The code is checking that the Ethernet frame will fit into a
netdev allocated socket buffer within the constraints of MTU size,
Ethernet header length plus VLAN header length.
The original code was checking rx->remaining each loop of the while
loop that processes multiple Ethernet frames per URB and/or Ethernet
frames that span across URBs. rx->remaining decreases per while loop
so there is no point in potentially checking multiple times that the
Ethernet frame (remaining part) will fit into the netdev socket buffer.
The modification checks that the size of the Ethernet frame will fit
the netdev socket buffer before allocating the netdev socket buffer.
This avoids grabbing memory and then deciding that the Ethernet frame
is too big and then freeing the memory.
Signed-off-by: Dean Jenkins <Dean_Jenkins@mentor.com>
Signed-off-by: Mark Craske <Mark_Craske@mentor.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Tidy-up the Data header 32-bit word synchronisation logic in
asix_rx_fixup_internal() by removing redundant logic tests.
The code is looking at the following cases of the Data header
32-bit word that is present before each Ethernet frame:
a) all 32 bits of the Data header word are in the URB socket buffer
b) first 16 bits of the Data header word are at the end of the URB
socket buffer
c) last 16 bits of the Data header word are at the start of the URB
socket buffer eg. split_head = true
Note that the lifetime of rx->split_head exists outside of the
function call and is accessed per processing of each URB. Therefore,
split_head being true acts on the next URB to be processed.
To check for b) the offset will be 16 bits (2 bytes) from the end of
the buffer then indicate split_head is true.
To check for c) split_head must be true because the first 16 bits
have been found.
To check for a) else c)
Note that the || logic of the old code included the state
(skb->len - offset == sizeof(u16) && rx->split_head) which is not
possible because the split_head cannot be true whilst checking for b).
This is because the split_head indicates that the first 16 bits have
been found and that is not possible whilst checking for the first 16
bits. Therefore simplify the logic.
Signed-off-by: Dean Jenkins <Dean_Jenkins@mentor.com>
Signed-off-by: Mark Craske <Mark_Craske@mentor.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The Data header synchronisation is easier to understand
if the variables "remaining" and "size" are renamed.
Therefore, the lifetime of the "remaining" variable exists
outside of asix_rx_fixup_internal() and is used to indicate
any remaining pending bytes of the Ethernet frame that need
to be obtained from the next socket buffer. This allows an
Ethernet frame to span across multiple socket buffers.
"size" is now local to asix_rx_fixup_internal() and contains
the size read from the Data header 32-bit word.
Add "copy_length" to hold the number of the Ethernet frame
bytes (maybe a part of a full frame) that are to be copied
out of the socket buffer.
Signed-off-by: Dean Jenkins <Dean_Jenkins@mentor.com>
Signed-off-by: Mark Craske <Mark_Craske@mentor.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The current ongoing effort to dump existing cBPF seccomp filters back
to user space requires to hold the pre-transformed instructions like
we do in case of socket filters from sk_attach_filter() side, so they
can be reloaded in original form at a later point in time by utilities
such as criu.
To prepare for this, simply extend the bpf_prog_create_from_user()
API to hold a flag that tells whether we should store the original
or not. Also, fanout filters could make use of that in future for
things like diag. While fanout filters already use bpf_prog_destroy(),
move seccomp over to them as well to handle original programs when
present.
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Cc: Tycho Andersen <tycho.andersen@canonical.com>
Cc: Pavel Emelyanov <xemul@parallels.com>
Cc: Kees Cook <keescook@chromium.org>
Cc: Andy Lutomirski <luto@amacapital.net>
Cc: Alexei Starovoitov <ast@plumgrid.com>
Tested-by: Tycho Andersen <tycho.andersen@canonical.com>
Acked-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This ethernet driver supports the Micorchip enc424j600/626j600 Ethernet
controller over a SPI bus interface. This driver makes use of the regmap API to
optimize access to registers by caching registers where possible.
