IF YOU WOULD LIKE TO GET AN ACCOUNT, please write an
email to Administrator. User accounts are meant only to access repo
and report issues and/or generate pull requests.
This is a purpose-specific Git hosting for
BaseALT
projects. Thank you for your understanding!
Только зарегистрированные пользователи имеют доступ к сервису!
Для получения аккаунта, обратитесь к администратору.
In the previous GC implementation, the shape of the inflight socket
graph was not expected to change while GC was in progress.
MSG_PEEK was tricky because it could install inflight fd silently
and transform the graph.
Let's say we peeked a fd, which was a listening socket, and accept()ed
some embryo sockets from it. The garbage collection algorithm would
have been confused because the set of sockets visited in scan_inflight()
would change within the same GC invocation.
That's why we placed spin_lock(&unix_gc_lock) and spin_unlock() in
unix_peek_fds() with a fat comment.
In the new GC implementation, we no longer garbage-collect the socket
if it exists in another queue, that is, if it has a bridge to another
SCC. Also, accept() will require the lock if it has edges.
Thus, we need not do the complicated lock dance.
Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Link: https://lore.kernel.org/r/20240401173125.92184-3-kuniyu@amazon.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
When we passed fds, we used to bump each file's refcount twice
in scm_fp_copy() and scm_fp_dup() before linking the socket to
gc_inflight_list.
This is because we incremented the inflight count of the socket
and linked it to the list in advance before passing skb to the
destination socket.
Otherwise, the inflight socket could have been garbage-collected
in a small race window between linking the socket to the list and
queuing skb:
CPU 1 : sendmsg(X) w/ A's fd CPU 2 : close(A)
----- -----
/* Here A's refcount is 1, and inflight count is 0 */
bump A's refcount to 2 in scm_fp_copy()
bump A's inflight count to 1
link A to gc_inflight_list
decrement A's refcount to 1
/* A's refcount == inflight count, thus A could be GC candidate */
start GC
mark A as candidate
purge A's receive queue
queue skb w/ A's fd to X
/* A is queued, but all data has been lost */
After commit 4090fa373f0e ("af_unix: Replace garbage collection
algorithm."), we increment the inflight count and link the socket
to the global list only when queuing the skb.
The race no longer exists, so let's not clone the fd nor bump
the count in unix_attach_fds().
Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Link: https://lore.kernel.org/r/20240401173125.92184-2-kuniyu@amazon.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Prior to this patch, what we can see by enabling trace_tcp_send is
only happening under two circumstances:
1) active rst mode
2) non-active rst mode and based on the full socket
That means the inconsistency occurs if we use tcpdump and trace
simultaneously to see how rst happens.
It's necessary that we should take into other cases into considerations,
say:
1) time-wait socket
2) no socket
...
By parsing the incoming skb and reversing its 4-tuple can
we know the exact 'flow' which might not exist.
Samples after applied this patch:
1. tcp_send_reset: skbaddr=XXX skaddr=XXX src=ip:port dest=ip:port
state=TCP_ESTABLISHED
2. tcp_send_reset: skbaddr=000...000 skaddr=XXX src=ip:port dest=ip:port
state=UNKNOWN
Note:
1) UNKNOWN means we cannot extract the right information from skb.
2) skbaddr/skaddr could be 0
Signed-off-by: Jason Xing <kernelxing@tencent.com>
Link: https://lore.kernel.org/r/20240401073605.37335-3-kerneljasonxing@gmail.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
For systems that use CPU isolation (via nohz_full), creating or destroying
a socket with SO_TIMESTAMP, SO_TIMESTAMPNS or SO_TIMESTAMPING with flag
SOF_TIMESTAMPING_RX_SOFTWARE will cause a static key to be enabled/disabled.
This in turn causes undesired IPIs to isolated CPUs.
So enable the static key unconditionally, if CPU isolation is enabled,
thus avoiding the IPIs.
Signed-off-by: Marcelo Tosatti <mtosatti@redhat.com>
Reviewed-by: Willem de Bruijn <willemb@google.com>
Link: https://lore.kernel.org/r/ZgrUiLLtbEUf9SFn@tpad
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
-Wflex-array-member-not-at-end is coming in GCC-14, and we are getting
ready to enable it globally.
There are currently a couple of objects in `struct smc_clc_msg_proposal_area`
that contain a couple of flexible structures:
struct smc_clc_msg_proposal_area {
...
struct smc_clc_v2_extension pclc_v2_ext;
...
struct smc_clc_smcd_v2_extension pclc_smcd_v2_ext;
...
};
So, in order to avoid ending up with a couple of flexible-array members
in the middle of a struct, we use the `struct_group_tagged()` helper to
separate the flexible array from the rest of the members in the flexible
structure:
struct smc_clc_smcd_v2_extension {
struct_group_tagged(smc_clc_smcd_v2_extension_fixed, fixed,
u8 system_eid[SMC_MAX_EID_LEN];
u8 reserved[16];
);
struct smc_clc_smcd_gid_chid gidchid[];
};
With the change described above, we now declare objects of the type of
the tagged struct without embedding flexible arrays in the middle of
another struct:
struct smc_clc_msg_proposal_area {
...
struct smc_clc_v2_extension_fixed pclc_v2_ext;
...
struct smc_clc_smcd_v2_extension_fixed pclc_smcd_v2_ext;
...
