Date: Tue, 25 Jun 2013 11:32:47 -0600 From: Warner Losh <imp@bsdimp.com> To: John Baldwin <jhb@freebsd.org> Cc: new-bus@freebsd.org Subject: Re: [PATCH] Change the PCI bus driver to free resources leaked by drivers Message-ID: <A777E5BC-56C0-4387-952F-0C916FE43A70@bsdimp.com> In-Reply-To: <201306251151.01717.jhb@freebsd.org> References: <201306241659.15119.jhb@freebsd.org> <201306250837.24093.jhb@freebsd.org> <BC40848E-5C75-451E-9B06-70F6B34E7950@bsdimp.com> <201306251151.01717.jhb@freebsd.org>
next in thread | previous in thread | raw e-mail | index | archive | help
On Jun 25, 2013, at 9:51 AM, John Baldwin wrote: > On Tuesday, June 25, 2013 10:15:00 am Warner Losh wrote: >>=20 >> On Jun 25, 2013, at 6:37 AM, John Baldwin wrote: >>=20 >>> On Tuesday, June 25, 2013 12:43:35 am Warner Losh wrote: >>>>=20 >>>> On Jun 24, 2013, at 2:59 PM, John Baldwin wrote: >>>>=20 >>>>> Currently our driver model trusts drivers to DTRT and properly = release any=20 >>>>> resources they allocated during probe() and attach(). I've added = a new >>>>> resource_list() helper method to release active resources on a = resource >>>>> list and used this to write a pci_child_detached() which cleans up = any >>>>> active resources when a device fails to probe or a driver finishes >>>>> detach. It also fixes an issue where we did not power down = devices when >>>>> the driver was detached (e.g. via kldunload). I've tested the = resource >>>>> bits by writing a dummy driver that intentionally attached to an = unattached >>>>> device and leaked a memory BAR and verified that the bus warned = about the >>>>> leak and cleaned it up. >>>>>=20 >>>>> http://www.FreeBSD.org/~jhb/patches/pci_clean_detach.patch >>>>=20 >>>> I think most of pci_child_detached() could be a generic thing = (except for the >>>> weird interaction with the msi-wart). This is likely fixable. >>>=20 >>> The existing design we've gone with for this sort of thing is to = provide >>> resource_list helpers, but let each bus driver decide which types of >>> resources it manages (see bus_print_child). Also, I think the order >>> matters (interrupts before memory & I/O). One thing the patch = doesn't do >>> currently is explicitly list the resource ranges being freed. >>=20 >> Yea, if ordering didn't matter, freeing them all would be a snap... >>=20 >>>> We don't tear down any interrupt handlers that the device = established. This >>>> is fixable, but the PCI bus would need to start tracking interrupts = that are >>>> established... >>>=20 >>> Eh, that is the part I don't like. That would be a lot of non-PCI = specific >>> crap in the PCI bus driver. >>=20 >> I'm not sure I understand this objection. We do (or did at one time) = this sort >> of thing for the cardbus code and it found a few bugs in a couple of = drivers... >=20 > I would rather solve this problem more generically so that if we want = to add > a child_detached method for other buses like ACPI then they can just = use a > library call (e.g. a bus_revoke_intr()) instead of having to do all = the same > tracking themselves. The "right" place for this seems to be in = whatever is > providing the IRQ resources and handles the actual bus_setup_intr = calls to > create cookies, etc. On x86 this is the nexus. On other platforms it = may > be in a nexus-like driver. I can work at prototyping something for = review as > a next step. I'd rather have that done at the interrupt controller level, which sadly = we don't model and put most, but not quite all, of the functionality in = the nexus driver. My experience is a bit colored from the PC Card and = CardBus stuff, since there we intercepted the setup_intr calls and did = the interrupt pass through directly for the child to cope with the = sudden removal cases where the bridge driver know the card was gone, so = the interrupt was dispatched to it, and only further dispatched to the = child if the bridge thought it was still there. I'd assumed we'd need to = that for proper support of hot-plug PCIe (including surprise removal = support), but since we don't have that, I guess the PCI code just passes = everything to the nexus. I think that's a long way of saying that this is a good path, and we may = have code on x86 that could benefit from it now that's doing ad-hoc = things... I'd love to see this next step, and the current patch you have = is sufficient for the more constrained problem it is trying to solve. Warner=
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?A777E5BC-56C0-4387-952F-0C916FE43A70>