Based on kernel version 4.16.1. Page generated on 2018-04-09 11:53 EST.
1 When the kernel unmaps or modified the attributes of a range of 2 memory, it has two choices: 3 1. Flush the entire TLB with a two-instruction sequence. This is 4 a quick operation, but it causes collateral damage: TLB entries 5 from areas other than the one we are trying to flush will be 6 destroyed and must be refilled later, at some cost. 7 2. Use the invlpg instruction to invalidate a single page at a 8 time. This could potentially cost many more instructions, but 9 it is a much more precise operation, causing no collateral 10 damage to other TLB entries. 11 12 Which method to do depends on a few things: 13 1. The size of the flush being performed. A flush of the entire 14 address space is obviously better performed by flushing the 15 entire TLB than doing 2^48/PAGE_SIZE individual flushes. 16 2. The contents of the TLB. If the TLB is empty, then there will 17 be no collateral damage caused by doing the global flush, and 18 all of the individual flush will have ended up being wasted 19 work. 20 3. The size of the TLB. The larger the TLB, the more collateral 21 damage we do with a full flush. So, the larger the TLB, the 22 more attractive an individual flush looks. Data and 23 instructions have separate TLBs, as do different page sizes. 24 4. The microarchitecture. The TLB has become a multi-level 25 cache on modern CPUs, and the global flushes have become more 26 expensive relative to single-page flushes. 27 28 There is obviously no way the kernel can know all these things, 29 especially the contents of the TLB during a given flush. The 30 sizes of the flush will vary greatly depending on the workload as 31 well. There is essentially no "right" point to choose. 32 33 You may be doing too many individual invalidations if you see the 34 invlpg instruction (or instructions _near_ it) show up high in 35 profiles. If you believe that individual invalidations being 36 called too often, you can lower the tunable: 37 38 /sys/kernel/debug/x86/tlb_single_page_flush_ceiling 39 40 This will cause us to do the global flush for more cases. 41 Lowering it to 0 will disable the use of the individual flushes. 42 Setting it to 1 is a very conservative setting and it should 43 never need to be 0 under normal circumstances. 44 45 Despite the fact that a single individual flush on x86 is 46 guaranteed to flush a full 2MB [1], hugetlbfs always uses the full 47 flushes. THP is treated exactly the same as normal memory. 48 49 You might see invlpg inside of flush_tlb_mm_range() show up in 50 profiles, or you can use the trace_tlb_flush() tracepoints. to 51 determine how long the flush operations are taking. 52 53 Essentially, you are balancing the cycles you spend doing invlpg 54 with the cycles that you spend refilling the TLB later. 55 56 You can measure how expensive TLB refills are by using 57 performance counters and 'perf stat', like this: 58 59 perf stat -e 60 cpu/event=0x8,umask=0x84,name=dtlb_load_misses_walk_duration/, 61 cpu/event=0x8,umask=0x82,name=dtlb_load_misses_walk_completed/, 62 cpu/event=0x49,umask=0x4,name=dtlb_store_misses_walk_duration/, 63 cpu/event=0x49,umask=0x2,name=dtlb_store_misses_walk_completed/, 64 cpu/event=0x85,umask=0x4,name=itlb_misses_walk_duration/, 65 cpu/event=0x85,umask=0x2,name=itlb_misses_walk_completed/ 66 67 That works on an IvyBridge-era CPU (i5-3320M). Different CPUs 68 may have differently-named counters, but they should at least 69 be there in some form. You can use pmu-tools 'ocperf list' 70 (https://github.com/andikleen/pmu-tools) to find the right 71 counters for a given CPU. 72 73 1. A footnote in Intel's SDM "4.10.4.2 Recommended Invalidation" 74 says: "One execution of INVLPG is sufficient even for a page 75 with size greater than 4 KBytes."