Skip to contentSkip to conversation

x86/lam: Allocate test page from AREA_LOW instead of AREA_NORMAL

PATCH Superseded yexun 2 participants 4 messages Jun 1 – Jun 2 KVM

v1 of 4 Superseded Latest is v4 →

lore
Share & tools
lore

yexun

author
#0 lore raw

PATCH 1/1 x86/lam: Allocate test page from AREA_LOW instead of AREA_NORMAL

The lam test does not set a guest memory size in x86/unittests.cfg, so QEMU falls back to the default of 128 MiB. AREA_NORMAL starts at PFN BIT(36-12), i.e. physical 64 GiB, which is never initialized in a 128 MiB guest. As a result, alloc_pages_flags(0, AREA_NORMAL) returns NULL and test_lam_user() ends up running its LAM checks against a NULL pointer, which is semantically meaningless even if the metadata-bit arithmetic happens to succeed.

Allocate from AREA_LOW instead. AREA_LOW_PFN is BIT(24-12) (16 MiB), well within a 128 MiB guest, and bits 63..47 of the resulting linear address are still zero, so the LAM48/LAM57 metadata-bit checks remain valid. Update the adjacent comment accordingly.

Fixes 0164d7595c85 ("x86: Add test cases for LAM_{U48,U57}")
---
 x86/lam.c | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)
x86/lam.c +3 -3
index 87efc5dd..b2dcf7f8 100644
@@ -231,13 +231,13 @@ static void test_lam_user(void)
 	bool has_lam = this_cpu_has(X86_FEATURE_LAM);
 
 	/*
-	 * The physical address of AREA_NORMAL is within 36 bits, so that using
+	 * The physical address of AREA_LOW is within 36 bits, so that using
 	 * identical mapping, the linear address will be considered as user mode
 	 * address from the view of LAM, and the metadata bits are not used as
 	 * address for both LAM48 and LAM57.
 	 */
-	vaddr = alloc_pages_flags(0, AREA_NORMAL);
-	static_assert((AREA_NORMAL_PFN & GENMASK(63, 47)) == 0UL);
+	vaddr = alloc_pages_flags(0, AREA_LOW);
+	static_assert((AREA_LOW_PFN & GENMASK(63, 47)) == 0UL);
 
 	/*
 	 * Note, LAM doesn't have a global control bit to turn on/off LAM
Signature · 1 line
2.54.0
#1 lore raw
Quoted 5 lines · yexun

The lam test does not set a guest memory size in x86/unittests.cfg, so QEMU falls back to the default of 128 MiB. AREA_NORMAL starts at PFN BIT(36-12), i.e. physical 64 GiB, which is never initialized in a 128 MiB guest. As a result, alloc_pages_flags(0, AREA_NORMAL) returns NULL

Oops, thanks for catching this.

Quoted 3 lines

and test_lam_user() ends up running its LAM checks against a NULL pointer, which is semantically meaningless even if the metadata-bit arithmetic happens to succeed.

From the LAM hardware perspective, NULL pointer is treated as a userspace virtual address, so the LAM hardware feature validation still works.

But it's not intended to use a NULL pointer, so it's a bug that needs to be fixed.

Quoted 20 lines

Allocate from AREA_LOW instead. AREA_LOW_PFN is BIT(24-12) (16 MiB), well within a 128 MiB guest, and bits 63..47 of the resulting linear address are still zero, so the LAM48/LAM57 metadata-bit checks remain valid. Update the adjacent comment accordingly.

Fixes 0164d7595c85 ("x86: Add test cases for LAM_{U48,U57}")
---
 x86/lam.c | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)
x86/lam.c +0 -0
index 87efc5dd..b2dcf7f8 100644
@@ -231,13 +231,13 @@ static void test_lam_user(void)
 	bool has_lam = this_cpu_has(X86_FEATURE_LAM);
 
 	/*

x86/lam.c static void test_lam_user(void) · line 240

Quoted 10 lines
x86/lam.c +3 -3
-	 * The physical address of AREA_NORMAL is within 36 bits, so that using
+	 * The physical address of AREA_LOW is within 36 bits, so that using
 	 * identical mapping, the linear address will be considered as user mode
 	 * address from the view of LAM, and the metadata bits are not used as
 	 * address for both LAM48 and LAM57.
 	 */
-	vaddr = alloc_pages_flags(0, AREA_NORMAL);
-	static_assert((AREA_NORMAL_PFN & GENMASK(63, 47)) == 0UL);
+	vaddr = alloc_pages_flags(0, AREA_LOW);
+	static_assert((AREA_LOW_PFN & GENMASK(63, 47)) == 0UL);

Although the error exists in the original code, the GENMASK(63, 47) is an address mask, it's not right to mask a PFN.

Quoted 3 lines
x86/lam.c +0 -0
 	/*
 	 * Note, LAM doesn't have a global control bit to turn on/off LAM
author
#2 lore raw
Quoted 8 lines · Binbin Wu
Quoted 7 lines · yexun
x86/lam.c +0 -0
index 87efc5dd..b2dcf7f8 100644
@@ -231,13 +231,13 @@ static void test_lam_user(void)
 	bool has_lam = this_cpu_has(X86_FEATURE_LAM);
 
Quoted 14 lines

x86/lam.c static void test_lam_user(void) · line 240

Quoted 11 lines
x86/lam.c +3 -3
 	/*
-	 * The physical address of AREA_NORMAL is within 36 bits, so that using
+	 * The physical address of AREA_LOW is within 36 bits, so that using
 	 * identical mapping, the linear address will be considered as user mode
 	 * address from the view of LAM, and the metadata bits are not used as
 	 * address for both LAM48 and LAM57.
 	 */
-	vaddr = alloc_pages_flags(0, AREA_NORMAL);
-	static_assert((AREA_NORMAL_PFN & GENMASK(63, 47)) == 0UL);
+	vaddr = alloc_pages_flags(0, AREA_LOW);
+	static_assert((AREA_LOW_PFN & GENMASK(63, 47)) == 0UL);

Although the error exists in the original code, the GENMASK(63, 47) is an address mask, it's not right to mask a PFN.

Thanks for your review. What about asserting '(vaddr & GENMASK(63, 47)) == 0UL'?

#3 lore raw
Quoted 24 lines · yexun
Quoted 8 lines · Binbin Wu
Quoted 7 lines · yexun
x86/lam.c +0 -0
index 87efc5dd..b2dcf7f8 100644
@@ -231,13 +231,13 @@ static void test_lam_user(void)
      bool has_lam = this_cpu_has(X86_FEATURE_LAM);
 
Quoted 14 lines

x86/lam.c static void test_lam_user(void) · line 240

Quoted 11 lines
x86/lam.c +3 -3
      /*
-     * The physical address of AREA_NORMAL is within 36 bits, so that using
+     * The physical address of AREA_LOW is within 36 bits, so that using
       * identical mapping, the linear address will be considered as user mode
       * address from the view of LAM, and the metadata bits are not used as
       * address for both LAM48 and LAM57.
       */
-    vaddr = alloc_pages_flags(0, AREA_NORMAL);
-    static_assert((AREA_NORMAL_PFN & GENMASK(63, 47)) == 0UL);
+    vaddr = alloc_pages_flags(0, AREA_LOW);
+    static_assert((AREA_LOW_PFN & GENMASK(63, 47)) == 0UL);

Although the error exists in the original code, the GENMASK(63, 47) is an address mask, it's not right to mask a PFN.

Quoted 1 line

Thanks for your review. What about asserting '(vaddr & GENMASK(63, 47)) == 0UL'?

LGTM.