[PATCH] livepatch: Avoid CPU hogging with cond_resched

When initializing a 'struct klp_object' in klp_init_object_loaded(), and
performing relocations in klp_resolve_symbols(), klp_find_object_symbol()
is invoked to look up the address of a symbol in an already-loaded module
(or vmlinux). This, in turn, calls kallsyms_on_each_symbol() or
module_kallsyms_on_each_symbol() to find the address of the symbol that is
being patched.

It turns out that symbol lookups often take up the most CPU time when
enabling and disabling a patch, and may hog the CPU and cause other tasks
on that CPU's runqueue to starve -- even in paths where interrupts are
enabled.  For example, under certain workloads, enabling a KLP patch with
many objects or functions may cause ksoftirqd to be starved, and thus for
interrupts to be backlogged and delayed. This may end up causing TCP
retransmits on the host where the KLP patch is being applied, and in
general, may cause any interrupts serviced by softirqd to be delayed while
the patch is being applied.

So as to ensure that kallsyms_on_each_symbol() does not end up hogging the
CPU, this patch adds a call to cond_resched() in kallsyms_on_each_symbol()
and module_kallsyms_on_each_symbol(), which are invoked when doing a symbol
lookup in vmlinux and a module respectively.  Without this patch, if a
live-patch is applied on a 36-core Intel host with heavy TCP traffic, a
~10x spike is observed in TCP retransmits while the patch is being applied.
Additionally, collecting sched events with perf indicates that ksoftirqd is
awakened ~1.3 seconds before it's eventually scheduled.  With the patch, no
increase in TCP retransmit events is observed, and ksoftirqd is scheduled
shortly after it's awakened.

Signed-off-by: David Vernet <void@manifault.com>
---
 kernel/kallsyms.c | 1 +
 kernel/module.c   | 2 ++
 2 files changed, 3 insertions(+)

diff --git a/kernel/kallsyms.c b/kernel/kallsyms.c
index 0ba87982d017..2a9afe484aec 100644
--- a/kernel/kallsyms.c
+++ b/kernel/kallsyms.c

@@ -223,6 +223,7 @@ int kallsyms_on_each_symbol(int (*fn)(void *, const char *, struct module *,
 		ret = fn(data, namebuf, NULL, kallsyms_sym_address(i));
 		if (ret != 0)
 			return ret;
+		cond_resched();
 	}
 	return 0;
 }

diff --git a/kernel/module.c b/kernel/module.c
index 40ec9a030eec..c96160f7f3f5 100644
--- a/kernel/module.c
+++ b/kernel/module.c

@@ -4462,6 +4462,8 @@ int module_kallsyms_on_each_symbol(int (*fn)(void *, const char *,
 				 mod, kallsyms_symbol_value(sym));
 			if (ret != 0)
 				goto out;
+
+			cond_resched();
 		}
 	}
 out:

-- 
2.30.2

On Wed, 29 Dec 2021, David Vernet wrote:

When initializing a 'struct klp_object' in klp_init_object_loaded(), and
performing relocations in klp_resolve_symbols(), klp_find_object_symbol()
is invoked to look up the address of a symbol in an already-loaded module
(or vmlinux). This, in turn, calls kallsyms_on_each_symbol() or
module_kallsyms_on_each_symbol() to find the address of the symbol that is
being patched.

It turns out that symbol lookups often take up the most CPU time when
enabling and disabling a patch, and may hog the CPU and cause other tasks
on that CPU's runqueue to starve -- even in paths where interrupts are
enabled.  For example, under certain workloads, enabling a KLP patch with
many objects or functions may cause ksoftirqd to be starved, and thus for
interrupts to be backlogged and delayed. This may end up causing TCP
retransmits on the host where the KLP patch is being applied, and in
general, may cause any interrupts serviced by softirqd to be delayed while
the patch is being applied.

