Fine-grain Page Table Lock¶
In old Linux or previous Lego, user page table operations, such as set, clear, are protected by mm->page_table_lock. This one single lock prohibits a lot parallelisms on big SMP machines. An ideal solution is to have finer-granularity locks, so that faults on different parts of the user address space can be handled with less contention.
But finer-granularity locks means you need more memory for the locks. This is a simple trade-off. Lego currently mimic the Linux x86 default setting1, where each PMD and PTE page table pages has their own lock. The lock is a spinlock embedded in the struct page. As illustrated below:
Both Processor and Memory managers are using the same mechanism to increase parallelism. And it is something that can improve performance a lot.
–
Yizhou Shan
Created: Mar 22, 2018
Last Updated: April 13, 2018