Date: Wed, 8 Feb 2023 07:27:53 -0700 From: Alan Somers <asomers@freebsd.org> To: Enji Cooper <yaneurabeya@gmail.com> Cc: FreeBSD Hackers <freebsd-hackers@freebsd.org> Subject: Re: RFC: GEOM and hard disk LEDs Message-ID: <CAOtMX2gCLqWRXZ8RWS9Cmpa2-1upg=PQWzH3g%2BSw%2BrJRo4mW%2Bg@mail.gmail.com> In-Reply-To: <0155CD68-849E-40D6-95A5-6AAD5E229A57@gmail.com> References: <CAOtMX2h5=xCH1CN3ravsOLfmeMq=1-BOw1aSFGKVP-js6g38GA@mail.gmail.com> <CAOtMX2hH5gEMj4tO2rq-x2DeRuB-o6RD83i-V1LB8F2T77GRCw@mail.gmail.com> <0155CD68-849E-40D6-95A5-6AAD5E229A57@gmail.com>
next in thread | previous in thread | raw e-mail | index | archive | help
On Tue, Feb 7, 2023 at 11:56 PM Enji Cooper <yaneurabeya@gmail.com> wrote: > > > > On Feb 7, 2023, at 3:30 PM, Alan Somers <asomers@FreeBSD.org> wrote: > > > > Most modern SES backplanes have two LEDs per hard disk. There's a > > "fault" LED and a "locate" LED. You can control either one with > > sesutil(8) or, with a little more work, sg_ses from > > sysutils/sg3_utils. They're very handy for tasks like replacing a > > failed disk, especially in large enclosures. However, there isn't any > > way to automatically control them. It would be very convenient if, > > for example, zfsd(8) could do it. Basically, it would just set the > > fault LED for any disk that has been kicked out of a ZFS pool, and > > clear it for any disk that is healthy or is being resilvered. But > > zfsd does not do that. Instead, users' only options are to write a > > custom daemon or to use sesutil by hand. Instead of forcing all of us > > to write our own custom daemons, why not train zfsd to do it? > > > > My proposal is to add boolean GEOM attributes for "fault" and > > "locate". A userspace program would be able to look up their values > > for any geom with DIOCGATTR. Setting them would require a new ioctl > > (DIOCSATTR?). The disk class would issue a ENCIOC_SETELMSTAT to > > actually change the LEDs whenever this attribute changes. GEOM > > transforms such as geli would simply pass the attribute through to > > lower layers. Many-to-one transforms like gmultipath would pass the > > attribute through to all lower layers. zfsd could then set all vdevs' > > fault attributes when it starts up, and adjust individual disk's as > > appropriate on an event-driven basis. > > > > Questions: > > > > * Are there any obvious flaws in this plan, any reasons why GEOM > > attributes can't be used this way? > > > > * For one-to-many transforms like gpart the correct behavior is less > > clear: what if a disk has two partitions in two different pools, and > > one of them is healthy but the other isn't? > > > > * Besides ZFS, are there any other systems that could take advantage? > > > > * SATA enclosures uses SGPIO instead of SES. SGPIO is too limited, > > IMHO, to be of almost any use at all. I suggest not even trying to > > make it work with this scheme. > > Out of curiosity, is there a way that a client of zfsd could instead reach out to zfsd via a pipe, etc, when an event triggered by devd occurs? > Just trying to think of a way that would result in a cleaner architecture than having zfsd doing a busy-loop of some kind waiting for an event to happen. > Thanks, > -Enji What do you mean a "client of zfsd"? The current architecture is that zfsd blocks waiting for events from devd. It reads from /var/run/devd.seqpacket.pipe.
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?CAOtMX2gCLqWRXZ8RWS9Cmpa2-1upg=PQWzH3g%2BSw%2BrJRo4mW%2Bg>