Datasheet:
http://ww1.microchip.com/downloads/en/DeviceDoc/39935b.pdf
Signed-off-by: Jon Ringle <jringle@gridpoint.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This commit allows installing a custom reg_update_bits function for cases where
the hardware provides a mechanism to set or clear register bits without a
read/modify/write cycle. Such is the case with the Microchip ENCX24J600.
Signed-off-by: Jon Ringle <jringle@gridpoint.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The current code invokes hang reset in case of error interrupt. We should
hang reset only in case of tx timeout. This because of the way hang reset
is implemented in firmware. Hang reset takes more firmware resources than
soft reset. Adaptor does not generate error interrupt in case of tx
timeout.
Hang reset only in case of tx timeout, in .ndo_tx_timeout. Do soft reset
otherwise. Introduce deferred work, enic_tx_hang_reset, to do hang reset.
Signed-off-by: Govindarajulu Varadarajan <_govind@gmx.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Some of the enic adaptors are know to generate spurious interrupts. When
error interrupt is generated, driver just resets the device. This patch
resets the device only when an error is occurred.
Signed-off-by: Govindarajulu Varadarajan <_govind@gmx.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Hariprasad Shenai says:
====================
cxgb4: Trivial fixes for cxgb4
Fixes the following issues
Don't read non existent T4/T5/T6 adapter registers for ethtool dump.
For T4, dont read mailbox control registers. Adds new devlog faility and
report correct link speed for unsupported ones.
This patch series has been created against net-next tree and includes
patches on cxgb4 driver.
We have included all the maintainers of respective drivers. Kindly review
the change and let us know in case of any review comments.
====================
Acked-by: Nicolas Dichtel <nicolas.dichtel@6wind.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
When we get garbage from the firmware with weird Port Speeds,
etc. we should emit a warning regarding unsupported speeds rather than
use the bogus default of "10Mbps" which isn't even an option in the
firmware Port Information message
Signed-off-by: Hariprasad Shenai <hariprasad@chelsio.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The firmware team added a new Device Log Facility FW_DEVLOG_FACILITY_CF,
but the driver has been decoding Device Log messages with that Facility as
"(NULL)", fixing it.
Signed-off-by: Hariprasad Shenai <hariprasad@chelsio.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
T4 doesn't have the Shadow copy of the register which we can read without
side effect. So don't read mbox control register for T4 adapter
Signed-off-by: Hariprasad Shenai <hariprasad@chelsio.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Update T4/T5/T6 adapter register ranges so that it doesn't read non
existent registers when dumped using ethtool
Signed-off-by: Hariprasad Shenai <hariprasad@chelsio.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Eric W. Biederman says:
====================
net: Pass net through ip fragmention
This is the next installment of my work to pass struct net through the
output path so the code does not need to guess how to figure out which
network namespace it is in, and ultimately routes can have output
devices in another network namespace.
This round focuses on passing net through ip fragmentation which we seem
to call from about everywhere. That is the main ip output paths, the
bridge netfilter code, and openvswitch. This has to happend at once
accross the tree as function pointers are involved.
First some prep work is done, then ipv4 and ipv6 are converted and then
temporary helper functions are removed.
====================
Acked-by: Nicolas Dichtel <nicolas.dichtel@6wind.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Sowmini Varadhan says:
====================
RDS: RDS-TCP perf enhancements
A 3-part patchset that (a) improves current RDS-TCP perf
by 2X-3X and (b) refactors earlier robustness code for
better observability/scaling.
Patch 1 is an enhancment of earlier robustness fixes
that had used separate sockets for client and server endpoints to
resolve race conditions. It is possible to have an equivalent
solution that does not use 2 sockets. The benefit of a
single socket solution is that it results in more predictable
and observable behavior for the underlying TCP pipe of an
RDS connection
Patches 2 and 3 are simple, straightforward perf bug fixes
that align the RDS TCP socket with other parts of the kernel stack.
v2: fix kbuild-test-robot warnings, comments from Sergei Shtylov
and Santosh Shilimkar.
