For some reason default_device_ops kept two exit method:
1) default_device_exit() is called for each netns being dismantled in
a cleanup_net() round. This acquires rtnl for each invocation.
2) default_device_exit_batch() is called once with the list of all netns
int the batch, allowing for a single rtnl invocation.
Get rid of the .exit() method to handle the logic from
default_device_exit_batch(), to decrease the number of rtnl acquisition
to one.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: David Ahern <dsahern@kernel.org>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
cleanup_net() is competing with other rtnl users.
Batching bond_net_exit() factorizes all rtnl acquistions
to a single one, giving chance for cleanup_net()
to progress much faster, holding rtnl a bit longer.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Jay Vosburgh <j.vosburgh@gmail.com>
Cc: Veaceslav Falico <vfalico@gmail.com>
Cc: Andy Gospodarek <andy@greyhouse.net>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
cleanup_net() is competing with other rtnl users.
Avoiding to acquire rtnl for each netns before calling
cgw_remove_all_jobs() gives chance for cleanup_net()
to progress much faster, holding rtnl a bit longer.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Acked-by: Oliver Hartkopp <socketcan@hartkopp.net>
Acked-by: Marc Kleine-Budde <mkl@pengutronix.de>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
cleanup_net() is competing with other rtnl users.
Avoiding to acquire rtnl for each netns before calling
ipmr_rules_exit() gives chance for cleanup_net()
to progress much faster, holding rtnl a bit longer.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: David Ahern <dsahern@kernel.org>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
cleanup_net() is competing with other rtnl users.
Avoiding to acquire rtnl for each netns before calling
ip6mr_rules_exit() gives chance for cleanup_net()
to progress much faster, holding rtnl a bit longer.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: David Ahern <dsahern@kernel.org>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
cleanup_net() is competing with other rtnl users.
fib6_rules_net_exit() seems a good candidate for exit_batch(),
as this gives chance for cleanup_net() to progress much faster,
holding rtnl a bit longer.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: David Ahern <dsahern@kernel.org>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
cleanup_net() is competing with other rtnl users.
Instead of acquiring rtnl at each fib_net_exit() invocation,
add fib_net_exit_batch() so that rtnl is acquired once.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: David Ahern <dsahern@kernel.org>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
cleanup_net() is competing with other rtnl users.
nexthop_net_exit() seems a good candidate for exit_batch(),
as this gives chance for cleanup_net() to progress much faster,
holding rtnl a bit longer.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: David Ahern <dsahern@kernel.org>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
IPv6 does not scale very well with the number of IPv6 addresses.
It uses a global (shared by all netns) hash table with 256 buckets.
Some functions like addrconf_verify_rtnl() and addrconf_ifdown()
have to iterate all addresses in the hash table.
I have seen addrconf_verify_rtnl() holding the cpu for 10ms or more.
Switch to the per netns hashtable (and spinlock) added
in prior patches.
This considerably speeds up netns dismantle times on hosts
with thousands of netns. This also has an impact
on regular (fast path) IPv6 processing.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: David Ahern <dsahern@kernel.org>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Next step for using per netns inet6_addr_lst
is to have per netns work item to ultimately
call addrconf_verify_rtnl() and addrconf_verify()
with a new 'struct net*' argument.
Everything is still using the global inet6_addr_lst[] table.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: David Ahern <dsahern@kernel.org>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Add a per netns hash table and a dedicated spinlock,
first step to get rid of the global inet6_addr_lst[] one.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: David Ahern <dsahern@kernel.org>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Song Liu says:
====================
Fix build errors reported by kernel test robot.
====================
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Two subtests in ksyms_module.c are not qualified as static, so these
subtests are exported as standalone tests in tests.h and lead to
confusion for the output of "./test_progs -t ksyms_module".
By using the following command ...
grep "^void \(serial_\)\?test_[a-zA-Z0-9_]\+(\(void\)\?)" \
tools/testing/selftests/bpf/prog_tests/*.c | \
awk -F : '{print $1}' | sort | uniq -c | awk '$1 != 1'
... one finds out that other tests also have a similar problem, so
fix these tests by marking subtests in these tests as static.
Signed-off-by: Hou Tao <houtao1@huawei.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Yonghong Song <yhs@fb.com>
Link: https://lore.kernel.org/bpf/20220208065444.648778-1-houtao1@huawei.com
In some cases, pages cannot be reused by i40e because the page is busy. Add
a counter for this event.
Busy page count is accessible via ethtool.
Signed-off-by: Joe Damato <jdamato@fastly.com>
Tested-by: Dave Switzer <david.switzer@intel.com>
Signed-off-by: Tony Nguyen <anthony.l.nguyen@intel.com>
In some cases, pages can not be reused because they are not associated with
the correct NUMA zone. Knowing how often pages are waived helps users to
understand the interaction between the driver's memory usage and their
system.
