[ Upstream commit a117f73dc2 ]
When mlx5_core is loaded it is expected to sync ports
with all vxlan devices so it can support vxlan encap/decap.
This is done via udp_tunnel_get_rx_info(). Currently this
call is set in mlx5e_nic_enable() and if the netdev is not in
NETREG_REGISTERED state it will not be called.
Normally on load the netdev state is not NETREG_REGISTERED
so udp_tunnel_get_rx_info() will not be called.
Moving udp_tunnel_get_rx_info() to mlx5e_open() so
it will be called on netdev UP event and allow encap/decap.
Fixes: 610e89e05c ("net/mlx5e: Don't sync netdev state when not registered")
Signed-off-by: Shahar Klein <shahark@mellanox.com>
Reviewed-by: Roi Dayan <roid@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
The vport admin original link state will be re-applied after returning
back to legacy mode, it is not right to change the admin link state value
when in switchdev mode.
Use direct vport commands to alter logical vport state in netdev
representor open/close flows rather than the administrative eswitch API.
Fixes: 20a1ea6747 ('net/mlx5e: Support VF vport link state control for SRIOV switchdev mode')
Signed-off-by: Jianbo Liu <jianbol@mellanox.com>
Reviewed-by: Roi Dayan <roid@mellanox.com>
Reviewed-by: Or Gerlitz <ogerlitz@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
[ Upstream commit 5807b22c91 ]
Enabling TSO can lead to abysmal performances when using seg6 in
encap mode, such as with the ixgbe driver. This patch adds a call to
iptunnel_handle_offloads() to remove the encapsulation bit if needed.
Before:
root@comp4-seg6bpf:~# iperf3 -c fc00::55
Connecting to host fc00::55, port 5201
[ 4] local fc45::4 port 36592 connected to fc00::55 port 5201
[ ID] Interval Transfer Bandwidth Retr Cwnd
[ 4] 0.00-1.00 sec 196 KBytes 1.60 Mbits/sec 47 6.66 KBytes
[ 4] 1.00-2.00 sec 304 KBytes 2.49 Mbits/sec 100 5.33 KBytes
[ 4] 2.00-3.00 sec 284 KBytes 2.32 Mbits/sec 92 5.33 KBytes
After:
root@comp4-seg6bpf:~# iperf3 -c fc00::55
Connecting to host fc00::55, port 5201
[ 4] local fc45::4 port 43062 connected to fc00::55 port 5201
[ ID] Interval Transfer Bandwidth Retr Cwnd
[ 4] 0.00-1.00 sec 1.03 GBytes 8.89 Gbits/sec 0 743 KBytes
[ 4] 1.00-2.00 sec 1.03 GBytes 8.87 Gbits/sec 0 743 KBytes
[ 4] 2.00-3.00 sec 1.03 GBytes 8.87 Gbits/sec 0 743 KBytes
Reported-by: Tom Herbert <tom@quantonium.net>
Fixes: 6c8702c60b ("ipv6: sr: add support for SRH encapsulation and injection with lwtunnels")
Signed-off-by: David Lebrun <dlebrun@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
[ Upstream commit 1489bbd10e ]
The NSP default buffer is a piece of NFP memory where additional
command data can be placed. Its format has been copied from
host buffer, but the PCIe selection bits do not make sense in
this case. If those get masked out from a NFP address - writes
to random place in the chip memory may be issued and crash the
device.
Even in the general NSP buffer case, it doesn't make sense to have the
PCIe selection bits there anymore. These are unused at the moment, and
when it becomes necessary, the PCIe selection bits should rather be
moved to another register to utilise more bits for the buffer address.
This has never been an issue because the buffer used to be
allocated in memory with less-than-38-bit-long address but that
is about to change.
Fixes: 1a64821c6a ("nfp: add support for service processor access")
Signed-off-by: Dirk van der Merwe <dirk.vandermerwe@netronome.com>
Reviewed-by: Jakub Kicinski <jakub.kicinski@netronome.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
[ Upstream commit af1607c37d ]
For NIC flows, the parsed attributes are not freed when we exit
successfully from mlx5e_configure_flower().
