I'm using libcurl in C++, and I'm calling curl_easy_perform
in a separate thread from my UI using Boost.Thread.
The main UI has a cancel button that I'd like to be perfectly responsive (i.e., when a user clicks on it, it should immediately react). I have read, write, and progress callbacks set up to read an atomic should_cancel
variable (as in this question), but there are two problems:
There's often a very small (but noticeable) delay from when cancel is pressed to when the curl operation completes.
Occasionally, there's a very long (sometimes interminable) delay. In this case, either:
a. the progress, read, and write callbacks simply aren't called for a long time, or
b. the progress callback is called, I return a nonzero value (meaning it should terminate), but the curl operation doesn't complete for a while longer (in fact, the progress function is called again in the meantime!)
So:
One possibility is to tell the UI that the cancel operation succeeded, but keep running the curl thread in the background until it cancels. The problem with this (I think) is that it forces the should_cancel
variable to be global, instead of scoped to the dialog box where the operation began.
I met the same problem with curl 7.21.6. When tring to abort smtp protocol. Returning CURL_READFUNC_ABORT from read callback stops the transfer but curl_easy_perform does not return for next 5+ minutes. Probably it waits for tcp timeout.
To walk it around i store socket used by curl (replace curl_opensocket_callback), and close this socket directly when needed.
curl_socket_t storeCurlSocket(SOCKET *data, curlsocktype purpose, struct curl_sockaddr *addr) {
SOCKET sock = socket(addr->family, addr->socktype, addr->protocol);
*data = sock;
return sock;
}
size_t abort_payload(void *ptr, size_t size, size_t nmemb, SOCKET *curl_socket) {
SOCKET l_socket = INVALID_SOCKET;
swap(l_socket, *curl_socket);
if (l_socket != INVALID_SOCKET) {
shutdown(l_socket, SD_BOTH);
closesocket(l_socket);
}
return CURL_READFUNC_ABORT;
}
...calling perform...
SOCKET curlSocket;
curet = curl_easy_setopt(curl, CURLOPT_OPENSOCKETFUNCTION, storeCurlSocket);
curet = curl_easy_setopt(curl, CURLOPT_OPENSOCKETDATA, &curlSocket);
curet = curl_easy_setopt(curl, CURLOPT_READFUNCTION, abort_payload);
curet = curl_easy_setopt(curl, CURLOPT_READDATA, &curlSocket);
curet = curl_easy_perform(curl);
if (curet == CURLE_ABORTED_BY_CALLBACK) {
//expected abort
}
...
Your basic idea is right. You should detach the curl operation from the UI. However, the implementation should be changed a bit. You shouldn't be using a global should_cancel
. Instead, you should have a global current_request
pointer, to an object of type Request
. This type should have an internal cancel
flag, and a public Cancel()
function. In response to the cancel button, you call Cancel
on the current_request
and then null it. A cancelled request is then repsonsible for its own cleanup at a later time (it's a thread, after all).
You'll need to be careful with your mutexes to prevent zombie objects. There's an inherent race condition between a cancel and a request completion.
The delay could happens because curl_easy_perform
is busy with the write or read callback, try to return 0 from the write callback or CURL_READFUNC_ABORT
from the read callback. According with the documentation, if the value returned by the write callback differs from the value received by the function the transfer will be aborted, and if the read callback return CURL_READFUNC_ABORT
, the transfer will be aborted too (valid from version 7.12.1 of the library).
If you love us? You can donate to us via Paypal or buy me a coffee so we can maintain and grow! Thank you!
Donate Us With