Re: [PATCH v8 0/5] arm64: support FEAT_BBM level 2 and large block mapping when rodata=full

From: Jinjiang Tu

Date: Wed Mar 18 2026 - 21:23:22 EST



在 2026/3/18 17:17, Ryan Roberts 写道:
On 18/03/2026 08:29, Jinjiang Tu wrote:
在 2026/3/17 17:07, Ryan Roberts 写道:
On 17/03/2026 02:06, Jinjiang Tu wrote:
在 2026/3/17 8:15, Yang Shi 写道:
On 3/16/26 8:47 AM, Ryan Roberts wrote:
Thanks for the report!

+ Kevin, who was looking at some adjacent issues and may have some ideas
for how
to fix.


On 16/03/2026 07:35, Jinjiang Tu wrote:
在 2025/9/18 3:02, Yang Shi 写道:
On systems with BBML2_NOABORT support, it causes the linear map to be mapped
with large blocks, even when rodata=full, and leads to some nice performance
improvements.
Hi,
Hi Jinjiang,

Thanks for reporting the problem.

I find this feature is incompatible with realm. The calltrace is as follows:

[    0.000000][    T0] ------------[ cut here ]------------
[    0.000000][    T0] WARNING: CPU: 0 PID: 0 at arch/arm64/mm/pageattr.c:56
pageattr_pmd_entry+0x60/0x78
[    0.000000][    T0] Modules linked in:
[    0.000000][    T0] CPU: 0 PID: 0 Comm: swapper/0 Not tainted 6.6.0 #16
[    0.000000][    T0] Hardware name: linux,dummy-virt (DT)
[    0.000000][    T0] pstate: 800000c5 (Nzcv daIF -PAN -UAO -TCO -DIT -SSBS
BTYPE=--)
[    0.000000][    T0] pc : pageattr_pmd_entry+0x60/0x78
[    0.000000][    T0] lr : walk_pmd_range.isra.0+0x170/0x1f0
[    0.000000][    T0] sp : ffffcb90a0f337d0
[    0.000000][    T0] x29: ffffcb90a0f337d0 x28: 0000000000000000 x27:
ffff0000035e0000
[    0.000000][    T0] x26: ffffcb90a0f338f8 x25: ffff00001fff60d0 x24:
ffff0000035d0000
[    0.000000][    T0] x23: 0400000000000001 x22: 0c00000000000001 x21:
ffff0000035dffff
[    0.000000][    T0] x20: ffffcb909fe3b7f0 x19: ffff0000035e0000 x18:
ffffffffffffffff
[    0.000000][    T0] x17: 7220303030303178 x16: 307e303030306435 x15:
ffffcb90a0f334c8
[    0.000000][    T0] x14: 0000000000000000 x13: 205d305420202020 x12:
5b5d303030303030
[    0.000000][    T0] x11: 00000000ffff7fff x10: 00000000ffff7fff x9 :
ffffcb909f1e27d8
[    0.000000][    T0] x8 : 00000000000bffe8 x7 : c0000000ffff7fff x6 :
0000000000000001
[    0.000000][    T0] x5 : 0000000000000001 x4 : 0078000083400705 x3 :
ffffcb90a0f338f8
[    0.000000][    T0] x2 : 0000000000010000 x1 : ffff0000035d0000 x0 :
ffff00001fff60d0
[    0.000000][    T0] Call trace:
[    0.000000][    T0]  pageattr_pmd_entry+0x60/0x78
[    0.000000][    T0]  walk_pud_range+0x124/0x190
[    0.000000][    T0]  walk_pgd_range+0x158/0x1b0
[    0.000000][    T0] walk_kernel_page_table_range_lockless+0x58/0x98
[    0.000000][    T0]  update_range_prot+0xb8/0x108
[    0.000000][    T0]  __change_memory_common+0x30/0x1a8
[    0.000000][    T0] __set_memory_enc_dec.part.0+0x170/0x260
[    0.000000][    T0]  realm_set_memory_decrypted+0x6c/0xb0
[    0.000000][    T0]  set_memory_decrypted+0x38/0x58
[    0.000000][    T0]  its_alloc_pages_node+0xc4/0x140
[    0.000000][    T0]  its_probe_one+0xbc/0x3c0
[    0.000000][    T0]  its_of_probe.isra.0+0x130/0x220
[    0.000000][    T0]  its_init+0x160/0x2f8
[    0.000000][    T0]  gic_init_bases+0x1fc/0x318
[    0.000000][    T0]  gic_of_init+0x2a0/0x300
[    0.000000][    T0]  of_irq_init+0x238/0x4b8
[    0.000000][    T0]  irqchip_init+0x20/0x50
[    0.000000][    T0]  init_IRQ+0x1c/0x100
[    0.000000][    T0]  start_kernel+0x1ec/0x4f0
[    0.000000][    T0]  __primary_switched+0xbc/0xd0
[    0.000000][    T0] ---[ end trace 0000000000000000 ]---
[    0.000000][    T0] ------------[ cut here ]------------
[    0.000000][    T0] Failed to decrypt memory, 16 pages will be leaked

