[PATCH] kasan: hw_tags: fix a false positive case of vrealloc in alloced size

Yeoreum Yun posted 1 patch 2 days, 9 hours ago
include/linux/kasan.h |  1 +
mm/kasan/hw_tags.c    | 11 +++++++++--
mm/vmalloc.c          |  1 +
3 files changed, 11 insertions(+), 2 deletions(-)
[PATCH] kasan: hw_tags: fix a false positive case of vrealloc in alloced size
Posted by Yeoreum Yun 2 days, 9 hours ago
When a memory region is allocated with vmalloc() and later expanded with
vrealloc() — while still within the originally allocated size —
KASAN may report a false positive because
it does not update the tags for the newly expanded portion of the memory.

A typical example of this pattern occurs in the BPF verifier,
and the following is a related false positive report:

[ 2206.486476] ==================================================================
[ 2206.486509] BUG: KASAN: invalid-access in __memcpy+0xc/0x30
[ 2206.486607] Write at addr f5ff800083765270 by task test_progs/205
[ 2206.486664] Pointer tag: [f5], memory tag: [fe]
[ 2206.486703]
[ 2206.486745] CPU: 4 UID: 0 PID: 205 Comm: test_progs Tainted: G           OE       6.18.0-rc7+ #145 PREEMPT(full)
[ 2206.486861] Tainted: [O]=OOT_MODULE, [E]=UNSIGNED_MODULE
[ 2206.486897] Hardware name:  , BIOS
[ 2206.486932] Call trace:
[ 2206.486961]  show_stack+0x24/0x40 (C)
[ 2206.487071]  __dump_stack+0x28/0x48
[ 2206.487182]  dump_stack_lvl+0x7c/0xb0
[ 2206.487293]  print_address_description+0x80/0x270
[ 2206.487403]  print_report+0x94/0x100
[ 2206.487505]  kasan_report+0xd8/0x150
[ 2206.487606]  __do_kernel_fault+0x64/0x268
[ 2206.487717]  do_bad_area+0x38/0x110
[ 2206.487820]  do_tag_check_fault+0x38/0x60
[ 2206.487936]  do_mem_abort+0x48/0xc8
[ 2206.488042]  el1_abort+0x40/0x70
[ 2206.488127]  el1h_64_sync_handler+0x50/0x118
[ 2206.488217]  el1h_64_sync+0xa4/0xa8
[ 2206.488303]  __memcpy+0xc/0x30 (P)
[ 2206.488412]  do_misc_fixups+0x4f8/0x1950
[ 2206.488528]  bpf_check+0x31c/0x840
[ 2206.488638]  bpf_prog_load+0x58c/0x658
[ 2206.488737]  __sys_bpf+0x364/0x488
[ 2206.488833]  __arm64_sys_bpf+0x30/0x58
[ 2206.488920]  invoke_syscall+0x68/0xe8
[ 2206.489033]  el0_svc_common+0xb0/0xf8
[ 2206.489143]  do_el0_svc+0x28/0x48
[ 2206.489249]  el0_svc+0x40/0xe8
[ 2206.489337]  el0t_64_sync_handler+0x84/0x140
[ 2206.489427]  el0t_64_sync+0x1bc/0x1c0

Here, 0xf5ff800083765000 is vmalloc()ed address for
env->insn_aux_data with the size of 0x268.
While this region is expanded size by 0x478 and initialise
increased region to apply patched instructions,
a false positive is triggered at the address 0xf5ff800083765270
because __kasan_unpoison_vmalloc() with KASAN_VMALLOC_PROT_NORMAL flag only
doesn't update the tag on increaed region.

To address this, introduces KASAN_VMALLOC_EXPAND flag which
is used to expand vmalloc()ed memory in range of real allocated size
to update tag for increased region.

