io: increase MAX_BUF from 16384 to 2MiB

[brxken128]

Motivation

While working on Spacedrive, we were noticing some oddities during file reading in regards to encryption and decryption.

We use a stream reader where 1MiB of the source file is read, encrypted, written and this repeats until the end. We arguably need to use read for this, as opposed to read_exact, as we need to know the exact amount of bytes read and guarantee that if the full 1MiB isn't read, the buffer still contains valid data.

I noticed some oddities while calling read, and that it was only reading 16KB at a time. This would have been fine, but this results in 64 read calls until the 1MiB buffer is populated.

Solution

The 64 read calls to populate a 1MiB buffer turned out to be extremely slow. By increasing the buffer size to 2MiB, we were able to bring encryption time down from 24~ seconds to 8.7~ seconds.

I have done some testing regarding the 2MiB value. 1MiB seemed to be a little too low, and gave us a time-to-encrypt of 15~ seconds. 4MiB gave us 8.7~ seconds also, but 2MiB should be cheaper and more universal.

This value could likely be configurable somewhere, but I'm not too sure how the API would work there.

On another note, 2MiB may be too large for some platforms and I'm not too sure if this is allocated on the stack or heap. In the event it's stack allocated, we may be better off increasing the default to 32/64KB and allowing it to be configurable.

[Darksonn]

It is heap allocated.

[Darksonn]

LGTM.