4 ms·
Perhaps a better (and simpler) workaround, then, could be to clamp the reported cache line size to 64 bytes. So even if the Samsung core reports 128-byte cache
by cesarb 10y ago
Perhaps a better (and simpler) workaround, then, could be to clamp the reported cache line size to 64 bytes. So even if the Samsung core reports 128-byte cache lines, the code would simply invalidate each line twice, and if it is migrated in the middle of the invalidation loop, it would still work correctly.
- gcp 10y agoThis is exactly the fix used in the kernel, yes. The trickiness and complexity comes from the cache line size probe having to be intercepted.