On 5/7/25 12:58 AM, Ryan Roberts wrote:

On 05/05/2025 22:39, Yang Shi wrote:

quoted

On 5/2/25 4:51 AM, Ryan Roberts wrote:

quoted

On 14/04/2025 22:24, Yang Shi wrote:

quoted

On 4/14/25 6:03 AM, Ryan Roberts wrote:

quoted

On 10/04/2025 23:00, Yang Shi wrote:

quoted

Hi Ryan,

I know you may have a lot of things to follow up after LSF/MM. Just gently
ping,
hopefully we can resume the review soon.

Hi, I'm out on holiday at the moment, returning on the 22nd April. But I'm very
keen to move this series forward so will come back to you next week. (although
TBH, I thought I was waiting for you to respond to me... :-| )

FWIW, having thought about it a bit more, I think some of the suggestions I
previously made may not have been quite right, but I'll elaborate next week.
I'm
keen to build a pgtable splitting primitive here that we can reuse with vmalloc
as well to enable huge mappings by default with vmalloc too.

Sounds good. I think the patches can support splitting vmalloc page table too.
Anyway we can discuss more after you are back. Enjoy your holiday.

Hi Yang,

Sorry I've taken so long to get back to you. Here's what I'm currently thinking:
I'd eventually like to get to the point where the linear map and most vmalloc
memory is mapped using the largest possible mapping granularity (i.e. block
mappings at PUD/PMD, and contiguous mappings at PMD/PTE level).

vmalloc has history with trying to do huge mappings by default; it ended up
having to be turned into an opt-in feature (instead of the original opt-out
approach) because there were problems with some parts of the kernel expecting
page mappings. I think we might be able to overcome those issues on arm64 with
BBML2.

arm64 can already support vmalloc PUD and PMD block mappings, and I have a
series (that should make v6.16) that enables contiguous PTE mappings in vmalloc
too. But these are currently limited to when VM_ALLOW_HUGE is specified. To be
able to use that by default, we need to be able to change permissions on
sub-regions of an allocation, which is where BBML2 and your series come in.
(there may be other things we need to solve as well; TBD).

I think the key thing we need is a function that can take a page-aligned kernel
VA, will walk to the leaf entry for that VA and if the VA is in the middle of
the leaf entry, it will split it so that the VA is now on a boundary. This will
work for PUD/PMD block entries and contiguous-PMD/contiguous-PTE entries. The
function can assume BBML2 is present. And it will return 0 on success, -EINVAL
if the VA is not mapped or -ENOMEM if it couldn't allocate a pgtable to perform
the split.

OK, the v3 patches already handled page table allocation failure with returning
-ENOMEM and BUG_ON if it is not mapped because kernel assumes linear mapping
should be always present. It is easy to return -EINVAL instead of BUG_ON.
However I'm wondering what usecases you are thinking about? Splitting vmalloc
area may run into unmapped VA?

I don't think BUG_ON is the right behaviour; crashing the kernel should be
discouraged. I think even for vmalloc under correct conditions we shouldn't see
any unmapped VA. But vmalloc does handle it gracefully today; see (e.g.)
vunmap_pmd_range() which skips the pmd if its none.

quoted

quoted

Then we can use that primitive on the start and end address of any range for
which we need exact mapping boundaries (e.g. when changing permissions on part
of linear map or vmalloc allocation, when freeing part of a vmalloc allocation,
etc). This way we only split enough to ensure the boundaries are precise, and
keep larger mappings inside the range.

Yeah, makes sense to me.

quoted

Next we need to reimplement __change_memory_common() to not use
apply_to_page_range(), because that assumes page mappings only. Dev Jain has
been working on a series that converts this to use walk_page_range_novma() so
that we can change permissions on the block/contig entries too. That's not
posted publicly yet, but it's not huge so I'll ask if he is comfortable with
posting an RFC early next week.

OK, so the new __change_memory_common() will change the permission of page
table, right?

It will change permissions of all the leaf entries in the range of VAs it is
passed. Currently it assumes that all the leaf entries are PTEs. But we will
generalize to support all the other types of leaf entries too.,

quoted

If I remember correctly, you suggested change permissions in
__create_pgd_mapping_locked() for v3. So I can disregard it?

Yes I did. I think this made sense (in my head at least) because in the context
of the linear map, all the PFNs are contiguous so it kind-of makes sense to
reuse that infrastructure. But it doesn't generalize to vmalloc because vmalloc
PFNs are not contiguous. So for that reason, I think it's preferable to have an
independent capability.

OK, sounds good to me.

quoted

The current code assumes the address range passed in by change_memory_common()
is *NOT* physically contiguous so __change_memory_common() handles page table
permission on page basis. I'm supposed Dev's patches will handle this then my
patch can safely assume the linear mapping address range for splitting is
physically contiguous too otherwise I can't keep large mappings inside the
range. Splitting vmalloc area doesn't need to worry about this.

I'm not sure I fully understand the point you're making here...

