I have 3 servers in production all recently upgraded to MariaDB 10.1 to, among other things, make use of InnoDB Page Compression.

All 3 use XFS on the data partition, over the top of hardware RAID10 with SSDs.

CentOS boxes are 16 cores/32 threads, 256GB RAM
Gentoo box is 20 cores/40 threads, 256GB RAM

I periodically need to run through and OPTIMIZE several of our InnoDB tables. Some due to heavy delete operations in production, others, if I need to restore from snapshot, tar doesn't preserve the sparseness of the InnoDB files so I run an optimize across them to get the space back.

However, when running a mass optimize, I kept running into hangs, with InnoDB bitching about semaphore locks etc in the error log.

I've since discovered if I disable innodb_use_mtflush then the stall doesn't occur, at all, across the entire DB (there's a few thousand tables using page compression).

I've attached an innodb status from when the semaphore hang occurs.

I've experimented with various values for adaptive hash index partitions, innodb thread concurrency, io read and write threads etc. The only thing that appears to make the problem happen or go away, is whether mtflush is enabled or not.

Apologies if mtflush is ONLY for FusionIO but it isn't readily apparent that you shouldn't, only that it was created to improve performance with FusionIO.