[RESEND PATCH v10 00/12] LUF(Lazy Unmap Flush) reducing tlb numbers over 90%

From: Byungchul Park
Date: Sun May 19 2024 - 22:18:08 EST

Hi everyone,

While I'm working with a tiered memory system e.g. CXL memory, I have
been facing migration overhead esp. tlb shootdown on promotion or
demotion between different tiers. Yeah.. most tlb shootdowns on
migration through hinting fault can be avoided thanks to Huang Ying's
work, commit 4d4b6d66db ("mm,unmap: avoid flushing tlb in batch if PTE
is inaccessible"). See the following link for more information:

https://lore.kernel.org/lkml/20231115025755.GA29979@xxxxxxxxxxxxxxxxxxx/

However, it's only for migration through hinting fault. I thought it'd
be much better if we have a general mechanism to reduce all the tlb
numbers that we can apply to any unmap code, that we normally believe
tlb flush should be followed.

I'm suggesting a new mechanism, LUF(Lazy Unmap Flush), defers tlb flush
until folios that have been unmapped and freed, eventually get allocated
again. It's safe for folios that had been mapped read-only and were
unmapped, since the contents of the folios don't change while staying in
pcp or buddy so we can still read the data through the stale tlb entries.

tlb flush can be defered when folios get unmapped as long as it
guarantees to perform tlb flush needed, before the folios actually
become used, of course, only if all the corresponding ptes don't have
write permission. Otherwise, the system will get messed up.

To achieve that:

1. For the folios that map only to non-writable tlb entries, prevent
tlb flush during unmapping but perform it just before the folios
actually become used, out of buddy or pcp.

2. When any non-writable ptes change to writable e.g. through fault
handler, give up luf mechanism and perform tlb flush required
right away.

3. When a writable mapping is created e.g. through mmap(), give up
luf mechanism and perform tlb flush required right away.

No matter what type of workload is used for performance evaluation, the
result would be positive thanks to the unconditional reduction of tlb
flushes, tlb misses and interrupts. For the test, I picked up one of
the most popular and heavy workload, llama.cpp that is a
LLM(Large Language Model) inference engine.

The result would depend on memory latency and how often reclaim runs,
which implies tlb miss overhead and how many times unmapping happens.
In my system, the result shows:

1. tlb flushes are reduced about 95%.
2. tlb misses(itlb) are reduced about 80%.
3. tlb misses(dtlb store) are reduced about 57%.
4. tlb misses(dtlb load) are reduced about 24%.
5. tlb shootdown interrupts are reduced about 95%.
6. The test program runtime is reduced about 5%.

The test environment and the result is like:

Machine: bare metal, x86_64, Intel(R) Xeon(R) Gold 6430
CPU: 1 socket 64 core with hyper thread on
Numa: 2 nodes (64 CPUs DRAM 42GB, no CPUs CXL expander 98GB)
Config: swap off, numa balancing tiering on, demotion enabled

The test set:

llama.cpp/main -m $(70G_model1) -p "who are you?" -s 1 -t 15 -n 20 &
llama.cpp/main -m $(70G_model2) -p "who are you?" -s 1 -t 15 -n 20 &
llama.cpp/main -m $(70G_model3) -p "who are you?" -s 1 -t 15 -n 20 &
wait

where -t: nr of threads, -s: seed used to make the runtime stable,
-n: nr of tokens that determines the runtime, -p: prompt to ask,
-m: LLM model to use.

Run the test set 10 times successively with caches dropped every run
via 'echo 3 > /proc/sys/vm/drop_caches'. Each inference prints its
runtime at the end of each.