Dev's series aims to use walk_page_range_novma() similar to riscv's
implementation so that it can walk a VA range and update the permissions on each
leaf entry it visits, regadless of which level the leaf entry is at. This
doesn't make any assumption of the physical contiguity of neighbouring leaf
entries in the page table.

So if we are changing permissions on the linear map, we have a range of VAs to
walk and convert all the leaf entries, regardless of their size. The same goes
for vmalloc... But for vmalloc, we will also want to change the underlying
permissions in the linear map, so we will have to figure out the contiguous
pieces of the linear map and call __change_memory_common() for each; there is
definitely some detail to work out there!

Yes, this is my point. When changing underlying linear map permission 
for vmalloc, the linear map address may be not contiguous. This is why 
change_memory_common() calls __change_memory_common() on page basis.

But how Dev's patch work should have no impact on how I implement the 
split primitive by thinking it further. It should be the caller's 
responsibility to make sure __create_pgd_mapping_locked() is called for 
contiguous linear map address range.

quoted

quoted

You'll still need to repaint the whole linear map with page mappings for the
case !BBML2 case, but I'm hoping __create_pgd_mapping_locked() (potentially with
minor modifications?) can do that repainting on the live mappings; similar to
how you are doing it in v3.

Yes, when repainting I need to split the page table all the way down to PTE
level. A simple flag should be good enough to tell __create_pgd_mapping_locked()
do the right thing off the top of my head.

Perhaps it may be sufficient to reuse the NO_BLOCK_MAPPINGS and NO_CONT_MAPPINGS
flags? For example, if you are find a leaf mapping and NO_BLOCK_MAPPINGS is set,
then you need to split it?

Yeah, sounds feasible. Anyway I will figure it out.

quoted

quoted

Miko's BBML2 series should hopefully get imminently queued for v6.16.

Great! Anyway my series is based on his advertising BBML2 patch.

quoted

So in summary, what I'm asking for your large block mapping the linear map
series is:
    - Paint linear map using blocks/contig if boot CPU supports BBML2
    - Repaint linear map using page mappings if secondary CPUs don't support BBML2

OK, I just need to add some simple tweak to split down to PTE level to v3.

quoted

    - Integrate Dev's __change_memory_common() series

OK, I think I have to do my patches on top of it. Because Dev's patch need
guarantee the linear mapping address range is physically contiguous.

quoted

    - Create primitive to ensure mapping entry boundary at a given page-aligned VA
    - Use primitive when changing permissions on linear map region

Sure.

quoted

This will be mergable on its own, but will also provide a great starting base
for adding huge-vmalloc-by-default.

What do you think?

Definitely makes sense to me.

If I remember correctly, we still have some unsolved comments/questions for v3
in my replies on March 17, particularly:
https://lore.kernel.org/linux-arm-kernel/2b715836-b566-4a9e-
b344-9401fa4c0feb@os.amperecomputing.com/

Ahh sorry about that. I'll take a look now...

No problem.

Thanks,
Yang

Thanks,
Ryan

quoted

Thanks,
Yang

quoted

Thanks,
Ryan

quoted

Thanks,
Yang

quoted

Thanks,
Ryan

quoted

Thanks,
Yang


On 3/13/25 10:40 AM, Yang Shi wrote:

quoted

On 3/13/25 10:36 AM, Ryan Roberts wrote:

quoted

On 13/03/2025 17:28, Yang Shi wrote:

quoted

Hi Ryan,

I saw Miko posted a new spin of his patches. There are some slight changes
that
have impact to my patches (basically check the new boot parameter). Do you
prefer I rebase my patches on top of his new spin right now then restart
review
from the new spin or review the current patches then solve the new review
comments and rebase to Miko's new spin together?

Hi Yang,

Sorry I haven't got to reviewing this version yet, it's in my queue!

I'm happy to review against v3 as it is. I'm familiar with Miko's series
and am
not too bothered about the integration with that; I think it's pretty
straight
forward. I'm more interested in how you are handling the splitting, which I
think is the bulk of the effort.

Yeah, sure, thank you.

quoted

I'm hoping to get to this next week before heading out to LSF/MM the
following
week (might I see you there?)

Unfortunately I can't make it this year. Have a fun!

Thanks,
Yang

quoted

Thanks,
Ryan

quoted

Thanks,
Yang


On 3/4/25 2:19 PM, Yang Shi wrote:

quoted

Changelog
=========
v3:
       * Rebased to v6.14-rc4.
       * Based on Miko's BBML2 cpufeature patch (https://lore.kernel.org/
linux-
arm-kernel/20250228182403.6269-3-miko.lenczewski@arm.com/).
         Also included in this series in order to have the complete
patchset.
       * Enhanced __create_pgd_mapping() to handle split as well per Ryan.
       * Supported CONT mappings per Ryan.
       * Supported asymmetric system by splitting kernel linear mapping if
such
         system is detected per Ryan. I don't have such system to test,
so the
         testing is done by hacking kernel to call linear mapping
repainting
         unconditionally. The linear mapping doesn't have any block and
cont
         mappings after booting.

