hubble/container: Properly deal with nil values in RingReader #11323

gandro · 2020-05-04T19:32:19Z

When the ring-buffer is not full, we must treat nil values the same as
reads outside of the readable indices. This commit ensures that read
treats nil values the same way readFrom does.

This has an impact on the RingReader, as we should not try to either
increase or decrease the index if the current value is unreadable.
If the current index is unreadable, then we have either read past the
last written value, or we have tried to access an index which has
already been overwritten.

This commit therefore adds an error return type to
RingReader.{Next,Prev} in order to determine which of the two cases
occured.

Most notably, Next may now return ErrInvalidData. This means this
means that the value it tried to access just got overwritten by the
writer from behind. In such cases we want to return an error, as
otherwise RingReader.Next would silently stop reading and
subsequentially terminate the request as if it had reached most
recent value in the ring buffer (which it has not).

The error is also added to Previous, however in that case
we only ever expect ErrInvalidData, as it traverses the ring buffer
in the opposite direction of the writer.

The check for concurrent writes is currently missing from readFrom,
which can suffer from the same problem. This will be added in a
follow-up PR.

Fixes: cilium/hubble#131

maintainer-s-little-helper · 2020-05-04T19:32:22Z

Commit 357e0a4b493be4c991d31b68f0a3b0f315366277 does not contain "Signed-off-by".

Please follow instructions provided in https://docs.cilium.io/en/stable/contributing/development/contributing_guide/#developer-s-certificate-of-origin

maintainer-s-little-helper · 2020-05-04T19:32:48Z

Commit 357e0a4b493be4c991d31b68f0a3b0f315366277 does not contain "Signed-off-by".

Please follow instructions provided in https://docs.cilium.io/en/stable/contributing/development/contributing_guide/#developer-s-certificate-of-origin

gandro · 2020-05-04T19:32:58Z

test-me-please

pkg/hubble/container/ring_reader_test.go

pkg/hubble/container/ring_reader.go

coveralls · 2020-05-04T19:56:04Z

Coverage increased (+0.03%) to 44.479% when pulling 4229ac8 on pr/gandro/ringreader-detect-concurrent-write into 5aeea78 on master.

gandro · 2020-05-04T20:18:39Z

test-me-please

glibsm · 2020-05-04T20:54:21Z

Are we expecting that the --since 10y is mostly fixed by this patch, or will those types of queries more often than not surface the concurrent writer error on busy nodes?

gandro · 2020-05-04T21:07:05Z

Are we expecting that the --since 10y is mostly fixed by this patch, or will those types of queries more often than not surface the concurrent writer error on busy nodes?

I believe they will suffer from the concurrent writer issue.

I'm currently thinking that the way to implement something like hubble observe --all (without changing the ring buffer implementation or introducing the lock) is to iterate the ring buffer from the most recent to the oldest entry. This way we can just stop once we hit the writer cursor.

gandro · 2020-05-04T21:22:21Z

I should also note: The concurrent write error only occurs if there really is a conflict, i.e. the events we were about to read got lost. If the reader is reading quick enough to stay ahead of the writer, it can actually "loop around", i.e. dump the ring buffer content which existed at the time the request was, followed by all flows that have been added since the request was started.

Me and @rolinh didn't consider this a bug, therefore I didn't "fix" it. It can however easily be avoided if we never read past the initial LastParallelWrite value. I would mean however that hubble observe will never the very most recent flows that occured after the request was started.

michi-covalent

now i see all the flows from ring buffer:

bash-5.0# hubble observe --last 10000 -o json | wc -l
4095
bash-5.0# hubble observe --since 1h -o json | wc -l
4095

sometimes i get more than len(ringbuffer) - 1 flows, but i guess this can happen if new flows are coming in as we read.

regarding @gandro 's comment, i think it's totally fine to stop at the initial LastParallelWrite index. you can always follow if you want to get more.

rolinh · 2020-05-05T07:58:15Z

sometimes i get more than len(ringbuffer) - 1 flows, but i guess this can happen if new flows are coming in as we read.

This feels non intuitive at first but it's definitely not a problem as long as the reader is able to eventually catch up with the writers.

regarding @gandro 's comment, i think it's totally fine to stop at the initial LastParallelWrite index. you can always follow if you want to get more.

The problem is not really to "not get enough" but that the caller will in effect miss the most recent flows. I'd be in favor of keeping the current behavior unless it's a source of bugs.

rolinh

Thanks for having invested time into this issue. I know the ring buffer is tricky to work with :)
I think this is definitely going to the right direction. That said, I have one concern with a suggested change. Could you have a look at it please?

pkg/hubble/container/ring_reader.go

gandro · 2020-05-05T12:47:19Z

test-me-please

gandro · 2020-05-05T12:50:17Z

pkg/hubble/container/ring_test.go

-	if !ok {
-		t.Errorf("Should be able to read position %x", lastWrite)
-	}
-	if entry != nil {
-		t.Errorf("Read Event should be %+v, got %+v instead", nil, entry)


Reviewer note: This test relied on read returning nil values. With this PR, we never allow nil values to be returned from read (same as readFrom, which had this restriction already for a while)

rolinh

I think we're almost there but I left a few comments. Note that the commit message is really hard to comprehend, eg:

If the current index is unreadable, then we have reached either end of the ring buffer.

