path: root/common/curl_request.cpp

#include <llmr/platform/platform.hpp>
#include <llmr/platform/request.hpp>
#include <llmr/util/uv.hpp>
#include <llmr/util/std.hpp>

#include <queue>
#include <boost/lockfree/queue.hpp>

#include <curl/curl.h>

// This file contains code from http://curl.haxx.se/libcurl/c/multi-uv.html:

/***************************************************************************
 *                                  _   _ ____  _
 *  Project                     ___| | | |  _ \| |
 *                             / __| | | | |_) | |
 *                            | (__| |_| |  _ <| |___
 *                             \___|\___/|_| \_\_____|
 *
 * Copyright (C) 1998 - 2013, Daniel Stenberg, <daniel@haxx.se>, et al.
 *
 * This software is licensed as described in the file COPYING, which
 * you should have received as part of this distribution. The terms
 * are also available at http://curl.haxx.se/docs/copyright.html.
 *
 * You may opt to use, copy, modify, merge, publish, distribute and/or sell
 * copies of the Software, and permit persons to whom the Software is
 * furnished to do so, under the terms of the COPYING file.
 *
 * This software is distributed on an "AS IS" basis, WITHOUT WARRANTY OF ANY
 * KIND, either express or implied.
 *
 ***************************************************************************/

/* Example application code using the multi socket interface to download
   multiple files at once, but instead of using curl_multi_perform and
   curl_multi_wait, which uses select(), we use libuv.
   It supports epoll, kqueue, etc. on unixes and fast IO completion ports on
   Windows, which means, it should be very fast on all platforms..

   Written by Clemens Gruber, based on an outdated example from uvbook and
   some tests from libuv.

   Requires libuv and (of course) libcurl.

   See http://nikhilm.github.com/uvbook/ for more information on libuv.
*/

namespace llmr {
namespace platform {
namespace request {

struct curl_context {
    uv_poll_t poll_handle;
    curl_socket_t sockfd;
};

// Handles the request thread + messaging to the thread.
static uv_thread_t thread;
static uv::once init_thread_once;
static uv_loop_t *loop = nullptr;
static uv_async_t async_add;
static uv_async_t async_cancel;

// Stores pointers (!) to shared_ptrs. We use shared_ptrs so that request objects don't get
// auto-destructed while they're in progress. The TileData object retains a weak_ptr to this
// request, so we have to use a shared_ptr here to ensure that this object stays alive.
static boost::lockfree::queue<std::shared_ptr<llmr::platform::Request> *> add_queue(8);
static boost::lockfree::queue<std::shared_ptr<llmr::platform::Request> *> cancel_queue(8);

// Used as the CURL timer function to periodically check for socket updates.
static uv_timer_t timeout;

// CURL multi handle that we use to request multiple URLs at the same time, without having to block
// and spawn threads.
static CURLM *curl_multi = nullptr;

// CURL share handles are used for sharing session state (e.g.)
static uv::mutex curl_share_mutex;
static CURLSH *curl_share = nullptr;

// A queue that we use for storing resuable CURL easy handles to avoid creating and destroying them
// all the time.
static std::queue<CURL *> curl_handle_cache;


class CURLRequest : public llmr::platform::Request {
public:
    CURLRequest(const std::string &url,
                std::function<void(llmr::platform::Response *)> callback,
                std::shared_ptr<uv::loop> loop)
        : Request(url, callback, loop) {}

    CURL *curl = nullptr;
};


// Implementation starts here.

// Locks the CURL share handle
void curl_share_lock(CURL *, curl_lock_data, curl_lock_access, void *) { curl_share_mutex.lock(); }

// Unlocks the CURL share handle
void curl_share_unlock(CURL *, curl_lock_data, void *) { curl_share_mutex.unlock(); }

curl_context *create_curl_context(curl_socket_t sockfd) {
    curl_context *context = new curl_context;
    context->sockfd = sockfd;

    uv_poll_init_socket(loop, &context->poll_handle, sockfd);
    context->poll_handle.data = context;

    return context;
}

void curl_close_cb(uv_handle_t *handle) {
    curl_context *context = (curl_context *)handle->data;
    free(context);
}

void destroy_curl_context(curl_context *context) {
    uv_close((uv_handle_t *)&context->poll_handle, curl_close_cb);
}

void remove_curl_handle(CURL *handle) {
    CURLMcode error = curl_multi_remove_handle(curl_multi, handle);
    if (error != CURLM_OK) {
        throw std::runtime_error(std::string("CURL multi error: ") + curl_multi_strerror(error));
    }

