#38149

Seems okayish to me with caveats, as I said to Paul in private email; it seems to me that we ought to, at a minimum:

1. be very very careful about how we document this, since it can wedge an entire server immediately if things go wrong, and perhaps
2. even supply a small test program (a little python snippet is plenty; I may even have put it in the original bug) to allow people to test an offline server, not running ceph, or at least not involving it, to find out if they are going to have issues.

I had also suggested looking for a way to use sg directly (or more likely through libsgutils2.so.2), but I can't find any healthy Python bindings, or even documentation (other than the header files). The only advantage is trial-by-fire; sg has been through a lot of buggy hardware, and with SCSI that's a lot of hardware.

I know it may seem overcautious, but the consequences are large, and the affected hardware is really unknown. We've seen one instance but that definitely doesn't mean there aren't others (and I've got past experience at Sun with a lot of others).

@dmick the test could even be cephadm itself - i.e. cephadm ceph-volume inventory --with-lsm --format=json

Yeah, probably so

https://pulpito.ceph.com/swagner-2021-02-22_13:53:31-rados:cephadm-wip-swagner3-testing-2021-02-22-1135-distro-basic-smithi/

@pcuzner @dmick @tbzatek

I had also suggested looking for a way to use sg directly (or more likely through libsgutils2.so.2), but I can't find any healthy
Python bindings, or even documentation (other than the header files). The only advantage is trial-by-fire; sg has been through a lot of buggy hardware, and with SCSI that's a lot of hardware.

If you want the benefit of trial-by-fire with regards to sg utils I believe you need to leverage the sg command line utilities. For example, looking at the header files for issuing an inquiry command sg_ll_inquiry the caller needs to specify the return buffer size. If you look at the sg_inq utility source code there is a good comment on the approach one should take when issuing an inquiry, e.g. initial inquiry should be limited to 36 bytes. Thus the acquired knowledge of safely talking to devices appears lost if one utilizes the sg library calls.

@tasleson good point. Perhaps opt-in with libstoragemgmt is our safest option at this point.

@sebastian-philipp where should this be documented? I was thinking cephadm/osd.rst ?

@sebastian-philipp where should this be documented? I was thinking cephadm/osd.rst ?

yeah, maybe somewhere around https://docs.ceph.com/en/latest/cephadm/osd/#list-devices ?

@sebastian-philipp @dmick docs added to cover the benefit and potential issues that turning libstoragemgmt introduces.

https://pulpito.ceph.com/swagner-2021-03-09_13:58:13-rados:cephadm-wip-swagner-testing-2021-03-09-1014-distro-basic-smithi/

• https://tracker.ceph.com/issues/49633
• dashboard e2e only runs on centos