From: Christian Couder Date: Tue, 19 May 2026 15:24:52 GMT Subject: Re: [PATCH v2 6/8] promisor-remote: trust known remotes matching acceptFromServerUrl Message-ID: In-Reply-To: <875x4yoys5.fsf@toon--20250203-5JQV3.mail-host-address-is-not-set> On Fri, May 8, 2026 at 3:45 PM Toon Claes wrote: > > Christian Couder writes: > > +static bool match_one_url(const struct url_info *pi, const struct url_info *ui) > > +{ > > + const char *pat = pi->url; > > + const char *url = ui->url; > > + char *p_str, *u_str; > > + bool res; > > + > > + /* > > + * Schemes must match exactly. They are case-folded by > > + * url_normalize(), so strncmp() suffices. > > + */ > > + if (pi->scheme_len != ui->scheme_len || strncmp(pat, url, pi->scheme_len)) > > + return false; > > + > > + /* > > + * Ports must match exactly. url_normalize() strips default > > + * ports (like 443 for https), so length and content > > + * comparisons are sufficient. > > + */ > > + if (pi->port_len != ui->port_len || > > + strncmp(pat + pi->port_off, url + ui->port_off, pi->port_len)) > > + return false; > > + > > + /* > > + * Match host and path separately to prevent a '*' in the host > > + * portion of the pattern from matching across the '/' > > + * boundary into the path. Use WM_PATHNAME for the host so '*' > > + * cannot cross '/' there, and 0 for the path so '*' can still > > + * match multi-level paths. > > + */ > > Do we actually need WM_PATHNAME, because we only xstrndup() the host > part anyway? Yeah, it's not really needed. On one hand it doesn't hurt either, and it conveys the intent, which is that no / boundary should be crossed. But on the other hand I agree it could be confusing and it's simpler to just remove it, so I have removed it in the v3 I will send very soon. > > + > > + p_str = xstrndup(pat + pi->host_off, pi->host_len); > > + u_str = xstrndup(url + ui->host_off, ui->host_len); > > + res = !wildmatch(p_str, u_str, WM_PATHNAME); > > + free(p_str); > > + free(u_str); > > + > > + if (!res) > > + return false; > > + > > + p_str = xstrndup(pat + pi->path_off, pi->path_len); > > + u_str = xstrndup(url + ui->path_off, ui->path_len); > > + res = !wildmatch(p_str, u_str, 0); > > + free(p_str); > > + free(u_str); > > Is it correct we intentionally do not compare the user and pass (at > `user_off` and `passwd_off`)? I assume so, because this allows the > server to update those? Yes, we ignore them intentionally. Using the existing `token` field should be prefered, but maybe some need a user and password part of the URL. Anyway I have documented that in v3. > > static int should_accept_remote(enum accept_promisor accept, > > struct promisor_info *advertised, > > + struct string_list *accept_urls, > > struct string_list *config_info) > > { > > struct promisor_info *p; > > @@ -771,9 +846,6 @@ static int should_accept_remote(enum accept_promisor accept, > > if (accept == ACCEPT_KNOWN_NAME) > > return all_fields_match(advertised, config_info, p); > > > > - if (accept != ACCEPT_KNOWN_URL) > > - BUG("Unhandled 'enum accept_promisor' value '%d'", accept); > > - > > if (strcmp(p->url, remote_url)) { > > warning(_("known remote named '%s' but with URL '%s' instead of '%s', " > > "ignoring this remote"), > > @@ -781,7 +853,21 @@ static int should_accept_remote(enum accept_promisor accept, > > return 0; > > } > > > > - return all_fields_match(advertised, config_info, p); > > + if (accept == ACCEPT_KNOWN_URL) > > + return all_fields_match(advertised, config_info, p); > > + > > + if (accept != ACCEPT_NONE) > > + BUG("Unhandled 'enum accept_promisor' value '%d'", accept); > > + > > + /* > > + * Even if accept == ACCEPT_NONE, we MUST trust this known > > + * remote to update its token or other such fields if its URL > > + * matches the acceptFromServerUrl allowlist! > > + */ > > + if (url_matches_accept_list(accept_urls, remote_url)) > > + return all_fields_match(advertised, config_info, p); > > I should verify in the following patches, but it seems to me only when > promisor.AcceptFromServer is set to None it will store the advertised > servers to the local .git/config, or not? Right, it's better to check if the URL is in the allowlist as soon as we can. So in the v3 I have moved as much as possible the `promisor.acceptFromServerUrl` related checks before the other checks. The idea is that `promisor.acceptFromServerUrl` takes precedence over `promisor.acceptFromServer`, so having the `promisor.acceptFromServerUrl` checks first makes sense. Note that we should still not accept an advertised remote with an URL that matches a pattern in `promisor.acceptFromServerUrl` if a remote with the same name but a different URL exist on the client, unless the user has explicitly set `promisor.AcceptFromServer` to either 'All' or 'knownName'. In the v3 I have also added some documentation to be explicit about this. > > +test_expect_success "clone with 'None' but URL allowlisted" ' > > + git -C server config promisor.advertise true && > > + test_when_finished "rm -rf client" && > > + > > + GIT_NO_LAZY_FETCH=0 git clone -c remote.lop.promisor=true \ > > + -c remote.lop.fetch="+refs/heads/*:refs/remotes/lop/*" \ > > + -c remote.lop.url="$TRASH_DIRECTORY_URL/lop" \ > > + -c promisor.acceptfromserver=None \ > > + -c promisor.acceptFromServerUrl="$ENCODED_TRASH_DIRECTORY_URL/*" \ > > + --no-local --filter="blob:limit=5k" server client && > > + > > + # Check that the largest object is still missing on the server > > + check_missing_objects server 1 "$oid" > > +' > > Why do some tests end with `initialize_server 1 "$oid"` and this one > not? Isn't it weird tests prepare for the next test? It's more cleaning up after themselves than preparing for the next test. Initializing the server with `initialize_server 1 "$oid"` is only needed when some large objects end up on the server where they should not be. If we wanted to be sure that the state of the server is always clean for the next test, then we should initialize the server at the end of every test, using something like: test_when_finished "initialize_server 1 \"$oid\"" just in case something went wrong and a large file was transferred to the server. But I think that's quite expensive for not much gain. It's also clearer for readers to have `initialize_server 1 "$oid"` only where we think it's really needed. So in the end I prefer to leave this as-is for now. We can still address this later in a separate series for all the tests in "t5710-promisor-remote-capability.sh" if we really think it's worth addressing. Thanks.