realm feature relies on rodata=full to dynamically update kernel page table
prot.

In init_IRQ(), realm_set_memory_decrypted() is called to update kernel page
table prot.
At this time, secondary cpus aren't booted, BBML2 noabort feature isn't
initializated,
and system_supports_bbml2_noabort() still returns false. As a result,
split_kernel_leaf_mapping() is skipped, leading to WARN_ON_ONCE((next -
addr) !=
PMD_SIZE)
in pageattr_pmd_entry().
If no secondary cpus are yet running, then it is technically safe to split
because we know all online cpus (i.e. just the boot cpu) supports
BBML2_NOABORT.
So we could explicitly only disallow splitting during the window between
booting
secondary cpus and finalizing the system caps. Feels a bit hacky though...
I think we can check whether system feature has been finalized or not. If it
has not been finalized yet, we just need to check whether the current cpu
(should be just boot cpu) supports BBML2_NOABORT or not. It sounds ok to me.
No I don't think that's sufficient; if the secondary cpus are started (even if
not running the code path doing the split) we have to assume the secondary cpus
are sharing the linear map pgtables, so if we split them on the boot cpu and the
secondary cpus don't support BBML2_NOABORT, things could break.

I think 2 options would be:

  - disallow split for the window between starting the secondary cpus and
finalizing the system caps.

  - Do the split in stop_machine() if any request for splitting is made between
starting the secondary cpus and finalizing the system caps.

Both feel pretty ugly. I'll have a chat with Catalin and try to guage opinons...


In the meantime, would you mind trying this (uncompiled, untested) patch? It's
attempting to implement option 1. TBH, I'm not sure if this is legal since we
will now try to get a mutex; is that allowed in early code that can't sleep? I
guess we only have a single thread running so there can't be any contention...
page table is allocated from buddy with GFP_PGTABLE_KERNEL. In init_IRQ(), the
buddy is initialized, but we shoudn't assume it?
Yes that's a fair point. So we also need to reject split requests made prior to
initializing the buddy. It looks like there is an optional mem_init() arch hook
which arm64 doesn't currently use which we may be able to use to signal "buddy
available"?

And, is GFP_PGTABLE_KERNEL
reasonable here? allocation may block.
I think in practice it should be fine; this early in boot we can't possibly be
out of memory so won't try to block.

But I was really just intending testing with this patch to validate that there
weren't any other issues, not proposing it as the final fix. Personally I'm
leaning more and more to initially mapping by PTE then collapsing once we know
the capabilities of the whole system.

I will test it ASAP when the test environment is available.


