[ndctl PATCH] test/cxl-poison.sh: test scanning past fully mapped partitions

[ndctl PATCH] test/cxl-poison.sh: test scanning past fully mapped partitions

[Alison Schofield]

Listing poison by memdev scans the unmapped tail of every partition.
When an earlier partition is fully mapped, its tail is zero length, but
the scan must continue to later partitions. A regression caused the
scan to stop at the first fully-mapped partition, leaving later
partitions unscanned.

Backstop that behavior with a test case that fully maps a memdev's RAM
partition so its unmapped tail is zero length, then injects poison into
the unmapped PMEM partition that follows. The PMEM poison is only
reported if the scan continues past the fully-mapped RAM partition.

Signed-off-by: Alison Schofield <alison.schofield@intel.com>
---
 test/cxl-poison.sh | 40 ++++++++++++++++++++++++++++++++++++++++
 1 file changed, 40 insertions(+)

diff --git a/test/cxl-poison.sh b/test/cxl-poison.sh
index 49aa1b68c5c1..a03e08084eb4 100644
--- a/test/cxl-poison.sh
+++ b/test/cxl-poison.sh
@@ -219,6 +219,45 @@ test_poison_by_region_offset_negative()
 	clear_poison "$region" "$large_offset" true
 }
 
+# Backstop a driver fix where a fully mapped partition prematurely
+# terminated the unmapped poison scan.
+test_poison_unmapped_later_partition()
+{
+	local decoder ram_size pmem_dpa
+
+	check_min_kver "7.3" || return 0
+
+	# Free the auto region to use the ram capacity
+	$CXL destroy-region -f -b "$CXL_TEST_BUS" all
+
+	find_memdev
+
+	# Fully map the ram partition so it has a zero-length unmapped tail
+	decoder=$($CXL list -b "$CXL_TEST_BUS" -D -d root -m "$memdev" |
+		  jq -r ".[] |
+		  select(.volatile_capable == true) |
+		  select(.nr_targets == 1) |
+		  .decoder")
+	[[ -n "$decoder" && "$decoder" != "null" ]] ||
+		do_skip "no x1 volatile decoder found"
+
+	ram_size=$($CXL list -m "$memdev" | jq -r ".[0].ram_size")
+	[[ -n "$ram_size" && "$ram_size" != "null" ]] || err "$LINENO"
+
+	region=$($CXL create-region -t ram -d "$decoder" -m "$memdev" \
+		 -s "$ram_size" | jq -r ".region")
+	[[ -n "$region" && "$region" != "null" ]] || err "$LINENO"
+
+	# Poison the unmapped pmem tail
+	pmem_dpa=$ram_size
+	inject_poison "$memdev" "$pmem_dpa"
+	validate_poison_found "-m $memdev" 1
+	clear_poison "$memdev" "$pmem_dpa"
+	validate_poison_found "-m $memdev" 0
+
+	$CXL destroy-region -f -b "$CXL_TEST_BUS" "$region"
+}
+
 is_unaligned() {
 	local region=$1
 	local hbiw=$2
@@ -332,6 +371,7 @@ run_poison_test()
 		do_skip "test cases requires inject by region kernel support"
 	test_poison_by_region_offset
 	test_poison_by_region_offset_negative
+	test_poison_unmapped_later_partition
 }
 
 modprobe -r cxl_test

base-commit: 15e932c4e1318a9608ad9b799ad83a32a8b5970d
-- 
2.37.3

Re: [ndctl PATCH] test/cxl-poison.sh: test scanning past fully mapped partitions

[Richard Cheng]

On Tue, Jun 30, 2026 at 09:42:02PM +0800, Alison Schofield wrote:
> Listing poison by memdev scans the unmapped tail of every partition.
> When an earlier partition is fully mapped, its tail is zero length, but
> the scan must continue to later partitions. A regression caused the
> scan to stop at the first fully-mapped partition, leaving later
> partitions unscanned.
> 
> Backstop that behavior with a test case that fully maps a memdev's RAM
> partition so its unmapped tail is zero length, then injects poison into
> the unmapped PMEM partition that follows. The PMEM poison is only
> reported if the scan continues past the fully-mapped RAM partition.
>

Hi Alison,

Thanks for this ! I ran it against cxl_test in the following ways.