Fixes: 23689e91fb22 ("kasan, vmalloc: add vmalloc tagging for HW_TAGS”)
Cc: <stable@vger.kernel.org>
Signed-off-by: Yeoreum Yun <yeoreum.yun@arm.com>
---
 include/linux/kasan.h |  1 +
 mm/kasan/hw_tags.c    | 11 +++++++++--
 mm/vmalloc.c          |  1 +
 3 files changed, 11 insertions(+), 2 deletions(-)

diff --git a/include/linux/kasan.h b/include/linux/kasan.h
index d12e1a5f5a9a..0608c5d4e6cf 100644
--- a/include/linux/kasan.h
+++ b/include/linux/kasan.h
@@ -28,6 +28,7 @@ typedef unsigned int __bitwise kasan_vmalloc_flags_t;
 #define KASAN_VMALLOC_INIT		((__force kasan_vmalloc_flags_t)0x01u)
 #define KASAN_VMALLOC_VM_ALLOC		((__force kasan_vmalloc_flags_t)0x02u)
 #define KASAN_VMALLOC_PROT_NORMAL	((__force kasan_vmalloc_flags_t)0x04u)
+#define KASAN_VMALLOC_EXPAND		((__force kasan_vmalloc_flags_t)0x08u)

 #define KASAN_VMALLOC_PAGE_RANGE 0x1 /* Apply exsiting page range */
 #define KASAN_VMALLOC_TLB_FLUSH  0x2 /* TLB flush */
diff --git a/mm/kasan/hw_tags.c b/mm/kasan/hw_tags.c
index 1c373cc4b3fa..d768c7360093 100644
--- a/mm/kasan/hw_tags.c
+++ b/mm/kasan/hw_tags.c
@@ -347,7 +347,7 @@ void *__kasan_unpoison_vmalloc(const void *start, unsigned long size,
 	 *
 	 * For non-VM_ALLOC allocations, page_alloc memory is tagged as usual.
 	 */
-	if (!(flags & KASAN_VMALLOC_VM_ALLOC)) {
+	if (!(flags & (KASAN_VMALLOC_VM_ALLOC | KASAN_VMALLOC_EXPAND))) {
 		WARN_ON(flags & KASAN_VMALLOC_INIT);
 		return (void *)start;
 	}
@@ -361,7 +361,14 @@ void *__kasan_unpoison_vmalloc(const void *start, unsigned long size,
 		return (void *)start;
 	}

-	tag = kasan_random_tag();
+	if (flags & KASAN_VMALLOC_EXPAND) {
+		size = round_up(size + ((unsigned long)start & KASAN_GRANULE_MASK),
+				KASAN_GRANULE_SIZE);
+		start = PTR_ALIGN_DOWN(start, KASAN_GRANULE_SIZE);
+		tag = get_tag(start);
+	} else
+		tag = kasan_random_tag();
+
 	start = set_tag(start, tag);

 	/* Unpoison and initialize memory up to size. */
diff --git a/mm/vmalloc.c b/mm/vmalloc.c
index 798b2ed21e46..6bfbf26fea3b 100644
--- a/mm/vmalloc.c
+++ b/mm/vmalloc.c
@@ -4176,6 +4176,7 @@ void *vrealloc_node_align_noprof(const void *p, size_t size, unsigned long align
 	 */
 	if (size <= alloced_size) {
 		kasan_unpoison_vmalloc(p + old_size, size - old_size,
+				       KASAN_VMALLOC_EXPAND |
 				       KASAN_VMALLOC_PROT_NORMAL);
 		/*
 		 * No need to zero memory here, as unused memory will have
--
LEVI:{C3F47F37-75D8-414A-A8BA-3980EC8A46D7}

Re: [PATCH] kasan: hw_tags: fix a false positive case of vrealloc in alloced size
Posted by Andrew Morton 2 days, 3 hours ago
On Sat, 29 Nov 2025 12:36:47 +0000 Yeoreum Yun <yeoreum.yun@arm.com> wrote:

