Skip to contentSkip to conversation

slab: store tagged freelist for off-slab slabmgmt

PATCH In review Qian Cai 2 participants 4 messages Apr 3, 2019 Memory management

lore
Share & tools
lore

Qian Cai

author
#0 lore raw

PATCH 1/1 slab: store tagged freelist for off-slab slabmgmt

The commit 51dedad06b5f ("kasan, slab: make freelist stored without tags") calls kasan_reset_tag() for off-slab slab management object leading to freelist being stored non-tagged. However, cache_grow_begin() -> alloc_slabmgmt() -> kmem_cache_alloc_node() which assigns a tag for the address and stores in the shadow address. As the result, it causes endless errors below during boot due to drain_freelist() -> slab_destroy() -> kasan_slab_free() which compares already untagged freelist against the stored tag in the shadow address. Since off-slab slab management object freelist is such a special case, so just store it tagged. Non-off-slab management object freelist is still stored untagged which has not been assigned a tag and should not cause any other troubles with this inconsistency.

BUG: KASAN: double-free or invalid-free in slab_destroy+0x84/0x88

Pointer tag: [ff], memory tag: [99]

CPU: 0 PID: 1376 Comm: kworker/0:4 Tainted: G        W
5.1.0-rc3+ #8
Hardware name: HPE Apollo 70             /C01_APACHE_MB         , BIOS
L50_5.13_1.0.6 07/10/2018
Workqueue: cgroup_destroy css_killed_work_fn
Call trace:
 dump_backtrace+0x0/0x450
 show_stack+0x20/0x2c
 dump_stack+0xe0/0x16c
 print_address_description+0x74/0x2a4
 kasan_report_invalid_free+0x80/0xc0
 __kasan_slab_free+0x204/0x208
 kasan_slab_free+0xc/0x18
 kmem_cache_free+0xe4/0x254
 slab_destroy+0x84/0x88
 drain_freelist+0xd0/0x104
 __kmem_cache_shrink+0x1ac/0x224
 __kmemcg_cache_deactivate+0x1c/0x28
 memcg_deactivate_kmem_caches+0xa0/0xe8
 memcg_offline_kmem+0x8c/0x3d4
 mem_cgroup_css_offline+0x24c/0x290
 css_killed_work_fn+0x154/0x618
 process_one_work+0x9cc/0x183c
 worker_thread+0x9b0/0xe38
 kthread+0x374/0x390
 ret_from_fork+0x10/0x18

Allocated by task 1625:
 __kasan_kmalloc+0x168/0x240
 kasan_slab_alloc+0x18/0x20
 kmem_cache_alloc_node+0x1f8/0x3a0
 cache_grow_begin+0x4fc/0xa24
 cache_alloc_refill+0x2f8/0x3e8
 kmem_cache_alloc+0x1bc/0x3bc
 sock_alloc_inode+0x58/0x334
 alloc_inode+0xb8/0x164
 new_inode_pseudo+0x20/0xec
 sock_alloc+0x74/0x284
 __sock_create+0xb0/0x58c
 sock_create+0x98/0xb8
 __sys_socket+0x60/0x138
 __arm64_sys_socket+0xa4/0x110
 el0_svc_handler+0x2c0/0x47c
 el0_svc+0x8/0xc

Freed by task 1625:
 __kasan_slab_free+0x114/0x208
 kasan_slab_free+0xc/0x18
 kfree+0x1a8/0x1e0
 single_release+0x7c/0x9c
 close_pdeo+0x13c/0x43c
 proc_reg_release+0xec/0x108
 __fput+0x2f8/0x784
 ____fput+0x1c/0x28
 task_work_run+0xc0/0x1b0
 do_notify_resume+0xb44/0x1278
 work_pending+0x8/0x10