====================
Acked-by: Santosh Shilimkar <santosh.shilimkar@oracle.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
For the same reasons as commit 2f53384424 ("tcp: allow splice() to
build full TSO packets") and commit 35f9c09fe9 ("tcp: tcp_sendpages()
should call tcp_push() once"), rds_tcp_xmit may have multiple pages to
send, so use the MSG_MORE and MSG_SENDPAGE_NOTLAST as hints to
tcp_sendpage()
Signed-off-by: Sowmini Varadhan <sowmini.varadhan@oracle.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Using the value of RDS_TCP_DEFAULT_BUFSIZE (128K)
clobbers efficient use of TSO because it inflates the size_goal
that is computed in tcp_sendmsg/tcp_sendpage and skews packet
latency, and the default values for these parameters actually
results in significantly better performance.
In request-response tests using rds-stress with a packet size of
100K with 16 threads (test parameters -q 100000 -a 256 -t16 -d16)
between a single pair of IP addresses achieves a throughput of
6-8 Gbps. Without this patch, throughput maxes at 2-3 Gbps under
equivalent conditions on these platforms.
Signed-off-by: Sowmini Varadhan <sowmini.varadhan@oracle.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Commit f711a6ae06 ("net/rds: RDS-TCP: Always create a new rds_sock
for an incoming connection.") modified rds-tcp so that an incoming SYN
would ignore an existing "client" TCP connection which had the local
port set to the transient port. The motivation for ignoring the existing
"client" connection in f711a6ae was to avoid race conditions and an
endless duel of reconnect attempts triggered by a restart/abort of one
of the nodes in the TCP connection.
However, having separate sockets for active and passive sides
is avoidable, and the simpler model of a single TCP socket for
both send and receives of all RDS connections associated with
that tcp socket makes for easier observability. We avoid the race
conditions from f711a6ae by attempting reconnects in rds_conn_shutdown
if, and only if, the (new) c_outgoing bit is set for RDS_TRANS_TCP.
The c_outgoing bit is initialized in __rds_conn_create().
A side-effect of re-using the client rds_connection for an incoming
SYN is the potential of encountering duelling SYNs, i.e., we
have an outgoing RDS_CONN_CONNECTING socket when we get the incoming
SYN. The logic to arbitrate this criss-crossing SYN exchange in
rds_tcp_accept_one() has been modified to emulate the BGP state
machine: the smaller IP address should back off from the connection attempt.
Signed-off-by: Sowmini Varadhan <sowmini.varadhan@oracle.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Tom Lendacky says:
====================
amd-xgbe: AMD XGBE driver updates 2015-09-30
The following patches are included in this driver update series:
- Remove unneeded semi-colon
- Follow the DT/ACPI precedence used by the device_ APIs
- Add ethtool support for getting and setting the msglevel
- Add ethtool support error and debug messages
- Simplify the hardware FIFO assignment calculations
- Add receive buffer unavailable statistic
- Use the device workqueue instead of the system workqueue
- Remove the use of a link state bit
This patch series is based on net-next.
====================
Signed-off-by: David S. Miller <davem@davemloft.net>
The XGBE_LINK bit is used just to determine whether to call the
netif_carrier_on/off functions. Rather than define and use this bit,
just call the functions. The netif_carrier_ok function can be used in
place of checking the XGBE_LINK bit in the future.
Signed-off-by: Tom Lendacky <thomas.lendacky@amd.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The driver creates, flushes and destroys a device workqueue but queues
work to the system workqueue. Switch from using the system workqueue to
the device workqueue.
Signed-off-by: Tom Lendacky <thomas.lendacky@amd.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Add a statistic that tracks how many times an interrupt is generated for
a receive buffer not being available to the hardware which prevents the
hardware from being able to DMA the received data.
