On 09/05/13 18:53, Juan Lang wrote:
[+wine-devel]
Hi Jacek, I've added the list so the discussion can take place in public.
On Tue, Sep 3, 2013 at 8:50 AM, Jacek Caban <jacek@codeweavers.com mailto:jacek@codeweavers.com> wrote:
I have a patch for crypt32. I'd appreciate your review before I submit it to Wine. It has high potential to be insecure... This should fix Wine bug 28004 as well as some websites that don't send full cert chain in SSL/TLS handshake, but instead just their own certs. For that we need two changes. First, we need to look for issuers in global stores (as in those that are not passed to CertGetCertificateChain). When I did that,I don't really see a security issue here. The global stores are implicitly trusted, so neglecting to check them was probably an oversight on my part. Your tests imply something is missing.
My only real question here is whether the test results are in any way impacted by some unknown caching happening in Windows. Has this chain been verified prior to running the test? I'm going to guess that the test bot machines are "clean" prior to running the tests, but if they download their test executables from winehq.org http://winehq.org, that won't be the case.
I'm not sure what's the state of testbot VMs, but I tested it extensively locally. When tests are ran on a fresh Windows snapshot (that doesn't have the Rapid SSL cert cached), tests work fine and I can see HTTP download of the cert in Wireshark. All subsequent runs also pass, but don't do any HTTP connections.
it came out that it's not enough for intermediate issuers. Those have to be downloaded if we have information about their location stored in validated cert. That's easy using cryptnet. Could you please review the attached patch?
- for(i=0; i < info->cAccDescr; i++) {
if(!strcmp(szOID_PKIX_CA_ISSUERS,info->rgAccDescr[i].pszAccessMethod)
&& info->rgAccDescr[i].AccessLocation.dwAltNameChoice== CERT_ALT_NAME_URL) {
You explicitly make use of the AIA extension to find the location of intermediate certs. This is reasonable, and it gets your test case to pass, but it won't cover end certs that don't make use of the AIA extension. As I said above, I wonder whether caching is impacting the test results. More:
The thing that would be also nice to have as follow up is to cache downloaded URLs. They are already cached in URL cache, but I believe that they should be also cached in some place like world collection or something like that. I'm not sure what would be appropriate. Do you have a suggestion?Yes. I suspect the most relevant bug is actually 27168. 28004 is probably a dupe of it,
That I do not believe is dupe of 27168, at least according to your comment 5 in bug 28004 (I didn't debug PartyPoker myself, I just found it in bugzilla, saw that it's probably the same problem I'm fixing and verified that it works with my patches).
except that I'm not certain that PartyPoker uses chromium. In brief, Microsoft's crypt32 appears to cache intermediate certs in a PCCERT_CONTEXT's hCertStore. What I should be doing in crypt32, but am not, is to have certs ref count the stores they're in, such that the last cert to be closed causes its store to be freed, if the store is closed prior to the cert being freed. This, combined with certificate chains being cached in Microsoft's implementation, would allow intermediate certs to be found without explicitly checking the AIA.
You can see how the certificate chain engine ought to support caching on MSDN: http://msdn.microsoft.com/en-us/library/windows/desktop/aa377184(v=vs.85).as... http://msdn.microsoft.com/en-us/library/windows/desktop/aa377184%28v=vs.85%29.aspx
I never implemented chain caching, as I expected it'd be a performance optimization and could be addressed at a later time, and never got back to it.
In addition to bug 27168, you can see how Chromium is making use of crypt32's certificate caching in net::X509Certificate::CreateOSCertChainForCert, both its comments and implementation: http://src.chromium.org/viewvc/chrome/trunk/src/net/cert/x509_certificate.h http://src.chromium.org/viewvc/chrome/trunk/src/net/cert/x509_certificate_wi...
Thanks for the pointers. Looking at 27168 is something I hope to look at at some point. I hit the problem already when working on schannel and even in my current patch I had to work around it in one place (and forgot to put a comment about it...).
Back to your patch: I'm not entirely opposed to the explicit AIA approach, but I think it's worth at least thinking about other methods. And, I'd prefer to see the code block more explicitly separated from the surrounding code, with either a comment or a clear function name to indicate what's going on. As it is, one has to read quite a bit of CRYPT_FindIssuer to realize that it's really looking for an AIA extension to find intermediate issuers.
Looking at this again, I could probably just use CryptGetObjectUrl, which will be more explicit and will avoid code duplication.
Last thing on the patch, some style nits. (You already remarked that it needs splitting, and I agree.)
- if(issuer) {
Elsewhere in crypt32 I indent prior to open parentheses, so I worry this will make it a little harder to maintain a consistent style. Please observe the existing style, here and elsewhere.
issuer = CertFindCertificateInStore(store,subject->dwCertEncodingType, 0, CERT_FIND_SUBJECT_NAME,&subject->pCertInfo->Issuer, prevIssuer);
issuer = CRYPT_FindIssuer(engine, subject, store,CERT_FIND_SUBJECT_NAME,
&subject->pCertInfo->Issuer, flags,prevIssuer);
It might not be easy to guess my indenting style for continuation lines, but I use a single additional space to indicate continuation through crypt32. I'd appreciate observing the same style.
hAdditionalStore, &chain);
hAdditionalStore,dwFlags, &chain);
Likewise, the indentation change here doesn't seem necessary.
Sure, I will change those.
Thanks very much for working on this. Hope my comments are helpful, let me know if I can be of further assistance.
Thanks for the review. I will send updated patch.
Thanks, Jacek