lkml.org 
[lkml]   [2022]   [Mar]   [29]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: Comments on new user events ABI
----- On Mar 28, 2022, at 8:29 PM, Beau Belgrave beaub@linux.microsoft.com wrote:

>> ----- On Mar 28, 2022, at 4:24 PM, Mathieu Desnoyers
>> mathieu.desnoyers@efficios.com wrote:
>> Hi Beau, Hi Steven,
>>
>> I've done a review of the trace events ABI, and I have a few comments.
>> Sorry for being late to the party, but I only noticed this new ABI recently.
>> Hopefully we can improve this ABI before the 5.18 release.
>>
>
> Welcome to the party :)
>
>> * user_events_status memory mapping
>>
>> As I understand it, one part of the user events ABI is a memory mapping
>> which contains "flags" which indicates whether a given event is enabled.
>> It is indexed by byte, and each byte has this bitwise meaning:
>>
>> /* Bits 0-6 are for known probe types, Bit 7 is for unknown probes */
>> #define EVENT_BIT_FTRACE 0
>> #define EVENT_BIT_PERF 1
>> #define EVENT_BIT_OTHER 7
>>
>> There are a few things I find odd here. First, to improve use of CPU cache,
>> I would have expected this memory mapping to expose enable flags as a
>> bitmap rather than an array of bytes, indexed bit-wise rather than byte-wise.
>> I also don't get what user-space is expected to do differently if FTRACE vs
>> PERF is enabled, considering that it gates a writev() to a file descriptor
>> associated with /sys/kernel/debug/tracing/user_events_data.
>>
>
> The intention wasn't for user-space to check the byte other than non-zero.
> User processes could do that, but it's more so administration tools can see
> where the events are registered if they cannot be closed and the state of the
> machine.

Administration tools could simply use a seqfile interface and parse text. I
don't expect those to be performance-sensitive at all. So we should not design
the overhead-sensitive mmap'd ABI (enable flags) for administration tools.

> Maybe you have a slicker way to do this, but it seems to check a bit in the
> page would involve at least a byte read followed by a mask or shift?

I suspect the minimum cost in terms of instruction would be to have the mask
prepared in advance (when the event is registered), and therefore the instructions
required on the user-space fast-path would be a load followed by a mask (&).

> That seems more expensive than checking a byte?

It is one single extra ALU instruction on the fast-path, indeed. But we have to
be careful about optimizing too much for a non-representative microbenchmark
and missing the big picture, which includes L1/L2 cache lines and TLB load overhead
when tracing a real-life workloads.

On modern x86 CPUs, an ALU instruction takes around 0.3ns, which is a few
orders of magnitude faster than a cache miss. Using bit-wise rather than
byte-wise indexing increases functional density of those enable flags by a
factor 8, which lessens the amount of cpu cache required by that same factor.

So as soon as the workload running on the system fills the CPU L1 or L2 caches,
when the number of user events registered scales up, and when the checks for
events enabled are done frequently enough, evicting fewer cache lines becomes
more important than the extra instruction required to apply the mask on the
fast-path.

>
>> I would have rather thought that tracers implemented in user-space could
>> register
>> themselves, and then there could be one
>> /sys/kernel/debug/tracing/user_events_status
>> per tracer. Considering that all kernel tracers use the same ABI to write an
>> event,
>> and then dispatch this event internally within the kernel to each registered
>> tracer, I would expect to have a single memory mapping for all those (e.g. a
>> /sys/kernel/debug/tracing/user_events_status/kernel_tracers file).
>>
>> Then eventually if we have other user-space tracers such as lttng-ust with its
>> their own user-space code performing tracing in a shared memory ring buffer, it
>> would make sense to allow it to register its own
>> /sys/kernel/debug/tracing/user_events_status/lttng_ust file, with its own
>> indexes.
>>
>
> I don't follow that. The intention is to get user processes to participate with
> trace_events and the built-in tooling. When would a user-space tracer be used
> instead of perf/ftrace?
>
> It seems like a feature request?

It can very well be out of scope for the user events, and I'm fine with that.
I was merely considering how the user events could be leveraged by tracers
implemented purely in user-space. But if the stated goal of this feature is
really to call into kernel tracers through a writev(), then I suspect that
supporting purely user-space tracers is indeed out of scope.