Signed-off-by: Tom Lendacky <thomas.lendacky@amd.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The calculation of the Tx and Rx fifo sizes can be calculated rather
than hardcoded in a switch statement. Additionally, the per-queue fifo
sizes can be calculated rather than hardcoded using if/else if statements
that can possibly underutilize the available fifo area.
Change the code to calculate the fifo sizes and the per-queue fifo sizes
to simplify the code and make best use of the available fifo.
Signed-off-by: Tom Lendacky <thomas.lendacky@amd.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Add error and dynamic debug messages to various ethtool functions in
the driver while also removing the DBGPR debug print calls. Also, change
the message level for some error messages from alert to err.
Signed-off-by: Tom Lendacky <thomas.lendacky@amd.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Provide the ethtool functions to support getting and setting the
msglevel for the driver.
Signed-off-by: Tom Lendacky <thomas.lendacky@amd.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Device tree presence takes precedence over ACPI in the device_* APIs.
The amd-xgbe driver should follow the same precedence. Update the check
on whether to use DT / ACPI to follow this.
Signed-off-by: Tom Lendacky <thomas.lendacky@amd.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Remove an unneeded semicolon at the end of a switch statement block.
Reported-by: kbuild test robot <fengguang.wu@intel.com>
Signed-off-by: Tom Lendacky <thomas.lendacky@amd.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
I accidentally cleared fastopenq.max_qlen in reqsk_queue_alloc()
while max_qlen can be set before listen() is called,
using TCP_FASTOPEN socket option for example.
Fixes: 0536fcc039 ("tcp: prepare fastopen code for upcoming listener changes")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Arnd Bergmann says:
====================
net: assorted y2038 changes
This is a set of changes for network drivers and core code to
get rid of the use of time_t and derived data structures.
I have a longer set of patches that enables me to build kernels
with the time_t definition removed completely as a help to find
y2038 overflow issues. This is the subset for networking that
contains all code that has a reasonable way of fixing at the
moment and that is either commonly used (in one of the defconfigs)
or that blocks building a whole subsystem.
Most of the patches in this series should be noncontroversial,
but the last two that I marked [RFC] are a bit tricky and
need input from people that are more familiar with the code than
I am. All 12 patches are independent of one another and can
be applied in any order, so feel free to pick all that look
good.
Patches that are not included here are:
- disabling less common device drivers that I don't have a fix
for yet, this includes
drivers/net/ethernet/brocade/bna/bfa_ioc.c
drivers/net/ethernet/qlogic/netxen/netxen_nic_hw.c
drivers/net/ethernet/tile/tilegx.c
drivers/net/hamradio/baycom_ser_fdx.c
drivers/net/wireless/ath/ath10k/core.h
drivers/net/wireless/ath/ath9k/
drivers/net/wireless/ath/ath9k/
drivers/net/wireless/atmel.c
drivers/net/wireless/prism54/isl_38xx.c
drivers/net/wireless/rt2x00/rt2x00debug.c
drivers/net/wireless/rtlwifi/
drivers/net/wireless/ti/wlcore/
drivers/staging/ozwpan/
net/atm/mpoa_caches.c
net/atm/mpoa_proc.c
net/dccp/probe.c
net/ipv4/tcp_probe.c
net/netfilter/nfnetlink_queue_core.c
net/netfilter/nfnetlink_queue_core.c
net/netfilter/xt_time.c
net/openvswitch/flow.c
net/sctp/probe.c
net/sunrpc/auth_gss/
net/sunrpc/svcauth_unix.c
net/vmw_vsock/af_vsock.c
We'll get there eventually, or we an add a dependency to ensure
they are not built on 32-bit kernels that need to survive
beyond 2038. Most of these should be really easy to fix.