};
We also use `container_of()` when we need to retrieve a pointer to the
flexible structures.
So, with these changes, fix the following warnings:
In file included from net/smc/af_smc.c:42:
net/smc/smc_clc.h:186:49: warning: structure containing a flexible array member is not at the end of another structure [-Wflex-array-member-not-at-end]
186 | struct smc_clc_v2_extension pclc_v2_ext;
| ^~~~~~~~~~~
net/smc/smc_clc.h:188:49: warning: structure containing a flexible array member is not at the end of another structure [-Wflex-array-member-not-at-end]
188 | struct smc_clc_smcd_v2_extension pclc_smcd_v2_ext;
| ^~~~~~~~~~~~~~~~
Reviewed-by: Kees Cook <keescook@chromium.org>
Signed-off-by: Gustavo A. R. Silva <gustavoars@kernel.org>
Reviewed-by: Wen Gu <guwen@linux.alibaba.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
UTILITY_NAME_LENGTH is 16. So better use the former when defining the
'utility_name' array. This makes the intent clearer when it is used around
line 260.
While at it, declare variable in reverse xmas tree style.
Signed-off-by: Christophe JAILLET <christophe.jaillet@wanadoo.fr>
Reviewed-by: Simon Horman <horms@kernel.org>
Link: https://lore.kernel.org/r/8c1160501f69b64bb2d45ce9f26f746eec80ac77.1711787352.git.christophe.jaillet@wanadoo.fr
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
For CONFIG_CPUMASK_OFFSTACK=y kernel, explicit allocation of cpumask
variable on stack is not recommended since it can cause potential stack
overflow.
Instead, kernel code should always use *cpumask_var API(s) to allocate
cpumask var in config-neutral way, leaving allocation strategy to
CONFIG_CPUMASK_OFFSTACK.
Use *cpumask_var API(s) to address it.
Signed-off-by: Dawei Li <dawei.li@shingroup.cn>
Reviewed-by: Alexandra Winter <wintera@linux.ibm.com>
Link: https://lore.kernel.org/r/20240331053441.1276826-2-dawei.li@shingroup.cn
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Now that the checks for direct recycling possibility live inside the
Page Pool core, reuse them when performing bulk recycling.
page_pool_put_page_bulk() can be called from process context as well,
page_pool_napi_local() takes care of this at the very beginning.
Under high .ndo_xdp_xmit() traffic load, the win is 2-3% Pps assuming
the sending driver uses xdp_return_frame_bulk() on Tx completion.
Signed-off-by: Alexander Lobakin <aleksander.lobakin@intel.com>
Link: https://lore.kernel.org/r/20240329165507.3240110-3-aleksander.lobakin@intel.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Since we have pool->p.napi (Jakub) and pool->cpuid (Lorenzo) to check
whether it's safe to use direct recycling, we can use both globally for
each page instead of relying solely on @allow_direct argument.
Let's assume that @allow_direct means "I'm sure it's local, don't waste
time rechecking this" and when it's false, try the mentioned params to
still recycle the page directly. If neither is true, we'll lose some
CPU cycles, but then it surely won't be hotpath. On the other hand,
paths where it's possible to use direct cache, but not possible to
safely set @allow_direct, will benefit from this move.
The whole propagation of @napi_safe through a dozen of skb freeing
functions can now go away, which saves us some stack space.
Signed-off-by: Alexander Lobakin <aleksander.lobakin@intel.com>
Link: https://lore.kernel.org/r/20240329165507.3240110-2-aleksander.lobakin@intel.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
No longer hold RTNL while calling inet6_dump_fib().
Also change return value for a completed dump,
so that NLMSG_DONE can be appended to current skb,
saving one recvmsg() system call.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: David Ahern <dsahern@kernel.org>
Link: https://lore.kernel.org/r/20240329183053.644630-1-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
genetlink.h is a shell of what used to be a combined uAPI
and kernel header over a decade ago. It has fewer than
10 lines of code. Merge it into net/genetlink.h.
In some ways it'd be better to keep the combined header
under linux/ but it would make looking through git history
harder.
Acked-by: Sven Eckelmann <sven@narfation.org>
Link: https://lore.kernel.org/r/20240329175710.291749-4-kuba@kernel.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
The only legit reason I could think of for net/genetlink.h
and linux/genetlink.h to be separate would be if one was
included by other headers and we wanted to keep it lightweight.
That is not the case, net/openvswitch/meter.h includes
linux/genetlink.h but for no apparent reason (for struct genl_family
perhaps? it's not necessary, types of externs do not need
to be known).
Link: https://lore.kernel.org/r/20240329175710.291749-3-kuba@kernel.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
There are things in linux/genetlink.h which are only used
under net/netlink/. Move them to a new local header.
A new header with just 2 externs isn't great, but alternative
would be to include af_netlink.h in genetlink.c which feels
even worse.