    curl_easy_reset(handle);
    curl_handle_cache.push(handle);
}

void curl_perform(uv_poll_t *req, int /*status*/, int events) {
    int running_handles;
    int flags = 0;
    curl_context *context;
    CURLMsg *message;
    int pending;

    uv_timer_stop(&timeout);

    if (events & UV_READABLE)
        flags |= CURL_CSELECT_IN;
    if (events & UV_WRITABLE)
        flags |= CURL_CSELECT_OUT;

    context = (curl_context *)req;

    curl_multi_socket_action(curl_multi, context->sockfd, flags, &running_handles);

    while ((message = curl_multi_info_read(curl_multi, &pending))) {
        switch (message->msg) {
        case CURLMSG_DONE: {
            std::shared_ptr<Request> *req = nullptr;
            curl_easy_getinfo(message->easy_handle, CURLINFO_PRIVATE, (char *)&req);

            // Add human-readable error code
            if (message->data.result != CURLE_OK) {
                (*req)->res->error_message = curl_easy_strerror(message->data.result);
                (*req)->res->code = -1;
            } else {
                curl_easy_getinfo(message->easy_handle, CURLINFO_RESPONSE_CODE, &(*req)->res->code);
            }

            // We're currently in the CURL request thread. We're going to schedule a uv_work request
            // that executes the background function in a threadpool, and tell it to call the
            // after callback back in the main uv loop.
            (*req)->complete();

            CURL *handle = message->easy_handle;
            remove_curl_handle(handle);

            // We're setting this to NULL because there might still be shared_ptrs around that could
            // be cancelled.
            ((CURLRequest *)req->get())->curl = nullptr;

            // Delete the shared_ptr pointer we created earlier.
            delete req;
            break;
        }

        default:
            // This should never happen, because there are no other message types.
            throw std::runtime_error("CURLMSG returned unknown message type");
        }
    }
}

int handle_socket(CURL * /*easy*/, curl_socket_t s, int action, void * /*userp*/, void *socketp) {
    curl_context *context = nullptr;

    if (socketp) {
        context = (curl_context *)socketp;
    } else if (action != CURL_POLL_REMOVE) {
        context = create_curl_context(s);
    }

    if (context) {
        curl_multi_assign(curl_multi, s, (void *)context);
        if (action == CURL_POLL_IN || action == CURL_POLL_INOUT) {
            uv_poll_start(&context->poll_handle, UV_READABLE, curl_perform);
        }
        if (action == CURL_POLL_OUT || action == CURL_POLL_INOUT) {
            uv_poll_start(&context->poll_handle, UV_WRITABLE, curl_perform);
        }
        if (action == CURL_POLL_REMOVE && socketp) {
            uv_poll_stop(&context->poll_handle);
            destroy_curl_context(context);
            curl_multi_assign(curl_multi, s, NULL);
        }
    }

    return 0;
}

void on_timeout(uv_timer_t * /*req*/) {
    int running_handles;
    CURLMcode error =
        curl_multi_socket_action(curl_multi, CURL_SOCKET_TIMEOUT, 0, &running_handles);
    if (error != CURLM_OK) {
        throw std::runtime_error(std::string("CURL multi error: ") + curl_multi_strerror(error));
    }
}

void start_timeout(CURLM * /*multi*/, long timeout_ms, void * /*userp*/) {
    if (timeout_ms <= 0) {
        on_timeout(&timeout);
    } else {
        uv_timer_start(&timeout, on_timeout, timeout_ms, 0);
    }
}

// This function is the first function called in the request thread. It sets up the CURL share/multi
// handles and runs the thread loop.
void thread_init(void * /*ptr*/) {
    uv_timer_init(loop, &timeout);

    CURLSHcode share_error;
    curl_share = curl_share_init();

    share_error = curl_share_setopt(curl_share, CURLSHOPT_LOCKFUNC, curl_share_lock);
    if (share_error != CURLSHE_OK) {
        throw std::runtime_error(std::string("CURL share error: ") + curl_share_strerror(share_error));
    }

    share_error = curl_share_setopt(curl_share, CURLSHOPT_UNLOCKFUNC, curl_share_unlock);
    if (share_error != CURLSHE_OK) {
        throw std::runtime_error(std::string("CURL share error: ") + curl_share_strerror(share_error));
    }

    CURLMcode multi_error;
    curl_multi = curl_multi_init();

    multi_error = curl_multi_setopt(curl_multi, CURLMOPT_SOCKETFUNCTION, handle_socket);
    if (multi_error != CURLM_OK) {
        throw std::runtime_error(std::string("CURL multi error: ") + curl_multi_strerror(multi_error));
    }
    multi_error = curl_multi_setopt(curl_multi, CURLMOPT_TIMERFUNCTION, start_timeout);
    if (multi_error != CURLM_OK) {
        throw std::runtime_error(std::string("CURL multi error: ") + curl_multi_strerror(multi_error));

