On Aug 20, 2019, at 9:05 AM, Song Liu songliubraving@fb.com wrote:
On Aug 20, 2019, at 7:18 AM, Dave Hansen dave.hansen@intel.com wrote:
On 8/20/19 7:14 AM, Song Liu wrote:
*But*, that shouldn't get hit on a Skylake CPU since those have PCIDs and shouldn't have a global kernel image. Could you confirm whether PCIDs are supported on this CPU?
Yes, pcid is listed in /proc/cpuinfo.
So what's going on? Could you confirm exactly which pti_clone_pgtable() is causing you problems? Do you have a theory as to why this manifests as a performance problem rather than a functional one?
A diff of these:
/sys/kernel/debug/page_tables/current_user /sys/kernel/debug/page_tables/current_kernel
before and after your patch might be helpful.
I believe the difference is from the following entries (7 PMDs)
Before the patch:
current_kernel: 0xffffffff81000000-0xffffffff81e04000 14352K ro GLB x pte efi: 0xffffffff81000000-0xffffffff81e04000 14352K ro GLB x pte kernel: 0xffffffff81000000-0xffffffff81e04000 14352K ro GLB x pte
After the patch:
current_kernel: 0xffffffff81000000-0xffffffff81e00000 14M ro PSE GLB x pmd efi: 0xffffffff81000000-0xffffffff81e00000 14M ro PSE GLB x pmd kernel: 0xffffffff81000000-0xffffffff81e00000 14M ro PSE GLB x pmd
current_kernel and kernel show same data though.
A little more details on how I got here.
We use huge page for hot text and thus reduces iTLB misses. As we benchmark 5.2 based kernel (vs. 4.16 based), we found ~2.5x more iTLB misses.
To figure out the issue, I use a debug patch that dumps page table for a pid. The following are information from the workload pid.
For the 4.16 based kernel:
host-4.16 # grep "x pmd" /sys/kernel/debug/page_tables/dump_pid 0x0000000000600000-0x0000000000e00000 8M USR ro PSE x pmd 0xffffffff81a00000-0xffffffff81c00000 2M ro PSE x pmd
For the 5.2 based kernel before this patch:
host-5.2-before # grep "x pmd" /sys/kernel/debug/page_tables/dump_pid 0x0000000000600000-0x0000000000e00000 8M USR ro PSE x pmd
The 8MB text in pmd is from user space. 4.16 kernel has 1 pmd for the irq entry table; while 4.16 kernel doesn't have it.
For the 5.2 based kernel after this patch:
host-5.2-after # grep "x pmd" /sys/kernel/debug/page_tables/dump_pid 0x0000000000600000-0x0000000000e00000 8M USR ro PSE x pmd 0xffffffff81000000-0xffffffff81e00000 14M ro PSE GLB x pmd
So after this patch, the 5.2 based kernel has 7 PMDs instead of 1 PMD in 4.16 kernel. This further reduces iTLB miss rate
Thanks, Song