There is possible double free for eswitch flows. If error is returned
from rhashtable_insert_fast(), the parse attrs will be freed in
mlx5e_tc_del_flow(), but they will be freed again before exiting
mlx5e_configure_flower().
To fix both issues we do the following:
(1) change the condition that determines if to issue the free call to
check if this flow is NIC flow, or it does not have encap action.
(2) reorder the code such that that the check and free calls are done
before we attempt to add into the hash table.
Fixes: 232c001398 ('net/mlx5e: Add support to neighbour update flow')
Signed-off-by: Jianbo Liu <jianbol@mellanox.com>
Reviewed-by: Or Gerlitz <ogerlitz@mellanox.com>
Reviewed-by: Roi Dayan <roid@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
[ Upstream commit 423c9db299 ]
Currently we use the global ipv6_stub var to access the ipv6 global
nd table. This practice gets us to troubles when the stub is only partially
set e.g when ipv6 is loaded under the disabled policy. In this case, as of commit
343d60aada ("ipv6: change ipv6_stub_impl.ipv6_dst_lookup to take net argument")
the stub is not null, but stub->nd_tbl is and we crash.
As we can access the ipv6 nd_tbl directly, the fix is just to avoid the
reference through the stub. There is one place in the code where we
issue ipv6 route lookup and keep doing it through the stub, but that
mentioned commit makes sure we get -EAFNOSUPPORT from the stack.
Fixes: 232c001398 ("net/mlx5e: Add support to neighbour update flow")
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
Reviewed-by: Aviv Heller <avivh@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
[ Upstream commit a9d48205d0 ]
We want to use dev_valid_name() to validate tunnel names,
so better use strnlen(name, IFNAMSIZ) than strlen(name) to make
sure to not upset KASAN.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
[ Upstream commit 9f5a90c107 ]
When dev_set_promiscuity(1) succeeds but dev_set_allmulti(1) fails,
dev_set_promiscuity(-1) should be done before going to the err path.
Otherwise, dev->promiscuity will leak.
Fixes: 7e1a1ac1fb ("bonding: Check return of dev_set_promiscuity/allmulti")
Signed-off-by: Xin Long <lucien.xin@gmail.com>
Acked-by: Andy Gospodarek <andy@greyhouse.net>
Signed-off-by: David S. Miller <davem@davemloft.net>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
[ Upstream commit ae42cc62a9 ]
Beniamino found a crash when adding vlan as slave of bond which is also
the parent link:
ip link add bond1 type bond
ip link set bond1 up
ip link add link bond1 vlan1 type vlan id 80
ip link set vlan1 master bond1
The call trace is as below:
[<ffffffffa850842a>] queued_spin_lock_slowpath+0xb/0xf
[<ffffffffa8515680>] _raw_spin_lock+0x20/0x30
[<ffffffffa83f6f07>] dev_mc_sync+0x37/0x80
[<ffffffffc08687dc>] vlan_dev_set_rx_mode+0x1c/0x30 [8021q]
[<ffffffffa83efd2a>] __dev_set_rx_mode+0x5a/0xa0
[<ffffffffa83f7138>] dev_mc_sync_multiple+0x78/0x80
[<ffffffffc084127c>] bond_enslave+0x67c/0x1190 [bonding]
[<ffffffffa8401909>] do_setlink+0x9c9/0xe50
[<ffffffffa8403bf2>] rtnl_newlink+0x522/0x880
[<ffffffffa8403ff7>] rtnetlink_rcv_msg+0xa7/0x260
[<ffffffffa8424ecb>] netlink_rcv_skb+0xab/0xc0
[<ffffffffa83fe498>] rtnetlink_rcv+0x28/0x30
[<ffffffffa8424850>] netlink_unicast+0x170/0x210
[<ffffffffa8424bf8>] netlink_sendmsg+0x308/0x420
[<ffffffffa83cc396>] sock_sendmsg+0xb6/0xf0
This is actually a dead lock caused by sync slave hwaddr from master when
the master is the slave's 'slave'. This dead loop check is actually done
by netdev_master_upper_dev_link. However, Commit 1f718f0f4f ("bonding:
populate neighbour's private on enslave") moved it after dev_mc_sync.