On a kernel with the fix [1], the test passes in both run_poison_test passes:
1 poison record found after inject at the first pmem DPA, 0 after clear.

On a kernel without the fix, it fails exactly at the "1 poison records expected, 0 found"
assertion, so it catches the regression as intended.

Just one small thing below.

 
> Signed-off-by: Alison Schofield <alison.schofield@intel.com>
> ---
>  test/cxl-poison.sh | 40 ++++++++++++++++++++++++++++++++++++++++
>  1 file changed, 40 insertions(+)
> 
> diff --git a/test/cxl-poison.sh b/test/cxl-poison.sh
> index 49aa1b68c5c1..a03e08084eb4 100644
> --- a/test/cxl-poison.sh
> +++ b/test/cxl-poison.sh
> @@ -219,6 +219,45 @@ test_poison_by_region_offset_negative()
>  	clear_poison "$region" "$large_offset" true
>  }
>  
> +# Backstop a driver fix where a fully mapped partition prematurely
> +# terminated the unmapped poison scan.
> +test_poison_unmapped_later_partition()
> +{
> +	local decoder ram_size pmem_dpa
> +

"region" is assigned below but not declared local like the others.
Does it have any reason for it to not be local var?
It might be harmless today but tests added after this one will inherit
a stale global region, if it ever uses one.

Tested-by: Richard Cheng <icheng@nvidia.com>
Reviewed-by: Richard Cheng <icheng@nvidia.com>

[1]:
https://lore.kernel.org/linux-cxl/20260708074228.43654-3-icheng@nvidia.com/

Best regards,
Richard Cheng.

> +	check_min_kver "7.3" || return 0
> +
> +	# Free the auto region to use the ram capacity
> +	$CXL destroy-region -f -b "$CXL_TEST_BUS" all
> +
> +	find_memdev
> +
> +	# Fully map the ram partition so it has a zero-length unmapped tail
> +	decoder=$($CXL list -b "$CXL_TEST_BUS" -D -d root -m "$memdev" |
> +		  jq -r ".[] |
> +		  select(.volatile_capable == true) |
> +		  select(.nr_targets == 1) |
> +		  .decoder")
> +	[[ -n "$decoder" && "$decoder" != "null" ]] ||
> +		do_skip "no x1 volatile decoder found"
> +
> +	ram_size=$($CXL list -m "$memdev" | jq -r ".[0].ram_size")
> +	[[ -n "$ram_size" && "$ram_size" != "null" ]] || err "$LINENO"
> +
> +	region=$($CXL create-region -t ram -d "$decoder" -m "$memdev" \
> +		 -s "$ram_size" | jq -r ".region")
> +	[[ -n "$region" && "$region" != "null" ]] || err "$LINENO"
> +
> +	# Poison the unmapped pmem tail
> +	pmem_dpa=$ram_size
> +	inject_poison "$memdev" "$pmem_dpa"
> +	validate_poison_found "-m $memdev" 1
> +	clear_poison "$memdev" "$pmem_dpa"
> +	validate_poison_found "-m $memdev" 0
> +
> +	$CXL destroy-region -f -b "$CXL_TEST_BUS" "$region"
> +}
> +
>  is_unaligned() {
>  	local region=$1
>  	local hbiw=$2
> @@ -332,6 +371,7 @@ run_poison_test()
>  		do_skip "test cases requires inject by region kernel support"
>  	test_poison_by_region_offset
>  	test_poison_by_region_offset_negative
> +	test_poison_unmapped_later_partition
>  }
>  
>  modprobe -r cxl_test
> 
> base-commit: 15e932c4e1318a9608ad9b799ad83a32a8b5970d
> -- 
> 2.37.3
>