- recvmmsg/sendmmsg system calls: patches have been sent out
as part of the syscall series, need a little more work and
review
- SIOCGSTAMP/SIOCGSTAMPNS/ ioctl calls: tricky, need to discuss
with some folks at kernel summit
- SO_RCVTIMEO/SO_SNDTIMEO/SO_TIMESTAMP/SO_TIMESTAMPNS socket
opt: similar and related to the ioctl
- mmapped packet socket: need to create v4 of the API, nontrivial
- pktgen: sends 32-bit timestamps over network, need to find out
if using unsigned stamps is good enough
- af_rxpc: similar to pktgen, uses 32-bit times for deadlines
- ppp ioctl: patch is being worked on, nontrivial but doable
====================
Signed-off-by: David S. Miller <davem@davemloft.net>
We want to avoid using time_t in the kernel because of the y2038
overflow problem. The use in sctp is not for storing seconds at
all, but instead uses microseconds and is passed as 32-bit
on all machines.
This patch changes the type to u32, which better fits the use.
Signed-off-by: Arnd Bergmann <arnd@arndb.de>
Cc: Vlad Yasevich <vyasevich@gmail.com>
Cc: Neil Horman <nhorman@tuxdriver.com>
Cc: linux-sctp@vger.kernel.org
Acked-by: Neil Horman <nhorman@tuxdriver.com>
Acked-by: Marcelo Ricardo Leitner <marcelo.leitner@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The ipv6 mip6 implementation is one of only a few users of the
skb_get_timestamp() function in the kernel, which is both unsafe
on 32-bit architectures because of the 2038 overflow, and slightly
less efficient than the skb_get_ktime() based approach.
This converts the function call and the mip6_report_rate_limiter
structure that stores the time stamp, eliminating all uses of
timeval in the ipv6 code.
Signed-off-by: Arnd Bergmann <arnd@arndb.de>
Cc: Alexey Kuznetsov <kuznet@ms2.inr.ac.ru>
Cc: James Morris <jmorris@namei.org>
Cc: Hideaki YOSHIFUJI <yoshfuji@linux-ipv6.org>
Cc: Patrick McHardy <kaber@trash.net>
Signed-off-by: David S. Miller <davem@davemloft.net>
The __build_packet_message function fills a nfulnl_msg_packet_timestamp
structure that uses 64-bit seconds and is therefore y2038 safe, but
it uses an intermediate 'struct timespec' which is not.
This trivially changes the code to use 'struct timespec64' instead,
to correct the result on 32-bit architectures.
Signed-off-by: Arnd Bergmann <arnd@arndb.de>
Cc: Pablo Neira Ayuso <pablo@netfilter.org>
Cc: Patrick McHardy <kaber@trash.net>
Cc: Jozsef Kadlecsik <kadlec@blackhole.kfki.hu>
Cc: netfilter-devel@vger.kernel.org
Cc: coreteam@netfilter.org
Acked-by: Pablo Neira Ayuso <pablo@netfilter.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
The zatm_t_hist structure is not used anywhere in the kernel, but is
exported to user space. As we are trying to eliminate uses of time_t
in the kernel for y2038 compatibility, the current definition triggers
checking tools because it contains 'struct timeval'.
As pointed out by Chas Williams, the only user of this structure was
the ZATM_GETHIST ioctl command that has been removed a long time ago,
and we can remove the structure as well without breaking any user
space.
Signed-off-by: Arnd Bergmann <arnd@arndb.de>
Cc: Chas Williams <3chas3@gmail.com>
Cc: linux-atm-general@lists.sourceforge.net
Signed-off-by: David S. Miller <davem@davemloft.net>
The mac80211 code uses ktime_get_ts to measure the connected time.
As this uses monotonic time, it is y2038 safe on 32-bit systems,
but we still want to deprecate the use of 'timespec' because most
other users are broken.
This changes the code to use ktime_get_seconds() instead, which
avoids the timespec structure and is slightly more efficient.
Signed-off-by: Arnd Bergmann <arnd@arndb.de>
Cc: Johannes Berg <johannes@sipsolutions.net>
Cc: linux-wireless@vger.kernel.org
Signed-off-by: David S. Miller <davem@davemloft.net>
mwifiex_get_random_ba_threshold() uses a complex homegrown implementation
to generate a pseudo-random number from the current time as returned
from do_gettimeofday().