What's the "end" of a ring structure?

Next may return ErrConcurrentRead

I think that the error name has changed.

rolinh · 2020-05-05T13:36:23Z

pkg/hubble/container/ring.go

+	// ErrConcurrentWrite indicates that a value was overwritten
+	ErrConcurrentWrite = errors.New("concurrent write detected")
+	// ErrEndOfBuffer is an alias for io.EOF
+	ErrEndOfBuffer = io.EOF


I don't think this alias is strictly necessary as io.EOF is widely accepted as meaning "end of whatever".
Besides, I'm still trying to understand what the "end" of a ring structure can be :)

rolinh · 2020-05-05T13:40:12Z

pkg/hubble/container/ring.go

@@ -24,6 +26,13 @@ import (
 	"github.com/cilium/cilium/pkg/lock"
 )

+var (
+	// ErrConcurrentWrite indicates that a value was overwritten
+	ErrConcurrentWrite = errors.New("concurrent write detected")


Your commit message mentions ErrConcurrentRead. I think that both names are misleading anyway as reading concurrently should not be a problem and a "concurrent write" error while reading is rather something unexpected. What about a more generic name like ErrInvalidRead for instance?

Oh, the ConcurrentRead is a typo in the commit messages. I'll think about the naming some more. I don't want it to be too generic, since the error is also presented to the user, and invalid read sounds like something is broken internally, when in fact they just tried to read stale data.

pkg/hubble/container/ring.go

pkg/hubble/container/ring_reader.go

rolinh · 2020-05-05T13:47:44Z

pkg/hubble/observer/local_observer.go

-		if e == nil {
+		e, err := reader.Previous()
+		if err != nil {
+			idx++ // we went backward 1 too far


we went backward 1 too far

Unless the error is io.EOF, no?

Kinda, I was not sure how to handle this properly. If Previous returns io.EOF, then something went horribly wrong. But yeah, we can check the error for ErrConcurrentWrite and propagate everything else.

gandro · 2020-05-05T17:00:28Z

test-me-please

pkg/hubble/container/ring.go

When the ring-buffer is not full, we must treat nil values the same as reads outside of the readable indices. This commit ensures that `read` treats nil values the same way `readFrom` does. This has an impact on the RingReader, as we should not try to either increase or decrease the index if the current value is unreadable. If the current index is unreadable, then we have either read past the last written value, or we have tried to access an index which has already been overwritten. This commit therefore adds an error return type to `RingReader.{Next,Prev}` in order to determine which of the two cases occured. Most notably, `Next` may now return ErrInvalidData. This means this means that the value it tried to access just got overwritten by the writer from behind. In such cases we want to return an error, as otherwise `RingReader.Next` would silently stop reading and subsequentially terminate the request as if it had reached most recent value in the ring buffer (which it has not). The error is also added to `Previous`, however in that case we only ever expect ErrInvalidData, as it traverses the ring buffer in the opposite direction of the writer. The check for concurrent writes is currently missing from `readFrom`, which can suffer from the same problem. This will be added in a follow-up PR. Fixes: cilium/hubble#131 Signed-off-by: Sebastian Wicki <[email protected]>

gandro · 2020-05-06T08:45:44Z

test-me-please

rolinh

Awesome, thanks for the great work! LGTM.

gandro added release-note/bug This PR fixes an issue in a previous release of Cilium. sig/hubble Impacts hubble server or relay labels May 4, 2020

gandro requested a review from a team May 4, 2020 19:32

maintainer-s-little-helper bot added the dont-merge/needs-sign-off The author needs to add signoff to their commits before merge. label May 4, 2020

gandro force-pushed the pr/gandro/ringreader-detect-concurrent-write branch from 357e0a4 to b06ee8f Compare May 4, 2020 19:32

glibsm reviewed May 4, 2020

View reviewed changes

pkg/hubble/container/ring_reader_test.go Outdated Show resolved Hide resolved

pkg/hubble/container/ring_reader.go Outdated Show resolved Hide resolved

gandro force-pushed the pr/gandro/ringreader-detect-concurrent-write branch from b06ee8f to 8835e85 Compare May 4, 2020 20:18

maintainer-s-little-helper bot removed the dont-merge/needs-sign-off The author needs to add signoff to their commits before merge. label May 4, 2020