So as to ensure that kallsyms_on_each_symbol() does not end up hogging the
CPU, this patch adds a call to cond_resched() in kallsyms_on_each_symbol()
and module_kallsyms_on_each_symbol(), which are invoked when doing a symbol
lookup in vmlinux and a module respectively.  Without this patch, if a
live-patch is applied on a 36-core Intel host with heavy TCP traffic, a
~10x spike is observed in TCP retransmits while the patch is being applied.
Additionally, collecting sched events with perf indicates that ksoftirqd is
awakened ~1.3 seconds before it's eventually scheduled.  With the patch, no
increase in TCP retransmit events is observed, and ksoftirqd is scheduled
shortly after it's awakened.

Signed-off-by: David Vernet <void@manifault.com>

Acked-by: Miroslav Benes <mbenes@suse.cz>

I had similar ideas Petr mentioned elsewhere, but I, also, have no strong 
opinion about it. Especially when livepatch is the only user of the said 
interface.

M

On Wed, Dec 29, 2021 at 1:57 PM David Vernet [off-list ref] wrote:

When initializing a 'struct klp_object' in klp_init_object_loaded(), and
performing relocations in klp_resolve_symbols(), klp_find_object_symbol()
is invoked to look up the address of a symbol in an already-loaded module
(or vmlinux). This, in turn, calls kallsyms_on_each_symbol() or
module_kallsyms_on_each_symbol() to find the address of the symbol that is
being patched.

It turns out that symbol lookups often take up the most CPU time when
enabling and disabling a patch, and may hog the CPU and cause other tasks
on that CPU's runqueue to starve -- even in paths where interrupts are
enabled.  For example, under certain workloads, enabling a KLP patch with
many objects or functions may cause ksoftirqd to be starved, and thus for
interrupts to be backlogged and delayed. This may end up causing TCP
retransmits on the host where the KLP patch is being applied, and in
general, may cause any interrupts serviced by softirqd to be delayed while
the patch is being applied.

So as to ensure that kallsyms_on_each_symbol() does not end up hogging the
CPU, this patch adds a call to cond_resched() in kallsyms_on_each_symbol()
and module_kallsyms_on_each_symbol(), which are invoked when doing a symbol
lookup in vmlinux and a module respectively.  Without this patch, if a
live-patch is applied on a 36-core Intel host with heavy TCP traffic, a
~10x spike is observed in TCP retransmits while the patch is being applied.
Additionally, collecting sched events with perf indicates that ksoftirqd is
awakened ~1.3 seconds before it's eventually scheduled.  With the patch, no
increase in TCP retransmit events is observed, and ksoftirqd is scheduled
shortly after it's awakened.

Signed-off-by: David Vernet <void@manifault.com>

Acked-by: Song Liu <song@kernel.org>

PS: Do we observe livepatch takes a longer time to load after this change?
(I believe longer time shouldn't be a problem at all. Just curious.)

On Thu 2022-01-06 16:21:18, Song Liu wrote:

On Wed, Dec 29, 2021 at 1:57 PM David Vernet [off-list ref] wrote:

quoted

When initializing a 'struct klp_object' in klp_init_object_loaded(), and
performing relocations in klp_resolve_symbols(), klp_find_object_symbol()
is invoked to look up the address of a symbol in an already-loaded module
(or vmlinux). This, in turn, calls kallsyms_on_each_symbol() or
module_kallsyms_on_each_symbol() to find the address of the symbol that is
being patched.

It turns out that symbol lookups often take up the most CPU time when
enabling and disabling a patch, and may hog the CPU and cause other tasks
on that CPU's runqueue to starve -- even in paths where interrupts are
enabled.  For example, under certain workloads, enabling a KLP patch with
many objects or functions may cause ksoftirqd to be starved, and thus for
interrupts to be backlogged and delayed. This may end up causing TCP
retransmits on the host where the KLP patch is being applied, and in
general, may cause any interrupts serviced by softirqd to be delayed while
the patch is being applied.