Thanks,
Ryan


---8<---
diff --git a/arch/arm64/mm/mmu.c b/arch/arm64/mm/mmu.c
index 8e1d80a7033e3..72790126db55c 100644
--- a/arch/arm64/mm/mmu.c
+++ b/arch/arm64/mm/mmu.c
@@ -779,7 +779,16 @@ int split_kernel_leaf_mapping(unsigned long start, unsigned
long end)
       * and let the permission change code raise a warning if not already
       * pte-mapped.
       */
-    if (!system_supports_bbml2_noabort())
+    if (system_capabilities_finalized() && !system_supports_bbml2_noabort())
+        return 0;
+
+    /*
+     * If system capabilities are not finalized and there is only 1 online
+     * cpu, then we must be running on the boot cpu during early boot before
+     * any secondaries have started. If the boot cpu supports bbml2, we can
+     * safely split.
+     */
+    if (num_online_cpus() > 1 || !cpu_supports_bbml2_noabort())
          return 0;

      /*
---8<---

Thanks,
Ryan



Before setup_system_features(), we don't know if all cpus support BBML2
noabort,
and we
couldn't split kernel page table, in case another cpu that doesn't support
BBML2
noabort
is running.

How could we fix this issue?

1. force pte mapping if realm feature is enabled? Although
force_pte_mapping()
return true if is_realm_world() return true, arm64_rsi_init() is called after
map_mem(). So is_realm_world() still return false during map_mem(). Thus
realm feature relies on rodata=full. If we fix by this solution, we need
to add a new cmdline to force pte mapping.
I don't quite get why is_realm_world() relies on rodata=full. I understand
realm needs PTE mapping if BBML2_NOABORT is not supported. But it doesn't mean
real relies on rodata=full.
https://lore.kernel.org/all/5aeb6f47-12be-40d5-be6f-847bb8ddc605@xxxxxxx/

This is the discussion why realm relies on rodata=full. The initization of realm
coudn't move to before map_mem(), so is_realm_world() is false. As a result,
realm
need rodata=full to indicate we need to make pages shared/protected at page
granularity.

I think we just need to make is_realm_world() work earlier in boot? I think
this
has been a known issue for a while. Not sure if there is any plan to fix it
though.

2. If we could try to split kernel page table before setup_system_features()?
Another option would be to initially map by pte then collapse to block
mappings
once we have determined that all cpus support BBML2_NOABORT. We originally
opted
not to do that because it's a tax on symetric systems. But we could throw
in the
towel if it's the least bad solution we can come up with for solving this. I
think it might help some of Kevin's use cases too?
May be an option too. When we discussed this there was no usecase for direct
mapping collapse. But if we can have multiple usecases, it may be worth it.
AFAICT, the ROX execmem cache may need this, which Will or someone else from
Google is going to work on.

Checking current cpu BBML2_NOABORT capability before system feature is
finalized seems like a fast way to stop bleeding IMHO before we find more
elegant long-term solution.

Thanks,
Yang

Thanks,
Ryan


Thanks.

Ryan tested v7 on an AmpereOne system (a VM with 12G RAM) in all 3 possible
modes by hacking the BBML2 feature detection code:

     - mode 1: All CPUs support BBML2 so the linear map uses large mappings
     - mode 2: Boot CPU does not support BBML2 so linear map uses pte
mappings
     - mode 3: Boot CPU supports BBML2 but secondaries do not so linear map
       initially uses large mappings but is then repainted to use pte
mappings

In all cases, mm selftests run and no regressions are observed. In all
cases,
ptdump of linear map is as expected. Because there are just some cleanups
between v7 and v8, so I kept using Ryan's test result:

Mode 1:
=======
---[ Linear Mapping start ]---
0xffff000000000000-0xffff000000200000           2M PMD RW NX SHD
AF        BLK UXN    MEM/NORMAL-TAGGED
0xffff000000200000-0xffff000000210000          64K PTE RW NX SHD AF
CON     UXN    MEM/NORMAL-TAGGED
0xffff000000210000-0xffff000000400000        1984K PTE ro NX SHD
AF            UXN    MEM/NORMAL
0xffff000000400000-0xffff000002400000          32M PMD ro NX SHD
AF        BLK UXN    MEM/NORMAL
0xffff000002400000-0xffff000002550000        1344K PTE ro NX SHD
AF            UXN    MEM/NORMAL
0xffff000002550000-0xffff000002600000         704K PTE RW NX SHD AF
CON     UXN    MEM/NORMAL-TAGGED
0xffff000002600000-0xffff000004000000          26M PMD RW NX SHD
AF        BLK UXN    MEM/NORMAL-TAGGED
0xffff000004000000-0xffff000040000000         960M PMD RW NX SHD AF
CON BLK UXN    MEM/NORMAL-TAGGED
0xffff000040000000-0xffff000140000000           4G PUD RW NX SHD
AF        BLK UXN    MEM/NORMAL-TAGGED
0xffff000140000000-0xffff000142000000          32M PMD RW NX SHD AF
CON BLK UXN    MEM/NORMAL-TAGGED
0xffff000142000000-0xffff000142120000        1152K PTE RW NX SHD AF
CON     UXN    MEM/NORMAL-TAGGED
0xffff000142120000-0xffff000142128000          32K PTE RW NX SHD
AF            UXN    MEM/NORMAL-TAGGED
0xffff000142128000-0xffff000142159000         196K PTE ro NX SHD
AF            UXN    MEM/NORMAL-TAGGED
0xffff000142159000-0xffff000142160000          28K PTE RW NX SHD
AF            UXN    MEM/NORMAL-TAGGED
0xffff000142160000-0xffff000142240000         896K PTE RW NX SHD AF
CON     UXN    MEM/NORMAL-TAGGED
0xffff000142240000-0xffff00014224e000          56K PTE RW NX SHD
AF            UXN    MEM/NORMAL-TAGGED
0xffff00014224e000-0xffff000142250000           8K PTE ro NX SHD
AF            UXN    MEM/NORMAL-TAGGED
0xffff000142250000-0xffff000142260000          64K PTE RW NX SHD
AF            UXN    MEM/NORMAL-TAGGED
0xffff000142260000-0xffff000142280000         128K PTE RW NX SHD AF
CON     UXN    MEM/NORMAL-TAGGED
0xffff000142280000-0xffff000142288000          32K PTE RW NX SHD
AF            UXN    MEM/NORMAL-TAGGED
0xffff000142288000-0xffff000142290000          32K PTE ro NX SHD
AF            UXN    MEM/NORMAL-TAGGED
0xffff000142290000-0xffff0001422a0000          64K PTE RW NX SHD
AF            UXN    MEM/NORMAL-TAGGED
0xffff0001422a0000-0xffff000142465000        1812K PTE ro NX SHD
AF            UXN    MEM/NORMAL-TAGGED
0xffff000142465000-0xffff000142470000          44K PTE RW NX SHD
AF            UXN    MEM/NORMAL-TAGGED
0xffff000142470000-0xffff000142600000        1600K PTE RW NX SHD AF
CON     UXN    MEM/NORMAL-TAGGED
0xffff000142600000-0xffff000144000000          26M PMD RW NX SHD
AF        BLK UXN    MEM/NORMAL-TAGGED
0xffff000144000000-0xffff000180000000         960M PMD RW NX SHD AF
CON BLK UXN    MEM/NORMAL-TAGGED
0xffff000180000000-0xffff000181a00000          26M PMD RW NX SHD
AF        BLK UXN    MEM/NORMAL-TAGGED
0xffff000181a00000-0xffff000181b90000        1600K PTE RW NX SHD AF
CON     UXN    MEM/NORMAL-TAGGED
0xffff000181b90000-0xffff000181b9d000          52K PTE RW NX SHD
AF            UXN    MEM/NORMAL-TAGGED
0xffff000181b9d000-0xffff000181c80000         908K PTE ro NX SHD
AF            UXN    MEM/NORMAL-TAGGED
0xffff000181c80000-0xffff000181c90000          64K PTE RW NX SHD
AF            UXN    MEM/NORMAL-TAGGED
0xffff000181c90000-0xffff000181ca0000          64K PTE RW NX SHD AF
CON     UXN    MEM/NORMAL-TAGGED
0xffff000181ca0000-0xffff000181dbd000        1140K PTE ro NX SHD
AF            UXN    MEM/NORMAL-TAGGED
0xffff000181dbd000-0xffff000181dc0000          12K PTE RW NX SHD
AF            UXN    MEM/NORMAL-TAGGED
0xffff000181dc0000-0xffff000181e00000         256K PTE RW NX SHD AF
CON     UXN    MEM/NORMAL-TAGGED
0xffff000181e00000-0xffff000182000000           2M PMD RW NX SHD
AF        BLK UXN    MEM/NORMAL-TAGGED
0xffff000182000000-0xffff0001c0000000         992M PMD RW NX SHD AF
CON BLK UXN    MEM/NORMAL-TAGGED
0xffff0001c0000000-0xffff000300000000           5G PUD RW NX SHD
AF        BLK UXN    MEM/NORMAL-TAGGED
0xffff000300000000-0xffff008000000000         500G PUD
0xffff008000000000-0xffff800000000000      130560G PGD
---[ Linear Mapping end ]---