Link: https://lore.kernel.org/r/20240329175710.291749-2-kuba@kernel.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
We can change inet_csk() to propagate its argument const qualifier,
thanks to container_of_const().
We have to fix few places that had mistakes, like tcp_bound_rto().
Signed-off-by: Eric Dumazet <edumazet@google.com>
Link: https://lore.kernel.org/r/20240329144931.295800-1-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
process_backlog() can batch increments of sd->input_queue_head,
saving some memory bandwidth.
Also add READ_ONCE()/WRITE_ONCE() annotations around
sd->input_queue_head accesses.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
input_queue_tail_incr_save() is incrementing the sd queue_tail
and save it in the flow last_qtail.
Two issues here :
- no lock protects the write on last_qtail, we should use appropriate
annotations.
- We can perform this write after releasing the per-cpu backlog lock,
to decrease this lock hold duration (move away the cache line miss)
Also move input_queue_head_incr() and rps helpers to include/net/rps.h,
while adding rps_ prefix to better reflect their role.
v2: Fixed a build issue (Jakub and kernel build bots)
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
We can remove a goto and a label by reversing a condition.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
If under extreme cpu backlog pressure enqueue_to_backlog() has
to drop a packet, it could do this without dirtying a cache line
and potentially slowing down the target cpu.
Move sd->dropped into a separate cache line, and make it atomic.
In non pressure mode, this field is not touched, no need to consume
valuable space in a hot cache line.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
If the device attached to the packet given to enqueue_to_backlog()
is not running, we drop the packet.
But we accidentally increase sd->dropped, giving false signals
to admins: sd->dropped should be reserved to cpu backlog pressure,
not to temporary glitches at device dismantles.
While we are at it, perform the netif_running() test before
we get the rps lock, and use REASON_DEV_READY
drop reason instead of NOT_SPECIFIED.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Move dev_xmit_recursion() and friends to net/core/dev.h
They are only used from net/core/dev.c and net/core/filter.c.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
kick_defer_list_purge() is defined in net/core/dev.c
and used from net/core/skubff.c
Because we need softnet_data, include <linux/netdevice.h>
from net/core/dev.h
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
In PFCP receive path set metadata needed by flower code to do correct
classification based on this metadata.
Signed-off-by: Michal Swiatkowski <michal.swiatkowski@linux.intel.com>
Signed-off-by: Marcin Szycik <marcin.szycik@linux.intel.com>
Reviewed-by: Simon Horman <horms@kernel.org>
Signed-off-by: Alexander Lobakin <aleksander.lobakin@intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Now that there are helpers for converting IP tunnel flags between the
old __be16 format and the bitmap format, make sure they work as expected
by adding a couple of tests to the networking testing suite. The helpers
are all inline, so no dependencies on the related CONFIG_* (or a
standalone module) are needed.
Cover three possible cases:
1. No bits past BIT(15) are set, VTI/SIT bits are not set. This
conversion is almost a direct assignment.
2. No bits past BIT(15) are set, but VTI/SIT bit is set. During the
conversion, it must be transformed into BIT(16) in the bitmap,
but still compatible with the __be16 format.
3. The bitmap has bits past BIT(15) set (not the VTI/SIT one). The
result will be truncated.
Note that currently __IP_TUNNEL_FLAG_NUM is 17 (incl. special),
which means that the result of this case is currently
semi-false-positive. When BIT(17) is finally here, it will be
adjusted accordingly.
Signed-off-by: Alexander Lobakin <aleksander.lobakin@intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Historically, tunnel flags like TUNNEL_CSUM or TUNNEL_ERSPAN_OPT
have been defined as __be16. Now all of those 16 bits are occupied
and there's no more free space for new flags.
It can't be simply switched to a bigger container with no
adjustments to the values, since it's an explicit Endian storage,
and on LE systems (__be16)0x0001 equals to
(__be64)0x0001000000000000.
We could probably define new 64-bit flags depending on the
Endianness, i.e. (__be64)0x0001 on BE and (__be64)0x00010000... on
LE, but that would introduce an Endianness dependency and spawn a
ton of Sparse warnings. To mitigate them, all of those places which
were adjusted with this change would be touched anyway, so why not
define stuff properly if there's no choice.
Define IP_TUNNEL_*_BIT counterparts as a bit number instead of the
value already coded and a fistful of <16 <-> bitmap> converters and
helpers. The two flags which have a different bit position are
SIT_ISATAP_BIT and VTI_ISVTI_BIT, as they were defined not as
__cpu_to_be16(), but as (__force __be16), i.e. had different
positions on LE and BE. Now they both have strongly defined places.
Change all __be16 fields which were used to store those flags, to
IP_TUNNEL_DECLARE_FLAGS() -> DECLARE_BITMAP(__IP_TUNNEL_FLAG_NUM) ->
unsigned long[1] for now, and replace all TUNNEL_* occurrences to
their bitmap counterparts. Use the converters in the places which talk
to the userspace, hardware (NFP) or other hosts (GRE header). The rest
must explicitly use the new flags only. This must be done at once,
otherwise there will be too many conversions throughout the code in
the intermediate commits.