michi-covalent approved these changes May 4, 2020

View reviewed changes

maintainer-s-little-helper bot added the ready-to-merge This PR has passed all tests and received consensus from code owners to merge. label May 4, 2020

rolinh requested changes May 5, 2020

View reviewed changes

pkg/hubble/container/ring_reader.go Outdated Show resolved Hide resolved

rolinh removed the ready-to-merge This PR has passed all tests and received consensus from code owners to merge. label May 5, 2020

gandro force-pushed the pr/gandro/ringreader-detect-concurrent-write branch from 8835e85 to bbbc6ba Compare May 5, 2020 12:46

gandro requested a review from rolinh May 5, 2020 12:47

gandro commented May 5, 2020

View reviewed changes

rolinh reviewed May 5, 2020

View reviewed changes

maintainer-s-little-helper bot added the ready-to-merge This PR has passed all tests and received consensus from code owners to merge. label May 5, 2020

gandro force-pushed the pr/gandro/ringreader-detect-concurrent-write branch from bbbc6ba to bee1bf7 Compare May 5, 2020 17:00

gandro requested a review from rolinh May 5, 2020 17:00

joestringer mentioned this pull request May 5, 2020

Contributor push does not invalidate the ready-to-merge flag cilium/github-actions#6

Closed

glibsm reviewed May 5, 2020

View reviewed changes

pkg/hubble/container/ring.go Outdated Show resolved Hide resolved

glibsm mentioned this pull request May 5, 2020

Add observe all cilium/hubble#253

Closed

aanm removed the ready-to-merge This PR has passed all tests and received consensus from code owners to merge. label May 6, 2020

gandro force-pushed the pr/gandro/ringreader-detect-concurrent-write branch from bee1bf7 to 4229ac8 Compare May 6, 2020 08:45

rolinh approved these changes May 6, 2020

View reviewed changes

maintainer-s-little-helper bot added the ready-to-merge This PR has passed all tests and received consensus from code owners to merge. label May 6, 2020

rolinh merged commit 8fab042 into master May 6, 2020

rolinh deleted the pr/gandro/ringreader-detect-concurrent-write branch May 6, 2020 12:12

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

hubble/container: Properly deal with nil values in RingReader #11323

hubble/container: Properly deal with nil values in RingReader #11323

gandro commented May 4, 2020 •

edited

Loading

maintainer-s-little-helper bot commented May 4, 2020

maintainer-s-little-helper bot commented May 4, 2020

gandro commented May 4, 2020

coveralls commented May 4, 2020 •

edited

Loading

gandro commented May 4, 2020

glibsm commented May 4, 2020

gandro commented May 4, 2020 •

edited

Loading

gandro commented May 4, 2020 •

edited

Loading

michi-covalent left a comment

rolinh commented May 5, 2020

rolinh left a comment

gandro commented May 5, 2020

gandro May 5, 2020

rolinh left a comment

rolinh May 5, 2020

rolinh May 5, 2020

gandro May 5, 2020 •

edited

Loading

rolinh May 5, 2020

gandro May 5, 2020 •

edited

Loading

gandro commented May 5, 2020

gandro commented May 6, 2020

rolinh left a comment

hubble/container: Properly deal with nil values in RingReader #11323

hubble/container: Properly deal with nil values in RingReader #11323

Conversation

gandro commented May 4, 2020 • edited Loading

maintainer-s-little-helper bot commented May 4, 2020

maintainer-s-little-helper bot commented May 4, 2020

gandro commented May 4, 2020

coveralls commented May 4, 2020 • edited Loading

gandro commented May 4, 2020

glibsm commented May 4, 2020

gandro commented May 4, 2020 • edited Loading

gandro commented May 4, 2020 • edited Loading

michi-covalent left a comment

Choose a reason for hiding this comment

rolinh commented May 5, 2020

rolinh left a comment

Choose a reason for hiding this comment

gandro commented May 5, 2020

gandro May 5, 2020

Choose a reason for hiding this comment

rolinh left a comment

Choose a reason for hiding this comment

rolinh May 5, 2020

Choose a reason for hiding this comment

rolinh May 5, 2020

Choose a reason for hiding this comment

gandro May 5, 2020 • edited Loading

Choose a reason for hiding this comment

rolinh May 5, 2020

Choose a reason for hiding this comment

gandro May 5, 2020 • edited Loading

Choose a reason for hiding this comment

gandro commented May 5, 2020

gandro commented May 6, 2020

rolinh left a comment

Choose a reason for hiding this comment

gandro commented May 4, 2020 •

edited

Loading

coveralls commented May 4, 2020 •

edited

Loading

gandro commented May 4, 2020 •

edited

Loading

gandro commented May 4, 2020 •

edited

Loading

gandro May 5, 2020 •

edited

Loading

gandro May 5, 2020 •

edited

Loading