>
>> If this facility is ever used by lttng-ust to enable user-space tracing, I would
>> not
>> want to take the overhead of calling writev for the sake of kernel tracers if
>> those are disabled.
>>
>
> If they were disabled the byte wouldn't be set, right? So no writev overhead.
>
> Seems I'm missing something.

I was wondering whether we could leverage the user events bit-enable ABI for
tracers which are not implemented within the Linux kernel (e.g. LTTng-UST).
But I suspect it might be better for me to re-implement this in user-space
over shared memory.

>
>> So perhaps in the short-term there is no need to implement the user-space tracer
>> registration ABI, but I would have expected a simple bitmap for
>> /sys/kernel/debug/tracing/user_events_data/kernel_tracers rather than the
>> bytewise index, because as far as the kernel tracers are concerned, providing
>> the bit to tell userspace instrumentation exactly which tracers are internally
>> enabled within the kernel does not appear to be of any use other than increasing
>> the footprint on the actively used cpu cache lines.
>>
>>
>> * user_events_data page faults
>>
>> If my understanding is correct, when the user-space program's memory containing
>> the payload passed to writev() to a user_events_data file descriptor is kicked
>> out from the page cache between fault_in_iov_iter_readable and its use by the
>> tracers due to high memory pressure, the writev() will fail with -EFAULT and
>> the data will be discarded unless user-space somehow handles this error (which
>> is not handled in the samples/user_events/sample.c example program). It is good
>> that the memory is faulted in immediately before calling the tracers, but
>> considering that it is not mlock'd, should we make more effort to ensure the
>> tracers are able to handle page faults ?
>>
>> Integration of the work done by Michael Jeanson and myself on faultable
>> tracepoint
>> would allow the tracepoint probes to take page faults. Then, further
>> modifications
>> in the kernel tracers would be needed to handle those page faults.
>>
>
> Is this something that can be done later or does it require ABI changes?
>
> I would love to never miss data due to page faults.

This is internal to the user events, tracepoint, and kernel tracers
implementation. I don't expect this to modify the user events ABI.

The only thing that might require some thinking ABI-wise is how the ring
buffers are exposed to user-space consumers, because we would want to
allow taking page faults between space reservation and commit.

The LTTng ring buffer has been supporting this out of the box for years
now, but this may be trickier for other kernel tracers, for which allowing
preemption between reserve and commit has never been a requirement until
now.

>
>>
>> * user_reg name_args and write_index vs purely user-space tracers
>>
>> That part of the user event registration (event layout and ID allocation)
>> appears
>> to be intrinsically tied to the kernel tracers and the expected event layout.
>> This
>> seems fine as long as the only users we consider are the kernel tracers, but it
>> appears to be less relevant for purely user-space tracers. Actually, tying the
>> mmap'd event enable mechanism with the event ID and description makes me wonder
>> whether it might be better to have LTTng-UST implement its own shared-memory
>> based
>> "fast-event-enabling" mechanism rather than use this user-event ABI. The other
>> advantage of doing all of this in user-space would be to allow many instances
>> of this bitmap to exist on a given system, e.g. one per container in a
>> multi-container
>> system, rather than requiring this to be a global kernel-wide singleton, and to
>> use
>> it from a non-privileged user.
>>
>
> We have some conversation going about using namespaces/cgroups to isolation
> containers with bitmaps/status pages. The main thing I personally want to be
> able to do is from the root namespace see all the events in the descendents
> easily via perf, eBPF or ftrace.
>
> Link:
> https://lore.kernel.org/linux-trace-devel/20220316232009.7952988633787ef1003f13b0@kernel.org/
>

I see that a notion close to "tracing namespaces" comes up in this thread. This is something
I brought forward at the last Linux Plumbers Conference aiming to facilitate system call tracing
for a given container (or from a process hierarchy). I suspect the tracing namespace could also
be tied to a set of kernel buffers, and to a user events "domain". I think this concept could
neatly solve many of our container-related isolation issues here. As you say, it should probably
be a hierarchical namespace.