    }

    // Main event loop. This will not return until the request loop is terminated.
    uv_run(loop, UV_RUN_DEFAULT);

    curl_multi_cleanup(curl_multi);
    curl_multi = nullptr;
    curl_share_cleanup(curl_share);
    curl_share = nullptr;

    // Clean up all the CURL easy handles that we kept around for potential future reuse.
    while (!curl_handle_cache.empty()) {
        curl_easy_cleanup(curl_handle_cache.front());
        curl_handle_cache.pop();
    }
}

// This function is called when we have new data for a request. We just append it to the string
// containing the previous data.
size_t curl_write_cb(void *contents, size_t size, size_t nmemb, void *userp) {
    ((std::string *)userp)->append((char *)contents, size * nmemb);
    return size * nmemb;
}

// This callback is called in the request event loop (on the request thread).
// It initializes newly queued up download requests and adds them to the CURL
// multi handle.
void async_add_cb(uv_async_t * /*async*/) {
    std::shared_ptr<Request> *req = nullptr;
    while (add_queue.pop(req)) {
        // Make sure that we're not starting requests that have been cancelled
        // already by async_cancel_cb.
        if ((*req)->cancelled) {
            delete req;
            continue;
        }

        // Obtain a curl handle (and try to reuse existing handles before creating new ones).
        CURL *handle = nullptr;
        if (!curl_handle_cache.empty()) {
            handle = curl_handle_cache.front();
            curl_handle_cache.pop();
        } else {
            handle = curl_easy_init();
        }

        ((CURLRequest *)req->get())->curl = handle;

        curl_easy_setopt(handle, CURLOPT_PRIVATE, req);
        curl_easy_setopt(handle, CURLOPT_CAINFO, "ca-bundle.crt");
        curl_easy_setopt(handle, CURLOPT_URL, (*req)->url.c_str());
        curl_easy_setopt(handle, CURLOPT_WRITEFUNCTION, curl_write_cb);
        curl_easy_setopt(handle, CURLOPT_WRITEDATA, &(*req)->res->body);
        curl_easy_setopt(handle, CURLOPT_ACCEPT_ENCODING, "deflate");
        curl_easy_setopt(handle, CURLOPT_SHARE, curl_share);
        curl_multi_add_handle(curl_multi, handle);
    }
}

void async_cancel_cb(uv_async_t * /*async*/) {
    std::shared_ptr<Request> *req = nullptr;
    while (cancel_queue.pop(req)) {
        // It is possible that the request has not yet been started, but that it already has been
        // added to the queue for scheduling new requests. In this case, the CURL handle is invalid
        // and we manually mark the Request as cancelled.
        CURL *handle = ((CURLRequest *)req->get())->curl;
        if (handle && !(*req)->cancelled) {
            remove_curl_handle(handle);
            ((CURLRequest *)req->get())->curl = nullptr;
        }
        (*req)->cancelled = true;

        delete req;
        req = nullptr;
    }
}

void thread_init_cb() {
    curl_global_init(CURL_GLOBAL_ALL);

    loop = uv_loop_new();
    uv_async_init(loop, &async_add, async_add_cb);
    uv_async_init(loop, &async_cancel, async_cancel_cb);
    uv_thread_create(&thread, thread_init, nullptr);
}
} // end namespace request
} // end namespace platform


std::shared_ptr<platform::Request>
platform::request_http(const std::string &url,
                       std::function<void(Response *)> callback,
                       std::shared_ptr<uv::loop> loop) {
    using namespace request;
    init_thread_once(thread_init_cb);
    std::shared_ptr<CURLRequest> req = std::make_shared<CURLRequest>(url, callback, loop);

    // Note that we are creating a new shared_ptr pointer(!) because the lockless queue can't store
    // objects with nontrivial destructors. We have to make absolutely sure that we manually delete
    // the shared_ptr when we pop it from the queue.
    add_queue.push(new std::shared_ptr<Request>(req));
    uv_async_send(&async_add);

    return req;
}

// Cancels an HTTP request.
void platform::cancel_request_http(const std::shared_ptr<Request> &req) {
    if (req) {
        using namespace request;

        // Note that we are creating a new shared_ptr pointer(!) because the lockless queue can't
        // store objects with nontrivial destructors. We have to make absolutely shure that we
        // manually delete the shared_ptr when we pop it from the queue.
        cancel_queue.push(new std::shared_ptr<Request>(req));
        uv_async_send(&async_cancel);
    }
}
} // end namespace llmr