This patch is to fix it by moving dev_mc_sync after master_upper_dev_link,
so that this loop check would be earlier than dev_mc_sync. It also moves
if (mode == BOND_MODE_8023AD) into if (!bond_uses_primary) clause as an
improvement.
Note team driver also has this issue, I will fix it in another patch.
Fixes: 1f718f0f4f ("bonding: populate neighbour's private on enslave")
Reported-by: Beniamino Galvani <bgalvani@redhat.com>
Signed-off-by: Xin Long <lucien.xin@gmail.com>
Acked-by: Andy Gospodarek <andy@greyhouse.net>
Signed-off-by: David S. Miller <davem@davemloft.net>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
[ Upstream commit 5c78f6bfae ]
vlan_vids_add_by_dev is called right after dev hwaddr sync, so on
the err path it should unsync dev hwaddr. Otherwise, the slave
dev's hwaddr will never be unsync when this err happens.
Fixes: 1ff412ad77 ("bonding: change the bond's vlan syncing functions with the standard ones")
Signed-off-by: Xin Long <lucien.xin@gmail.com>
Reviewed-by: Nikolay Aleksandrov <nikolay@cumulusnetworks.com>
Acked-by: Andy Gospodarek <andy@greyhouse.net>
Signed-off-by: David S. Miller <davem@davemloft.net>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
[ Upstream commit 82dd0d2a9a ]
Miguel reported an skb use after free / double free in vrf_finish_output
when neigh_output returns an error. The vrf driver should return after
the call to neigh_output as it takes over the skb on error path as well.
Patch is a simplified version of Miguel's patch which was written for 4.9,
and updated to top of tree.
Fixes: 8f58336d3f ("net: Add ethernet header for pass through VRF device")
Signed-off-by: Miguel Fadon Perlines <mfadon@teldat.com>
Signed-off-by: David Ahern <dsahern@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
[ Upstream commit ec1d8ccb07 ]
Just like function ethtool_get_ts_info(), we should also consider the
phy_driver ts_info call back. For example, driver dp83640.
Fixes: 37dd9255b2 ("vlan: Pass ethtool get_ts_info queries to real device.")
Acked-by: Richard Cochran <richardcochran@gmail.com>
Signed-off-by: Hangbin Liu <liuhangbin@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
[ Upstream commit 19c9ea363a ]
pci_set_drvdata() is called only after registering the net_device,
therefore we could run into a NPE if one of the functions using
driver_data is called before it's set.
Fix this by calling pci_set_drvdata() before registering the
net_device.
This fix is a candidate for stable. As far as I can see the
bug has been there in kernel version 3.2 already, therefore
I can't provide a reference which commit is fixed by it.
The fix may need small adjustments per kernel version because
due to other changes the label which is jumped to if
register_netdev() fails has changed over time.
Reported-by: David Miller <davem@davemloft.net>
Signed-off-by: Heiner Kallweit <hkallweit1@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
[ Upstream commit 734549eb55 ]
Fixes a bug in the tcf_dump_walker function that can cause some actions
to not be reported when dumping a large number of actions. This issue
became more aggrevated when cookies feature was added. In particular
this issue is manifest when large cookie values are assigned to the
actions and when enough actions are created that the resulting table
must be dumped in multiple batches.
The number of actions returned in each batch is limited by the total
number of actions and the memory buffer size. With small cookies
the numeric limit is reached before the buffer size limit, which avoids
the code path triggering this bug. When large cookies are used buffer
fills before the numeric limit, and the erroneous code path is hit.
For example after creating 32 csum actions with the cookie
aaaabbbbccccdddd
$ tc actions ls action csum
total acts 26
action order 0: csum (tcp) action continue
index 1 ref 1 bind 0
cookie aaaabbbbccccdddd
.....
action order 25: csum (tcp) action continue
index 26 ref 1 bind 0
cookie aaaabbbbccccdddd
total acts 6
action order 0: csum (tcp) action continue
index 28 ref 1 bind 0
cookie aaaabbbbccccdddd
......
action order 5: csum (tcp) action continue
index 32 ref 1 bind 0
cookie aaaabbbbccccdddd
Note that the action with index 27 is omitted from the report.
