aboutsummaryrefslogtreecommitdiff
path: root/lib/sqfs/data_writer/pthread.c
AgeCommit message (Collapse)Author
2019-12-12Don't kick off the threads until the queue is fullDavid Oberhollenzer
The idea is as follows: Initially let the user submit blocks until the queue is filled, then kick of the threads. Every thread will end up getting a block without any waits until they completely deplete the queue. Assuming the threads take longer to process the data than it takes the main thread to do I/O and submit new blocks, the queue should stay mostly filled with minimal wait times. Signed-off-by: David Oberhollenzer <david.oberhollenzer@sigma-star.at>
2019-12-12Fix thread pool queue accountingDavid Oberhollenzer
- ONLY manipulate the back log counter in the main thread. - Fix the order of operations when submitting blocks. Signed-off-by: David Oberhollenzer <david.oberhollenzer@sigma-star.at>
2019-12-05Minor data writer cleanupDavid Oberhollenzer
Move "do block" function over to the rest of the block related code and internalizie the pthread worker structure. Signed-off-by: David Oberhollenzer <david.oberhollenzer@sigma-star.at>
2019-12-05Fix pthread data writer interfering with signal handlingDavid Oberhollenzer
If a signal is sent to a process, the POSIX spec says that ANY thread could be picked ARBITRARILY to handle the signal. In our case this could be one of the internal worker threads. The problem here is that an unsuspecting user of the library might suddenly have their signal handler run in parallel to their main thread and run into weird concurrency issues, because they didn't expect that to happen. In fact, the libsquashfs API tries to be transparent about whether or not it uses a thread pool internally and does everything other than number crunching (e.g. I/O that may happen through user supplied callbacks) in the same thread as the one that submitts the blocks. Unfortunately, pthread doesn't have a way to set a signal mask for the new thread and setting it inside the thread is racy (i.e. a signal might be delivered before the worker thread sets the mask). The only portable and non-racy way to do this, is to disable all signals in the calling thread that sets up the data writer, create the threads (which will inherit the mask) and then resetore the previous signal mask, hoping for the best. The downside to this is that signals may be lost in that short time. Signed-off-by: David Oberhollenzer <david.oberhollenzer@sigma-star.at>
2019-09-28Fix missing mutex unlick in data writer error pathDavid Oberhollenzer
Signed-off-by: David Oberhollenzer <david.oberhollenzer@sigma-star.at>
2019-09-26Add file API stub to sqfs data writerDavid Oberhollenzer
Basically move the state tracking from the old data writer over to the new one. Signed-off-by: David Oberhollenzer <david.oberhollenzer@sigma-star.at>
2019-09-25Rename block processor to sqfs_data_writer_tDavid Oberhollenzer
Signed-off-by: David Oberhollenzer <david.oberhollenzer@sigma-star.at>