1. Runtime from the output of llama.cpp:

BEFORE
------
llama_print_timings: total time = 1002461.95 ms / 24 tokens
llama_print_timings: total time = 1044978.38 ms / 24 tokens
llama_print_timings: total time = 1000653.09 ms / 24 tokens
llama_print_timings: total time = 1047104.80 ms / 24 tokens
llama_print_timings: total time = 1069430.36 ms / 24 tokens
llama_print_timings: total time = 1068201.16 ms / 24 tokens
llama_print_timings: total time = 1078092.59 ms / 24 tokens
llama_print_timings: total time = 1073200.45 ms / 24 tokens
llama_print_timings: total time = 1067136.00 ms / 24 tokens
llama_print_timings: total time = 1076442.56 ms / 24 tokens
llama_print_timings: total time = 1004142.64 ms / 24 tokens
llama_print_timings: total time = 1042942.65 ms / 24 tokens
llama_print_timings: total time = 999933.76 ms / 24 tokens
llama_print_timings: total time = 1046548.83 ms / 24 tokens
llama_print_timings: total time = 1068671.48 ms / 24 tokens
llama_print_timings: total time = 1068285.76 ms / 24 tokens
llama_print_timings: total time = 1077789.63 ms / 24 tokens
llama_print_timings: total time = 1071558.93 ms / 24 tokens
llama_print_timings: total time = 1066181.55 ms / 24 tokens
llama_print_timings: total time = 1076767.53 ms / 24 tokens
llama_print_timings: total time = 1004065.63 ms / 24 tokens
llama_print_timings: total time = 1044522.13 ms / 24 tokens
llama_print_timings: total time = 999725.33 ms / 24 tokens
llama_print_timings: total time = 1047510.77 ms / 24 tokens
llama_print_timings: total time = 1068010.27 ms / 24 tokens
llama_print_timings: total time = 1068999.31 ms / 24 tokens
llama_print_timings: total time = 1077648.05 ms / 24 tokens
llama_print_timings: total time = 1071378.96 ms / 24 tokens
llama_print_timings: total time = 1066326.32 ms / 24 tokens
llama_print_timings: total time = 1077088.92 ms / 24 tokens

AFTER
-----
llama_print_timings: total time = 988522.03 ms / 24 tokens
llama_print_timings: total time = 997204.52 ms / 24 tokens
llama_print_timings: total time = 996605.86 ms / 24 tokens
llama_print_timings: total time = 991985.50 ms / 24 tokens
llama_print_timings: total time = 1035143.31 ms / 24 tokens
llama_print_timings: total time = 993660.18 ms / 24 tokens
llama_print_timings: total time = 983082.14 ms / 24 tokens
llama_print_timings: total time = 990431.36 ms / 24 tokens
llama_print_timings: total time = 992707.09 ms / 24 tokens
llama_print_timings: total time = 992673.27 ms / 24 tokens
llama_print_timings: total time = 989285.43 ms / 24 tokens
llama_print_timings: total time = 996710.06 ms / 24 tokens
llama_print_timings: total time = 996534.64 ms / 24 tokens
llama_print_timings: total time = 991344.17 ms / 24 tokens
llama_print_timings: total time = 1035210.84 ms / 24 tokens
llama_print_timings: total time = 994714.13 ms / 24 tokens
llama_print_timings: total time = 984184.15 ms / 24 tokens
llama_print_timings: total time = 990909.45 ms / 24 tokens
llama_print_timings: total time = 991881.48 ms / 24 tokens
llama_print_timings: total time = 993918.03 ms / 24 tokens
llama_print_timings: total time = 990061.34 ms / 24 tokens
llama_print_timings: total time = 998076.69 ms / 24 tokens
llama_print_timings: total time = 997082.59 ms / 24 tokens
llama_print_timings: total time = 990677.58 ms / 24 tokens
llama_print_timings: total time = 1036054.94 ms / 24 tokens
llama_print_timings: total time = 994125.93 ms / 24 tokens
llama_print_timings: total time = 982467.01 ms / 24 tokens
llama_print_timings: total time = 990191.60 ms / 24 tokens
llama_print_timings: total time = 993319.24 ms / 24 tokens
llama_print_timings: total time = 992540.57 ms / 24 tokens

2. tlb shootdowns from 'cat /proc/interrupts':

BEFORE
------
TLB:
125553646 141418810 161932620 176853972 186655697 190399283
192143823 196414038 192872439 193313658 193395617 192521416
190788161 195067598 198016061 193607347 194293972 190786732
191545637 194856822 191801931 189634535 190399803 196365922
195268398 190115840 188050050 193194908 195317617 190820190
190164820 185556071 226797214 229592631 216112464 209909495
205575979 205950252 204948111 197999795 198892232 205287952
199344631 195015158 195869844 198858745 195692876 200961904
203463252 205921722 199850838 206145986 199613202 199961345
200129577 203020521 207873649 203697671 197093386 204243803
205993323 200934664 204193128 194435376 TLB shootdowns

AFTER
-----
TLB:
5648092 6610142 7032849 7882308 8088518 8352310
8656536 8705136 8647426 8905583 8985408 8704522
8884344 9026261 8929974 8869066 8877575 8810096
8770984 8754503 8801694 8865925 8787524 8656432
8755912 8682034 8773935 8832925 8797997 8515777
8481240 8891258 10595243 10285973 9756935 9573681
9398968 9069244 9242984 8899009 9310690 9029095
9069758 9105825 9092703 9270202 9460287 9258546
9180415 9232723 9270611 9175020 9490420 9360316
9420818 9057663 9525631 9310152 9152242 8654483
9181804 9050847 8919916 8883856 TLB shootdowns

