Fix tectonic-typesetting#5: use thread local client

Reqwest client should be destroyed before forking to avoid threading
issue. This is done by using thread local storage, that is initialized
and used by geturl and cleared by texpresso.

Alternative: use pthread_atfork in reqwest to properly handle fork.

Author: let-def

Cargo.lock

@@ -11,11 +11,37 @@ use reqwest::{
 };
 use tectonic_errors::{anyhow::bail, Result};
 use tectonic_status_base::{tt_note, StatusBackend};
+use std::cell::RefCell;
 
 use crate::{GetUrlBackend, RangeReader};
 
 const MAX_HTTP_REDIRECTS_ALLOWED: usize = 10;
 
+thread_local!(static SHARED_CLIENT: RefCell<Option<Client>> = RefCell::new(None));
+
+fn get_shared_client() -> Client {
+    SHARED_CLIENT.with(|rclient| {
+        let mut oclient = rclient.borrow_mut();
+        match oclient.to_owned() {
+            Some(client) => client,
+            None => {
+                let client = Client::new();
+                *oclient = Some(client.clone());
+                client
+            }
+        }
+    })
+}
+
+/// Drop the shared client.
+/// It maintains a connection pool that causes problem when forking.
+pub fn clear_shared_client() {
+    SHARED_CLIENT.with(|rclient| {
+        let mut oclient = rclient.borrow_mut();
+        *oclient = None
+    })
+}
+
 /// URL-get backend implemented using the `reqwest` crate.
 #[derive(Debug, Default)]
 pub struct ReqwestBackend {}
@@ -25,7 +51,7 @@ impl GetUrlBackend for ReqwestBackend {
     type RangeReader = ReqwestRangeReader;
 
     fn get_url(&mut self, url: &str, _status: &mut dyn StatusBackend) -> Result<Response> {
-        let res = Client::new().get(url).send()?;
+        let res = get_shared_client().get(url).send()?;
         if !res.status().is_success() {
             bail!(
                 "unexpected HTTP response code {} for URL {}",
@@ -114,14 +140,14 @@ impl GetUrlBackend for ReqwestBackend {
 #[derive(Debug)]
 pub struct ReqwestRangeReader {
     url: String,
-    client: Client,
+    //client: Rc<Client>,
 }
 
 impl ReqwestRangeReader {
     fn new(url: &str) -> ReqwestRangeReader {
         ReqwestRangeReader {
             url: url.to_owned(),
-            client: Client::new(),
+            //client: Rc::new(Client::new()),
         }
     }
 }
@@ -136,7 +162,8 @@ impl RangeReader for ReqwestRangeReader {
         let mut headers = HeaderMap::new();
         headers.insert(RANGE, header_val);
 
-        let res = self.client.get(&self.url).headers(headers).send()?;
+        let client = get_shared_client();
+        let res = client.get(&self.url).headers(headers).send()?;
 
         if res.status() != StatusCode::PARTIAL_CONTENT {
             bail!(

crates/geturl/src/reqwest.rs

@@ -23,6 +23,7 @@ sha2 = "^0.10" # for digest computations
 thiserror = "1.0"
 tectonic_errors = { path = "../errors", version =">=0.1.0,<1"}
 tectonic_status_base = { path = "../status_base", version =">=0.1.0,<1"}
+tectonic_geturl = { path = "../geturl", version =">=0.3.0,<1" }
 texpresso_protocol = { path = "../texpresso_protocol", version = "0.1.0" }
 
 [package.metadata.internal_dep_versions]

crates/io_base/Cargo.toml

@@ -115,6 +115,7 @@ impl io::Read for TexpressoReader {
                     None => {
                         io.client.flush();
                         io.gen += 1;
+                        tectonic_geturl::reqwest::clear_shared_client();
                         let child = unsafe { io.client.fork() };
                         if child == 0 {
                             io.client.child(unsafe{libc::getpid()})