The buggy address belongs to the object at ffff809681b89e00
 which belongs to the cache kmalloc-128 of size 128
The buggy address is located 0 bytes inside of
 128-byte region [ffff809681b89e00, ffff809681b89e80)
The buggy address belongs to the page:
page:ffff7fe025a06e00 count:1 mapcount:0 mapping:01ff80082000fb00
index:0xffff809681b8fe04
flags: 0x17ffffffc000200(slab)
raw: 017ffffffc000200 ffff7fe025a06d08 ffff7fe022ef7b88 01ff80082000fb00
raw: ffff809681b8fe04 ffff809681b80000 00000001000000e0 0000000000000000
page dumped because: kasan: bad access detected
page allocated via order 0, migratetype Unmovable, gfp_mask
0x2420c0(__GFP_IO|__GFP_FS|__GFP_NOWARN|__GFP_COMP|__GFP_THISNODE)
 prep_new_page+0x4e0/0x5e0
 get_page_from_freelist+0x4ce8/0x50d4
 __alloc_pages_nodemask+0x738/0x38b8
 cache_grow_begin+0xd8/0xa24
 ____cache_alloc_node+0x14c/0x268
 __kmalloc+0x1c8/0x3fc
 ftrace_free_mem+0x408/0x1284
 ftrace_free_init_mem+0x20/0x28
 kernel_init+0x24/0x548
 ret_from_fork+0x10/0x18

Memory state around the buggy address:
 ffff809681b89c00: fe fe fe fe fe fe fe fe fe fe fe fe fe fe fe fe
 ffff809681b89d00: fe fe fe fe fe fe fe fe fe fe fe fe fe fe fe fe
>ffff809681b89e00: 99 99 99 99 99 99 99 99 fe fe fe fe fe fe fe fe
                   ^
 ffff809681b89f00: 43 43 43 43 43 fe fe fe fe fe fe fe fe fe fe fe
 ffff809681b8a000: 6d fe fe fe fe fe fe fe fe fe fe fe fe fe fe fe
Fixes 51dedad06b5f ("kasan, slab: make freelist stored without tags")
---
 mm/slab.c | 1 -
 1 file changed, 1 deletion(-)
mm/slab.c +0 -1
index 329bfe67f2ca..46a6e084222b 100644
@@ -2374,7 +2374,6 @@ static void *alloc_slabmgmt(struct kmem_cache *cachep,
 		/* Slab management obj is off-slab. */
 		freelist = kmem_cache_alloc_node(cachep->freelist_cache,
 					      local_flags, nodeid);
-		freelist = kasan_reset_tag(freelist);
 		if (!freelist)
 			return NULL;
 	} else {
Signature · 1 line
2.17.2 (Apple Git-113)
#1 lore raw
Quoted 13 lines · Qian Cai

The commit 51dedad06b5f ("kasan, slab: make freelist stored without tags") calls kasan_reset_tag() for off-slab slab management object leading to freelist being stored non-tagged. However, cache_grow_begin() -> alloc_slabmgmt() -> kmem_cache_alloc_node() which assigns a tag for the address and stores in the shadow address. As the result, it causes endless errors below during boot due to drain_freelist() -> slab_destroy() -> kasan_slab_free() which compares already untagged freelist against the stored tag in the shadow address. Since off-slab slab management object freelist is such a special case, so just store it tagged. Non-off-slab management object freelist is still stored untagged which has not been assigned a tag and should not cause any other troubles with this inconsistency.

Hi Qian,

Could you share the config (or other steps) you used to reproduce this?

Thanks!

