While at it, remove the error return from BrowserOption, to make it
consistent with all the other option func types.
We don't have a good test for this feature yet, but at least we check
that it doesn't crash or error via an example.
Fixes#292.
It's Run that actually starts a Browser, not NewContext. If the browser
is closed or crashes, the browser handler will fail to read from the
websocket, and its goroutines will stop.
However, the target handler's goroutines may not stop. The browser
handler uses a separate cancel function to stop itself when encountering
a websocket error, so that doesn't propagate to the original context
children, like the target handler.
To fix this, make it so that NewContext can keep the cancel function
around, for Run to use it in this scenario. And add a test case that
tests this very edge case, which used to time out before the fix.
Fixes#289.
This fixes the data race uncovered by the recent refactor to run all
tests as tabs under the same browser.
The problem was that a write on the pages map could be done from the
goroutine calling NewContext to create a new map, while other goroutines
could similarly read or write the same map.
Instead of adding a lock around the map, make one of the Browser's
goroutines be the sole user of the map. To make that extra obvious and
avoid potential races in the future, declare the map inside the
goroutine's scope.
For some reason, this makes the Attributes tests flakier than before.
For now, add short sleeps; we can investigate that separately, now that
the data races are gone.
That way, we avoid the racy map access via Browser.executorForTarget. If
a context is attached to a target, the Target field must be non-nil.
The Browser.pages map is still racy, since multiple tabs can be created
concurrently; we'll fix this other data race in another commit.
First, collapse Browser.Start with NewBrowser. There's no reason to
split them up.
Second, unexport Browser.userDataDir, since it's only needed for a test.
It's also a bad precedent, as only the ExecAllocator will control the
user data directory.
Third, export Context.Browser, since we were already exporting
Context.Allocator.
Finally, remove the Executor interface, a duplicate of cdp.Executor.
We hadn't noticed a few uncaught exceptions being received from the
browser, because the events were ignored. Start printing them via the
error logger.
The ones we were getting were caused by testAllocate running Navigate
actions when the path argument was empty. Navigating to "testdata/"
causes JS exceptions, as it's not a valid page.
Instead, leave the new target pointing at a blank document.
Use a single websocket connection per browser, removing the need for an
extra websocket connection per target.
This is thanks to the Target.sendMessageToTarget command to send
messages to each target, and the Target.receivedMessageFromTarget event
to receive messages back.
The browser handles activity via a single worker goroutine, and the same
technique is used for each target. This means that commands and events
are dealt with in order, and we can do away with some complexity like
mutexes and extra go statements.
First, we want all of the functionality in a single package; this means
collapsing whatever is useful into the root chromedp package.
The runner package is being replaced by the Allocator interface, with a
default implementation which starts browser processes.
The client package doesn't really have a place in the new design. The
context, allocator, and browser types will handle the connection with
each browser.
Finally, the new API is context-based, hence the addition of context.go.
The tests have been modified to build and run against the new API.
