…-linux

…he compressor.

…ified.

…ession/decompression operation back to userspace (for now, recycling the struct crypt_op.iv member, to avoid any big changes or compatibility breaks with other cryptodev structures).

…or some yet unknown reason, are surprisingly slow (major bottleneck) when HW-acceleration is available.

…ng crypto_op.iv, simply return it back in crypto_op.dlen.

…(e.g. uncompressing an empty compression stream).

cryptodev_compr_init could leak memory if an allocation error failed within the initialization routine.
This refactors the initialization routine to make sure this doesn't happen.

…tion failure

…to avoid undesired different behaviors depending on the system's page sizes.

This commit extends compression so that the user can compress or decompress data in multiple separate chunks (compressed bitstreams) with a single ioctl system call. This avoids reducing the overhead if the user wants to compress the input in many small chunks, which previously required a separate ioctl for each chunk, involving substantial overhead.

… destionation buffers are page-aligned and the chunk sizes are multiples of the page size.

…r-module instance.

…llback less costly, and even allow some zerocopy even if the user buffer isn't perfectly aligned. It also allows leaking uninitialized kernel memory through the inter-chunk gaps like it was doing before.

…kernel very fast.

…re copying it there.

…tly in order to be able to distinguish between compressible (ret=0) and incompressible(ret=-ENOSPC) chunks on compression.

…specify that the chunk is not present (i.e. should be skipped). This is useful in the presence of uncompressible chunks, which can be transmitted uncompressed, which ends up creating 'gaps' between compressible chunks.

…tra header that makes it incompatible with the other implementations.

…ession (for decompression the NX842 driver has the same problem internally but it will fix it itself without our help).

…-linux

… for why.

…-linux into master

Pull changes from joanbm