Commit graph

2 commits

Author SHA1 Message Date
Jiahao XU
db45f2fb7f
Refactor binstalk-downloader APIs: Remove cancellation_future plus optimizations (#591)
- Refactor: Mv fn `utils::asyncify` into mod `utils`
 - Improve err msg for task failure in `utils::asyncify`
 - Make sure `asyncify` always returns the same annoymous type
   that implements `Future` if the `T` is same.
 - Rewrite `extract_bin` to avoid `block_in_place`
   support cancellation by dropping
 - Rm unused dep scopeguard from binstalk-downloader
 - Rewrite `extract_tar_based_stream` so that it is cancellable by dropping
 - Unbox `extract_future` in `async_extracter::extract_zip`
 - Refactor `Download` API: Remove `CancellationFuture` as param

   since all futures returned by `Download::and_*` does not call
   `block_in_place`, so they can be cancelled by drop instead of using this
   cumbersome hack.
 - Fix exports from mod `async_tar_visitor`
 - Make `signal::{ignore_signals, wait_on_cancellation_signal}` private
 - Rm the global variable `CANCELLED` in `wait_on_cancellation_signal`
   and rm fn `wait_on_cancellation_signal_inner`
 - Optimize `wait_on_cancellation_signal`: Avoid `tokio::select!` on `not(unix)`
 - Rm unnecessary `tokio::select!` in `wait_on_cancellation_signal` on unix
   Since `unix::wait_on_cancellation_signal_unix` already waits for ctrl + c signal.
 - Optimize `extract_bin`: Send `Bytes` to blocking thread for zero-copy
 - Optimize `extract_with_blocking_decoder`: Avoid dup monomorphization
 - Box fut of `fetch_crate_cratesio` in `PackageInfo::resolve`
 - Optimize `extract_zip_entry`: Spawn only one blocking task per fn call

   by using a mspc queue for the data to be written to the `outfile`.

   This would improve efficiency as using `tokio::fs::File` is expensive:
   It spawns a new blocking task, which needs one heap allocation and then
   pushed to a mpmc queue, and then wait for it to be done on every loop.

   This also fix a race condition where the unix permission is set before
   the whole file is written, which might be used by attackers.
 - Optimize `extract_zip`: Use one `BytesMut` for entire extraction process
   To avoid frequent allocation and deallocation.
 - Optimize `extract_zip_entry`: Inc prob of reusing alloc in `BytesMut`

   Performs the reserve before sending the buf over mpsc queue to
   increase the possibility of reusing the previous allocation.

   NOTE: `BytesMut` only reuses the previous allocation if it is the
   only one holds the reference to it, which is either on the first
   allocation or all the `Bytes` in the mpsc queue has been consumed,
   written to the file and dropped.

   Since reading from entry would have to wait for external file I/O,
   this would give the blocking thread some time to flush `Bytes`
   out.
 - Disable unused feature fs of dep tokio

Signed-off-by: Jiahao XU <Jiahao_XU@outlook.com>
2022-12-12 03:15:30 +00:00
Jiahao XU
3b1a7f2c78
Optimize extract_zip: Use async_zip::read::stream::ZipFileReader to avoid temporary file (#590)
* Add new dep async_zip v0.0.9 to binstalk-downloader
   with features "gzip", "zstd", "xz", "bzip2", "tokio".
* Refactor: Simplify `async_extracter::extract_*` API
* Refactor: Create newtype wrapper of `ZipError`
   so that the zip can be upgraded without affecting API of this crate.
* Enable feature fs of dep tokio in binstalk-downloader
* Rewrite `extract_zip` to use `async_zip::read::stream::ZipFileReader`
   which avoids writing the zip file to a temporary file and then read it
   back into memory.
* Refactor: Impl new fn `await_on_option` and use it
* Optimize `tokio::select!`: Make them biased and check for cancellation first
  to make cancellation takes effect ASAP.
* Rm unused dep zip from binstalk-downloader

Signed-off-by: Jiahao XU <Jiahao_XU@outlook.com>
2022-12-12 00:44:34 +00:00