On 30/11/2019 07:29, Anton Ivanov wrote:
On 29/11/2019 23:12, Daniel Borkmann wrote:
On 11/29/19 12:54 PM, Anton Ivanov wrote:
On 29/11/2019 09:15, Daniel Borkmann wrote:
On 11/28/19 6:44 PM, anton.ivanov@xxxxxxxxxxxxxxxxxx wrote:
From: Anton Ivanov <anton.ivanov@xxxxxxxxxxxxxxxxxx>
This fixes a possible hang in bpf firmware loading in the
UML vector io drivers due to use of GFP_KERNEL while holding
a spinlock.
Based on a prposed fix by weiyongjun1@xxxxxxxxxx and suggestions for
improving it by dan.carpenter@xxxxxxxxxx
Signed-off-by: Anton Ivanov <anton.ivanov@xxxxxxxxxxxxxxxxxx>
Any reason why this BPF firmware loading mechanism in UML vector
driver that was
recently added [0] is plain old classic BPF? Quoting your commit log
[0]:
It will allow whatever is allowed by sockfilter. Looking at the
sockfilter implementation in the kernel it takes eBPF, however even
the kernel docs still state BPF.
You are using SO_ATTACH_FILTER in uml_vector_attach_bpf() which is the
old classic
BPF (and not eBPF). The kernel internally moves that over to eBPF
insns, but you'll
be constrained forever with the abilities of cBPF. The later added
SO_ATTACH_BPF is
the one for eBPF where you pass the prog fd from bpf().
I will switch to that in the next version.
All vector drivers now allow a BPF program to be loaded and
associated with the RX socket in the host kernel.
1. The program can be loaded as an extra kernel command line
option to any of the vector drivers.
2. The program can also be loaded as "firmware", using the
ethtool flash option. It is possible to turn this facility
on or off using a command line option.
A simplistic wrapper for generating the BPF firmware for the raw
socket driver out of a tcpdump/libpcap filter expression can be
found at: https://github.com/kot-begemot-uk/uml_vector_utilities/
... it tells what it does but /nothing/ about the original rationale
/ use case
why it is needed. So what is the use case? And why is this only
classic BPF? Is
there any discussion to read up that lead you to this decision of
only implementing
handling for classic BPF?
Moving processing out of the GUEST onto the HOST using a safe
language. The firmware load is on the GUEST and your BPF is your
virtual NIC "firmware" which runs on the HOST (in the host kernel in
fact).
It is identical as an idea to what Netronome cards do in hardware.
I'm asking because classic BPF is /legacy/ stuff that is on feature
freeze and
only very limited in terms of functionality compared to native
(e)BPF which is
why you need this weird 'firmware' loader [1] which wraps around
tcpdump to
parse the -ddd output into BPF insns ...
Because there is no other mechanism of retrieving it after it is
compiled by libpcap in any of the common scripting languages.
The pcap Perl, Python, Go (or whatever else) wrappers do not give you
access to the compiled code after the filter has been compiled.
Why is that ingenious design - you have to take it with their
maintainers.
So if you want to start with pcap/tcpdump syntax and you do not want
to rewrite that part of tcpdump as a dumper in C you have no other
choice.
The starting point is chosen because the idea is at some point to
replace the existing and very aged pcap network transport in UML.
That takes pcap syntax on the kernel command line.
I admit it is a kludge, I will probably do the "do not want" bit and
rewrite that in C.
Yeah, it would probably be about the same # of LOC in C.
In any case - the "loader" is only an example, you can compile BPF
using LLVM or whatever else you like.
But did you try that with the code you have? Seems not, which is
perhaps why there are some
wrong assumptions.
All of my tests were done using bpf generated by tcpdump out of a pcap
expression. So the answer is no - I did not try LLVM because I did not
need to for what I was aiming to achieve.
The pcap route matches 1:1 existing functionality in the uml pcap driver
as well as existing functionality in the vector drivers for the cases
where they need to avoid seeing their own xmits and cannot use features
like QDISC_BYPASS.
You can't use LLVM's BPF backend here since you only allow to pass in
cBPF, and LLVM emits
an object file with native eBPF insns (you could use libbpf (in-tree
under tools/lib/bpf/)
for loading that).
My initial aim was the same feature sets as pcap and achieve it using a
virtual analogue of what cards like Netronome do - via the firmware route.
Switching to SO_ATTACH_BPF will come in the next revision.
A.
A.
_______________________________________________
linux-um mailing list
linux-um@xxxxxxxxxxxxxxxxxxx
http://lists.infradead.org/mailman/listinfo/linux-um
After reviewing what is needed for switching from SOCK_FILTER to
SOCK_BPF, IMHO it will have to wait for a while.
1. I am not sticking yet another direct host syscall invocation into the
userspace portion of the uml kernel and we cannot add extra userspace
libraries like libbpf at present because it is not supported by kbuild.
I have a patch in the queue for that, but it will need to be approved by
the kernel build people and merged before this can be done.
2. On top of that, in order to make use of eBPF for vNIC firmware
properly, I will need to figure out the correct abstractions. The
"program" part is quite clear - an eBPF program fits exactly into the
role of virtual nic firmware - it is identical to classic BPF and the
way it is used at present.
The maps, however, and how do they go along with the "program firmware"
is something which will need to be figured out. It may require a more
complex load mechanisms and a proper (not 5 liner wrapper around pcap or
tcpdump) firmware packer/unpacker.
Once I have figured it out and it can fit into the kbuild, I will send
the next revision. I suspect that it will happen at about the same time
I will finish the AF_XDP UML vNIC transport (it has the same
requirements, needs the same calls and uses the same libraries).
--
Anton R. Ivanov
Cambridgegreys Limited. Registered in England. Company Number 10273661
https://www.cambridgegreys.com/
