linux.git/include/net/checksum.h, branch v6.1

powerpc/net: Implement powerpc specific csum_shift() to remove branch

2022-03-11T10:57:22+00:00

Today's implementation of csum_shift() leads to branching based on
parity of 'offset'

	000002f8 :
	     2f8:	70 a5 00 01 	andi.   r5,r5,1
	     2fc:	41 a2 00 08 	beq     304 
	     300:	54 84 c0 3e 	rotlwi  r4,r4,24
	     304:	7c 63 20 14 	addc    r3,r3,r4
	     308:	7c 63 01 94 	addze   r3,r3
	     30c:	4e 80 00 20 	blr

Use first bit of 'offset' directly as input of the rotation instead of
branching.

	000002f8 :
	     2f8:	54 a5 1f 38 	rlwinm  r5,r5,3,28,28
	     2fc:	20 a5 00 20 	subfic  r5,r5,32
	     300:	5c 84 28 3e 	rotlw   r4,r4,r5
	     304:	7c 63 20 14 	addc    r3,r3,r4
	     308:	7c 63 01 94 	addze   r3,r3
	     30c:	4e 80 00 20 	blr

And change to left shift instead of right shift to skip one more
instruction. This has no impact on the final sum.

	000002f8 :
	     2f8:	54 a5 1f 38 	rlwinm  r5,r5,3,28,28
	     2fc:	5c 84 28 3e 	rotlw   r4,r4,r5
	     300:	7c 63 20 14 	addc    r3,r3,r4
	     304:	7c 63 01 94 	addze   r3,r3
	     308:	4e 80 00 20 	blr

Seems like only powerpc benefits from a branchless implementation.
Other main architectures like ARM or X86 get better code with
the generic implementation and its branch.

Signed-off-by: Christophe Leroy 
Signed-off-by: David S. Miller

openvswitch: Fix setting ipv6 fields causing hw csum failure

2022-02-24T17:16:21+00:00

Ipv6 ttl, label and tos fields are modified without first
pulling/pushing the ipv6 header, which would have updated
the hw csum (if available). This might cause csum validation
when sending the packet to the stack, as can be seen in
the trace below.

Fix this by updating skb->csum if available.

Trace resulted by ipv6 ttl dec and then sending packet
to conntrack [actions: set(ipv6(hlimit=63)),ct(zone=99)]:
[295241.900063] s_pf0vf2: hw csum failure
[295241.923191] Call Trace:
[295241.925728]  
[295241.927836]  dump_stack+0x5c/0x80
[295241.931240]  __skb_checksum_complete+0xac/0xc0
[295241.935778]  nf_conntrack_tcp_packet+0x398/0xba0 [nf_conntrack]
[295241.953030]  nf_conntrack_in+0x498/0x5e0 [nf_conntrack]
[295241.958344]  __ovs_ct_lookup+0xac/0x860 [openvswitch]
[295241.968532]  ovs_ct_execute+0x4a7/0x7c0 [openvswitch]
[295241.979167]  do_execute_actions+0x54a/0xaa0 [openvswitch]
[295242.001482]  ovs_execute_actions+0x48/0x100 [openvswitch]
[295242.006966]  ovs_dp_process_packet+0x96/0x1d0 [openvswitch]
[295242.012626]  ovs_vport_receive+0x6c/0xc0 [openvswitch]
[295242.028763]  netdev_frame_hook+0xc0/0x180 [openvswitch]
[295242.034074]  __netif_receive_skb_core+0x2ca/0xcb0
[295242.047498]  netif_receive_skb_internal+0x3e/0xc0
[295242.052291]  napi_gro_receive+0xba/0xe0
[295242.056231]  mlx5e_handle_rx_cqe_mpwrq_rep+0x12b/0x250 [mlx5_core]
[295242.062513]  mlx5e_poll_rx_cq+0xa0f/0xa30 [mlx5_core]
[295242.067669]  mlx5e_napi_poll+0xe1/0x6b0 [mlx5_core]
[295242.077958]  net_rx_action+0x149/0x3b0
[295242.086762]  __do_softirq+0xd7/0x2d6
[295242.090427]  irq_exit+0xf7/0x100
[295242.093748]  do_IRQ+0x7f/0xd0
[295242.096806]  common_interrupt+0xf/0xf
[295242.100559]  
[295242.102750] RIP: 0033:0x7f9022e88cbd
[295242.125246] RSP: 002b:00007f9022282b20 EFLAGS: 00000246 ORIG_RAX: ffffffffffffffda
[295242.132900] RAX: 0000000000000005 RBX: 0000000000000010 RCX: 0000000000000000
[295242.140120] RDX: 00007f9022282ba8 RSI: 00007f9022282a30 RDI: 00007f9014005c30
[295242.147337] RBP: 00007f9014014d60 R08: 0000000000000020 R09: 00007f90254a8340
[295242.154557] R10: 00007f9022282a28 R11: 0000000000000246 R12: 0000000000000000
[295242.161775] R13: 00007f902308c000 R14: 000000000000002b R15: 00007f9022b71f40