3. tlb numbers from 'perf stat' per test set:

BEFORE
------
3163679332 dTLB-load-misses
2017751856 dTLB-store-misses
327092903 iTLB-load-misses
1357543886 tlb:tlb_flush

AFTER
-----
2394694609 dTLB-load-misses
861144167 dTLB-store-misses
64055579 iTLB-load-misses
69175002 tlb:tlb_flush

---

Changes from v9:

1. Expand the candidate to apply this mechanism:

BEFORE - The souce folios at any type of migration.
AFTER - Any folios that have been unmapped and freed.

2. Change the workload for test:

BEFORE - XSBench
AFTER - llama.cpp (one of the most popluar real workload)

3. Change the test environment:

BEFORE - qemu machine, too small DRAM(1GB), large remote mem
AFTER - bare metal, real CXL memory, practical memory size

4. Rename the mechanism from MIGRC(Migration Read Copy) to
LUF(Lazy Unmap Flush) to reflect the current version of the
mechanism can be applied not only to unmap during migration
but any unmap code e.g. unmap in shrink_folio_list().

5. Fix build error for riscv. (feedbacked by kernel test bot)

6. Supplement commit messages to describe what this mechanism is
for, especially in the patches for arch code. (feedbacked by
Thomas Gleixner)

7. Clean up some trivial things.

Changes from v8:

1. Rebase on akpm/mm.git mm-unstable as of April 18, 2024.
2. Supplement comments and commit message.
3. Change the candidate to apply migrc mechanism:

BEFORE - The source folios at demotion and promotion.
AFTER - The souce folios at any type of migration.

4. Change how migrc mechanism works:

BEFORE - Reduce tlb flushes by deferring folio_free() for
source folios during demotion and promotion.
AFTER - Reduce tlb flushes by deferring tlb flush until they
actually become used, out of pcp or buddy. The
current version of migrc does *not* defer calling
folio_free() but let it go as it is as the same as
vanilla kernel, with the folios marked kind of 'need
to tlb flush'. And then handle the flush when the
page exits from pcp or buddy so as to prevent
changing vm stats e.g. free pages.

Changes from v7:

1. Rewrite cover letter to explain what 'migrc' mechasism is.
(feedbacked by Andrew Morton)
2. Supplement the commit message of a patch 'mm: Add APIs to
free a folio directly to the buddy bypassing pcp'.
(feedbacked by Andrew Morton)

Changes from v6:

1. Fix build errors in case of
CONFIG_ARCH_WANT_BATCHED_UNMAP_tlb_FLUSH disabled by moving
migrc_flush_{start,end}() calls from arch code to
try_to_unmap_flush() in mm/rmap.c.

Changes from v5:

1. Fix build errors in case of CONFIG_MIGRATION disabled or
CONFIG_HWPOISON_INJECT moduled. (feedbacked by kernel test
bot and Raymond Jay Golo)
2. Organize migrc code with two kconfigs, CONFIG_MIGRATION and
CONFIG_ARCH_WANT_BATCHED_UNMAP_tlb_FLUSH.

Changes from v4:

1. Rebase on v6.7.
2. Fix build errors in arm64 that is doing nothing for tlb flush
but has CONFIG_ARCH_WANT_BATCHED_UNMAP_tlb_FLUSH. (reported
by kernel test robot)
3. Don't use any page flag. So the system would give up migrc
mechanism more often but it's okay. The final improvement is
good enough.
4. Instead, optimize full tlb flush(arch_tlbbatch_flush()) by
avoiding redundant CPUs from tlb flush.