Pass rx_stats through to i40e_can_reuse_rx_page to allow tracking when
pages are waived.
The page waive count is accessible via ethtool.
Signed-off-by: Joe Damato <jdamato@fastly.com>
Tested-by: Dave Switzer <david.switzer@intel.com>
Signed-off-by: Tony Nguyen <anthony.l.nguyen@intel.com>
rx page reuse was already being tracked by the i40e driver per RX ring.
Aggregate the counts and make them accessible via ethtool.
Signed-off-by: Joe Damato <jdamato@fastly.com>
Tested-by: Dave Switzer <david.switzer@intel.com>
Signed-off-by: Tony Nguyen <anthony.l.nguyen@intel.com>
Page reuse was being tracked from two locations:
- i40e_reuse_rx_page (via 40e_clean_rx_irq), and
- i40e_alloc_mapped_page
Remove the double count and only count reuse from i40e_alloc_mapped_page
when the page is about to be reused.
Signed-off-by: Joe Damato <jdamato@fastly.com>
Tested-by: Dave Switzer <david.switzer@intel.com>
Signed-off-by: Tony Nguyen <anthony.l.nguyen@intel.com>
Guillaume Nault says:
====================
inet: Separate DSCP from ECN bits using new dscp_t type
The networking stack currently doesn't clearly distinguish between DSCP
and ECN bits. The entire DSCP+ECN bits are stored in u8 variables (or
structure fields), and each part of the stack handles them in their own
way, using different macros. This has created several bugs in the past
and some uncommon code paths are still unfixed.
Such bugs generally manifest by selecting invalid routes because of ECN
bits interfering with FIB routes and rules lookups (more details in the
LPC 2021 talk[1] and in the RFC of this series[2]).
This patch series aims at preventing the introduction of such bugs (and
detecting existing ones), by introducing a dscp_t type, representing
"sanitised" DSCP values (that is, with no ECN information), as opposed
to plain u8 values that contain both DSCP and ECN information. dscp_t
makes it clear for the reader what we're working on, and Sparse can
flag invalid interactions between dscp_t and plain u8.
This series converts only a few variables and structures:
* Patch 1 converts the tclass field of struct fib6_rule. It
effectively forbids the use of ECN bits in the tos/dsfield option
of ip -6 rule. Rules now match packets solely based on their DSCP
bits, so ECN doesn't influence the result any more. This contrasts
with the previous behaviour where all 8 bits of the Traffic Class
field were used. It is believed that this change is acceptable as
matching ECN bits wasn't usable for IPv4, so only IPv6-only
deployments could be depending on it. Also the previous behaviour
made DSCP-based ip6-rules fail for packets with both a DSCP and an
ECN mark, which is another reason why any such deploy is unlikely.
* Patch 2 converts the tos field of struct fib4_rule. This one too
effectively forbids defining ECN bits, this time in ip -4 rule.
Before that, setting ECN bit 1 was accepted, while ECN bit 0 was
rejected. But even when accepted, the rule would never match, as
the packets would have their ECN bits cleared before doing the
rule lookup.
* Patch 3 converts the fc_tos field of struct fib_config. This is
equivalent to patch 2, but for IPv4 routes. Routes using a
tos/dsfield option with any ECN bit set is now rejected. Before
this patch, they were accepted but, as with ip4 rules, these routes
couldn't match any packet, since their ECN bits are cleared before
the lookup.
* Patch 4 converts the fa_tos field of struct fib_alias. This one is
pure internal u8 to dscp_t conversion. While patches 1-3 had user
facing consequences, this patch shouldn't have any side effect and
is there to give an overview of what future conversion patches will
look like. Conversions are quite mechanical, but imply some code
churn, which is the price for the extra clarity a possibility of
type checking.
To summarise, all the behaviour changes required for the dscp_t type
approach to work should be contained in patches 1-3. These changes are
edge cases of ip-route and ip-rule that don't currently work properly.
So they should be safe. Also, a kernel selftest is added for each of
them.
Finally, this work also paves the way for allowing the usage of the 3
high order DSCP bits in IPv4 (a few call paths already handle them, but
in general the stack clears them before IPv4 rule and route lookups).
References:
[1] LPC 2021 talk:
- https://linuxplumbersconf.org/event/11/contributions/943/
- Direct link to slide deck:
https://linuxplumbersconf.org/event/11/contributions/943/attachments/901/1780/inet_tos_lpc2021.pdf
[2] RFC version of this series:
- https://lore.kernel.org/netdev/cover.1638814614.git.gnault@redhat.com/
====================
Link: https://lore.kernel.org/r/cover.1643981839.git.gnault@redhat.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Use the new dscp_t type to replace the fa_tos field of fib_alias. This
ensures ECN bits are ignored and makes the field compatible with the
fc_dscp field of struct fib_config.