Fixes: 3fdbd1ce11e5 ("openvswitch: add ipv6 'set' action")
Signed-off-by: Paul Blakey 
Link: https://lore.kernel.org/r/20220223163416.24096-1-paulb@nvidia.com
Signed-off-by: Jakub Kicinski

net: Force inlining of checksum functions in net/checksum.h

2022-02-19T16:07:12+00:00

All functions defined as static inline in net/checksum.h are
meant to be inlined for performance reason.

But since commit ac7c3e4ff401 ("compiler: enable
CONFIG_OPTIMIZE_INLINING forcibly") the compiler is allowed to
uninline functions when it wants.

Fair enough in the general case, but for tiny performance critical
checksum helpers that's counter-productive.

The problem mainly arises when selecting CONFIG_CC_OPTIMISE_FOR_SIZE,
Those helpers being 'static inline' in header files you suddenly find
them duplicated many times in the resulting vmlinux.

Here is a typical exemple when building powerpc pmac32_defconfig
with CONFIG_CC_OPTIMISE_FOR_SIZE. csum_sub() appears 4 times:

	c04a23cc :
	c04a23cc:	7c 84 20 f8 	not     r4,r4
	c04a23d0:	7c 63 20 14 	addc    r3,r3,r4
	c04a23d4:	7c 63 01 94 	addze   r3,r3
	c04a23d8:	4e 80 00 20 	blr
		...
	c04a2ce8:	4b ff f6 e5 	bl      c04a23cc 
		...
	c04a2d2c:	4b ff f6 a1 	bl      c04a23cc 
		...
	c04a2d54:	4b ff f6 79 	bl      c04a23cc 
		...
	c04a754c :
	c04a754c:	7c 84 20 f8 	not     r4,r4
	c04a7550:	7c 63 20 14 	addc    r3,r3,r4
	c04a7554:	7c 63 01 94 	addze   r3,r3
	c04a7558:	4e 80 00 20 	blr
		...
	c04ac930:	4b ff ac 1d 	bl      c04a754c 
		...
	c04ad264:	4b ff a2 e9 	bl      c04a754c 
		...
	c04e3b08 :
	c04e3b08:	7c 84 20 f8 	not     r4,r4
	c04e3b0c:	7c 63 20 14 	addc    r3,r3,r4
	c04e3b10:	7c 63 01 94 	addze   r3,r3
	c04e3b14:	4e 80 00 20 	blr
		...
	c04e5788:	4b ff e3 81 	bl      c04e3b08 
		...
	c04e65c8:	4b ff d5 41 	bl      c04e3b08 
		...
	c0512d34 :
	c0512d34:	7c 84 20 f8 	not     r4,r4
	c0512d38:	7c 63 20 14 	addc    r3,r3,r4
	c0512d3c:	7c 63 01 94 	addze   r3,r3
	c0512d40:	4e 80 00 20 	blr
		...
	c0512dfc:	4b ff ff 39 	bl      c0512d34 
		...
	c05138bc:	4b ff f4 79 	bl      c0512d34 
		...

Restore the expected behaviour by using __always_inline for all
functions defined in net/checksum.h

vmlinux size is even reduced by 256 bytes with this patch:

	   text	   data	    bss	    dec	    hex	filename
	6980022	2515362	 194384	9689768	 93daa8	vmlinux.before
	6979862	2515266	 194384	9689512	 93d9a8	vmlinux.now

Fixes: ac7c3e4ff401 ("compiler: enable CONFIG_OPTIMIZE_INLINING forcibly")
Cc: Masahiro Yamada 
Cc: Nick Desaulniers 
Cc: Andrew Morton 
Signed-off-by: Christophe Leroy 
Signed-off-by: David S. Miller

net: fix recent csum changes

2021-12-07T00:26:46+00:00

Vladimir reported csum issues after my recent change in skb_postpull_rcsum()

Issue here is the following:

initial skb->csum is the csum of

[part to be pulled][rest of packet]

Old code:
 skb->csum = csum_sub(skb->csum, csum_partial(pull, pull_length, 0));

New code:
 skb->csum = ~csum_partial(pull, pull_length, ~skb->csum);

This is broken if the csum of [pulled part]
happens to be equal to skb->csum, because end
result of skb->csum is 0 in new code, instead
of being 0xffffffff

David Laight suggested to use

skb->csum = -csum_partial(pull, pull_length, -skb->csum);

I based my patches on existing code present in include/net/seg6.h,
update_csum_diff4() and update_csum_diff16() which might need
a similar fix.

