> On Nov 12, 2022, at 4:04 AM, yangerkun <yangerkun@xxxxxxxxxx> wrote: > > On 2022/11/12 13:01, yangerkun wrote: >> Hi, Chuck Lever, >> CVE-2022-43945(https://nvd.nist.gov/vuln/detail/CVE-2022-43945) describe that a normal request header ended with garbage data can trigger the nfsd overflow since nfsd share the request and response with the same pages array. >> It seems that the patchset(https://lore.kernel.org/linux-nfs/166204973526.1435.6068003336048840051.stgit@xxxxxxxxxxxxxxxxxxxxx/T/#t) has solved NFSv2/NFSv3, but leave NFSv4 still vulnerably? I asked the folks who reported this issue to check NFSv4 as well. They were not able to exploit NFSv4 in the same way. For now we believe this vulnerability does not impact the NFSv4 code paths. >> Another question, for stable branch like lts-5.10, since NFSv2/NFSv3 did not switch to xdr_stream, the nfs_request_too_big in nfsd_dispatch will reject the request like READ/READDIR with too large request. So it seems branch without that "switch" seems ok for NFSv2/NFSv3, but NFSv3 still vulnerably. right? >> Looking forward to your reply! > > Sorry, notice that 76ce4dcec0dc ("NFSD: Cap rsize_bop result based on send buffer size") fix same problem for NFSv4. 76ce4dcec0dc is a defensive fix. But, as I stated above, we haven't yet found that NFSD's NFSv4 implementation is vulnerable to this issue. > So, for the stable branch like lts-5.10 which NFSv2/NFSv3 do not switch to xdr_stream, it seems we only need 76ce4dcec0dc"NFSD: Cap rsize_bop result based on send buffer size"). Right? At this time we don't believe 76ce4dcec0dc is required. But if you want it applied to v5.10 (or any LTS kernel) please first test that it does not result in a regression, and then make a request to the usual stable maintainers. -- Chuck Lever