>>
>> Some comments about the implementation:
[...]
>
>>
>> return user_events_write_core(file, &i);
>> }
>>
>> include/uapi/linux/user_events.h:
>>
>> struct user_reg {
>>
>> /* Input: Size of the user_reg structure being used */
>> __u32 size;
>>
>> /* Input: Pointer to string with event name, description and flags */
>> __u64 name_args;
>>
>> /* Output: Byte index of the event within the status page */
>> __u32 status_index;
>>
>> /* Output: Index of the event to use when writing data */
>> __u32 write_index;
>> };
>>
>> As this structure is expected to grow, and the user-space sample program uses
>> "sizeof()"
>> to figure out its size (which includes padding), I would be more comfortable if
>> this was
>> a packed structure rather than non-packed, because as fields are added, it's
>> tricky to
>> figure out from the kernel perspective whether the size received are fields that
>> user-space
>> is aware of, or if this is just padding.
>>
>
> I think that would be a good idea, Steven?

[leaving this for Steven to answer]

>
>> include/uapi/linux/user_events.h:
>>
>> struct user_bpf_iter {
>>
>> /* Offset of the data within the first iovec */
>> __u32 iov_offset;
>>
>> /* Number of iovec structures */
>> __u32 nr_segs;
>>
>> /* Pointer to iovec structures */
>> const struct iovec *iov;
>>
>> ^ a pointer in a uapi header is usually a no-go. This should be a u64.
>> };
>>
>> include/uapi/linux/user_events.h:
>>
>> struct user_bpf_context {
>>
>> /* Data type being passed (see union below) */
>> __u32 data_type;
>>
>> /* Length of the data */
>> __u32 data_len;
>>
>> /* Pointer to data, varies by data type */
>> union {
>> /* Kernel data (data_type == USER_BPF_DATA_KERNEL) */
>> void *kdata;
>>
>> /* User data (data_type == USER_BPF_DATA_USER) */
>> void *udata;
>>
>> /* Direct iovec (data_type == USER_BPF_DATA_ITER) */
>> struct user_bpf_iter *iter;
>>
>> ^ likewise for the 3 pointers above. Should be u64 in uapi headers.
>> };
>> };
>>
>
> The bpf structs are only used within a BPF program. At that point the pointer
> sizes should all align, right?

I must admit I do not know enough about the eBPF uapi practices to answer this.
[CCing Alexei on this]

It's just that seeing a pointer in a uapi header raises red flags in terms of
32 vs 64 bit compatibility, so I am bringing this up.

>
> I wanted to ensure libbpf type scenarios have easy access to the BPF argument
> structures.
>
> I guess I could move them to the kernel side, but then users would have to get
> access to them for BPF programs.
>
>> kernel/trace/trace_events_user.c:
>>
>> static long user_reg_get(struct user_reg __user *ureg, struct user_reg *kreg)
>> {
>> u32 size;
>> long ret;
>>
>> ret = get_user(size, &ureg->size);
>>
>> if (ret)
>> return ret;
>>
>>
>> if (size > PAGE_SIZE)
>> return -E2BIG;
>>
>> ^ here I would be tempted to validate that the structure size at least provides
>> room
>> for the "v0" ABI, e.g.:
>>
>> if (size < offsetofend(struct user_reg, write_index))
>> return -EINVAL;
>>
>> return copy_struct_from_user(kreg, sizeof(*kreg), ureg, size);
>>
>> ^ I find it odd that the kernel copy of struct user_reg may contain a
>> size field which contents differs from the size fetched by get_user().
>> This can happen if a buggy or plainly hostile user-space attempts to
>> confuse the kernel about the size of this structure. Fortunately, the
>> size field does not seem to be used afterwards, but I would think it
>> safer to copy back the "size" fetched by get_user into the reg->size
>> after copy_struct_from_user in case future changes in the code end up
>> relying on a consistent size field.
>> }
>>
>
> Any size that the user doesn't fill in will zero out, if they enter a large
> size then we error out. I tried to get the data out from the user once and use
> it consistently from the kernel side. I didn't want users changing their values
> at key places trying to get the kernel to do bad things.
>
> This is why the size isn't used beyond the first copy (and why the data in the
> write paths is only validated after paging in and copy the data to a buffer
> that the user process no longer can modify mid-write).

As long as the kreg->size value is not used afterwards, it's technically fine, but
having it there and inconsistent with the size value loaded by get_user() is error-prone.
I would recommend to either overwrite this kreg->size value with the size loaded by
get_user(), or document that this field should never be used because it is inconsistent
with the size used for copy_struct_from_user().

