Hi Zhijian,
On 8/26/2025 11:30 PM, Zhijian Li (Fujitsu) wrote:
All,
I have confirmed that in the !CXL_REGION configuration, the same environment
may fail to fall back to hmem.(Your new patch cannot resolve this issue)
In my environment:
- There are two CXL memory devices corresponding to:
```
5d0000000-6cffffff : CXL Window 0
6d0000000-7cffffff : CXL Window 1
```
- E820 table contains a 'soft reserved' entry:
```
[ 0.000000] BIOS-e820: [mem 0x00000005d0000000-0x00000007cfffffff] soft
reserved
```
However, since my ACPI SRAT doesn't describe the CXL memory devices (the
point), `acpi/hmat.c` won't allocate memory targets for them. This prevents the
call chain:
```c
hmat_register_target_devices() // for each SRAT-described target
-> hmem_register_resource()
-> insert entry into "HMEM devices" resource
```
Therefore, for successful fallback to hmem in this environment: `dax_hmem.ko`
and `kmem.ko` must request resources BEFORE `cxl_acpi.ko` inserts 'CXL Window X'
However the kernel cannot guarantee this initialization order.
When cxl_acpi runs before dax_kmem/kmem:
```
(built-in) CXL_REGION=n
driver/dax/hmem/device.c cxl_acpi.ko dax_hmem.ko kmem.ko
(1) Add entry '15d0000000-7cfffffff'
(2) Traverse "HMEM devices"
Insert to iomem:
5d0000000-7cffffff : Soft
Reserved
(3) Insert CXL Window 0/1
/proc/iomem shows:
5d0000000-7cffffff : Soft Reserved
5d0000000-6cffffff : CXL Window 0
6d0000000-7cffffff : CXL Window 1
(4) Create dax device
(5)
request_mem_region() fails
for
5d0000000-7cffffff
Reason:
Children of 'Soft Reserved'
(CXL
Windows 0/1) don't cover full range
```
Thanks for confirming the failure point. I was thinking of two possible
ways forward here, and I would like to get feedback from others:
[1] Teach dax_hmem to split when the parent claim fails:
If __request_region() fails for the top-level Soft Reserved range
because IORES_DESC_CXL children already exist, dax_hmem could iterate
those windows and register each one individually. The downside is that
it adds some complexity and feels a bit like papering over the fact that
CXL should eventually own all of this memory. As Dan mentioned, the
long-term plan is for Linux to not need the soft-reserve fallback at
all, and simply ignore Soft Reserve for CXL Windows because the CXL
subsystem will handle it.
[2] Always unconditionally load CXL early..
Call request_module("cxl_acpi"); request_module("cxl_pci"); from
dax_hmem_init() (without the IS_ENABLED(CONFIG_DEV_DAX_CXL) guard). If
those are y/m, they’ll be present; if n, it’s a no-op. Then in
hmem_register_device() drop the IS_ENABLED(CONFIG_DEV_DAX_CXL) gate and do:
if (region_intersects(res->start, resource_size(res),
IORESOURCE_MEM, IORES_DESC_CXL) !=REGION_DISJOINT)
/* defer to CXL */;
and defer to CXL if windows are present. This makes Soft Reserved
unavailable once CXL Windows have been discovered, even if CXL_REGION is
disabled. That aligns better with the idea that “CXL should win”
whenever a window is visible (This also needs to be considered alongside
patch 6/6 in my series.)
With CXL_REGION=n there would be no devdax and no kmem for that range;
proc/iomem would show only the windows something like below
850000000-284fffffff : CXL Window 0
2850000000-484fffffff : CXL Window 1
4850000000-684fffffff : CXL Window 2
That means the memory is left unclaimed/unavailable.. (no System RAM, no
/dev/dax). Is that acceptable when CXL_REGION is disabled?
Thanks
Smita
---------------------
In my another environment where ACPI SRAT has separate entries per CXL device:
1. `acpi/hmat.c` inserts two entries into "HMEM devices":
- 5d0000000-6cffffff
- 6d0000000-7cffffff
2. Regardless of module order, dax/kmem requests per-device resources,
resulting in:
```
5d0000000-7cffffff : Soft Reserved
5d0000000-6cffffff : CXL Window 0
5d0000000-6cffffff : dax0.0
5d0000000-6cffffff : System RAM (kmem)
6d0000000-7cffffff : CXL Window 1
6d0000000-7cffffff : dax1.0
6d0000000-7cffffff : System RAM (kmem)
```
Thanks,
Zhijian
On 25/08/2025 15:50, Li Zhijian wrote:
On 22/08/2025 11:56, Koralahalli Channabasappa, Smita wrote:
```
3. When CXL_REGION is disabled, there is a failure to fallback to dax_hmem, in
which case only CXL Window X is visible.
Haven't tested !CXL_REGION yet.
When CXL_REGION is disabled, DEV_DAX_CXL will also be disabled. So dax_hmem
should handle it.
Yes, falling back to dax_hmem/kmem is the result we expect.
I haven't figured out the root cause of the issue yet, but I can tell you that
in my QEMU environment,
there is currently a certain probability that it cannot fall back to
dax_hmem/kmem.
Upon its failure, I observed the following warnings and errors (with my local
fixup kernel).
[ 12.203254] kmem dax0.0: mapping0: 0x5d0000000-0x7cfffffff could not reserve
region
[ 12.203437] kmem dax0.0: probe with driver kmem failed with error -16
I was able to fallback to dax_hmem. But let me know if I'm missing something.
config DEV_DAX_CXL
tristate "CXL DAX: direct access to CXL RAM regions"
depends on CXL_BUS && CXL_REGION && DEV_DAX
..
On failure:
```
100000000-27ffffff : System RAM
5c0001128-5c00011b7 : port1
5c0011128-5c00111b7 : port2
5d0000000-6cffffff : CXL Window 0
6d0000000-7cffffff : CXL Window 1
7000000000-700000ffff : PCI Bus 0000:0c
7000000000-700000ffff : 0000:0c:00.0
7000001080-70000010d7 : mem1
```
On success:
```
5d0000000-7cffffff : dax0.0
5d0000000-7cffffff : System RAM (kmem)
5d0000000-6cffffff : CXL Window 0
6d0000000-7cffffff : CXL Window 1
```