> When a memory region is allocated with vmalloc() and later expanded with
> vrealloc() — while still within the originally allocated size —
> KASAN may report a false positive because
> it does not update the tags for the newly expanded portion of the memory.
> 
> A typical example of this pattern occurs in the BPF verifier,
> and the following is a related false positive report:
> 
> [ 2206.486476] ==================================================================
> [ 2206.486509] BUG: KASAN: invalid-access in __memcpy+0xc/0x30
> [ 2206.486607] Write at addr f5ff800083765270 by task test_progs/205
> [ 2206.486664] Pointer tag: [f5], memory tag: [fe]
> [ 2206.486703]
> [ 2206.486745] CPU: 4 UID: 0 PID: 205 Comm: test_progs Tainted: G           OE       6.18.0-rc7+ #145 PREEMPT(full)
> [ 2206.486861] Tainted: [O]=OOT_MODULE, [E]=UNSIGNED_MODULE
> [ 2206.486897] Hardware name:  , BIOS
> [ 2206.486932] Call trace:
> [ 2206.486961]  show_stack+0x24/0x40 (C)
> [ 2206.487071]  __dump_stack+0x28/0x48
> [ 2206.487182]  dump_stack_lvl+0x7c/0xb0
> [ 2206.487293]  print_address_description+0x80/0x270
> [ 2206.487403]  print_report+0x94/0x100
> [ 2206.487505]  kasan_report+0xd8/0x150
> [ 2206.487606]  __do_kernel_fault+0x64/0x268
> [ 2206.487717]  do_bad_area+0x38/0x110
> [ 2206.487820]  do_tag_check_fault+0x38/0x60
> [ 2206.487936]  do_mem_abort+0x48/0xc8
> [ 2206.488042]  el1_abort+0x40/0x70
> [ 2206.488127]  el1h_64_sync_handler+0x50/0x118
> [ 2206.488217]  el1h_64_sync+0xa4/0xa8
> [ 2206.488303]  __memcpy+0xc/0x30 (P)
> [ 2206.488412]  do_misc_fixups+0x4f8/0x1950
> [ 2206.488528]  bpf_check+0x31c/0x840
> [ 2206.488638]  bpf_prog_load+0x58c/0x658
> [ 2206.488737]  __sys_bpf+0x364/0x488
> [ 2206.488833]  __arm64_sys_bpf+0x30/0x58
> [ 2206.488920]  invoke_syscall+0x68/0xe8
> [ 2206.489033]  el0_svc_common+0xb0/0xf8
> [ 2206.489143]  do_el0_svc+0x28/0x48
> [ 2206.489249]  el0_svc+0x40/0xe8
> [ 2206.489337]  el0t_64_sync_handler+0x84/0x140
> [ 2206.489427]  el0t_64_sync+0x1bc/0x1c0
> 
> Here, 0xf5ff800083765000 is vmalloc()ed address for
> env->insn_aux_data with the size of 0x268.
> While this region is expanded size by 0x478 and initialise
> increased region to apply patched instructions,
> a false positive is triggered at the address 0xf5ff800083765270
> because __kasan_unpoison_vmalloc() with KASAN_VMALLOC_PROT_NORMAL flag only
> doesn't update the tag on increaed region.
> 
> To address this, introduces KASAN_VMALLOC_EXPAND flag which
> is used to expand vmalloc()ed memory in range of real allocated size
> to update tag for increased region.

Thanks.

> Fixes: 23689e91fb22 ("kasan, vmalloc: add vmalloc tagging for HW_TAGS”)
> Cc: <stable@vger.kernel.org>

Unfortunately this is changing the same code as "mm/kasan: fix
incorrect unpoisoning in vrealloc for KASAN",
(https://lkml.kernel.org/r/20251128111516.244497-1-jiayuan.chen@linux.dev)
which is also cc:stable.

So could you please take a look at the code in mm.git's
mm-hotfixes-unstable branch
(git://git.kernel.org/pub/scm/linux/kernel/git/akpm/mm) and base the
fix upon that?  This way everything should merge and backport nicely.


Re: [PATCH] kasan: hw_tags: fix a false positive case of vrealloc in alloced size
Posted by Yeoreum Yun 1 day, 22 hours ago
Hi Andrew,