This currently requires two 32-bit divisions plus a couple of other
computations that are eventually discarded as only eight bits of
the microsecond portion are used at all.
We could replace this with a call to get_random_bytes(), but that
might drain the entropy pool too fast if this is called for each
packet.
Instead, this patch converts it to use ktime_get_ns(), which is a
bit faster than do_gettimeofday(), and then uses a similar algorithm
as before, but in a way that takes both the nanosecond and second
portion into account for slightly-more-but-still-not-very-random
pseudorandom number.
Signed-off-by: Arnd Bergmann <arnd@arndb.de>
Cc: Amitkumar Karwar <akarwar@marvell.com>
Cc: Nishant Sarmukadam <nishants@marvell.com>
Cc: Kalle Valo <kvalo@codeaurora.org>
Cc: linux-wireless@vger.kernel.org
Signed-off-by: David S. Miller <davem@davemloft.net>
The mwifiex_11n_aggregate_pkt() function creates a ktime_t from
a timeval returned by do_gettimeofday, which is slow and causes
an overflow in 2038 on 32-bit architectures.
This solves both problems by using the appropriate ktime_get_real()
function.
Signed-off-by: Arnd Bergmann <arnd@arndb.de>
Cc: Amitkumar Karwar <akarwar@marvell.com>
Cc: Nishant Sarmukadam <nishants@marvell.com>
Cc: Kalle Valo <kvalo@codeaurora.org>
Cc: linux-wireless@vger.kernel.org
Signed-off-by: David S. Miller <davem@davemloft.net>
We want to deprecate the use of 'struct timespec' on 32-bit
architectures, as it is will overflow in 2038. The igb
driver uses it to read the current time, and can simply
be changed to use ktime_get_real_ts64() instead.
Because of hardware limitations, there is still an overflow
in year 2106, which we cannot really avoid, but this documents
the overflow.
Signed-off-by: Arnd Bergmann <arnd@arndb.de>
Cc: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
Cc: intel-wired-lan@lists.osuosl.org
Reviewed-by: Richard Cochran <richardcochran@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
We want to deprecate the use of 'struct timespec' on 32-bit
architectures, as it is will overflow in 2038. The stmmac
driver uses it to read the current time, and can simply
be changed to use ktime_get_real_ts64() instead.
Because of hardware limitations, there is still an overflow
in year 2106, which we cannot really avoid, but this documents
the overflow.
Signed-off-by: Arnd Bergmann <arnd@arndb.de>
Cc: Giuseppe Cavallaro <peppe.cavallaro@st.com>
Cc: Richard Cochran <richardcochran@gmail.com>
Acked-by: Richard Cochran <richardcochran@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The fec_ptp_enable_pps uses an open-coded implementation of ns_to_timespec,
which will be removed eventually as it is not y2038-safe on 32-bit
architectures. Two more instances of the same code in this file were
already converted to use the safe ns_to_timespec64 in commit 6630514fce
("ptp: fec: use helpers for converting ns to timespec"), this changes
the last one as well.
The seconds portion here is actually unused and we could just remove the
timespec variable, but using ns_to_timespec64 can still be better as the
implementation can be hand-optimized in the future.
Signed-off-by: Arnd Bergmann <arnd@arndb.de>
Cc: Richard Cochran <richardcochran@gmail.com>
Cc: Fugang Duan <b38611@freescale.com>
Cc: Luwei Zhou <b45643@freescale.com>
Cc: Frank Li <Frank.Li@freescale.com>
Acked-by: Richard Cochran <richardcochran@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Peter Nørlund says:
====================
ipv4: Hash-based multipath routing
When the routing cache was removed in 3.6, the IPv4 multipath algorithm changed
from more or less being destination-based into being quasi-random per-packet
scheduling. This increases the risk of out-of-order packets and makes it
impossible to use multipath together with anycast services.