Fixes: 4b3550ef53 ("[NET_SCHED]: Use nla_nest_start/nla_nest_end")"
Signed-off-by: Craig Dillabaugh <cdillaba@mojatatu.com>
Acked-by: Jamal Hadi Salim <jhs@mojatatu.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
[ Upstream commit 7880287981 ]
KMSAN reports use of uninitialized memory in the case when |alen| is
smaller than sizeof(struct sockaddr_nl), and therefore |nladdr| isn't
fully copied from the userspace.
Signed-off-by: Alexander Potapenko <glider@google.com>
Fixes: 1da177e4c3 ("Linux-2.6.12-rc2")
Reviewed-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
[ Upstream commit 71a1c91523 ]
At the end of ip6_forward(), IPSTATS_MIB_OUTFORWDATAGRAMS and
IPSTATS_MIB_OUTOCTETS are incremented immediately before the NF_HOOK call
for NFPROTO_IPV6 / NF_INET_FORWARD. As a result, these counters get
incremented regardless of whether or not the netfilter hook allows the
packet to continue being processed. This change increments the counters
in ip6_forward_finish() so that it will not happen if the netfilter hook
chooses to terminate the packet, which is similar to how IPv4 works.
Signed-off-by: Jeff Barnhill <0xeffeff@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
[ Upstream commit b6cdbc8523 ]
Donald reported that IPv6 route leaking between VRFs is not working.
The root cause is the strict argument in the call to rt6_lookup when
validating the nexthop spec.
ip6_route_check_nh validates the gateway and device (if given) of a
route spec. It in turn could call rt6_lookup (e.g., lookup in a given
table did not succeed so it falls back to a full lookup) and if so
sets the strict argument to 1. That means if the egress device is given,
the route lookup needs to return a result with the same device. This
strict requirement does not work with VRFs (IPv4 or IPv6) because the
oif in the flow struct is overridden with the index of the VRF device
to trigger a match on the l3mdev rule and force the lookup to its table.
The right long term solution is to add an l3mdev index to the flow
struct such that the oif is not overridden. That solution will not
backport well, so this patch aims for a simpler solution to relax the
strict argument if the route spec device is an l3mdev slave. As done
in other places, use the FLOWI_FLAG_SKIP_NH_OIF to know that the
RT6_LOOKUP_F_IFACE flag needs to be removed.
Fixes: ca254490c8 ("net: Add VRF support to IPv6 stack")
Reported-by: Donald Sharp <sharpd@cumulusnetworks.com>
Signed-off-by: David Ahern <dsahern@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
[ Upstream commit 2d2d99ec13 ]
Description:
Crash was reported with syzkaller pointing to lan78xx_write_reg routine.
Root-cause:
Proper cleanup of workqueues and init/setup routines was not happening
in failure conditions.
Fix:
Handled the error conditions by cleaning up the queues and init/setup
routines.
Fixes: 55d7de9de6 ("Microchip's LAN7800 family USB 2/3 to 10/100/1000 Ethernet device driver")
Reported-by: Andrey Konovalov <andreyknvl@google.com>
Signed-off-by: Raghuram Chary J <raghuramchary.jallipalli@microchip.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
[ Upstream commit 58b35f2768 ]
arp_filter performs an ip_route_output search for arp source address and
checks if output device is the same where the arp request was received,
if it is not, the arp request is not answered.
This route lookup is always done on main route table so l3slave devices
never find the proper route and arp is not answered.
Passing l3mdev_master_ifindex_rcu(dev) return value as oif fixes the
lookup for l3slave devices while maintaining same behavior for non
l3slave devices as this function returns 0 in that case.
Fixes: 613d09b30f ("net: Use VRF device index for lookups on TX")
Signed-off-by: Miguel Fadon Perlines <mfadon@teldat.com>
Acked-by: David Ahern <dsa@cumulusnetworks.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
commit bb8c13d61a upstream.
Emanuel reported an issue with a hang during microcode update because my
dumb idea to use one atomic synchronization variable for both rendezvous
- before and after update - was simply bollocks:
microcode: microcode_reload_late: late_cpus: 4
microcode: __reload_late: cpu 2 entered
microcode: __reload_late: cpu 1 entered
microcode: __reload_late: cpu 3 entered
microcode: __reload_late: cpu 0 entered
microcode: __reload_late: cpu 1 left
microcode: Timeout while waiting for CPUs rendezvous, remaining: 1
CPU1 above would finish, leave and the others will still spin waiting for
it to join.