> On Sat, 29 Nov 2025 12:36:47 +0000 Yeoreum Yun <yeoreum.yun@arm.com> wrote:
>
> > When a memory region is allocated with vmalloc() and later expanded with
> > vrealloc() — while still within the originally allocated size —
> > KASAN may report a false positive because
> > it does not update the tags for the newly expanded portion of the memory.
> >
> > A typical example of this pattern occurs in the BPF verifier,
> > and the following is a related false positive report:
> >
> > [ 2206.486476] ==================================================================
> > [ 2206.486509] BUG: KASAN: invalid-access in __memcpy+0xc/0x30
> > [ 2206.486607] Write at addr f5ff800083765270 by task test_progs/205
> > [ 2206.486664] Pointer tag: [f5], memory tag: [fe]
> > [ 2206.486703]
> > [ 2206.486745] CPU: 4 UID: 0 PID: 205 Comm: test_progs Tainted: G           OE       6.18.0-rc7+ #145 PREEMPT(full)
> > [ 2206.486861] Tainted: [O]=OOT_MODULE, [E]=UNSIGNED_MODULE
> > [ 2206.486897] Hardware name:  , BIOS
> > [ 2206.486932] Call trace:
> > [ 2206.486961]  show_stack+0x24/0x40 (C)
> > [ 2206.487071]  __dump_stack+0x28/0x48
> > [ 2206.487182]  dump_stack_lvl+0x7c/0xb0
> > [ 2206.487293]  print_address_description+0x80/0x270
> > [ 2206.487403]  print_report+0x94/0x100
> > [ 2206.487505]  kasan_report+0xd8/0x150
> > [ 2206.487606]  __do_kernel_fault+0x64/0x268
> > [ 2206.487717]  do_bad_area+0x38/0x110
> > [ 2206.487820]  do_tag_check_fault+0x38/0x60
> > [ 2206.487936]  do_mem_abort+0x48/0xc8
> > [ 2206.488042]  el1_abort+0x40/0x70
> > [ 2206.488127]  el1h_64_sync_handler+0x50/0x118
> > [ 2206.488217]  el1h_64_sync+0xa4/0xa8
> > [ 2206.488303]  __memcpy+0xc/0x30 (P)
> > [ 2206.488412]  do_misc_fixups+0x4f8/0x1950
> > [ 2206.488528]  bpf_check+0x31c/0x840
> > [ 2206.488638]  bpf_prog_load+0x58c/0x658
> > [ 2206.488737]  __sys_bpf+0x364/0x488
> > [ 2206.488833]  __arm64_sys_bpf+0x30/0x58
> > [ 2206.488920]  invoke_syscall+0x68/0xe8
> > [ 2206.489033]  el0_svc_common+0xb0/0xf8
> > [ 2206.489143]  do_el0_svc+0x28/0x48
> > [ 2206.489249]  el0_svc+0x40/0xe8
> > [ 2206.489337]  el0t_64_sync_handler+0x84/0x140
> > [ 2206.489427]  el0t_64_sync+0x1bc/0x1c0
> >
> > Here, 0xf5ff800083765000 is vmalloc()ed address for
> > env->insn_aux_data with the size of 0x268.
> > While this region is expanded size by 0x478 and initialise
> > increased region to apply patched instructions,
> > a false positive is triggered at the address 0xf5ff800083765270
> > because __kasan_unpoison_vmalloc() with KASAN_VMALLOC_PROT_NORMAL flag only
> > doesn't update the tag on increaed region.
> >
> > To address this, introduces KASAN_VMALLOC_EXPAND flag which
> > is used to expand vmalloc()ed memory in range of real allocated size
> > to update tag for increased region.
>
> Thanks.
>
> > Fixes: 23689e91fb22 ("kasan, vmalloc: add vmalloc tagging for HW_TAGS”)
> > Cc: <stable@vger.kernel.org>
>
> Unfortunately this is changing the same code as "mm/kasan: fix
> incorrect unpoisoning in vrealloc for KASAN",
> (https://lkml.kernel.org/r/20251128111516.244497-1-jiayuan.chen@linux.dev)
> which is also cc:stable.
>
> So could you please take a look at the code in mm.git's
> mm-hotfixes-unstable branch
> (git://git.kernel.org/pub/scm/linux/kernel/git/akpm/mm) and base the
> fix upon that?  This way everything should merge and backport nicely.

Thanks for sharing this :)
But I think the patch from Jiayuan still has a problem
since vrealloc() can pass the "unaligned address" with KASAN_GRANULE_SIZE
and this will trigger WARN_ON() in kasan_unpoison().

Except that I don't have a strong opinion whether adding a new
interface or a new flag.
But, I want to receive a comment from KASAN maintainers to
check what is better.