RFC v2:
       * Used allowlist to advertise BBM lv2 on the CPUs which can
handle TLB
         conflict gracefully per Will Deacon
       * Rebased onto v6.13-rc5
       * https://lore.kernel.org/linux-arm-kernel/20250103011822.1257189-1-
yang@os.amperecomputing.com/

RFC v1: https://lore.kernel.org/lkml/20241118181711.962576-1-
yang@os.amperecomputing.com/

Description
===========
When rodata=full kernel linear mapping is mapped by PTE due to arm's
break-before-make rule.

A number of performance issues arise when the kernel linear map is using
PTE entries due to arm's break-before-make rule:
       - performance degradation
       - more TLB pressure
       - memory waste for kernel page table

These issues can be avoided by specifying rodata=on the kernel command
line but this disables the alias checks on page table permissions and
therefore compromises security somewhat.

With FEAT_BBM level 2 support it is no longer necessary to invalidate the
page table entry when changing page sizes.  This allows the kernel to
split large mappings after boot is complete.

This patch adds support for splitting large mappings when FEAT_BBM level 2
is available and rodata=full is used. This functionality will be used
when modifying page permissions for individual page frames.

Without FEAT_BBM level 2 we will keep the kernel linear map using PTEs
only.

If the system is asymmetric, the kernel linear mapping may be repainted
once
the BBML2 capability is finalized on all CPUs.  See patch #6 for more
details.

We saw significant performance increases in some benchmarks with
rodata=full without compromising the security features of the kernel.

Testing
=======
The test was done on AmpereOne machine (192 cores, 1P) with 256GB
memory and
4K page size + 48 bit VA.

Function test (4K/16K/64K page size)
       - Kernel boot.  Kernel needs change kernel linear mapping
permission at
         boot stage, if the patch didn't work, kernel typically didn't
boot.
       - Module stress from stress-ng. Kernel module load change permission
for
         linear mapping.
       - A test kernel module which allocates 80% of total memory via
vmalloc(),
         then change the vmalloc area permission to RO, this also change
linear
         mapping permission to RO, then change it back before vfree(). Then
launch
         a VM which consumes almost all physical memory.
       - VM with the patchset applied in guest kernel too.
       - Kernel build in VM with guest kernel which has this series
applied.
       - rodata=on. Make sure other rodata mode is not broken.
       - Boot on the machine which doesn't support BBML2.

Performance
===========
Memory consumption
Before:
MemTotal:       258988984 kB
MemFree:        254821700 kB

After:
MemTotal:       259505132 kB
MemFree:        255410264 kB

Around 500MB more memory are free to use.  The larger the machine, the
more memory saved.

Performance benchmarking
* Memcached
We saw performance degradation when running Memcached benchmark with
rodata=full vs rodata=on.  Our profiling pointed to kernel TLB pressure.
With this patchset we saw ops/sec is increased by around 3.5%, P99
latency is reduced by around 9.6%.
The gain mainly came from reduced kernel TLB misses.  The kernel TLB
MPKI is reduced by 28.5%.

The benchmark data is now on par with rodata=on too.

* Disk encryption (dm-crypt) benchmark
Ran fio benchmark with the below command on a 128G ramdisk (ext4) with
disk
encryption (by dm-crypt).
fio --directory=/data --random_generator=lfsr --norandommap --
randrepeat 1 \
         --status-interval=999 --rw=write --bs=4k --loops=1 --
ioengine=sync \
         --iodepth=1 --numjobs=1 --fsync_on_close=1 --group_reporting --
thread \
         --name=iops-test-job --eta-newline=1 --size 100G

The IOPS is increased by 90% - 150% (the variance is high, but the worst
number of good case is around 90% more than the best number of bad case).
The bandwidth is increased and the avg clat is reduced proportionally.

* Sequential file read
Read 100G file sequentially on XFS (xfs_io read with page cache
populated).
The bandwidth is increased by 150%.


Mikołaj Lenczewski (1):
           arm64: Add BBM Level 2 cpu feature

Yang Shi (5):
           arm64: cpufeature: add AmpereOne to BBML2 allow list
           arm64: mm: make __create_pgd_mapping() and helpers non-void
           arm64: mm: support large block mapping when rodata=full
           arm64: mm: support split CONT mappings
           arm64: mm: split linear mapping if BBML2 is not supported on
secondary
CPUs

      arch/arm64/Kconfig                  |  11 +++++
      arch/arm64/include/asm/cpucaps.h    |   2 +
      arch/arm64/include/asm/cpufeature.h |  15 ++++++
      arch/arm64/include/asm/mmu.h        |   4 ++
      arch/arm64/include/asm/pgtable.h    |  12 ++++-
      arch/arm64/kernel/cpufeature.c      |  95 ++++++++++++++++++++++++
++++++
+++++++
      arch/arm64/mm/mmu.c                 | 397 ++++++++++++++++++++++++
++++++
++++
+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
+++++
++++++++++++++++++++++-------------------
      arch/arm64/mm/pageattr.c            |  37 ++++++++++++---
      arch/arm64/tools/cpucaps            |   1 +
      9 files changed, 518 insertions(+), 56 deletions(-)