Re: [PATCH 4/5] locking/lockdep: Reuse free chain_hlocks entries

From: Peter Zijlstra
Date: Fri Dec 13 2019 - 05:50:54 EST


On Fri, Dec 13, 2019 at 11:25:25AM +0100, Peter Zijlstra wrote:

> void push_block(struct chain_block **bp, struct chain_block *b)
> {
> b->next = *bp;
> *bp = b;
> }
>
> /* could contemplate ilog2() buckets */
> int size2bucket(int size)
> {
> return size >= MAX_BUCKET ? 0 : size;
> }

If you make the allocation granularity 4 entries, then you can have
push_block() store the chain_block * at the start of every free entry,
which would enable merging adjecent blocks.

That is, if I'm not mistaken these u16 chain entries encode the class
index (per lock_chain_get_class()). And since the lock_classes array is
MAX_LOCKDEP_KEYS, or 13 bits, big, we have the MSB of each entry spare.

union {
struct {
u16 hlock[4];
}
u64 addr;
} ponies;

So if we make the rule that:

!(idx % 4) && (((union ponies *)chain_hlock[idx])->addr & BIT_ULL(63))

encodes a free block at said addr, then we should be able to detect if:

chain_hlock[b->base + b->size]

is also free and merge the blocks.

(we just have to fix up the MSB of the address, not all arches have
negative addresses for kernel space)