Doing some refactoring resulted in a kernel splat when reading sock_ops
fields.
Patch 1, has the details and proposed fix for sock_ops sk field access.
Patch 2, has the details and proposed fix for reading sock_ops->sk field
Patch 3, Gives a reproducer and test to verify the fix. I used the netcnt
program to test this because I wanted a splat to be generated which can
only be done if we have real traffic exercising the code.
Patch 4, Is an optional patch. While doing above I wanted to also verify
loads were OK. The code looked good, but I wanted some xlated code to
review as well. It seems like a good idea to add it here or at least
shouldn't hurt. I could push it into bpf-next if folks want.
Patch 5, Add reproducers for reading scok_ops->sk field.
I split Patch1 and Patch2 into two two patches because they have different
fixes tags. Seems like this will help with backporting. They could be
squashed though if folks want.
For selftests I was fairly verbose creating three patches each with the
associated xlated code to handle each of the three cases. My hope is this
helps the reader understand issues and review fixes. Its more or less
how I debugged the issue and created reproducers so it at least helped
me to have them logically different patches.
v2->v3: Updated commit msg in patch1 to include ommited line of asm
output, per Daniels comment.
v1->v2: Added fix sk access case
---
John Fastabend (5):
bpf: sock_ops ctx access may stomp registers in corner case
bpf: sock_ops sk access may stomp registers when dst_reg = src_reg
bpf, selftests: Add tests for ctx access in sock_ops with single register
bpf, selftests: Add tests for sock_ops load with r9,r8.r7 registers
bpf, selftests: Add tests to sock_ops for loading sk
net/core/filter.c | 75 +++++++++++++++++---
.../testing/selftests/bpf/progs/test_tcpbpf_kern.c | 41 +++++++++++
2 files changed, 103 insertions(+), 13 deletions(-)
--
Signature
