|
From: | Corey Minyard |
Subject: | Re: [Qemu-devel] [PATCH v2] Sort the fw_cfg file list |
Date: | Tue, 15 Mar 2016 11:36:39 -0500 |
User-agent: | Mozilla/5.0 (X11; Linux x86_64; rv:38.0) Gecko/20100101 Thunderbird/38.6.0 |
On 03/15/2016 07:45 AM, Michael S. Tsirkin wrote:
On Tue, Mar 15, 2016 at 07:38:43AM -0500, Corey Minyard wrote:On 03/15/2016 04:37 AM, Michael S. Tsirkin wrote:On Tue, Mar 15, 2016 at 09:45:22AM +0100, Gerd Hoffmann wrote:Depends on how you code it up. We have a list, we look each file there and sort accordingly. Fine. New devices will not be on this list, I guess you can just ignore them and guests will not see them. OK but I think it is better to make old machine types see them.Not a new fw_cfg file. It's existing smbios file which gets new records added by a new device. So when initializing it early (old order) it doesn't (yet) contain the new records. When initializing it late it has them, but also has a different place in the fw_cfg directory. So old machine types initialize smbios early (for compatibility).I see. So in this model, we'd have to somehow keep track of the old initialization order forever, and add hacks whenever we change it. IMHO That would just be too hard to maintain. I have an alternative proposal.New machine types initialize smbios late (so guests see the new records).So here is what I propose instead: - always initialize it late - sort late, a machine done, not when inserting entries - figure out what the order of existing entries is currently, and fill an array listing them in this order. for old machine types, insert the existing entries in this specific order by using a sorting function: qsort(....., fw_cfg_cmp);
I've hit a bit of a snag here. For 0.11 and before, PCI option ROMs were loaded via fw_cfg, not in the PCI ROM BAR. This causes two issues: * The order depends on the device initialization order, which I'm not sure is quantifiable. I believe it depends on how they are listed on the command line. * Users can load their own romfile with their own name, which means it can't be in the list. Also, for the ISA VGA ROMs, their order will also depend on the device list order. Outside of that, I have an order of file names. I think if I treat the device ROMs separately and handle them in init order, and then stick that device list in the proper location, that will work. Does that sound reasonable? Thanks, -corey
where: fw_cfg_find(a) { for (index = 0; index < fw_cfg_legacy_array_size; ++index) if (!strcmp(a, ...)) break; return index; } fw_cfg_cmp(a, b) { in cmp; if (legacy_fw_cfg_order) { int list1 = find(a); int list2 = find(b); if (list1 < list2) return -1; if (list1 > list2) return 1; } return strcmp(a, b); }Last night I had an idea something like this. Sorting by filename may not work because the user may pass in the file from the command line and you wouldn't be able to track the file name that way.command line files must all have a consistent prefix, so we can skip sorting them. I'll need to look at the code - don't they already? If not we IMHO absolutely must fix that before release and give them consistent prefixes.Instead, you could add a "legacy_order" parameter to the fw_cfg_add functions. Then figure out the current order add the numeric order to each call. Then sort by the numeric order. As long as you don't reorder things with the same numeric value I think that would work and be fairly simple to implement. New calls could pass in NO_FW_CFG_LEGACY_ORDER or something like that and be pasted onto the end in legacy mode. -coreyOK but it's a much larger change and less well contained.While mucking with the file ordering anyway: Good opportunity to make new machine types also sort the fw_cfg directory entries, so they get a fixed order independent from the order they are created, and we will not face this problem again. cheers, GerdWhat exactly do you mean by directory entries here?
[Prev in Thread] | Current Thread | [Next in Thread] |