I guess that my tests, mostly pulling 40 bytes of IPv6 header
were not providing enough entropy to hit this bug.

v2: added wsum_negate() to make sparse happy.

Fixes: 29c3002644bd ("net: optimize skb_postpull_rcsum()")
Fixes: 0bd28476f636 ("gro: optimize skb_gro_postpull_rcsum()")
Signed-off-by: Eric Dumazet 
Reported-by: Vladimir Oltean 
Suggested-by: David Laight 
Cc: David Lebrun 
Tested-by: Vladimir Oltean 
Link: https://lore.kernel.org/r/20211204045356.3659278-1-eric.dumazet@gmail.com
Signed-off-by: Jakub Kicinski

csum_and_copy_to_iter(): massage into form closer to csum_and_copy_from_iter()

2021-06-10T15:45:14+00:00

Namely, have off counted starting from 0 rather than from csstate->off.
To compensate we need to shift the initial value (csstate->sum) (rotate
by 8 bits, as usual for csum) and do the same after we are finished adding
the pieces up.

What we get out of that is a bit more redundancy in our variables - from
is always equal to addr + off, which will be useful several commits down
the road.

Signed-off-by: Al Viro

saner calling conventions for csum_and_copy_..._user()

2020-08-20T19:45:15+00:00

All callers of these primitives will
	* discard anything we might've copied in case of error
	* ignore the csum value in case of error
	* always pass 0xffffffff as the initial sum, so the
resulting csum value (in case of success, that is) will never be 0.

That suggest the following calling conventions:
	* don't pass err_ptr - just return 0 on error.
	* don't bother with zeroing destination, etc. in case of error
	* don't pass the initial sum - just use 0xffffffff.

This commit does the minimal conversion in the instances of csum_and_copy_...();
the changes of actual asm code behind them are done later in the series.
Note that this asm code is often shared with csum_partial_copy_nocheck();
the difference is that csum_partial_copy_nocheck() passes 0 for initial
sum while csum_and_copy_..._user() pass 0xffffffff.  Fortunately, we are
free to pass 0xffffffff in all cases and subsequent patches will use that
freedom without any special comments.

A part that could be split off: parisc and uml/i386 claimed to have
csum_and_copy_to_user() instances of their own, but those were identical
to the generic one, so we simply drop them.  Not sure if it's worth
a separate commit...

Signed-off-by: Al Viro

csum_partial_copy_nocheck(): drop the last argument

2020-08-20T19:45:14+00:00

It's always 0.  Note that we theoretically could use ~0U as well -
result will be the same modulo 0xffff, _if_ the damn thing did the
right thing for any value of initial sum; later we'll make use of
that when convenient.

However, unlike csum_and_copy_..._user(), there are instances that
did not work for arbitrary initial sums; c6x is one such.

Signed-off-by: Al Viro

unify generic instances of csum_partial_copy_nocheck()

2020-08-20T19:45:14+00:00

quite a few architectures have the same csum_partial_copy_nocheck() -
simply memcpy() the data and then return the csum of the copy.

hexagon, parisc, ia64, s390, um: explicitly spelled out that way.

arc, arm64, csky, h8300, m68k/nommu, microblaze, mips/GENERIC_CSUM, nds32,
nios2, openrisc, riscv, unicore32: end up picking the same thing spelled
out in lib/checksum.h (with varying amounts of perversions along the way).

everybody else (alpha, arm, c6x, m68k/mmu, mips/!GENERIC_CSUM, powerpc,
sh, sparc, x86, xtensa) have non-generic variants.  For all except c6x
the declaration is in their asm/checksum.h.  c6x uses the wrapper
from asm-generic/checksum.h that would normally lead to the lib/checksum.h
instance, but in case of c6x we end up using an asm function from arch/c6x
instead.

Screw that mess - have architectures with private instances define
_HAVE_ARCH_CSUM_AND_COPY in their asm/checksum.h and have the default
one right in net/checksum.h conditional on _HAVE_ARCH_CSUM_AND_COPY
*not* defined.

Signed-off-by: Al Viro

default csum_and_copy_to_user(): don't bother with access_ok()

2020-05-29T20:11:50+00:00

Signed-off-by: Al Viro

take the dummy csum_and_copy_from_user() into net/checksum.h

2020-05-29T20:11:50+00:00

now that can be done conveniently - all non-trivial cases have
_HAVE_ARCH_COPY_AND_CSUM_FROM_USER defined, so the fallback in
net/checksum.h is used only for dummy (copy_from_user, then
csum_partial) implementation.  Allowing us to get rid of all
dummy instances, both of csum_and_copy_from_user() and
csum_partial_copy_from_user().

Signed-off-by: Al Viro