This patch series replaces the old implementation with flow-based load
balancing based on a hash over the source and destination addresses.
Distribution of the hash is done with thresholds as described in RFC 2992.
This reduces the disruption when a path is added/remove when having more than
two paths.
To futher the chance of successful usage in conjuction with anycast, ICMP
error packets are hashed over the inner IP addresses. This ensures that PMTU
will work together with anycast or load-balancers such as IPVS.
Port numbers are not considered since fragments could cause problems with
anycast and IPVS. Relying on the DF-flag for TCP packets is also insufficient,
since ICMP inspection effectively extracts information from the opposite
flow which might have a different state of the DF-flag. This is also why the
RSS hash is not used. These are typically based on the NDIS RSS spec which
mandates TCP support.
Measurements of the additional overhead of a two-path multipath
(p_mkroute_input excl. __mkroute_input) on a Xeon X3550 (4 cores, 2.66GHz):
Original per-packet: ~394 cycles/packet
L3 hash: ~76 cycles/packet
Changes in v5:
- Fixed compilation error
Changes in v4:
- Functions take hash directly instead of func ptr
- Added inline hash function
- Added dummy macros to minimize ifdefs
- Use upper 31 bits of hash instead of lower
Changes in v3:
- Multipath algorithm is no longer configurable (always L3)
- Added random seed to hash
- Moved ICMP inspection to isolated function
- Ignore source quench packets (deprecated as per RFC 6633)
Changes in v2:
- Replaced 8-bit xor hash with 31-bit jenkins hash
- Don't scale weights (since 31-bit)
- Avoided unnecesary renaming of variables
- Rely on DF-bit instead of fragment offset when checking for fragmentation
- upper_bound is now inclusive to avoid overflow
- Use a callback to postpone extracting flow information until necessary
- Skipped ICMP inspection entirely with L4 hashing
- Handle newly added sysctl ignore_routes_with_linkdown
====================
Signed-off-by: David S. Miller <davem@davemloft.net>
ICMP packets are inspected to let them route together with the flow they
belong to, minimizing the chance that a problematic path will affect flows
on other paths, and so that anycast environments can work with ECMP.
Signed-off-by: Peter Nørlund <pch@ordbogen.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Replaces the per-packet multipath with a hash-based multipath using
source and destination address.
Signed-off-by: Peter Nørlund <pch@ordbogen.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
inet_reqsk_alloc() is used to allocate a temporary request
in order to generate a SYNACK with a cookie. Then later,
syncookie validation also uses a temporary request.
These paths already took a reference on listener refcount,
we can avoid a couple of atomic operations.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
SYN_RECV & TIMEWAIT sockets are not full blown, they do not have a
sk_dst_cache pointer.
Fixes: ca6fb06518 ("tcp: attach SYNACK messages to request sockets instead of listener")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
SYN_RECV & TIMEWAIT sockets are not full blown, they do not have a pinet6
pointer.
Fixes: ca6fb06518 ("tcp: attach SYNACK messages to request sockets instead of listener")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
SYN_RECV & TIMEWAIT sockets are not full blown,
do not even try to call ip_sk_use_pmtu() on them.
Fixes: ca6fb06518 ("tcp: attach SYNACK messages to request sockets instead of listener")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
There are multiple races that need fixes :
1) skb_get() + queue skb + kfree_skb() is racy
An accept() can be done on another cpu, data consumed immediately.
tcp_recvmsg() uses __kfree_skb() as it is assumed all skb found in
socket receive queue are private.
Then the kfree_skb() in tcp_rcv_state_process() uses an already freed skb
2) tcp_reqsk_record_syn() needs to be done before tcp_try_fastopen()
for the same reasons.
3) We want to send the SYNACK before queueing child into accept queue,
otherwise we might reintroduce the ooo issue fixed in
commit 7c85af8810 ("tcp: avoid reorders for TFO passive connections")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
