Previously, we used the existence of a `prior*` field in the Resource object as an indication for whether we should consult the cache or not. However, this is prone to error, since a failed cache lookup won't set any prior fields. Therefore, we manually set `priorExpires` to 0. This in turn triggered another bug where generated wrong expiration timestamps when the server response we got was expired (or expired between sending and receiving).
This commit changes the flags so that we can now explicitly request CacheOnly/NetworkOnly (or All) loading methods, rather than the implicit Optional/Required naming scheme.
In https://github.com/mapbox/mapbox-gl-native/pull/9670, we implemented support for the Cache-Control: must-revalidate header. While we now respect this header and don't show resources that are stale and have this header set, the optional cache request never completes. This means we're also never going to try to actually get a fresh tile and just never show this tile anymore.
This commit fixes this by responding with a Not Found error when the resource is unusable (= stale and has must-revalidate set). Since we actually still have the data (but can't use it due to caching rules), we're responding with the data as well. To avoid a second cache hit, tile_loader_impl.hpp now passes on the data from the Optional to the Required request so that it can be reused when we get a 304 Not Modified response from the server.
Port of https://github.com/mapbox/mapbox-gl-js/pull/5101: adds a new render pass `Pass3D` before any other rendering wherein we render layers with 3D passes (fill-extrusion layers) to offscreen framebuffers, sharing a depth renderbuffer between those layers in order to render 3D space correctly. Those framebuffers are saved on the RenderLayers and copied back to the map during the translucent pass. Rendering to offscreen framebuffers before we do any clear + draw means we can avoid expensive framebuffer restores.
Previously, when we started a worker operation that eventually throws an exception (e.g. due to the tile not being parseable), and then enqueue another worker operation while the first one is processing, we treated the worker as idle once the first operation's error callback fired, even though the second operation was still in progress. Due to our use of coalescing, I was unable to come up with a reliable test since we'd need to reproduce the behavior described above, which is timing dependent.
Since 9a9408e811, we marked tiles as non-renderable if an error occured. This lead to situations where a tile was loaded + parsed successfully, then a revalidation attempt occured (e.g. because the resource was stale) which failed. In this case, we used to mark the tile as non-renderable although we could've used the perfectly parsed (stale) resource.
When a texture is deleted, it is as if all texture units to which that texture object is bound are rebound to texture object zero. Therefore it's the *bindings* that need to be dirtied, not the active texture unit.
- AnnotationData.updateData() was not completely thread safe. Calling it from the renderer thread instead of the main thread fixes this
- Since we moved the coalescing of updates out of the map and into the renderer frontend implementations, the updateData was called too much, resulting in degraded performance. Moving it into the renderer makes sure it gets coalesced again.