Converting old *tos variables and fields to dscp_t allows sparse to
flag incorrect uses of DSCP and ECN bits. This patch is entirely about
type annotation and shouldn't change any existing behaviour.
Signed-off-by: Guillaume Nault <gnault@redhat.com>
Acked-by: David Ahern <dsahern@kernel.org>
Reviewed-by: Toke Høiland-Jørgensen <toke@redhat.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Use the new dscp_t type to replace the fc_tos field of fib_config, to
ensure IPv4 routes aren't influenced by ECN bits when configured with
non-zero rtm_tos.
Before this patch, IPv4 routes specifying an rtm_tos with some of the
ECN bits set were accepted. However they wouldn't work (never match) as
IPv4 normally clears the ECN bits with IPTOS_RT_MASK before doing a FIB
lookup (although a few buggy code paths don't).
After this patch, IPv4 routes specifying an rtm_tos with any ECN bit
set is rejected.
Note: IPv6 routes ignore rtm_tos altogether, any rtm_tos is accepted,
but treated as if it were 0.
Signed-off-by: Guillaume Nault <gnault@redhat.com>
Acked-by: David Ahern <dsahern@kernel.org>
Reviewed-by: Toke Høiland-Jørgensen <toke@redhat.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Use the new dscp_t type to replace the tos field of struct fib4_rule,
so that fib4-rules consistently ignore ECN bits.
Before this patch, fib4-rules did accept rules with the high order ECN
bit set (but not the low order one). Also, it relied on its callers
masking the ECN bits of ->flowi4_tos to prevent those from influencing
the result. This was brittle and a few call paths still do the lookup
without masking the ECN bits first.
After this patch fib4-rules only compare the DSCP bits. ECN can't
influence the result anymore, even if the caller didn't mask these
bits. Also, fib4-rules now must have both ECN bits cleared or they will
be rejected.
Signed-off-by: Guillaume Nault <gnault@redhat.com>
Acked-by: David Ahern <dsahern@kernel.org>
Reviewed-by: Toke Høiland-Jørgensen <toke@redhat.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Define a dscp_t type and its appropriate helpers that ensure ECN bits
are not taken into account when handling DSCP.
Use this new type to replace the tclass field of struct fib6_rule, so
that fib6-rules don't get influenced by ECN bits anymore.
Before this patch, fib6-rules didn't make any distinction between the
DSCP and ECN bits. Therefore, rules specifying a DSCP (tos or dsfield
options in iproute2) stopped working as soon a packets had at least one
of its ECN bits set (as a work around one could create four rules for
each DSCP value to match, one for each possible ECN value).
After this patch fib6-rules only compare the DSCP bits. ECN doesn't
influence the result anymore. Also, fib6-rules now must have the ECN
bits cleared or they will be rejected.
Signed-off-by: Guillaume Nault <gnault@redhat.com>
Acked-by: David Ahern <dsahern@kernel.org>
Reviewed-by: Toke Høiland-Jørgensen <toke@redhat.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Reading the PTP clock is a simple operation requiring only 3 register
reads. Under a PREEMPT_RT kernel, protecting those reads by a spin_lock is
counter-productive: if the 2nd task preempting the 1st has a higher prio
but needs to read time as well, it will require 2 context switches, which
will pretty much always be more costly than just disabling preemption for
the duration of the reads. Moreover, with the code logic recently added
to get_systime(), disabling preemption is not even required anymore:
reads and writes just need to be protected from each other, to prevent a
clock read while the clock is being updated.
Improve the above situation by replacing the PTP spinlock by a rwlock, and
using read_lock for PTP clock reads so simultaneous reads do not block
each other.
Signed-off-by: Yannick Vignon <yannick.vignon@nxp.com>
Link: https://lore.kernel.org/r/20220204135545.2770625-1-yannick.vignon@oss.nxp.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
This place also uses signed min_t and passes this singed int to
copy_to_user (which accepts unsigned argument). I don't think
there is an issue, but let's be consistent.
Fixes: 7855e0db15 ("bpf: test_run: add xdp_shared_info pointer in bpf_test_finish signature")
Signed-off-by: Stanislav Fomichev <sdf@google.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Acked-by: Yonghong Song <yhs@fb.com>
Link: https://lore.kernel.org/bpf/20220204235849.14658-2-sdf@google.com
Song Liu says:
====================
Changes v8 => v9:
1. Fix an error with multi function program, in 4/9.
Changes v7 => v8:
1. Rebase and fix conflicts.
2. Lock text_mutex for text_poke_copy. (Daniel)
Changes v6 => v7:
1. Redesign the interface between generic and arch logic, based on feedback
from Alexei and Ilya.
