Can you please help us understand how to reproduce this? I've read the code a few times now and I can't see how it can log that message and not return an error... can you?

This is a tricky one to reproduce in lab, but happens regularly in practice; the other end needs to close the connection after the initial check for dead connection but before the upload completes.
I'm reading the code now and I think this may happen when we enter this:

Gist of code used to reproduce:

CURL *curl;
curl = curl_easy_init();
curl_easy_setopt(curl, CURLOPT_URL, "https://s3.amazonaws.com");
curl_easy_setopt(curl, CURLOPT_VERBOSE, 1L);
curl_easy_perform(curl);
Sleep(5950); // experimental delay that usually results in the server closing connection while we upload
struct curl_slist *headers = NULL;
headers = curl_slist_append( headers, "Expect:");
curl_easy_setopt(curl, CURLOPT_HTTPHEADER, headers);
curl_easy_setopt(curl, CURLOPT_UPLOAD, 1L);
curl_easy_setopt(curl, CURLOPT_INFILESIZE, 1L);
curl_easy_setopt(curl, CURLOPT_READFUNCTION, read_callback_dummy );
CURLcode result;
result = curl_easy_perform(curl);
printf("%d",(int)result); // result: 0

With read callback:

size_t read_callback_dummy(void *ptr, size_t size, size_t nmemb, void *stream)
{
static int i = 1;
if(i) {i=0;return 1;}
return 0;
}

Also my comment above was apparently right; the following patch gives me CURLE_SEND_FAIL_REWIND:
djelinski@dfa8181

I don't understand the code enough to know if this is the right place for this fix or not; let me know if I should file a pull request.

Thanks! I think it looks like a good fix. Please submit a PR for it!