Quoted 115 lines
BUG: KASAN: double-free or invalid-free in slab_destroy+0x84/0x88

Pointer tag: [ff], memory tag: [99]

CPU: 0 PID: 1376 Comm: kworker/0:4 Tainted: G        W
5.1.0-rc3+ #8
Hardware name: HPE Apollo 70             /C01_APACHE_MB         , BIOS
L50_5.13_1.0.6 07/10/2018
Workqueue: cgroup_destroy css_killed_work_fn
Call trace:
 dump_backtrace+0x0/0x450
 show_stack+0x20/0x2c
 dump_stack+0xe0/0x16c
 print_address_description+0x74/0x2a4
 kasan_report_invalid_free+0x80/0xc0
 __kasan_slab_free+0x204/0x208
 kasan_slab_free+0xc/0x18
 kmem_cache_free+0xe4/0x254
 slab_destroy+0x84/0x88
 drain_freelist+0xd0/0x104
 __kmem_cache_shrink+0x1ac/0x224
 __kmemcg_cache_deactivate+0x1c/0x28
 memcg_deactivate_kmem_caches+0xa0/0xe8
 memcg_offline_kmem+0x8c/0x3d4
 mem_cgroup_css_offline+0x24c/0x290
 css_killed_work_fn+0x154/0x618
 process_one_work+0x9cc/0x183c
 worker_thread+0x9b0/0xe38
 kthread+0x374/0x390
 ret_from_fork+0x10/0x18

Allocated by task 1625:
 __kasan_kmalloc+0x168/0x240
 kasan_slab_alloc+0x18/0x20
 kmem_cache_alloc_node+0x1f8/0x3a0
 cache_grow_begin+0x4fc/0xa24
 cache_alloc_refill+0x2f8/0x3e8
 kmem_cache_alloc+0x1bc/0x3bc
 sock_alloc_inode+0x58/0x334
 alloc_inode+0xb8/0x164
 new_inode_pseudo+0x20/0xec
 sock_alloc+0x74/0x284
 __sock_create+0xb0/0x58c
 sock_create+0x98/0xb8
 __sys_socket+0x60/0x138
 __arm64_sys_socket+0xa4/0x110
 el0_svc_handler+0x2c0/0x47c
 el0_svc+0x8/0xc

Freed by task 1625:
 __kasan_slab_free+0x114/0x208
 kasan_slab_free+0xc/0x18
 kfree+0x1a8/0x1e0
 single_release+0x7c/0x9c
 close_pdeo+0x13c/0x43c
 proc_reg_release+0xec/0x108
 __fput+0x2f8/0x784
 ____fput+0x1c/0x28
 task_work_run+0xc0/0x1b0
 do_notify_resume+0xb44/0x1278
 work_pending+0x8/0x10

The buggy address belongs to the object at ffff809681b89e00
 which belongs to the cache kmalloc-128 of size 128
The buggy address is located 0 bytes inside of
 128-byte region [ffff809681b89e00, ffff809681b89e80)
The buggy address belongs to the page:
page:ffff7fe025a06e00 count:1 mapcount:0 mapping:01ff80082000fb00
index:0xffff809681b8fe04
flags: 0x17ffffffc000200(slab)
raw: 017ffffffc000200 ffff7fe025a06d08 ffff7fe022ef7b88 01ff80082000fb00
raw: ffff809681b8fe04 ffff809681b80000 00000001000000e0 0000000000000000
page dumped because: kasan: bad access detected
page allocated via order 0, migratetype Unmovable, gfp_mask
0x2420c0(__GFP_IO|__GFP_FS|__GFP_NOWARN|__GFP_COMP|__GFP_THISNODE)
 prep_new_page+0x4e0/0x5e0
 get_page_from_freelist+0x4ce8/0x50d4
 __alloc_pages_nodemask+0x738/0x38b8
 cache_grow_begin+0xd8/0xa24
 ____cache_alloc_node+0x14c/0x268
 __kmalloc+0x1c8/0x3fc
 ftrace_free_mem+0x408/0x1284
 ftrace_free_init_mem+0x20/0x28
 kernel_init+0x24/0x548
 ret_from_fork+0x10/0x18

