I can reproduce this, and I can reproduce it back to 5.0 (haven’t tried any release before that). I couldn’t find a definite reason for why it breaks (curl_clean_state() is called because curl reports CURLMSG_DONE, freeing a socket, but then curl_multi_do() is called again for that socket, resulting in a use-after-free – but I don’t know why curl_multi_do() is invoked after CURLMSG_DONE).
Because I remembered a similar situation where the curl driver suddenly failed (and then failed for every qemu release until that point), and where it turned out a change in libcurl broke our driver, I tried bisecting libcurl, but it turned out that when I build it myself and use it via LD_PRELOAD, I don’t get a crash. I’ve tried building it with different options and in different versions, but consistently I see that using the system libcurl results in a crash, and using one I built myself does not. (Tested on Fedora and Arch.)
That isn’t to say the bug isn’t in our curl driver, but to find out where it is exactly, it seems necessary to find out what the difference between the system libcurl and the one I built is... So far, I have no idea. :/
I can reproduce this, and I can reproduce it back to 5.0 (haven’t tried any release before that). I couldn’t find a definite reason for why it breaks (curl_clean_state() is called because curl reports CURLMSG_DONE, freeing a socket, but then curl_multi_do() is called again for that socket, resulting in a use-after-free – but I don’t know why curl_multi_do() is invoked after CURLMSG_DONE).
Because I remembered a similar situation where the curl driver suddenly failed (and then failed for every qemu release until that point), and where it turned out a change in libcurl broke our driver, I tried bisecting libcurl, but it turned out that when I build it myself and use it via LD_PRELOAD, I don’t get a crash. I’ve tried building it with different options and in different versions, but consistently I see that using the system libcurl results in a crash, and using one I built myself does not. (Tested on Fedora and Arch.)
That isn’t to say the bug isn’t in our curl driver, but to find out where it is exactly, it seems necessary to find out what the difference between the system libcurl and the one I built is... So far, I have no idea. :/