--
Sincerely,
Yeoreum Yun
Re: [PATCH] kasan: hw_tags: fix a false positive case of vrealloc in alloced size
Posted by Yeoreum Yun 1 day, 22 hours ago
On Sat, Nov 29, 2025 at 11:26:25PM +0000, Yeoreum Yun wrote:
> Hi Andrew,
>
> > On Sat, 29 Nov 2025 12:36:47 +0000 Yeoreum Yun <yeoreum.yun@arm.com> wrote:
> >
> > > When a memory region is allocated with vmalloc() and later expanded with
> > > vrealloc() — while still within the originally allocated size —
> > > KASAN may report a false positive because
> > > it does not update the tags for the newly expanded portion of the memory.
> > >
> > > A typical example of this pattern occurs in the BPF verifier,
> > > and the following is a related false positive report:
> > >
> > > [ 2206.486476] ==================================================================
> > > [ 2206.486509] BUG: KASAN: invalid-access in __memcpy+0xc/0x30
> > > [ 2206.486607] Write at addr f5ff800083765270 by task test_progs/205
> > > [ 2206.486664] Pointer tag: [f5], memory tag: [fe]
> > > [ 2206.486703]
> > > [ 2206.486745] CPU: 4 UID: 0 PID: 205 Comm: test_progs Tainted: G           OE       6.18.0-rc7+ #145 PREEMPT(full)
> > > [ 2206.486861] Tainted: [O]=OOT_MODULE, [E]=UNSIGNED_MODULE
> > > [ 2206.486897] Hardware name:  , BIOS
> > > [ 2206.486932] Call trace:
> > > [ 2206.486961]  show_stack+0x24/0x40 (C)
> > > [ 2206.487071]  __dump_stack+0x28/0x48
> > > [ 2206.487182]  dump_stack_lvl+0x7c/0xb0
> > > [ 2206.487293]  print_address_description+0x80/0x270
> > > [ 2206.487403]  print_report+0x94/0x100
> > > [ 2206.487505]  kasan_report+0xd8/0x150
> > > [ 2206.487606]  __do_kernel_fault+0x64/0x268
> > > [ 2206.487717]  do_bad_area+0x38/0x110
> > > [ 2206.487820]  do_tag_check_fault+0x38/0x60
> > > [ 2206.487936]  do_mem_abort+0x48/0xc8
> > > [ 2206.488042]  el1_abort+0x40/0x70
> > > [ 2206.488127]  el1h_64_sync_handler+0x50/0x118
> > > [ 2206.488217]  el1h_64_sync+0xa4/0xa8
> > > [ 2206.488303]  __memcpy+0xc/0x30 (P)
> > > [ 2206.488412]  do_misc_fixups+0x4f8/0x1950
> > > [ 2206.488528]  bpf_check+0x31c/0x840
> > > [ 2206.488638]  bpf_prog_load+0x58c/0x658
> > > [ 2206.488737]  __sys_bpf+0x364/0x488
> > > [ 2206.488833]  __arm64_sys_bpf+0x30/0x58
> > > [ 2206.488920]  invoke_syscall+0x68/0xe8
> > > [ 2206.489033]  el0_svc_common+0xb0/0xf8
> > > [ 2206.489143]  do_el0_svc+0x28/0x48
> > > [ 2206.489249]  el0_svc+0x40/0xe8
> > > [ 2206.489337]  el0t_64_sync_handler+0x84/0x140
> > > [ 2206.489427]  el0t_64_sync+0x1bc/0x1c0
> > >
> > > Here, 0xf5ff800083765000 is vmalloc()ed address for
> > > env->insn_aux_data with the size of 0x268.
> > > While this region is expanded size by 0x478 and initialise
> > > increased region to apply patched instructions,
> > > a false positive is triggered at the address 0xf5ff800083765270
> > > because __kasan_unpoison_vmalloc() with KASAN_VMALLOC_PROT_NORMAL flag only
> > > doesn't update the tag on increaed region.
> > >
> > > To address this, introduces KASAN_VMALLOC_EXPAND flag which
> > > is used to expand vmalloc()ed memory in range of real allocated size
> > > to update tag for increased region.
> >
> > Thanks.
> >
> > > Fixes: 23689e91fb22 ("kasan, vmalloc: add vmalloc tagging for HW_TAGS”)
> > > Cc: <stable@vger.kernel.org>
> >
> > Unfortunately this is changing the same code as "mm/kasan: fix
> > incorrect unpoisoning in vrealloc for KASAN",
> > (https://lkml.kernel.org/r/20251128111516.244497-1-jiayuan.chen@linux.dev)
> > which is also cc:stable.
> >
> > So could you please take a look at the code in mm.git's
> > mm-hotfixes-unstable branch
> > (git://git.kernel.org/pub/scm/linux/kernel/git/akpm/mm) and base the
> > fix upon that?  This way everything should merge and backport nicely.
>
> Thanks for sharing this :)
> But I think the patch from Jiayuan still has a problem
> since vrealloc() can pass the "unaligned address" with KASAN_GRANULE_SIZE
> and this will trigger WARN_ON() in kasan_unpoison().

Ah, I missed his patch change the pointer from p + old_size to p
for the kasan_unpoison_vrealloc().
Then the patch seems almost identical to fix the same problem.
So, you can drop my patch.

Sorry to make noise :\

[...]


--
Sincerely,
Yeoreum Yun