Finally, disable the old __be16 flags for use in the kernel code
(except for the two 'irregular' flags mentioned above), to prevent
any accidental (mis)use of them. For the userspace, nothing is
changed, only additions were made.
Most noticeable bloat-o-meter difference (.text):
vmlinux: 307/-1 (306)
gre.ko: 62/0 (62)
ip_gre.ko: 941/-217 (724) [*]
ip_tunnel.ko: 390/-900 (-510) [**]
ip_vti.ko: 138/0 (138)
ip6_gre.ko: 534/-18 (516) [*]
ip6_tunnel.ko: 118/-10 (108)
[*] gre_flags_to_tnl_flags() grew, but still is inlined
[**] ip_tunnel_find() got uninlined, hence such decrease
The average code size increase in non-extreme case is 100-200 bytes
per module, mostly due to sizeof(long) > sizeof(__be16), as
%__IP_TUNNEL_FLAG_NUM is less than %BITS_PER_LONG and the compilers
are able to expand the majority of bitmap_*() calls here into direct
operations on scalars.
Reviewed-by: Simon Horman <horms@kernel.org>
Signed-off-by: Alexander Lobakin <aleksander.lobakin@intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Unlike IPv6 tunnels which use purely-kernel __ip6_tnl_parm structure
to store params inside the kernel, IPv4 tunnel code uses the same
ip_tunnel_parm which is being used to talk with the userspace.
This makes it difficult to alter or add any fields or use a
different format for whatever data.
Define struct ip_tunnel_parm_kern, a 1:1 copy of ip_tunnel_parm for
now, and use it throughout the code. Define the pieces, where the copy
user <-> kernel happens, as standalone functions, and copy the data
there field-by-field, so that the kernel-side structure could be easily
modified later on and the users wouldn't have to care about this.
Reviewed-by: Simon Horman <horms@kernel.org>
Signed-off-by: Alexander Lobakin <aleksander.lobakin@intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
There are, especially with multi-attr arrays, many cases
of needing to iterate all attributes of a specific type
in a netlink message or a nested attribute. Add specific
macros to support that case.
Also convert many instances using this spatch:
@@
iterator nla_for_each_attr;
iterator name nla_for_each_attr_type;
identifier nla;
expression head, len, rem;
expression ATTR;
type T;
identifier x;
@@
-nla_for_each_attr(nla, head, len, rem)
+nla_for_each_attr_type(nla, ATTR, head, len, rem)
{
<... T x; ...>
-if (nla_type(nla) == ATTR) {
...
-}
}
@@
identifier nla;
iterator nla_for_each_nested;
iterator name nla_for_each_nested_type;
expression attr, rem;
expression ATTR;
type T;
identifier x;
@@
-nla_for_each_nested(nla, attr, rem)
+nla_for_each_nested_type(nla, ATTR, attr, rem)
{
<... T x; ...>
-if (nla_type(nla) == ATTR) {
...
-}
}
@@
iterator nla_for_each_attr;
iterator name nla_for_each_attr_type;
identifier nla;
expression head, len, rem;
expression ATTR;
type T;
identifier x;
@@
-nla_for_each_attr(nla, head, len, rem)
+nla_for_each_attr_type(nla, ATTR, head, len, rem)
{
<... T x; ...>
-if (nla_type(nla) != ATTR) continue;
...
}
@@
identifier nla;
iterator nla_for_each_nested;
iterator name nla_for_each_nested_type;
expression attr, rem;
expression ATTR;
type T;
identifier x;
@@
-nla_for_each_nested(nla, attr, rem)
+nla_for_each_nested_type(nla, ATTR, attr, rem)
{
<... T x; ...>
-if (nla_type(nla) != ATTR) continue;
...
}
Although I had to undo one bad change this made, and
I also adjusted some other code for whitespace and to
use direct variable initialization now.
Signed-off-by: Johannes Berg <johannes.berg@intel.com>
Link: https://lore.kernel.org/r/20240328203144.b5a6c895fb80.I1869b44767379f204998ff44dd239803f39c23e0@changeid
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
While looking at UDP receive performance, I saw sk_wake_async()
was no longer inlined.
This matters at least on AMD Zen1-4 platforms (see SRSO)
This might be because rcu_read_lock() and rcu_read_unlock()
are no longer nops in recent kernels ?
Add sk_wake_async_rcu() variant, which must be called from
contexts already holding rcu lock.
As SOCK_FASYNC is deprecated in modern days, use unlikely()
to give a hint to the compiler.
sk_wake_async_rcu() is properly inlined from
__udp_enqueue_schedule_skb() and sock_def_readable().
Signed-off-by: Eric Dumazet <edumazet@google.com>
Link: https://lore.kernel.org/r/20240328144032.1864988-5-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
sock_def_readable() is quite expensive (particularly
when ep_poll_callback() is in the picture).
We must call sk->sk_data_ready() when :
- receive queue was empty, or
- SO_PEEK_OFF is enabled on the socket, or
- sk->sk_data_ready is not sock_def_readable.
We still need to call sk_wake_async().
