[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index] Re: [Xen-devel] [PATCH v3 1/2] libxl: Implement the handler to handle unrecoverable AER errors [and 1 more messages]
On 2018-04-03 17:51:50 +0100, Ian Jackson wrote: > Venu Busireddy writes ("Re: [PATCH v3 1/2] libxl: Implement the handler to > handle unrecoverable AER errors [and 1 more messages]"): > > On 2018-04-03 16:06:17 +0100, Ian Jackson wrote: > > > Ian Jackson writes ("Re: [PATCH v3 1/2] libxl: Implement the handler to > > > handle unrecoverable AER errors"): > > > > I'm afraid that I still have reservations about the design questions. > > > > Evidently I didn't make my questions clear enough. > > > > > > > > [ 64 lines of detailed discussion elided ] > > > > > > I haven't seen a reply to that. > > > > Reply to that is the v5 patch. Your concern in v4 was, "why is this > > error handling done only in some cases?" Meaning, the error handling > > happens only for guests created using xl, but it does not happen for > > guests created using libvirt. I addressed that in the v5 patch. Please > > see below for more details. > > Oh. I see. > > > > I'm confused by the responses in the thread which relate to libvirt. > > > ISTM that a libvirt patch is also required. Do you mean that in v5 > > > there is also a libvirt patch ? > > > > libvirt ends up calling do_domain_create() in tools/libxl/libxl_create.c, > > and that is where I am registering the error handler. That change takes > > care of guests created using xl command as well as libvirt. Hence there > > is no change in libvirt. > > I'm sorry to say that this is completely wrong. I didn't spot that > hunk in the v5 2/2 patch. I don't think your description in your v4 > to v5 changes summary really highlights the substantial design change. > > I think it would have been better to reply to my prose email. We > would have been able to explore the design possibilities. > > What you have done is wrong because: > > * You have removed the libxl__aer_watch from the > libxl_reg_aer_events_handler API which means the effect is now > global for the ctx. This is not correct for a libxl event > generation request function. (Although this isn't one.) > > * Not all callers of libxl will necessarily retain the process, or > the ctx, in which they called libxl_domain_create. I think libvirt > does (but I'm not sure), and xl usually does, but it's not > guaranteed. > > * It's quite unclear why this function is a public one. > > The entire approach is wrong, I'm afraid. > > We need to go back to the design. Please would you reply to my mail > from September. Sure. I will reply to your email from September. Venu > Thanks, > Ian. _______________________________________________ Xen-devel mailing list Xen-devel@xxxxxxxxxxxxxxxxxxxx https://lists.xenproject.org/mailman/listinfo/xen-devel
|
Lists.xenproject.org is hosted with RackSpace, monitoring our |