Mode 3:
=======
---[ Linear Mapping start ]---
0xffff000000000000-0xffff000000210000        2112K PTE RW NX SHD
AF            UXN    MEM/NORMAL-TAGGED
0xffff000000210000-0xffff000000400000        1984K PTE ro NX SHD
AF            UXN    MEM/NORMAL
0xffff000000400000-0xffff000002400000          32M PMD ro NX SHD
AF        BLK UXN    MEM/NORMAL
0xffff000002400000-0xffff000002550000        1344K PTE ro NX SHD
AF            UXN    MEM/NORMAL
0xffff000002550000-0xffff000143a61000     5264452K PTE RW NX SHD
AF            UXN    MEM/NORMAL-TAGGED
0xffff000143a61000-0xffff000143c61000           2M PTE ro NX SHD
AF            UXN    MEM/NORMAL-TAGGED
0xffff000143c61000-0xffff000181b9a000     1015012K PTE RW NX SHD
AF            UXN    MEM/NORMAL-TAGGED
0xffff000181b9a000-0xffff000181d9a000           2M PTE ro NX SHD
AF            UXN    MEM/NORMAL-TAGGED
0xffff000181d9a000-0xffff000300000000     6261144K PTE RW NX SHD
AF            UXN    MEM/NORMAL-TAGGED
0xffff000300000000-0xffff008000000000         500G PUD
0xffff008000000000-0xffff800000000000      130560G PGD
---[ Linear Mapping end ]---


Performance Testing
===================
* Memory use after boot
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.

* 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%.

Additionally Ryan also ran this through a random selection of benchmarks on
AmpereOne. None show any regressions, and various benchmarks show
statistically
significant improvement. I'm just showing those improvements here:

+----------------------
+----------------------------------------------------------
+-------------------------+
| Benchmark            | Result
Class                                             | Improvement vs 6.17-
rc1 |
+======================+==========================================================+=========================+
| micromm/vmalloc      | full_fit_alloc_test: p:1, h:0, l:500000
(usec)           |              (I) -9.00% |
|                      | kvfree_rcu_1_arg_vmalloc_test: p:1, h:0, l:500000
(usec) |              (I) -6.93% |
|                      | kvfree_rcu_2_arg_vmalloc_test: p:1, h:0, l:500000
(usec) |              (I) -6.77% |
|                      | pcpu_alloc_test: p:1, h:0, l:500000
(usec)               |              (I) -4.63% |
+----------------------
+----------------------------------------------------------
+-------------------------+
| mmtests/hackbench    | process-sockets-30
(seconds)                             |              (I) -2.96% |
+----------------------
+----------------------------------------------------------
+-------------------------+
| mmtests/kernbench    | syst-192
(seconds) |             (I) -12.77% |
+----------------------
+----------------------------------------------------------
+-------------------------+
| pts/perl-benchmark   | Test: Interpreter
(Seconds)                              |              (I) -4.86% |
+----------------------
+----------------------------------------------------------
+-------------------------+
| pts/pgbench          | Scale: 1 Clients: 1 Read Write
(TPS)                     |               (I) 5.07% |
|                      | Scale: 1 Clients: 1 Read Write - Latency
(ms)            |              (I) -4.72% |
|                      | Scale: 100 Clients: 1000 Read Write
(TPS)                |               (I) 2.58% |
|                      | Scale: 100 Clients: 1000 Read Write - Latency
(ms)       |              (I) -2.52% |
+----------------------
+----------------------------------------------------------
+-------------------------+
| pts/sqlite-speedtest | Timed Time - Size 1,000
(Seconds)                        |              (I) -2.68% |
+----------------------
+----------------------------------------------------------
+-------------------------+

