forked from Minki/linux
89ca4e126a
When switching processes, currently all user SLBEs are cleared, and a few (exec_base, pc, and stack) are preloaded. In trivial testing with small apps, this tends to miss the heap and low 256MB segments, and it will also miss commonly accessed segments on large memory workloads. Add a simple round-robin preload cache that just inserts the last SLB miss into the head of the cache and preloads those at context switch time. Every 256 context switches, the oldest entry is removed from the cache to shrink the cache and require fewer slbmte if they are unused. Much more could go into this, including into the SLB entry reclaim side to track some LRU information etc, which would require a study of large memory workloads. But this is a simple thing we can do now that is an obvious win for common workloads. With the full series, process switching speed on the context_switch benchmark on POWER9/hash (with kernel speculation security masures disabled) increases from 140K/s to 178K/s (27%). POWER8 does not change much (within 1%), it's unclear why it does not see a big gain like POWER9. Booting to busybox init with 256MB segments has SLB misses go down from 945 to 69, and with 1T segments 900 to 21. These could almost all be eliminated by preloading a bit more carefully with ELF binary loading. Signed-off-by: Nicholas Piggin <npiggin@gmail.com> Signed-off-by: Michael Ellerman <mpe@ellerman.id.au> |
||
|---|---|---|
| .. | ||
| alpha | ||
| arc | ||
| arm | ||
| arm64 | ||
| c6x | ||
| h8300 | ||
| hexagon | ||
| ia64 | ||
| m68k | ||
| microblaze | ||
| mips | ||
| nds32 | ||
| nios2 | ||
| openrisc | ||
| parisc | ||
| powerpc | ||
| riscv | ||
| s390 | ||
| sh | ||
| sparc | ||
| um | ||
| unicore32 | ||
| x86 | ||
| xtensa | ||
| .gitignore | ||
| Kconfig | ||