So do two synchronization atomics instead, which makes the code a lot more
straightforward.
Also, since the update is serialized and it also takes quite some time per
microcode engine, increase the exit timeout by the number of CPUs on the
system.
That's ok because the moment all CPUs are done, that timeout will be cut
short.
Furthermore, panic when some of the CPUs timeout when returning from a
microcode update: we can't allow a system with not all cores updated.
Also, as an optimization, do not do the exit sync if microcode wasn't
updated.
Reported-by: Emanuel Czirai <xftroxgpx@protonmail.com>
Signed-off-by: Borislav Petkov <bp@suse.de>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Tested-by: Emanuel Czirai <xftroxgpx@protonmail.com>
Tested-by: Ashok Raj <ashok.raj@intel.com>
Tested-by: Tom Lendacky <thomas.lendacky@amd.com>
Link: https://lkml.kernel.org/r/20180314183615.17629-2-bp@alien8.de
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
commit a5321aec64 upstream.
Original idea by Ashok, completely rewritten by Borislav.
Before you read any further: the early loading method is still the
preferred one and you should always do that. The following patch is
improving the late loading mechanism for long running jobs and cloud use
cases.
Gather all cores and serialize the microcode update on them by doing it
one-by-one to make the late update process as reliable as possible and
avoid potential issues caused by the microcode update.
[ Borislav: Rewrite completely. ]
Co-developed-by: Borislav Petkov <bp@suse.de>
Signed-off-by: Ashok Raj <ashok.raj@intel.com>
Signed-off-by: Borislav Petkov <bp@suse.de>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Tested-by: Tom Lendacky <thomas.lendacky@amd.com>
Tested-by: Ashok Raj <ashok.raj@intel.com>
Reviewed-by: Tom Lendacky <thomas.lendacky@amd.com>
Cc: Arjan Van De Ven <arjan.van.de.ven@intel.com>
Link: https://lkml.kernel.org/r/20180228102846.13447-8-bp@alien8.de
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
commit 019cd46984 upstream.
Most crypto drivers involving kernel mode NEON take care to put the code
that actually touches the NEON register file in a separate compilation
unit, to prevent the compiler from reordering code that preserves or
restores the NEON context with code that may corrupt it. This is
necessary because we currently have no way to express the restrictions
imposed upon use of the NEON in kernel mode in a way that the compiler
understands.
However, in the case of aes-ce-cipher, it did not seem unreasonable to
deviate from this rule, given how it does not seem possible for the
compiler to reorder cross object function calls with asm blocks whose
in- and output constraints reflect that it reads from and writes to
memory.
Now that LTO is being proposed for the arm64 kernel, it is time to
revisit this. The link time optimization may replace the function
calls to kernel_neon_begin() and kernel_neon_end() with instantiations
of the IR that make up its implementation, allowing further reordering
with the asm block.
So let's clean this up, and move the asm() blocks into a separate .S
file.
Signed-off-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Reviewed-By: Nick Desaulniers <ndesaulniers@google.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Cc: Matthias Kaehlcke <mka@google.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
commit 3c1f05835c upstream.
Since the ORC unwinder was made the default on x86_64, Clang-built
defconfig kernels have triggered some new objtool warnings:
drivers/gpu/drm/i915/i915_gpu_error.o: warning: objtool: i915_error_printf()+0x6c: return with modified stack frame
drivers/gpu/drm/i915/intel_display.o: warning: objtool: pipe_config_err()+0xa6: return with modified stack frame
The problem is that objtool has never seen clang-built binaries before.
Shockingly enough, objtool is apparently able to follow the code flow
mostly fine, except for one instruction sequence. Instead of a LEAVE
instruction, clang restores RSP and RBP the long way:
67c: 48 89 ec mov %rbp,%rsp
67f: 5d pop %rbp
Teach objtool about this new code sequence.