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Willem de Bruijn <willemb@google.com>
Acked-by: Paolo Abeni <pabeni@redhat.com>
Link: https://lore.kernel.org/r/20240328144032.1864988-4-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
sk->sk_rcvbuf is read locklessly twice, while other threads
could change its value.
Use a READ_ONCE() to annotate the race.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Link: https://lore.kernel.org/r/20240328144032.1864988-2-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
If we find a dead SCC during iteration, we call unix_collect_skb()
to splice all skb in the SCC to the global sk_buff_head, hitlist.
After iterating all SCC, we unlock unix_gc_lock and purge the queue.
Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Acked-by: Paolo Abeni <pabeni@redhat.com>
Link: https://lore.kernel.org/r/20240325202425.60930-15-kuniyu@amazon.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
When iterating SCC, we call unix_vertex_dead() for each vertex
to check if the vertex is close()d and has no bridge to another
SCC.
If both conditions are true for every vertex in SCC, we can
execute garbage collection for all skb in the SCC.
The actual garbage collection is done in the following patch,
replacing the old implementation.
Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Acked-by: Paolo Abeni <pabeni@redhat.com>
Link: https://lore.kernel.org/r/20240325202425.60930-14-kuniyu@amazon.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
The definition of the lowlink in Tarjan's algorithm is the
smallest index of a vertex that is reachable with at most one
back-edge in SCC. This is not useful for a cross-edge.
If we start traversing from A in the following graph, the final
lowlink of D is 3. The cross-edge here is one between D and C.
A -> B -> D D = (4, 3) (index, lowlink)
^ | | C = (3, 1)
| V | B = (2, 1)
`--- C <--' A = (1, 1)
This is because the lowlink of D is updated with the index of C.
In the following patch, we detect a dead SCC by checking two
conditions for each vertex.
1) vertex has no edge directed to another SCC (no bridge)
2) vertex's out_degree is the same as the refcount of its file
If 1) is false, there is a receiver of all fds of the SCC and
its ancestor SCC.
To evaluate 1), we need to assign a unique index to each SCC and
assign it to all vertices in the SCC.
This patch changes the lowlink update logic for cross-edge so
that in the example above, the lowlink of D is updated with the
lowlink of C.
A -> B -> D D = (4, 1) (index, lowlink)
^ | | C = (3, 1)
| V | B = (2, 1)
`--- C <--' A = (1, 1)
Then, all vertices in the same SCC have the same lowlink, and we
can quickly find the bridge connecting to different SCC if exists.
However, it is no longer called lowlink, so we rename it to
scc_index. (It's sometimes called lowpoint.)
Also, we add a global variable to hold the last index used in DFS
so that we do not reset the initial index in each DFS.
This patch can be squashed to the SCC detection patch but is
split deliberately for anyone wondering why lowlink is not used
as used in the original Tarjan's algorithm and many reference
implementations.
Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Acked-by: Paolo Abeni <pabeni@redhat.com>
Link: https://lore.kernel.org/r/20240325202425.60930-13-kuniyu@amazon.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Once a cyclic reference is formed, we need to run GC to check if
there is dead SCC.
However, we do not need to run Tarjan's algorithm if we know that
the shape of the inflight graph has not been changed.
If an edge is added/updated/deleted and the edge's successor is
inflight, we set false to unix_graph_grouped, which means we need
to re-classify SCC.
Once we finalise SCC, we set true to unix_graph_grouped.
While unix_graph_grouped is true, we can iterate the grouped
SCC using vertex->scc_entry in unix_walk_scc_fast().
list_add() and list_for_each_entry_reverse() uses seem weird, but
they are to keep the vertex order consistent and make writing test
easier.
Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Acked-by: Paolo Abeni <pabeni@redhat.com>
Link: https://lore.kernel.org/r/20240325202425.60930-12-kuniyu@amazon.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
We do not need to run GC if there is no possible cyclic reference.
We use unix_graph_maybe_cyclic to decide if we should run GC.
If a fd of an AF_UNIX socket is passed to an already inflight AF_UNIX
socket, they could form a cyclic reference. Then, we set true to
unix_graph_maybe_cyclic and later run Tarjan's algorithm to group
them into SCC.
Once we run Tarjan's algorithm, we are 100% sure whether cyclic
references exist or not. If there is no cycle, we set false to
unix_graph_maybe_cyclic and can skip the entire garbage collection
next time.
When finalising SCC, we set true to unix_graph_maybe_cyclic if SCC
consists of multiple vertices.
Even if SCC is a single vertex, a cycle might exist as self-fd passing.
Given the corner case is rare, we detect it by checking all edges of
the vertex and set true to unix_graph_maybe_cyclic.
With this change, __unix_gc() is just a spin_lock() dance in the normal
usage.
Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Acked-by: Paolo Abeni <pabeni@redhat.com>
Link: https://lore.kernel.org/r/20240325202425.60930-11-kuniyu@amazon.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Before starting Tarjan's algorithm, we need to mark all vertices
as unvisited. We can save this O(n) setup by reserving two special
indices (0, 1) and using two variables.