Changes from v3:

1. Don't use the kconfig, CONFIG_MIGRC, and remove sysctl knob,
migrc_enable. (feedbacked by Nadav)
2. Remove the optimization skipping CPUs that have already
performed tlb flushes needed by any reason when performing
tlb flushes by migrc because I can't tell the performance
difference between w/ the optimization and w/o that.
(feedbacked by Nadav)
3. Minimize arch-specific code. While at it, move all the migrc
declarations and inline functions from include/linux/mm.h to
mm/internal.h (feedbacked by Dave Hansen, Nadav)
4. Separate a part making migrc paused when the system is in
high memory pressure to another patch. (feedbacked by Nadav)
5. Rename:
a. arch_tlbbatch_clean() to arch_tlbbatch_clear(),
b. tlb_ubc_nowr to tlb_ubc_ro,
c. migrc_try_flush_free_folios() to migrc_flush_free_folios(),
d. migrc_stop to migrc_pause.
(feedbacked by Nadav)
6. Use ->lru list_head instead of introducing a new llist_head.
(feedbacked by Nadav)
7. Use non-atomic operations of page-flag when it's safe.
(feedbacked by Nadav)
8. Use stack instead of keeping a pointer of 'struct migrc_req'
in struct task, which is for manipulating it locally.
(feedbacked by Nadav)
9. Replace a lot of simple functions to inline functions placed
in a header, mm/internal.h. (feedbacked by Nadav)
10. Add additional sufficient comments. (feedbacked by Nadav)
11. Remove a lot of wrapper functions. (feedbacked by Nadav)

Changes from RFC v2:

1. Remove additional occupation in struct page. To do that,
unioned with lru field for migrc's list and added a page
flag. I know page flag is a thing that we don't like to add
but no choice because migrc should distinguish folios under
migrc's control from others. Instead, I force migrc to be
used only on 64 bit system to mitigate you guys from getting
angry.
2. Remove meaningless internal object allocator that I
introduced to minimize impact onto the system. However, a ton
of tests showed there was no difference.
3. Stop migrc from working when the system is in high memory
pressure like about to perform direct reclaim. At the
condition where the swap mechanism is heavily used, I found
the system suffered from regression without this control.
4. Exclude folios that pte_dirty() == true from migrc's interest
so that migrc can work simpler.
5. Combine several patches that work tightly coupled to one.
6. Add sufficient comments for better review.
7. Manage migrc's request in per-node manner (from globally).
8. Add tlb miss improvement in commit message.
9. Test with more CPUs(4 -> 16) to see bigger improvement.

Changes from RFC:

1. Fix a bug triggered when a destination folio at the previous
migration becomes a source folio at the next migration,
before the folio gets handled properly so that the folio can
play with another migration. There was inconsistency in the
folio's state. Fixed it.
2. Split the patch set into more pieces so that the folks can
review better. (Feedbacked by Nadav Amit)
3. Fix a wrong usage of barrier e.g. smp_mb__after_atomic().
(Feedbacked by Nadav Amit)
4. Tried to add sufficient comments to explain the patch set
better. (Feedbacked by Nadav Amit)

Byungchul Park (12):
x86/tlb: add APIs manipulating tlb batch's arch data
arm64: tlbflush: add APIs manipulating tlb batch's arch data
riscv, tlb: add APIs manipulating tlb batch's arch data
x86/tlb, riscv/tlb, mm/rmap: separate arch_tlbbatch_clear() out of
arch_tlbbatch_flush()
mm: buddy: make room for a new variable, ugen, in struct page
mm: add folio_put_ugen() to deliver unmap generation number to pcp or
buddy
mm: add a parameter, unmap generation number, to free_unref_folios()
mm/rmap: recognize read-only tlb entries during batched tlb flush
mm: implement LUF(Lazy Unmap Flush) defering tlb flush when folios get
unmapped
mm: separate move/undo parts from migrate_pages_batch()
mm, migrate: apply luf mechanism to unmapping during migration
mm, vmscan: apply luf mechanism to unmapping during folio reclaim

arch/arm64/include/asm/tlbflush.h | 18 ++
arch/riscv/include/asm/tlbflush.h | 21 ++
arch/riscv/mm/tlbflush.c | 1 -
arch/x86/include/asm/tlbflush.h | 18 ++
arch/x86/mm/tlb.c | 2 -
include/linux/mm.h | 22 ++
include/linux/mm_types.h | 40 +++-
include/linux/rmap.h | 7 +-
include/linux/sched.h | 11 +
mm/compaction.c | 10 +
mm/internal.h | 115 +++++++++-
mm/memory.c | 8 +
mm/migrate.c | 184 ++++++++++------
mm/mmap.c | 8 +
mm/page_alloc.c | 157 +++++++++++---
mm/page_isolation.c | 6 +
mm/page_reporting.c | 10 +
mm/rmap.c | 345 +++++++++++++++++++++++++++++-
mm/swap.c | 18 +-
mm/vmscan.c | 29 ++-
20 files changed, 904 insertions(+), 126 deletions(-)


base-commit: f52bcd4a9f6058704a6f6b6b50418f579defd4fe
--
2.17.1