2. Split 6/7 of v6 to 7/9 and 8/9 in v7, for cleaner logic.
3. Add bpf_arch_text_copy in 6/9.
Changes v5 => v6:
1. Make jit_hole_buffer 128 byte long. Only fill the first and last 128
bytes of header with INT3. (Alexei)
2. Use kvmalloc for temporary buffer. (Alexei)
3. Rename tmp_header/tmp_image => rw_header/rw_image. Remove tmp_image from
x64_jit_data. (Alexei)
4. Change fall back round_up_to in bpf_jit_binary_alloc_pack() from
BPF_PROG_MAX_PACK_PROG_SIZE to PAGE_SIZE.
Changes v4 => v5:
1. Do not use atomic64 for bpf_jit_current. (Alexei)
Changes v3 => v4:
1. Rename text_poke_jit() => text_poke_copy(). (Peter)
2. Change comment style. (Peter)
Changes v2 => v3:
1. Fix tailcall.
Changes v1 => v2:
1. Use text_poke instead of writing through linear mapping. (Peter)
2. Avoid making changes to non-x86_64 code.
Most BPF programs are small, but they consume a page each. For systems
with busy traffic and many BPF programs, this could also add significant
pressure to instruction TLB. High iTLB pressure usually causes slow down
for the whole system, which includes visible performance degradation for
production workloads.
This set tries to solve this problem with customized allocator that pack
multiple programs into a huge page.
Patches 1-6 prepare the work. Patch 7 contains key logic of bpf_prog_pack
allocator. Patch 8 contains bpf_jit_binary_pack_alloc logic on top of
bpf_prog_pack allocator. Patch 9 uses this allocator in x86_64 jit.
====================
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Use bpf_jit_binary_pack_alloc in x86_64 jit. The jit engine first writes
the program to the rw buffer. When the jit is done, the program is copied
to the final location with bpf_jit_binary_pack_finalize.
Note that we need to do bpf_tail_call_direct_fixup after finalize.
Therefore, the text_live = false logic in __bpf_arch_text_poke is no
longer needed.
Signed-off-by: Song Liu <songliubraving@fb.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Link: https://lore.kernel.org/bpf/20220204185742.271030-10-song@kernel.org
This is the jit binary allocator built on top of bpf_prog_pack.
bpf_prog_pack allocates RO memory, which cannot be used directly by the
JIT engine. Therefore, a temporary rw buffer is allocated for the JIT
engine. Once JIT is done, bpf_jit_binary_pack_finalize is used to copy
the program to the RO memory.
bpf_jit_binary_pack_alloc reserves 16 bytes of extra space for illegal
instructions, which is small than the 128 bytes space reserved by
bpf_jit_binary_alloc. This change is necessary for bpf_jit_binary_hdr
to find the correct header. Also, flag use_bpf_prog_pack is added to
differentiate a program allocated by bpf_jit_binary_pack_alloc.
Signed-off-by: Song Liu <songliubraving@fb.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Link: https://lore.kernel.org/bpf/20220204185742.271030-9-song@kernel.org
Most BPF programs are small, but they consume a page each. For systems
with busy traffic and many BPF programs, this could add significant
pressure to instruction TLB. High iTLB pressure usually causes slow down
for the whole system, which includes visible performance degradation for
production workloads.
Introduce bpf_prog_pack allocator to pack multiple BPF programs in a huge
page. The memory is then allocated in 64 byte chunks.
Memory allocated by bpf_prog_pack allocator is RO protected after initial
allocation. To write to it, the user (jit engine) need to use text poke
API.
Signed-off-by: Song Liu <song@kernel.org>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Link: https://lore.kernel.org/bpf/20220204185742.271030-8-song@kernel.org
This enables module_alloc() to allocate huge page for 2MB+ requests.
To check the difference of this change, we need enable config
CONFIG_PTDUMP_DEBUGFS, and call module_alloc(2MB). Before the change,
/sys/kernel/debug/page_tables/kernel shows pte for this map. With the
change, /sys/kernel/debug/page_tables/ show pmd for thie map.
Signed-off-by: Song Liu <songliubraving@fb.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Link: https://lore.kernel.org/bpf/20220204185742.271030-2-song@kernel.org
On changing the RX ring parameters igb uses a hack to avoid a warning
when calling xdp_rxq_info_reg via igb_setup_rx_resources. It just
clears the struct xdp_rxq_info content.
Instead, change this to unregister if we're already registered. Align
code to the igc code.
Fixes: 9cbc948b5a ("igb: add XDP support")
Signed-off-by: Corinna Vinschen <vinschen@redhat.com>
Acked-by: Vinicius Costa Gomes <vinicius.gomes@intel.com>
Tested-by: Sandeep Penigalapati <sandeep.penigalapati@intel.com>
Signed-off-by: Tony Nguyen <anthony.l.nguyen@intel.com>