The first time we link a vertex to unix_unvisited_vertices, we set
unix_vertex_unvisited_index to index.
During DFS, we can see that the index of unvisited vertices is the
same as unix_vertex_unvisited_index.
When we finalise SCC later, we set unix_vertex_grouped_index to each
vertex's index.
Then, we can know (i) that the vertex is on the stack if the index
of a visited vertex is >= 2 and (ii) that it is not on the stack and
belongs to a different SCC if the index is unix_vertex_grouped_index.
After the whole algorithm, all indices of vertices are set as
unix_vertex_grouped_index.
Next time we start DFS, we know that all unvisited vertices have
unix_vertex_grouped_index, and we can use unix_vertex_unvisited_index
as the not-on-stack marker.
To use the same variable in __unix_walk_scc(), we can swap
unix_vertex_(grouped|unvisited)_index at the end of Tarjan's
algorithm.
Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Acked-by: Paolo Abeni <pabeni@redhat.com>
Link: https://lore.kernel.org/r/20240325202425.60930-10-kuniyu@amazon.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
To garbage collect inflight AF_UNIX sockets, we must define the
cyclic reference appropriately. This is a bit tricky if the loop
consists of embryo sockets.
Suppose that the fd of AF_UNIX socket A is passed to D and the fd B
to C and that C and D are embryo sockets of A and B, respectively.
It may appear that there are two separate graphs, A (-> D) and
B (-> C), but this is not correct.
A --. .-- B
X
C <-' `-> D
Now, D holds A's refcount, and C has B's refcount, so unix_release()
will never be called for A and B when we close() them. However, no
one can call close() for D and C to free skbs holding refcounts of A
and B because C/D is in A/B's receive queue, which should have been
purged by unix_release() for A and B.
So, here's another type of cyclic reference. When a fd of an AF_UNIX
socket is passed to an embryo socket, the reference is indirectly held
by its parent listening socket.
.-> A .-> B
| `- sk_receive_queue | `- sk_receive_queue
| `- skb | `- skb
| `- sk == C | `- sk == D
| `- sk_receive_queue | `- sk_receive_queue
| `- skb +---------' `- skb +-.
| |
`---------------------------------------------------------'
Technically, the graph must be denoted as A <-> B instead of A (-> D)
and B (-> C) to find such a cyclic reference without touching each
socket's receive queue.
.-> A --. .-- B <-.
| X | == A <-> B
`-- C <-' `-> D --'
We apply this fixup during GC by fetching the real successor by
unix_edge_successor().
When we call accept(), we clear unix_sock.listener under unix_gc_lock
not to confuse GC.
Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Acked-by: Paolo Abeni <pabeni@redhat.com>
Link: https://lore.kernel.org/r/20240325202425.60930-9-kuniyu@amazon.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
This is a prep patch for the following change, where we need to
fetch the listening socket from the successor embryo socket
during GC.
We add a new field to struct unix_sock to save a pointer to a
listening socket.
We set it when connect() creates a new socket, and clear it when
accept() is called.
Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Acked-by: Paolo Abeni <pabeni@redhat.com>
Link: https://lore.kernel.org/r/20240325202425.60930-8-kuniyu@amazon.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
In the new GC, we use a simple graph algorithm, Tarjan's Strongly
Connected Components (SCC) algorithm, to find cyclic references.
The algorithm visits every vertex exactly once using depth-first
search (DFS).
DFS starts by pushing an input vertex to a stack and assigning it
a unique number. Two fields, index and lowlink, are initialised
with the number, but lowlink could be updated later during DFS.
If a vertex has an edge to an unvisited inflight vertex, we visit
it and do the same processing. So, we will have vertices in the
stack in the order they appear and number them consecutively in
the same order.
If a vertex has a back-edge to a visited vertex in the stack,
we update the predecessor's lowlink with the successor's index.
After iterating edges from the vertex, we check if its index
equals its lowlink.
If the lowlink is different from the index, it shows there was a
back-edge. Then, we go backtracking and propagate the lowlink to
its predecessor and resume the previous edge iteration from the
next edge.
If the lowlink is the same as the index, we pop vertices before
and including the vertex from the stack. Then, the set of vertices
is SCC, possibly forming a cycle. At the same time, we move the
vertices to unix_visited_vertices.
When we finish the algorithm, all vertices in each SCC will be
linked via unix_vertex.scc_entry.
Let's take an example. We have a graph including five inflight
vertices (F is not inflight):
A -> B -> C -> D -> E (-> F)
^ |
`---------'
Suppose that we start DFS from C. We will visit C, D, and B first
and initialise their index and lowlink. Then, the stack looks like
this:
> B = (3, 3) (index, lowlink)
D = (2, 2)
C = (1, 1)
When checking B's edge to C, we update B's lowlink with C's index
and propagate it to D.
B = (3, 1) (index, lowlink)
> D = (2, 1)
C = (1, 1)
Next, we visit E, which has no edge to an inflight vertex.
> E = (4, 4) (index, lowlink)
B = (3, 1)
D = (2, 1)
C = (1, 1)
When we leave from E, its index and lowlink are the same, so we
pop E from the stack as single-vertex SCC. Next, we leave from
B and D but do nothing because their lowlink are different from
their index.