So as to ensure that kallsyms_on_each_symbol() does not end up hogging the
CPU, this patch adds a call to cond_resched() in kallsyms_on_each_symbol()
and module_kallsyms_on_each_symbol(), which are invoked when doing a symbol
lookup in vmlinux and a module respectively.  Without this patch, if a
live-patch is applied on a 36-core Intel host with heavy TCP traffic, a
~10x spike is observed in TCP retransmits while the patch is being applied.
Additionally, collecting sched events with perf indicates that ksoftirqd is
awakened ~1.3 seconds before it's eventually scheduled.  With the patch, no
increase in TCP retransmit events is observed, and ksoftirqd is scheduled
shortly after it's awakened.

Signed-off-by: David Vernet <void@manifault.com>

Acked-by: Song Liu <song@kernel.org>

PS: Do we observe livepatch takes a longer time to load after this change?
(I believe longer time shouldn't be a problem at all. Just curious.)

It should depend on the load of the system and the number of patched
symbols. The module is typically loaded with a normal priority
process.

The commit message talks about 1.3 seconds delay of ksoftirq. In
principle, the change caused that this 1.3 sec of a single CPU time
was interleaved with other scheduled tasks on the same CPU. I would
expect that it prolonged the load just by a couple of seconds in
the described use case.

Note that the change has effect only with voluntary scheduling.
Well, it is typically used on servers where the livepatching
makes sense.

Best Regards,
Petr

On Wed 2021-12-29 13:56:47, David Vernet wrote:

When initializing a 'struct klp_object' in klp_init_object_loaded(), and
performing relocations in klp_resolve_symbols(), klp_find_object_symbol()
is invoked to look up the address of a symbol in an already-loaded module
(or vmlinux). This, in turn, calls kallsyms_on_each_symbol() or
module_kallsyms_on_each_symbol() to find the address of the symbol that is
being patched.

It turns out that symbol lookups often take up the most CPU time when
enabling and disabling a patch, and may hog the CPU and cause other tasks
on that CPU's runqueue to starve -- even in paths where interrupts are
enabled.  For example, under certain workloads, enabling a KLP patch with
many objects or functions may cause ksoftirqd to be starved, and thus for
interrupts to be backlogged and delayed. This may end up causing TCP
retransmits on the host where the KLP patch is being applied, and in
general, may cause any interrupts serviced by softirqd to be delayed while
the patch is being applied.

So as to ensure that kallsyms_on_each_symbol() does not end up hogging the
CPU, this patch adds a call to cond_resched() in kallsyms_on_each_symbol()
and module_kallsyms_on_each_symbol(), which are invoked when doing a symbol
lookup in vmlinux and a module respectively.  Without this patch, if a
live-patch is applied on a 36-core Intel host with heavy TCP traffic, a
~10x spike is observed in TCP retransmits while the patch is being applied.
Additionally, collecting sched events with perf indicates that ksoftirqd is
awakened ~1.3 seconds before it's eventually scheduled.  With the patch, no
increase in TCP retransmit events is observed, and ksoftirqd is scheduled
shortly after it's awakened.

Signed-off-by: David Vernet <void@manifault.com>

OK, there was not any strong pushback.

I have committed the patch into livepatch.git, branch for-5.17/kallsyms.

Best Regards,
Petr

On 12/29/21 4:56 PM, David Vernet wrote:

For example, under certain workloads, enabling a KLP patch with
many objects or functions may cause ksoftirqd to be starved, and thus for
interrupts to be backlogged and delayed.

Just curious, approximately how many objects/functions does it take to
hit this condition?  While the livepatching kselftests are more about
API and kernel correctness, this sounds like an interesting test case
for some of the other (out of tree) test suites.

Thanks,
-- 
Joe

On Fri, Jan 7, 2022 at 6:13 AM Joe Lawrence [off-list ref] wrote:

On 12/29/21 4:56 PM, David Vernet wrote:

quoted

For example, under certain workloads, enabling a KLP patch with
many objects or functions may cause ksoftirqd to be starved, and thus for
interrupts to be backlogged and delayed.

