|
@@ -0,0 +1,75 @@
|
|
|
+When the kernel unmaps or modified the attributes of a range of
|
|
|
+memory, it has two choices:
|
|
|
+ 1. Flush the entire TLB with a two-instruction sequence. This is
|
|
|
+ a quick operation, but it causes collateral damage: TLB entries
|
|
|
+ from areas other than the one we are trying to flush will be
|
|
|
+ destroyed and must be refilled later, at some cost.
|
|
|
+ 2. Use the invlpg instruction to invalidate a single page at a
|
|
|
+ time. This could potentialy cost many more instructions, but
|
|
|
+ it is a much more precise operation, causing no collateral
|
|
|
+ damage to other TLB entries.
|
|
|
+
|
|
|
+Which method to do depends on a few things:
|
|
|
+ 1. The size of the flush being performed. A flush of the entire
|
|
|
+ address space is obviously better performed by flushing the
|
|
|
+ entire TLB than doing 2^48/PAGE_SIZE individual flushes.
|
|
|
+ 2. The contents of the TLB. If the TLB is empty, then there will
|
|
|
+ be no collateral damage caused by doing the global flush, and
|
|
|
+ all of the individual flush will have ended up being wasted
|
|
|
+ work.
|
|
|
+ 3. The size of the TLB. The larger the TLB, the more collateral
|
|
|
+ damage we do with a full flush. So, the larger the TLB, the
|
|
|
+ more attrative an individual flush looks. Data and
|
|
|
+ instructions have separate TLBs, as do different page sizes.
|
|
|
+ 4. The microarchitecture. The TLB has become a multi-level
|
|
|
+ cache on modern CPUs, and the global flushes have become more
|
|
|
+ expensive relative to single-page flushes.
|
|
|
+
|
|
|
+There is obviously no way the kernel can know all these things,
|
|
|
+especially the contents of the TLB during a given flush. The
|
|
|
+sizes of the flush will vary greatly depending on the workload as
|
|
|
+well. There is essentially no "right" point to choose.
|
|
|
+
|
|
|
+You may be doing too many individual invalidations if you see the
|
|
|
+invlpg instruction (or instructions _near_ it) show up high in
|
|
|
+profiles. If you believe that individual invalidations being
|
|
|
+called too often, you can lower the tunable:
|
|
|
+
|
|
|
+ /sys/debug/kernel/x86/tlb_single_page_flush_ceiling
|
|
|
+
|
|
|
+This will cause us to do the global flush for more cases.
|
|
|
+Lowering it to 0 will disable the use of the individual flushes.
|
|
|
+Setting it to 1 is a very conservative setting and it should
|
|
|
+never need to be 0 under normal circumstances.
|
|
|
+
|
|
|
+Despite the fact that a single individual flush on x86 is
|
|
|
+guaranteed to flush a full 2MB [1], hugetlbfs always uses the full
|
|
|
+flushes. THP is treated exactly the same as normal memory.
|
|
|
+
|
|
|
+You might see invlpg inside of flush_tlb_mm_range() show up in
|
|
|
+profiles, or you can use the trace_tlb_flush() tracepoints. to
|
|
|
+determine how long the flush operations are taking.
|
|
|
+
|
|
|
+Essentially, you are balancing the cycles you spend doing invlpg
|
|
|
+with the cycles that you spend refilling the TLB later.
|
|
|
+
|
|
|
+You can measure how expensive TLB refills are by using
|
|
|
+performance counters and 'perf stat', like this:
|
|
|
+
|
|
|
+perf stat -e
|
|
|
+ cpu/event=0x8,umask=0x84,name=dtlb_load_misses_walk_duration/,
|
|
|
+ cpu/event=0x8,umask=0x82,name=dtlb_load_misses_walk_completed/,
|
|
|
+ cpu/event=0x49,umask=0x4,name=dtlb_store_misses_walk_duration/,
|
|
|
+ cpu/event=0x49,umask=0x2,name=dtlb_store_misses_walk_completed/,
|
|
|
+ cpu/event=0x85,umask=0x4,name=itlb_misses_walk_duration/,
|
|
|
+ cpu/event=0x85,umask=0x2,name=itlb_misses_walk_completed/
|
|
|
+
|
|
|
+That works on an IvyBridge-era CPU (i5-3320M). Different CPUs
|
|
|
+may have differently-named counters, but they should at least
|
|
|
+be there in some form. You can use pmu-tools 'ocperf list'
|
|
|
+(https://github.com/andikleen/pmu-tools) to find the right
|
|
|
+counters for a given CPU.
|
|
|
+
|
|
|
+1. A footnote in Intel's SDM "4.10.4.2 Recommended Invalidation"
|
|
|
+ says: "One execution of INVLPG is sufficient even for a page
|
|
|
+ with size greater than 4 KBytes."
|