Reported-and-test-by: Matthias Kaehlcke <mka@chromium.org>
Signed-off-by: Josh Poimboeuf <jpoimboe@redhat.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthias Kaehlcke <mka@chromium.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Link: http://lkml.kernel.org/r/fce88ce81c356eedcae7f00ed349cfaddb3363cc.1521741586.git.jpoimboe@redhat.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
[ Upstream commit 810b8153c4 ]
If we cannot setup a cmd because we run out of ring space
or global pages release the blocks before sleeping. This
prevents a deadlock where dev0 has waiting_blocks set and
needs N blocks, but dev1 to devX have each allocated N / X blocks
and also hit the global block limit so they went to sleep.
find_free_blocks is not able to take the sleeping dev's
blocks becaause their waiting_blocks is set and even
if it was not the block returned by find_last_bit could equal
dbi_max. The latter will probably never happen because
DATA_BLOCK_BITS is so high but in the next patches
DATA_BLOCK_BITS and TCMU_GLOBAL_MAX_BLOCKS will be settable so
it might be lower and could happen.
Signed-off-by: Mike Christie <mchristi@redhat.com>
Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
Signed-off-by: Sasha Levin <alexander.levin@microsoft.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
[ Upstream commit 148b974dee ]
While testing other changes, I discovered that gcc-7.2.1 produces badly
optimized code for aes_encrypt/aes_decrypt. This is especially true when
CONFIG_UBSAN_SANITIZE_ALL is enabled, where it leads to extremely
large stack usage that in turn might cause kernel stack overflows:
crypto/aes_generic.c: In function 'aes_encrypt':
crypto/aes_generic.c:1371:1: warning: the frame size of 4880 bytes is larger than 2048 bytes [-Wframe-larger-than=]
crypto/aes_generic.c: In function 'aes_decrypt':
crypto/aes_generic.c:1441:1: warning: the frame size of 4864 bytes is larger than 2048 bytes [-Wframe-larger-than=]
I verified that this problem exists on all architectures that are
supported by gcc-7.2, though arm64 in particular is less affected than
the others. I also found that gcc-7.1 and gcc-8 do not show the extreme
stack usage but still produce worse code than earlier versions for this
file, apparently because of optimization passes that generally provide
a substantial improvement in object code quality but understandably fail
to find any shortcuts in the AES algorithm.
Possible workarounds include
a) disabling -ftree-pre and -ftree-sra optimizations, this was an earlier
patch I tried, which reliably fixed the stack usage, but caused a
serious performance regression in some versions, as later testing
found.
b) disabling UBSAN on this file or all ciphers, as suggested by Ard
Biesheuvel. This would lead to massively better crypto performance in
UBSAN-enabled kernels and avoid the stack usage, but there is a concern
over whether we should exclude arbitrary files from UBSAN at all.
c) Forcing the optimization level in a different way. Similar to a),
but rather than deselecting specific optimization stages,
this now uses "gcc -Os" for this file, regardless of the
CONFIG_CC_OPTIMIZE_FOR_PERFORMANCE/SIZE option. This is a reliable
workaround for the stack consumption on all architecture, and I've
retested the performance results now on x86, cycles/byte (lower is
better) for cbc(aes-generic) with 256 bit keys:
-O2 -Os
gcc-6.3.1 14.9 15.1
gcc-7.0.1 14.7 15.3
gcc-7.1.1 15.3 14.7
gcc-7.2.1 16.8 15.9
gcc-8.0.0 15.5 15.6
This implements the option c) by enabling forcing -Os on all compiler
versions starting with gcc-7.1. As a workaround for PR83356, it would
only be needed for gcc-7.2+ with UBSAN enabled, but since it also shows
better performance on gcc-7.1 without UBSAN, it seems appropriate to
use the faster version here as well.
Side note: during testing, I also played with the AES code in libressl,
which had a similar performance regression from gcc-6 to gcc-7.2,
but was three times slower overall. It might be interesting to
investigate that further and possibly port the Linux implementation
into that.
Link: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=83356
Link: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=83651
Cc: Richard Biener <rguenther@suse.de>
Cc: Jakub Jelinek <jakub@gcc.gnu.org>
Cc: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Signed-off-by: Arnd Bergmann <arnd@arndb.de>
Acked-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: Sasha Levin <alexander.levin@microsoft.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
