Re: [PATCH v2 0/4] Speed up boot with faster linear map creation

From: David Hildenbrand
Date: Tue Apr 09 2024 - 07:23:03 EST


On 09.04.24 12:13, Itaru Kitayama wrote:


On Apr 9, 2024, at 19:04, Ryan Roberts <ryan.roberts@xxxxxxx> wrote:

On 09/04/2024 01:10, Itaru Kitayama wrote:
Hi Ryan,

On Apr 8, 2024, at 16:30, Ryan Roberts <ryan.roberts@xxxxxxx> wrote:

On 06/04/2024 11:31, Itaru Kitayama wrote:
Hi Ryan,

On Sat, Apr 06, 2024 at 09:32:34AM +0100, Ryan Roberts wrote:
Hi Itaru,

On 05/04/2024 08:39, Itaru Kitayama wrote:
On Thu, Apr 04, 2024 at 03:33:04PM +0100, Ryan Roberts wrote:
Hi All,

It turns out that creating the linear map can take a significant proportion of
the total boot time, especially when rodata=full. And most of the time is spent
waiting on superfluous tlb invalidation and memory barriers. This series reworks
the kernel pgtable generation code to significantly reduce the number of those
TLBIs, ISBs and DSBs. See each patch for details.

The below shows the execution time of map_mem() across a couple of different
systems with different RAM configurations. We measure after applying each patch
and show the improvement relative to base (v6.9-rc2):

| Apple M2 VM | Ampere Altra| Ampere Altra| Ampere Altra
| VM, 16G | VM, 64G | VM, 256G | Metal, 512G
---------------|-------------|-------------|-------------|-------------
| ms (%) | ms (%) | ms (%) | ms (%)
---------------|-------------|-------------|-------------|-------------
base | 153 (0%) | 2227 (0%) | 8798 (0%) | 17442 (0%)
no-cont-remap | 77 (-49%) | 431 (-81%) | 1727 (-80%) | 3796 (-78%)
batch-barriers | 13 (-92%) | 162 (-93%) | 655 (-93%) | 1656 (-91%)
no-alloc-remap | 11 (-93%) | 109 (-95%) | 449 (-95%) | 1257 (-93%)
lazy-unmap | 6 (-96%) | 61 (-97%) | 257 (-97%) | 838 (-95%)

This series applies on top of v6.9-rc2. All mm selftests pass. I've compile and
boot tested various PAGE_SIZE and VA size configs.

---

Changes since v1 [1]
====================

- Added Tested-by tags (thanks to Eric and Itaru)
- Renamed ___set_pte() -> __set_pte_nosync() (per Ard)
- Reordered patches (biggest impact & least controversial first)
- Reordered alloc/map/unmap functions in mmu.c to aid reader
- pte_clear() -> __pte_clear() in clear_fixmap_nosync()
- Reverted generic p4d_index() which caused x86 build error. Replaced with
unconditional p4d_index() define under arm64.


[1] https://lore.kernel.org/linux-arm-kernel/20240326101448.3453626-1-ryan.roberts@xxxxxxx/<https://lore.kernel.org/linux-arm-kernel/20240326101448.3453626-1-ryan.roberts@xxxxxxx/>

Thanks,
Ryan


Ryan Roberts (4):
arm64: mm: Don't remap pgtables per-cont(pte|pmd) block
arm64: mm: Batch dsb and isb when populating pgtables
arm64: mm: Don't remap pgtables for allocate vs populate
arm64: mm: Lazily clear pte table mappings from fixmap

arch/arm64/include/asm/fixmap.h | 5 +-
arch/arm64/include/asm/mmu.h | 8 +
arch/arm64/include/asm/pgtable.h | 13 +-
arch/arm64/kernel/cpufeature.c | 10 +-
arch/arm64/mm/fixmap.c | 11 +
arch/arm64/mm/mmu.c | 377 +++++++++++++++++++++++--------
6 files changed, 319 insertions(+), 105 deletions(-)

--
2.25.1


I've build and boot tested the v2 on FVP, base is taken from your
linux-rr repo. Running run_vmtests.sh on v2 left some gup longterm not oks, would you take a look at it? The mm ksefltests used is from your linux-rr repo too.