Memory state around the buggy address:
 ffff809681b89c00: fe fe fe fe fe fe fe fe fe fe fe fe fe fe fe fe
 ffff809681b89d00: fe fe fe fe fe fe fe fe fe fe fe fe fe fe fe fe
>ffff809681b89e00: 99 99 99 99 99 99 99 99 fe fe fe fe fe fe fe fe
                   ^
 ffff809681b89f00: 43 43 43 43 43 fe fe fe fe fe fe fe fe fe fe fe
 ffff809681b8a000: 6d fe fe fe fe fe fe fe fe fe fe fe fe fe fe fe
Fixes 51dedad06b5f ("kasan, slab: make freelist stored without tags")
---
 mm/slab.c | 1 -
 1 file changed, 1 deletion(-)
mm/slab.c +0 -2
index 329bfe67f2ca..46a6e084222b 100644
@@ -2374,7 +2374,6 @@ static void *alloc_slabmgmt(struct kmem_cache *cachep,
                /* Slab management obj is off-slab. */
                freelist = kmem_cache_alloc_node(cachep->freelist_cache,
                                              local_flags, nodeid);
-               freelist = kasan_reset_tag(freelist);
                if (!freelist)
                        return NULL;
        } else {
--

2.17.2 (Apple Git-113)

author
#2 lore raw
Quoted 15 lines · Andrey Konovalov
Quoted 13 lines · Qian Cai

The commit 51dedad06b5f ("kasan, slab: make freelist stored without tags") calls kasan_reset_tag() for off-slab slab management object leading to freelist being stored non-tagged. However, cache_grow_begin() -> alloc_slabmgmt() -> kmem_cache_alloc_node() which assigns a tag for the address and stores in the shadow address. As the result, it causes endless errors below during boot due to drain_freelist() -> slab_destroy() -> kasan_slab_free() which compares already untagged freelist against the stored tag in the shadow address. Since off-slab slab management object freelist is such a special case, so just store it tagged. Non-off-slab management object freelist is still stored untagged which has not been assigned a tag and should not cause any other troubles with this inconsistency.

Quoted 3 lines

Hi Qian,

Could you share the config (or other steps) you used to reproduce this?

https://git.sr.ht/~cai/linux-debug/blob/master/config

Additional command-line option to boot:

page_poison=on crashkernel=768M earlycon page_owner=on numa_balancing=enable systemd.unified_cgroup_hierarchy=1 debug_guardpage_minorder=1

Reviewed-by
#3 lore raw
Quoted 19 lines · Qian Cai
Quoted 17 lines · Andrey Konovalov
Quoted 13 lines · Qian Cai

The commit 51dedad06b5f ("kasan, slab: make freelist stored without tags") calls kasan_reset_tag() for off-slab slab management object leading to freelist being stored non-tagged. However, cache_grow_begin() -> alloc_slabmgmt() -> kmem_cache_alloc_node() which assigns a tag for the address and stores in the shadow address. As the result, it causes endless errors below during boot due to drain_freelist() -> slab_destroy() -> kasan_slab_free() which compares already untagged freelist against the stored tag in the shadow address. Since off-slab slab management object freelist is such a special case, so just store it tagged. Non-off-slab management object freelist is still stored untagged which has not been assigned a tag and should not cause any other troubles with this inconsistency.

Hi Qian,

Quoted 8 lines
Quoted 1 line

Could you share the config (or other steps) you used to reproduce this?

https://git.sr.ht/~cai/linux-debug/blob/master/config

Additional command-line option to boot:

page_poison=on crashkernel=768M earlycon page_owner=on numa_balancing=enable systemd.unified_cgroup_hierarchy=1 debug_guardpage_minorder=1

Reproduced, thanks!

As far as my understanding of how SLAB works goes, this change looks good to me.

Thanks, Qian!

Quoted 7 lines

--
You received this message because you are subscribed to the Google Groups "kasan-dev" group.
To unsubscribe from this group and stop receiving emails from it, send an email to [email protected]. To post to this group, send email to [email protected].
To view this discussion on the web visit https://groups.google.com/d/msgid/kasan-dev/1554296870.26196.32.camel%40lca.pw.
For more options, visit https://groups.google.com/d/optout.