Just curious, approximately how many objects/functions does it take to
hit this condition?  While the livepatching kselftests are more about
API and kernel correctness, this sounds like an interesting test case
for some of the other (out of tree) test suites.

Not many patched functions. We only do small fixes at the moment. In the recent
example, we hit the issue with ~10 patched functions. Another version
with 2 to 3
patched function seems fine.

Yes, I think this is an important test case.

Song

Petr Mladek [off-list ref] wrote on Fri [2022-Jan-07 09:17:52 +0100]:

On Thu 2022-01-06 16:21:18, Song Liu wrote:

quoted

PS: Do we observe livepatch takes a longer time to load after this change?
(I believe longer time shouldn't be a problem at all. Just curious.)

It should depend on the load of the system and the number of patched
symbols. The module is typically loaded with a normal priority
process.

The commit message talks about 1.3 seconds delay of ksoftirq. In
principle, the change caused that this 1.3 sec of a single CPU time
was interleaved with other scheduled tasks on the same CPU. I would
expect that it prolonged the load just by a couple of seconds in
the described use case.

Just to add a data point here for the record, I didn't observe any increase
in latency when applying a livepatch with this change. It took roughly ~2
+/- ~.5 seconds both with and without the change. Obviously that number
doesn't mean much without knowing the specifics of the patch and what
workloads were running on the host, but in general the invocations to
cond_resched() patch did not seem to materially affect the overhead of
livepatching.

The only time I would expect it to cause longer livepatches is when there
are a lot of tasks that need the CPU, which is dependent on system load as
Petr said. I'd argue that in this case, the patch would be working as
intended as hogging the CPU for 1 or more seconds seems like a recipe for
problems.

On 1/7/22 11:46 AM, Song Liu wrote:

On Fri, Jan 7, 2022 at 6:13 AM Joe Lawrence [off-list ref] wrote:

quoted

On 12/29/21 4:56 PM, David Vernet wrote:

quoted

For example, under certain workloads, enabling a KLP patch with
many objects or functions may cause ksoftirqd to be starved, and thus for
interrupts to be backlogged and delayed.

Just curious, approximately how many objects/functions does it take to
hit this condition?  While the livepatching kselftests are more about
API and kernel correctness, this sounds like an interesting test case
for some of the other (out of tree) test suites.

Not many patched functions. We only do small fixes at the moment. In the recent
example, we hit the issue with ~10 patched functions. Another version
with 2 to 3
patched function seems fine.

Yes, I think this is an important test case.

Thanks, Song.  If you can share any test setup details, I'll pass those
along to our internal QE group.  And once merged, we'll be adding this
one to the list of backports for our distro.

-- 
Joe

On Mon, Jan 10, 2022 at 8:17 AM Joe Lawrence [off-list ref] wrote:

On 1/7/22 11:46 AM, Song Liu wrote:

quoted

On Fri, Jan 7, 2022 at 6:13 AM Joe Lawrence [off-list ref] wrote:

quoted

On 12/29/21 4:56 PM, David Vernet wrote:

quoted

For example, under certain workloads, enabling a KLP patch with
many objects or functions may cause ksoftirqd to be starved, and thus for
interrupts to be backlogged and delayed.

Just curious, approximately how many objects/functions does it take to
hit this condition?  While the livepatching kselftests are more about
API and kernel correctness, this sounds like an interesting test case
for some of the other (out of tree) test suites.

Not many patched functions. We only do small fixes at the moment. In the recent
example, we hit the issue with ~10 patched functions. Another version
with 2 to 3
patched function seems fine.

Yes, I think this is an important test case.

Thanks, Song.  If you can share any test setup details, I'll pass those
along to our internal QE group.  And once merged, we'll be adding this
one to the list of backports for our distro.

We get this on shadow production traffic of our web service, which we cannot
recreate out of our data centers.
I tried to use iperf to generate traffic (among a few hosts), but it
doesn't work.
At the moment, I am not sure what's the easiest way to repro this.

Thanks,
Song