[Bug 1921749] [NEW] nautilus: ceph radosgw beast frontend coroutine stack corruption

[Mauricio Faria de Oliveira]

Public bug reported:

[Impact]

The radosgw beast frontend in ceph nautilus might hit coroutine stack
corruption on startup and requests.

This is usually observed right at the startup of the ceph-radosgw systemd unit; sometimes 1 minute later.
But it might occur any time handling requests, depending on coroutine/request's function path/stack size.

The symptoms are usually a crash with stack trace listing TCMalloc (de)allocate/release to central cache,
but less rare signs are large allocs in the _terabytes_ range (pointer to stack used as allocation size)
and stack traces showing function return addresses (RIP) that are actually pointers to an stack address.

This is not widely hit in Ubuntu as most deployments use the ceph-radosgw charm that hardcodes 'civetweb'
as rgw frontend, which is _not_ affected; custom/cephadm deployments that choose 'beast' might hit this.

  @ charm-ceph-radosgw/templates/ceph.conf
        rgw frontends = civetweb port={{ port }}

Let's report this LP bug for documentation and tracking purposes until
UCA gets the fixes.

[Fix]

This has been reported by an Ubuntu Advantage user, and another user in ceph tracker #47910 [1].
This had been reported and fixed in Octopus [2] (confirmed by UA user; no longer affected.)

The Nautilus backport has recently been merged [3, 4] and should be
available in v14.2.19.

[Test Case]

The conditions to trigger the bug aren't clear, but apparently related to EC pools w/ very large buckets,
and of course the radosgw frontend beast being enabled (civetweb is not affected.)

[Where problems could occur]

The fixes are restricted to the beast frontend, specifically to the coroutines used to handle requests.
So problems would probably be seen in request handling only with the beast frontend.
Workarounds thus include switching back to the civetweb frontend.

This changes core/base parts of the RGW beast frontend code, but are in place from Octopus released.
The other user/reporter in the ceph tracker has been using the patches for weeks with no regression;
the ceph tests have passed and likely serious issues would be caught by ceph CI upstream.

[1] https://tracker.ceph.com/issues/47910 report tracker (nautilus)
[2] https://tracker.ceph.com/issues/43739 master tracker (octopus)
[3] https://tracker.ceph.com/issues/43921 backport tracker (nautilus)
[4] https://github.com/ceph/ceph/pull/39947 github PR

** Affects: cloud-archive
     Importance: Undecided
         Status: Fix Released

** Affects: cloud-archive/train
     Importance: Undecided
         Status: Confirmed

** Affects: ceph (Ubuntu)
     Importance: Undecided
         Status: Fix Released

** Changed in: ceph (Ubuntu)
       Status: New => Fix Released

** Also affects: cloud-archive
   Importance: Undecided
       Status: New

** Also affects: cloud-archive/train
   Importance: Undecided
       Status: New

** Changed in: cloud-archive
       Status: New => Fix Released

** Changed in: cloud-archive/train
       Status: New => Confirmed

** Summary changed:

- nautilus: ceph radosgw beast frontend might hit coroutine stack corruption
+ nautilus: ceph radosgw beast frontend coroutine stack corruption

-- 
You received this bug notification because you are a member of Ubuntu
OpenStack, which is subscribed to ceph in Ubuntu.
https://bugs.launchpad.net/bugs/1921749

Title:
  nautilus: ceph radosgw beast frontend coroutine stack corruption

Status in Ubuntu Cloud Archive:
  Fix Released
Status in Ubuntu Cloud Archive train series:
  Confirmed
Status in ceph package in Ubuntu:
  Fix Released

Bug description:
  [Impact]

  The radosgw beast frontend in ceph nautilus might hit coroutine stack
  corruption on startup and requests.

  This is usually observed right at the startup of the ceph-radosgw systemd unit; sometimes 1 minute later.
  But it might occur any time handling requests, depending on coroutine/request's function path/stack size.

  The symptoms are usually a crash with stack trace listing TCMalloc (de)allocate/release to central cache,
  but less rare signs are large allocs in the _terabytes_ range (pointer to stack used as allocation size)
  and stack traces showing function return addresses (RIP) that are actually pointers to an stack address.

  This is not widely hit in Ubuntu as most deployments use the ceph-radosgw charm that hardcodes 'civetweb'
  as rgw frontend, which is _not_ affected; custom/cephadm deployments that choose 'beast' might hit this.

    @ charm-ceph-radosgw/templates/ceph.conf
          rgw frontends = civetweb port={{ port }}

  Let's report this LP bug for documentation and tracking purposes until
  UCA gets the fixes.

  [Fix]

  This has been reported by an Ubuntu Advantage user, and another user in ceph tracker #47910 [1].
  This had been reported and fixed in Octopus [2] (confirmed by UA user; no longer affected.)

  The Nautilus backport has recently been merged [3, 4] and should be
  available in v14.2.19.

  [Test Case]

  The conditions to trigger the bug aren't clear, but apparently related to EC pools w/ very large buckets,
  and of course the radosgw frontend beast being enabled (civetweb is not affected.)

  [Where problems could occur]

  The fixes are restricted to the beast frontend, specifically to the coroutines used to handle requests.
  So problems would probably be seen in request handling only with the beast frontend.
  Workarounds thus include switching back to the civetweb frontend.

  This changes core/base parts of the RGW beast frontend code, but are in place from Octopus released.
  The other user/reporter in the ceph tracker has been using the patches for weeks with no regression;
  the ceph tests have passed and likely serious issues would be caught by ceph CI upstream.

  [1] https://tracker.ceph.com/issues/47910 report tracker (nautilus)
  [2] https://tracker.ceph.com/issues/43739 master tracker (octopus)
  [3] https://tracker.ceph.com/issues/43921 backport tracker (nautilus)
  [4] https://github.com/ceph/ceph/pull/39947 github PR

To manage notifications about this bug go to:
https://bugs.launchpad.net/cloud-archive/+bug/1921749/+subscriptions