Changes since v7 [1]
====================
- Rebased on v6.17-rc6 and Shijie's rodata series (https://
git.kernel.org/pub/
scm/linux/kernel/git/arm64/linux.git/commit/?id=bfbbb0d3215f)
     which has been picked up by Will.
- Patch 1: Fixed pmd_leaf/pud_leaf issue since the code may need to change
     permission for invalid entries per Jinjiang Tu.
- Patch 1: Removed pageattr_pgd_entry and pageattr_p4d_entry per Ryan.
- Used (-1ULL) instead of -1 per Catalin.
- Added comment about arm64 lazy mmu allow sleeping per Ryan.
- Squashed patch #4 in v7 into patch #3.
- Squashed patch #6 in v7 into patch #4.
- Added patch #5 to fix a arm64 kprobes bug. It guarantees set_memory_rox()
     is called before vfree(). It can go into separately or with this series
     together.
- Collected all the R-bs and A-bs.

Changes since v6 [2]
====================
- Patch 1: Minor refactor to implement walk_kernel_page_table_range() in
terms
     of walk_kernel_page_table_range_lockless(). Also lead to adding *pmd
argument
     to the lockless variant for consistency (per Catalin).
- Misc function/variable renames to improve clarity and consistency.
- Share same syncrhonization flag between idmap_kpti_install_ng_mappings and
     wait_linear_map_split_to_ptes, which allows removal of bbml2_ptes[] to
save
     ~20K from kernel image.
- Only take pgtable_split_lock and enter lazy mmu mode once for both splits.
- Only walk the pgtable once for the common "split single page" case.
- Bypass split to contpmd and contpte when spllitting linear map to ptes.

[1] https://lore.kernel.org/linux-arm-kernel/20250829115250.2395585-1-
ryan.roberts@xxxxxxx/
[2] https://lore.kernel.org/linux-arm-kernel/20250805081350.3854670-1-
ryan.roberts@xxxxxxx/


Dev Jain (1):
         arm64: Enable permission change on arm64 kernel block mappings

Ryan Roberts (1):
         arm64: mm: split linear mapping if BBML2 unsupported on
secondary CPUs

Yang Shi (3):
         arm64: cpufeature: add AmpereOne to BBML2 allow list
         arm64: mm: support large block mapping when rodata=full
         arm64: kprobes: call set_memory_rox() for kprobe page

    arch/arm64/include/asm/cpufeature.h |   2 +
    arch/arm64/include/asm/mmu.h        |   3 +
    arch/arm64/include/asm/pgtable.h    |   5 ++
    arch/arm64/kernel/cpufeature.c      |  12 +++-
    arch/arm64/kernel/probes/kprobes.c  |  12 ++++
    arch/arm64/mm/mmu.c                 | 422 ++++++++++++++++++++++++++++++
++++
+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
+----
    arch/arm64/mm/pageattr.c            | 123 +++++++++++++++++++++++
+---------
    arch/arm64/mm/proc.S                |  27 ++++++--
    include/linux/pagewalk.h            |   3 +
    mm/pagewalk.c                       |  36 ++++++----
    10 files changed, 581 insertions(+), 64 deletions(-)