Avoiding devices' cache data duplication wasting ram

[eddy89]

Hi all,
I don't know very well how this works, but as far as I know, block devices have its own block cache and buffers, as long as they are opened, and filesystems have it's own file cache, and some virtual devices (e.g. loop) have their cache too. I wonder if the same data, when I read a file from the filesystem, is cached at least twice in the block cache and in the file cache...
In my case I have some particular configurations where this can become huge and very memory wasting:

-A raspberry pi (256Mb of ram) where the path from physical layer to filesystem is:
mmcblk0p1 (block cache) -> block2mtd (mtd cache?) -> ubi -> ubifs (FS cache) -> files

Another simple example is using loop:
sda (block cache) -> filesystem (file cache) -> loop device (another block cache) -> filesystem (another file cache) -> file
Here we have the same data occupying 4 times the memory it should take, without any benefit.

So, where can I check how much buffer/cache is taking a device in ram?
And where can I disable some of them (/proc? /sys? sysctl?)

Thank you for your help

[HeissFuss]

Take a look at slabtop, it may be in there.

I wouldn't expect block device buffers to be very large though, but memory for things like MD devices can be fairly large (compared to 256MB total).

The filesystem buffers would get dropped first if you started hitting memory constraints.