VSwap testing

Each container is assigned RAM and SWAP limits. When a container hits the RAM limit a (slow) swapout starts. When it hits the SWAP limit an OOM killer is invoked for that container.

Note, that when a container hits its RAM limit, but there is enough RAM on the host pages do not go to *real* swap, but instead are cached in host RAM. Real swapout starts when there is no more real RAM. The same is fair for FS page cache - clean cache pages are moved to the host RAM instead of being dropped when a container hits its RAM limit so that no pagein will happen on the next read provided there is enough free RAM on the host.

The base functionality will be available in

• vzkernel-3.10.0-123.1.2.vz7.5.17
• vcmmd-7.0.6

How to test

[originally written by xemul@ for vswap]

We need the following set of atomic tests

• RAM limit works
• SWAP limit works
• OOM killer works
• Pages sent to swap can be read back again
• No real swapout starts when there's enough RAM on the node
• No leaks happen after container stop

To do this the following set of atomic tests is required: 1) RAM limit works

• Start container with unlimited RAM and SWAP.
• Note the physpages.held value
• Set RAM = physpages.held + X
• Run memeater, that will map and touch RAM pages
• _Before_ memeater exits check, that physpages.held didn't grow above the .limit

2) SWAP limit works

• Start container with unlimited RAM and SWAP.
• Note the physpages.held value
• Set RAM = physpages.held + X
• Set SWAP = Y
• Run memeater, that will map and touch RAM + SWAP pages
• Check that none of physpages.held and swappages.held grew above the .limit (memeater _may_ be killed during this test)

3) OOM killer works

• Start two containers with unlimited RAM and SWAP
• Set RAM and SWAP for both like in previous examples
• Run memeater in one of them, that consumes infinite amount of RAM.
• Check, that memeater is killed in the 1st container
• Check, that no tasks in the 2nd one are killed

(the above three can probably be made of vzt-slm test) 4) Pages sent to swap can be read back again

• Start container with unlimited RAM and SWAP.
• Note the physpages.held value
• Set RAM = physpages.held + X
• Run memeater, that will map and write to 2xRAM pages
• Check, that swappages.held is not zero
• Make memeater read the whole memory again and check that the values written are preserved

5) No real swapout starts when there's enough RAM on the node

• Take a node with non-zero swap partition
• Start a container with unlimited RAM and SWAP
• Note the physpages.held value
• Check that real swap is not used
• Set RAM = physpages.held + X
• Run memeater, that will eat 2xRAM pages
• Check, that memeater is not dead
• Before memeater exits check, that swappages.held is not zero
• Check that real swap is still not used

6) No leaks happen after container stop. This will be done using kmemleak - see PSBM-34183

More details will come later.