B = (3, 1) (index, lowlink)
D = (2, 1)
> C = (1, 1)
Then, we leave from C, whose index and lowlink are the same, so
we pop B, D and C as SCC.
Last, we do DFS for the rest of vertices, A, which is also a
single-vertex SCC.
Finally, each unix_vertex.scc_entry is linked as follows:
A -. B -> C -> D E -.
^ | ^ | ^ |
`--' `---------' `--'
We use SCC later to decide whether we can garbage-collect the
sockets.
Note that we still cannot detect SCC properly if an edge points
to an embryo socket. The following two patches will sort it out.
Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Acked-by: Paolo Abeni <pabeni@redhat.com>
Link: https://lore.kernel.org/r/20240325202425.60930-7-kuniyu@amazon.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
The new GC will use a depth first search graph algorithm to find
cyclic references. The algorithm visits every vertex exactly once.
Here, we implement the DFS part without recursion so that no one
can abuse it.
unix_walk_scc() marks every vertex unvisited by initialising index
as UNIX_VERTEX_INDEX_UNVISITED and iterates inflight vertices in
unix_unvisited_vertices and call __unix_walk_scc() to start DFS from
an arbitrary vertex.
__unix_walk_scc() iterates all edges starting from the vertex and
explores the neighbour vertices with DFS using edge_stack.
After visiting all neighbours, __unix_walk_scc() moves the visited
vertex to unix_visited_vertices so that unix_walk_scc() will not
restart DFS from the visited vertex.
Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Acked-by: Paolo Abeni <pabeni@redhat.com>
Link: https://lore.kernel.org/r/20240325202425.60930-6-kuniyu@amazon.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Currently, we track the number of inflight sockets in two variables.
unix_tot_inflight is the total number of inflight AF_UNIX sockets on
the host, and user->unix_inflight is the number of inflight fds per
user.
We update them one by one in unix_inflight(), which can be done once
in batch. Also, sendmsg() could fail even after unix_inflight(), then
we need to acquire unix_gc_lock only to decrement the counters.
Let's bulk update the counters in unix_add_edges() and unix_del_edges(),
which is called only for successfully passed fds.
Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Acked-by: Paolo Abeni <pabeni@redhat.com>
Link: https://lore.kernel.org/r/20240325202425.60930-5-kuniyu@amazon.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
As with the previous patch, we preallocate to skb's scm_fp_list an
array of struct unix_edge in the number of inflight AF_UNIX fds.
There we just preallocate memory and do not use immediately because
sendmsg() could fail after this point. The actual use will be in
the next patch.
When we queue skb with inflight edges, we will set the inflight
socket's unix_sock as unix_edge->predecessor and the receiver's
unix_sock as successor, and then we will link the edge to the
inflight socket's unix_vertex.edges.
Note that we set NULL to cloned scm_fp_list.edges in scm_fp_dup()
so that MSG_PEEK does not change the shape of the directed graph.
Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Acked-by: Paolo Abeni <pabeni@redhat.com>
Link: https://lore.kernel.org/r/20240325202425.60930-3-kuniyu@amazon.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
We will replace the garbage collection algorithm for AF_UNIX, where
we will consider each inflight AF_UNIX socket as a vertex and its file
descriptor as an edge in a directed graph.
This patch introduces a new struct unix_vertex representing a vertex
in the graph and adds its pointer to struct unix_sock.
When we send a fd using the SCM_RIGHTS message, we allocate struct
scm_fp_list to struct scm_cookie in scm_fp_copy(). Then, we bump
each refcount of the inflight fds' struct file and save them in
scm_fp_list.fp.
After that, unix_attach_fds() inexplicably clones scm_fp_list of
scm_cookie and sets it to skb. (We will remove this part after
replacing GC.)
Here, we add a new function call in unix_attach_fds() to preallocate
struct unix_vertex per inflight AF_UNIX fd and link each vertex to
skb's scm_fp_list.vertices.
When sendmsg() succeeds later, if the socket of the inflight fd is
still not inflight yet, we will set the preallocated vertex to struct
unix_sock.vertex and link it to a global list unix_unvisited_vertices
under spin_lock(&unix_gc_lock).
If the socket is already inflight, we free the preallocated vertex.
This is to avoid taking the lock unnecessarily when sendmsg() could
fail later.
In the following patch, we will similarly allocate another struct
per edge, which will finally be linked to the inflight socket's
unix_vertex.edges.
And then, we will count the number of edges as unix_vertex.out_degree.
Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Acked-by: Paolo Abeni <pabeni@redhat.com>
Link: https://lore.kernel.org/r/20240325202425.60930-2-kuniyu@amazon.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
TCP ehash table is often sparsely populated.
inet_twsk_purge() spends too much time calling cond_resched().
This patch can reduce time spent in inet_twsk_purge() by 20x.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Link: https://lore.kernel.org/r/20240327191206.508114-1-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
smc_hash_sk and smc_unhash_sk are only used in af_smc.c, so make them
static and remove the output symbol. They can be called under the path
.prot->hash()/unhash().
