had a setup
where these fixes helped him mitigate the issue and not cause
the crash.
The idea here is to not alter the data-path with additional
locks or smb()/rmb() barriers to avoid racy assignments but
to create a new device that has really low MTU that has
.ndo_start_xmit essentially a kfree_skb(). Make use of this
device instead of 'lo' when marking the dst dead.
First patch implements the blackhole device and second
patch uses it in IPv4 and IPv6 stack while the third patch
is the self test that ensures the sanity of this device.
v1->v2
fixed the self-test patch to handle the conflict
v2 -> v3
fixed Kconfig text/string.
====================
Signed-off-by: David S. Miller <davem@davemloft.net>
Since this is not really a device with all capabilities, this test
ensures that it has *enough* to make it through the data path
without causing unwanted side-effects (read crash!).
Signed-off-by: Mahesh Bandewar <maheshb@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Shay Agroskin