On 3/12/24 18:32, Andrei Matei wrote:
On Tue, Mar 12, 2024 at 1:16 PM Kui-Feng Lee <sinquersw@xxxxxxxxx> wrote:
On 3/8/24 07:43, Andrei Matei wrote:
On Thu, Mar 7, 2024 at 6:02 PM Kui-Feng Lee <sinquersw@xxxxxxxxx> wrote:
On 3/5/24 15:53, Song Liu wrote:
On Tue, Mar 5, 2024 at 9:18 AM Jiri Olsa <olsajiri@xxxxxxxxx> wrote:
On Fri, Mar 01, 2024 at 11:39:03AM -0800, Kui-Feng Lee wrote:
On 2/29/24 06:39, Jiri Olsa wrote:
One of uprobe pain points is having slow execution that involves
two traps in worst case scenario or single trap if the original
instruction can be emulated. For return uprobes there's one extra
trap on top of that.
My current idea on how to make this faster is to follow the optimized
kprobes and replace the normal uprobe trap instruction with jump to
user space trampoline that:
- executes syscall to call uprobe consumers callbacks
- executes original instructions
- jumps back to continue with the original code
There are of course corner cases where above will have trouble or
won't work completely, like:
- executing original instructions in the trampoline is tricky wrt
rip relative addressing
- some instructions we can't move to trampoline at all
- the uprobe address is on page boundary so the jump instruction to
trampoline would span across 2 pages, hence the page replace won't
be atomic, which might cause issues
- ... ? many others I'm sure
Still with all the limitations I think we could be able to speed up
some amount of the uprobes, which seems worth doing.
Just a random idea related to this.
Could we also run jit code of bpf programs in the user space to collect
information instead of going back to the kernel every time?
I was thinking about a similar idea. I guess these user space BPF
programs will have limited features that we can probably use them
update bpf maps. For this limited scope, we still need bpf_arena.
Otherwise, the user space bpf program will need to update the bpf
maps with sys_bpf(), which adds the same overhead as triggering
That is true. However, even without bpf_arena, it still works with
some workarounds without going through sys_bpf().
Anything making uprobes faster would be very welcomed for my project. The
biggest performance problem for us is the cost of bpf_probe_read_user()
relative to raw memory access. Every call to this helper walks the process'
"raw memory access"? Do you mean not going through any helper function,
reading from a pointer directly?
Right.
I recognize that, as long as bpf runs "in the kernel", one cannot simply
dereference a user-space pointer since the kernel is a different virtual memory
space (*). Still, I wish there bpf_probe_read_user() were faster.
(*) Or, is it indeed a different memory space or is the kernel's virtual
address space mapped into every process? Did this change through KPTI? I would
be curious to read a good resource on what exactly it means to switch from
user-space to the kernel and back, if such a thing exists.
FYI! This is architecture dependent. AFAIK, with x86 platforms, kernel
can access the memory of the user space directly if it is in a
process/task context. But, you should not relies on it.
If you look into bpf_probe_read_user(), it eventually do something like
"rep movsb" on x86 platforms. Access user space memory directly with
some extra checks. So, the bottleneck here can be the extra checks and
memory copying. If you access small chunks like what you said bellow,
the overhead of checks could be expensive.
page table to check that the access would not cause a fault (I think); this is
very slow. I wonder if there's some other option that would keep the safety
requirement for the memory access -- I'm imagining an optimistic mode where the
raw access is performed (in the target process' memory space) and, in the rare
case when a fault happens, the kernel would somehow recover from the fault and
I am not very familiar with this part. I read the implementation of
bpf_probe_read_user() a little bit. It does what you mentioned here. It
would cause page faults, however, the handler will skip the instruction
leaving the counter non-zero. By checking the counter, it knows the
instruction is not completed, and returns an error.
I am curious about what your access pattern looks like. Does it access a
large number of small chunks of data? Or, does it access a small number
of big chunks of data?
My access pattern looks like a lot of small reads. Some of these reads could be
done at the same time if we had a vectorized API (i.e. some of the pointers are
known in advance); for others there are data dependencies (i.e. we need to
dereference a pointer to know what we'll want to read next). Specifically, the
use case is a debugger of sorts which uses BPF uprobes for poking around in the
target process' memory, rather than the more traditional ptrace-based
techniques (ptrace being very slow). This debugger needs to walk a lot of
thread stacks by following stack pointers or by using DWARF unwind information,
and then it further reads data structures from the target process' stacks and
heaps, chasing pointers recursively.
A related information. You may already know that bpf_probe_read_user()
can fail if a page fault happens. A vectorized API probably doesn't
change it. It is a limitation of non-sleepable BPF programs. Sleepable
BPF programs should be able to overcome it.
fail the bpf_probe_read_user() helper. Would something like that be technically
feasible / has there been any prior interest in faster access to user memory
A more limited option that might be helpful would be a vectorized version of
bpf_probe_read_user() that verifies many pointers at once.
the program with a syscall.
sorry for late reply, do you mean like ubpf? the scope of this change
is to speed up the generic uprobe, ebpf is just one of the consumers
I guess this means we need a new syscall?
Thanks,
Song
