The reported problem of segmentation faults in mod_http2 during a graceful restart seems to boil down to the fact that the document root is updated at the same time. On my OS X I can reproduce the SIGBUS when GETting the same 10 MB file again and again and copying over it.
When I disable mmap for the file buckets, the crash does not occur and the request fails as expected. So, my understanding is that apr_bucket_read() hands out a pointer to the mmap'ed memory and when I copy that data somewhere else, and the underlying file is gone, a SIGBUS occurs as the OS has no other means of telling me about the problem. Interestingly, when I send the same buckets down to the core output filters, the problem does not happen. Since this seems to involve writes using iovec's the kernel might handle this properly with a return code? I am not sure. Or maybe it's the sendfile() that takes care of it... Disabling mmap in general seems not really a good solution, but it is not as bad as http2 sends file out in <= 16KB chunks anyway. Since for TLS, I need to assemble frame data anyway, reading directly from file into a properly allocated frame buffer might be better anyway. However I'd like to understand the problem and my options here: what would I need to do to use mmap'ed files properly? Catch SIGBUS during the copy? Advice appreciated. Cheers, -Stefan