Signed-off-by: Zhengchao Shao <shaozhengchao@huawei.com>
Reviewed-by: Wen Gu <guwen@linux.alibaba.com>
Reviewed-by: Simon Horman <horms@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
TC filters come in 3 variants:
- no flag (try to process in hardware, but fallback to software))
- skip_hw (do not process filter by hardware)
- skip_sw (do not process filter by software)
However skip_sw is implemented so that the skip_sw
flag can first be checked, after it has been matched.
IMHO it's common when using skip_sw, to use it on all rules.
So if all filters in a block is skip_sw filters, then
we can bail early, we can thus avoid having to match
the filters, just to check for the skip_sw flag.
This patch adds a bypass, for when only TC skip_sw rules
are used. The bypass is guarded by a static key, to avoid
harming other workloads.
There are 3 ways that a packet from a skip_sw ruleset, can
end up in the kernel path. Although the send packets to a
non-existent chain way is only improved a few percents, then
I believe it's worth optimizing the trap and fall-though
use-cases.
+----------------------------+--------+--------+--------+
| Test description | Pre- | Post- | Rel. |
| | kpps | kpps | chg. |
+----------------------------+--------+--------+--------+
| basic forwarding + notrack | 3589.3 | 3587.9 | 1.00x |
| switch to eswitch mode | 3081.8 | 3094.7 | 1.00x |
| add ingress qdisc | 3042.9 | 3063.6 | 1.01x |
| tc forward in hw / skip_sw |37024.7 |37028.4 | 1.00x |
| tc forward in sw / skip_hw | 3245.0 | 3245.3 | 1.00x |
+----------------------------+--------+--------+--------+
| tests with only skip_sw rules below: |
+----------------------------+--------+--------+--------+
| 1 non-matching rule | 2694.7 | 3058.7 | 1.14x |
| 1 n-m rule, match trap | 2611.2 | 3323.1 | 1.27x |
| 1 n-m rule, goto non-chain | 2886.8 | 2945.9 | 1.02x |
| 5 non-matching rules | 1958.2 | 3061.3 | 1.56x |
| 5 n-m rules, match trap | 1911.9 | 3327.0 | 1.74x |
| 5 n-m rules, goto non-chain| 2883.1 | 2947.5 | 1.02x |
| 10 non-matching rules | 1466.3 | 3062.8 | 2.09x |
| 10 n-m rules, match trap | 1444.3 | 3317.9 | 2.30x |
| 10 n-m rules,goto non-chain| 2883.1 | 2939.5 | 1.02x |
| 25 non-matching rules | 838.5 | 3058.9 | 3.65x |
| 25 n-m rules, match trap | 824.5 | 3323.0 | 4.03x |
| 25 n-m rules,goto non-chain| 2875.8 | 2944.7 | 1.02x |
| 50 non-matching rules | 488.1 | 3054.7 | 6.26x |
| 50 n-m rules, match trap | 484.9 | 3318.5 | 6.84x |
| 50 n-m rules,goto non-chain| 2884.1 | 2939.7 | 1.02x |
+----------------------------+--------+--------+--------+
perf top (25 n-m skip_sw rules - pre patch):
20.39% [kernel] [k] __skb_flow_dissect
16.43% [kernel] [k] rhashtable_jhash2
10.58% [kernel] [k] fl_classify
10.23% [kernel] [k] fl_mask_lookup
4.79% [kernel] [k] memset_orig
2.58% [kernel] [k] tcf_classify
1.47% [kernel] [k] __x86_indirect_thunk_rax
1.42% [kernel] [k] __dev_queue_xmit
1.36% [kernel] [k] nft_do_chain
1.21% [kernel] [k] __rcu_read_lock
perf top (25 n-m skip_sw rules - post patch):
5.12% [kernel] [k] __dev_queue_xmit
4.77% [kernel] [k] nft_do_chain
3.65% [kernel] [k] dev_gro_receive
3.41% [kernel] [k] check_preemption_disabled
3.14% [kernel] [k] mlx5e_skb_from_cqe_mpwrq_nonlinear
2.88% [kernel] [k] __netif_receive_skb_core.constprop.0
2.49% [kernel] [k] mlx5e_xmit
2.15% [kernel] [k] ip_forward
1.95% [kernel] [k] mlx5e_tc_restore_tunnel
1.92% [kernel] [k] vlan_gro_receive
Test setup:
DUT: Intel Xeon D-1518 (2.20GHz) w/ Nvidia/Mellanox ConnectX-6 Dx 2x100G
Data rate measured on switch (Extreme X690), and DUT connected as
a router on a stick, with pktgen and pktsink as VLANs.
Pktgen-dpdk was in range 36.6-37.7 Mpps 64B packets across all tests.
Full test data at https://files.fiberby.net/ast/2024/tc_skip_sw/v2_tests/
Signed-off-by: Asbjørn Sloth Tønnesen <ast@fiberby.net>
Reviewed-by: Simon Horman <horms@kernel.org>
Reviewed-by: Marcelo Ricardo Leitner <marcelo.leitner@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