[...]

>> kernel/trace/trace_events_user.c:
>>
>> static int user_events_release(struct inode *node, struct file *file)
>> {
>> [...]
>> /*
>> * Ensure refs cannot change under any situation by taking the
>> * register mutex during the final freeing of the references.
>> */
>> mutex_lock(&reg_mutex);
>> [...]
>> mutex_unlock(&reg_mutex);
>>
>> kfree(refs);
>>
>> ^ AFAIU, the user_events_write() does not rely on reg_mutex to ensure mutual
>> exclusion.
>> Doing so would be prohibitive performance-wise. But I suspect that freeing
>> "refs" here
>> without waiting for a RCU grace period can be an issue if user_events_write_core
>> is using
>> refs concurrently with file descriptor close.
>>
>
> The refs is per-file, so if user_events_write() is running release cannot be
> called for that file instance, since it's being used. Did I miss something?

I'm possibly the one missing something here, but what prevents 2 threads
from doing user_event_write() and close() concurrently on the same file descriptor ?

>
>> kernel/trace/trace_events_user.c:
>>
>> static bool user_field_match(struct ftrace_event_field *field, int argc,
>> const char **argv, int *iout)
>> [...]
>>
>> for (; i < argc; ++i) {
>> [...]
>> pos += snprintf(arg_name + pos, len - pos, argv[i]);
>>
>> ^ what happens if strlen(argv[i]) > (len - pos) ? Based on lib/vsprintf.c:
>>
>> * The return value is the number of characters which would be
>> * generated for the given input, excluding the trailing null,
>> * as per ISO C99. If the return is greater than or equal to
>> * @size, the resulting string is truncated.
>>
>> So the "pos" returned by the first call to sprintf would be greater than
>> MAX_FIELD_ARG_NAME.
>> Then the second call to snprintf passes a @size argument of "len - pos" using
>> the pos value
>> which is larger than len... which is a negative integer passed as argument to a
>> size_t (unsigned).
>> So it expects a very long string. And the @buf argument is out-of-bound
>> (field_name + pos).
>> Is this pattern for using snprintf() used elsewhere ? From a quick grep, I find
>> this pattern in
>> a few places where AFAIU the input is not user-controlled (as it seems to be the
>> case here), but
>> still it might be worth looking into:
>>
>> kernel/cgroup/cgroup.c:show_delegatable_files()
>> kernel/time/clocksource.c:available_clocksource_show()
>>
>> Also, passing a copy of a userspace string (argv[i]) as format string argument
>> to
>> snprintf can be misused to leak kernel data to user-space.
>>
>> The same function also appear to have similar issues with its use of the
>> field->name userspace input
>> string.
>>
>
> The strings are copied first before coming down these paths, so they cannot
> be forced to changed mid-reads, etc.

That's correct. But their combined length can be larger than MAX_FIELD_ARG_NAME.

> I think simply checking if any arg is beyond the MAX_FIELD_ARG_NAME would
> prevent these types of issues, right?

No. The issue is when the combined length is larger than the max. If I look at
other tracing code, I notice that it often uses a pattern where snprintf is first
called on the input with a @size=0 to calculate the string length, then the
output string is dynamically allocated with the correct length, and the entire
code performing the snprintf formatting is called again with a nonzero size,
thus effectively populating the output string. There are other ways to prevent
these issues, but I would be tempted to use the same pattern used elsewhere in
the tracing code. See trace_events_user.c:user_event_set_print_fmt() as an example.

And don't forget that passing a copy of a user input as snprintf format string is
bad, because it allows user-space to provide event/fields names with e.g. "%s" or
"%p" and control the behavior of snprintf in unwanted ways, typically leading to
a kernel memory information leak to user-space.

Thanks,

Mathieu

>
>> Unfortunately this is all the time I have for review right now, but it is at
>> least a good starting
>> point for discussion.
>>
>> Thanks,
>>
>> Mathieu
>>
>> --
>> Mathieu Desnoyers
>> EfficiOS Inc.
>> http://www.efficios.com
>
> Thanks,
> -Beau

--
Mathieu Desnoyers
EfficiOS Inc.
http://www.efficios.com

\
 
 \ /
  Last update: 2022-03-29 18:18    [W:0.137 / U:0.428 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site