> On Mar 3, 2021, at 1:27 PM, Chuck Lever <chuck.lever@xxxxxxxxxx> wrote: > > > >> On Mar 3, 2021, at 1:25 PM, Bruce Fields <bfields@xxxxxxxxxxxx> wrote: >> >> On Wed, Mar 03, 2021 at 06:19:33PM +0000, Chuck Lever wrote: >>> >>> >>>> On Mar 3, 2021, at 11:52 AM, Bruce Fields <bfields@xxxxxxxxxxxx> wrote: >>>> >>>> On Wed, Mar 03, 2021 at 03:43:28PM +0000, Chuck Lever wrote: >>>>> Why would that not be OK? the next call to xdr_get_next_encode_buffer() >>>>> should do the right thing and bounce the new encoded data from the >>>>> next page into this one again. >>>>> >>>>> So far I have not encountered any problems. Would such a problem show >>>>> up with some frequency under normal use, or would it be especially >>>>> subtle? >>>> >>>> I mainly just want to make sure we've got a coherent idea what this code >>>> is doing.... >>> >>> Agreed, that's a good thing. >> >> I'm also a little vague on what exactly the problem is you're running >> into. (Probably because I haven't really looked at the v3 readdir >> encoding.) >> >> Is it approaching the end of a page, or is it running out of buflen? >> How exactly does it fail? > > I don't recall exactly, it was a late last summer when I wrote all these. > > Approaching the end of a page, IIRC, the unpatched code would leave > a gap in the directory entry stream. Well, when I converted the entry encoders to use xdr_stream, it would have a problem around the end of a page. The existing encoders are open-coded to deal with this case. >> --b. >> >>> >>> >>>> For testing: large replies that aren't just read data are readdir and >>>> getacl. So reading large directories with lots of variably-named files >>>> might be good. Also pynfs could easily send a single compound with lots >>>> of variable-sized reads, that might be interesting. >>> >>> I've run the full git regression suite over NFSv3 many many times with >>> this patch applied. That includes unpacking via tar, a full build from >>> scratch, and then running thousands of individual tests. >>> >>> So that doesn't exercise a particular corner case, but it does reflect >>> a broad variety of directory operations. >>> >>> >>>> Constructing a compound that will result in xdr_reserve_space calls that >>>> exactly hit the various corner cases may be hard. But maybe if we just >>>> send a bunch of compounds that vary over some range we can still >>>> guarantee hitting those cases. >>>> >>>> --b. >>> >>> -- >>> Chuck Lever > > -- > Chuck Lever -- Chuck Lever