Thanks for taking a look at this.

I can't reproduce your issue unfortunately; steps as follows on Apple M2 VM:

Config: arm64 defconfig + the following:

# Squashfs for snaps, xfs for large file folios.
./scripts/config --enable CONFIG_SQUASHFS_LZ4
./scripts/config --enable CONFIG_SQUASHFS_LZO
./scripts/config --enable CONFIG_SQUASHFS_XZ
./scripts/config --enable CONFIG_SQUASHFS_ZSTD
./scripts/config --enable CONFIG_XFS_FS

# For general mm debug.
./scripts/config --enable CONFIG_DEBUG_VM
./scripts/config --enable CONFIG_DEBUG_VM_MAPLE_TREE
./scripts/config --enable CONFIG_DEBUG_VM_RB
./scripts/config --enable CONFIG_DEBUG_VM_PGFLAGS
./scripts/config --enable CONFIG_DEBUG_VM_PGTABLE
./scripts/config --enable CONFIG_PAGE_TABLE_CHECK

# For mm selftests.
./scripts/config --enable CONFIG_USERFAULTFD
./scripts/config --enable CONFIG_TEST_VMALLOC
./scripts/config --enable CONFIG_GUP_TEST

Running on VM with 12G memory, split across 2 (emulated) NUMA nodes (needed by
some mm selftests), with kernel command line to reserve hugetlbs and other
features required by some mm selftests:

"
transparent_hugepage=madvise earlycon root=/dev/vda2 secretmem.enable
hugepagesz=1G hugepages=0:2,1:2 hugepagesz=32M hugepages=0:2,1:2
default_hugepagesz=2M hugepages=0:64,1:64 hugepagesz=64K hugepages=0:2,1:2
"

Ubuntu userspace running off XFS rootfs. Build and run mm selftests from same
git tree.


Although I don't think any of this config should make a difference to gup_longterm.

Looks like your errors are all "ftruncate() failed". I've seen this problem on
our CI system. There it is due to running the tests from NFS file system. What
filesystem are you using? Perhaps you are sharing into the FVP using 9p? That
might also be problematic.

That was it. This time I booted up the kernel including your series on
QEMU on my M1 and executed the gup_longterm program without the ftruncate
failures. When testing your kernel on FVP, I was executing the script from the FVP's host filesystem using 9p.

I'm not sure exactly what the root cause is. Perhaps there isn't enough space on
the disk? It might be worth enhancing the error log to provide the errno in
tools/testing/selftests/mm/gup_longterm.c.


Attached is the strace’d gup_longterm executiong log on your
pgtable-boot-speedup-v2 kernel.

Sorry are you saying that it only fails with the pgtable-boot-speedup-v2 patch
set applied? I thought we previously concluded that it was independent of that?
I was under the impression that it was filesystem related and not something that
I was planning to investigate.

No, irrespective of the kernel, if using 9p on FVP the test program fails.
It is indeed 9p filesystem related, as I switched to using NFS all the issues are gone.

Did it never work on 9p? If so, we might have to SKIP that test.

openat(AT_FDCWD, "gup_longterm.c_tmpfile_BLboOt", O_RDWR|O_CREAT|O_EXCL, 0600) = 3
unlinkat(AT_FDCWD, "gup_longterm.c_tmpfile_BLboOt", 0) = 0
fstatfs(3, 0xffffe505a840) = -1 EOPNOTSUPP (Operation not supported)
ftruncate(3, 4096) = -1 ENOENT (No such file or directory)


fstatfs() fails and makes get_fs_type() simply say "0" -- IOW "I don't know",
which should be fine here, as it will make fs_is_unknown() trigger for relevant
cases where the type matters.

ftruncate() failing with ENOENT seems to be the problem.

But that error is a bit weird.

The man page says "ENOENT The named file does not exist.", which should only apply to
truncate() but not ftruncate().

Sound weird, but maybe that's the way to say here "not supported" ?

--
Cheers,

David / dhildenb