On Mon, Sep 28, 2015 at 11:05:32PM +0200, Laszlo Ersek wrote: > On 09/28/15 22:56, Laszlo Ersek wrote: > > On 09/28/15 22:00, Eric Blake wrote: > >> On 09/28/2015 01:51 PM, Gabriel L. Somlo wrote: > >>> On Mon, Sep 28, 2015 at 01:46:33PM -0600, Eric Blake wrote: > >>>> On 09/28/2015 07:30 AM, Laszlo Ersek wrote: > >>>> > >>>>>> > >>>>>> +Small enough items may be provided directly as strings on the command > >>>>>> +line, using the syntax: > >>>>>> + > >>>>>> + -fw_cfg [name=]<item_name>,content=<string> > >>>>>> + > >>>>> > >>>>> Please consider spelling out that these blobs will NOT be NUL-terminated > >>>>> when viewed on the guest. (It kinda follows from all the other fw_cfg > >>>>> things, but once we leave host-side files for qemu command line strings, > >>>>> it might become non-obvious to users.) > >>>> > >>>> Or else GUARANTEE that it will be NUL-terminated (and the only way to > >>>> get blobs that are not NUL terminated is to use files rather than > >>>> content=). > >>> > >>> I went with the first suggestion (leave out the trailing '\0' from the > >>> blob payload, and say so in docs/specs/fw_cfg.txt) in v2 of the patch. > >>> > >>> Do you feel strongly about including the \0 ? Otherwise, we're already > >>> there :) > >> > >> I don't know what users are more likely to want to push through. A > >> trailing NUL implies a binary file (as text files cannot contain \0), > >> but even without a trailing NUL, a file is not a text file (per the > >> POSIX definition) unless it is either empty or ends in a newline. Use > >> of content=... is unlikely to have users remembering to place a newline > >> there if examples don't suggest it. And I don't know if guests are > >> expecting text data from these blobs, or are okay with binary blobs. > > > > fw_cfg blobs are considered binary, unless a specific selector key or > > fw_cfg file name makes different arrangements. (Described in QEMU docs, > > or elsewhere.) See more below. > > > >> That's a long-winded way of stating that omitting the NUL is fine by me, > >> as long as you document it, and as long as you are catering to the most > >> common user usage of the feature. > > > > The main consumer of the -fw_cfg switch is guest firmware (and, perhaps > > soon, the guest kernel too); the idea is to pass down firmware config > > items without QEMU being aware of their actual meaning. Therefore I'd > > like to see as little smarts as possible in QEMU wrt. the content passed > > down with -fw_cfg. > > > >> > >> Either that, or it's my way of dreaming about alternative 3: guarantee a > >> trailing newline (rather than NUL), so that 'content="abc"' on the > >> command line results in the 4-byte blob "abc\n" in the guest. > >> > > > > Please don't :) > > > > The current client code in OVMF (in effect for two specific fw_cfg file > > names) recognizes the following content pattern: > > > > [0nN1yY](\n|\r\n)? > > > > E.g., QEMU may pass in a simple host-side file as an fw_cfg entry on a > > Windows host too. If you edited that file with "notepad.exe", it'll have > > \r\n, or maybe no line terminator at all. Other (really binary) blobs > > (passed in with file=...) may have embedded \0 characters. > > > > I think such flexibility is best left to the firmware, or else should be > > restricted in specifications living outside of QEMU, and QEMU should be > > dumb and transparent here, in accordance with the original goal of this > > feature. > > > > Re: policy vs. mechanism, the opt/ prefix is also strongly recommended > > (for the names), but we don't enforce it. > > ... This made me think of the following language that Gabriel added in > v2 (at my request, and to my acceptance): > > > Both <item_name> and, if applicable, the content <string> are assumed > > to consist exclusively of plain ASCII characters. > > Now I think that this could be improved. I think we should say "should > consist" rather than "are assumed to consist", because neither the QEMU > nor the firmware(s) "assume" anything in general here -- that would be > policy --, we just want to help the user avoid shooting himself in the > foot (and reporting a bug), lest he pass non-ASCII UTF-8 on the command > line, and the firmware do surprising things. > > Maybe I should even retract my request for spelling out ASCII... That's > really not a requirement, just a high-level recommendation for humans > who develop guest code for this interface, to save their sanity.
Maybe something like this, then: Both <item_name> and, if applicable, the content <string> will be passed through by QEMU without any interpretation, expansion, or further processing. Any such processing (potentially performed by e.g., the shell) is outside QEMU's responsibility; as such, using plain ASCII characters is recommended. Let me know what you think. Thanks, --Gabriel