Skip to contentSkip to conversation

iommu/amd: fix a warning in increase_address_space

PATCH In review Qian Cai 3 participants 6 messages Oct 16, 2019 – Oct 18, 2019

Subsystems (2)
lore
Share & tools
lore

Qian Cai

author
#0 lore raw

PATCH 1/1 iommu/amd: fix a warning in increase_address_space

After the commit 754265bcab78 ("iommu/amd: Fix race in increase_address_space()"), it could still possible trigger a race condition under some heavy memory pressure below. The race to trigger a warning is,

CPU0: CPU1:
in alloc_pte(): in increase_address_space():
while (address > PM_LEVEL_SIZE(domain->mode)) [1]

				spin_lock_irqsave(&domain->lock
				domain->mode    += 1;
				spin_unlock_irqrestore(&domain->lock

in increase_address_space():
spin_lock_irqsave(&domain->lock
if (WARN_ON_ONCE(domain->mode == PAGE_MODE_6_LEVEL))

[1] domain->mode = 5

It is unclear the triggering of the warning is the root cause of the smartpqi offline yet, but let's fix it first by lifting the locking.

WARNING: CPU: 57 PID: 124314 at drivers/iommu/amd_iommu.c:1474
iommu_map_page+0x718/0x7e0

smartpqi 0000:23:00.0: AMD-Vi: Event logged [IO_PAGE_FAULT domain=0x0000 address=0xffffffffffec0000 flags=0x0010]
smartpqi 0000:23:00.0: AMD-Vi: Event logged [IO_PAGE_FAULT domain=0x0000 address=0xffffffffffec1000 flags=0x0010]

CPU: 57 PID: 124314 Comm: oom01 Tainted: G           O
Hardware name: HPE ProLiant DL385 Gen10/ProLiant DL385 Gen10, BIOS A40
07/10/2019
RIP: 0010:iommu_map_page+0x718/0x7e0
Code: 88 a5 70 ff ff ff e9 5d fa ff ff 48 8b b5 70 ff ff ff 4c 89 ef e8

08 32 2f 00 41 80 fc 01 0f 87 b7 3d 00 00 41 83 e4 01 eb be <0f> 0b 48 8b b5 70 ff ff ff 4c 89 ef e8 e7 31 2f 00 eb dd 0f 0b 48
RSP: 0018:ffff888da4816cb8 EFLAGS: 00010046
RAX: 0000000000000000 RBX: ffff8885fe689000 RCX: ffffffff96f4a6c4
RDX: 0000000000000007 RSI: dffffc0000000000 RDI: ffff8885fe689124
RBP: ffff888da4816da8 R08: ffffed10bfcd120e R09: ffffed10bfcd120e
R10: ffffed10bfcd120d R11: ffff8885fe68906b R12: 0000000000000000 smartpqi 0000:23:00.0: AMD-Vi: Event logged [IO_PAGE_FAULT domain=0x0000 address=0xffffffffffec1a00 flags=0x0010]
R13: ffff8885fe689068 R14: ffff8885fe689124 R15: 0000000000000000 smartpqi 0000:23:00.0: AMD-Vi: Event logged [IO_PAGE_FAULT domain=0x0000 address=0xffffffffffec1e00 flags=0x0010]
FS: 00007f29722ba700(0000) GS:ffff88902f880000(0000) knlGS:0000000000000000

CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
CR2: 00007f27f82d8000 CR3: 000000102ed9c000 CR4: 00000000003406e0
Call Trace:

smartpqi 0000:23:00.0: AMD-Vi: Event logged [IO_PAGE_FAULT domain=0x0000 address=0xffffffffffec2000 flags=0x0010]
map_sg+0x1ce/0x2f0
smartpqi 0000:23:00.0: AMD-Vi: Event logged [IO_PAGE_FAULT domain=0x0000 address=0xffffffffffec2400 flags=0x0010]

 scsi_dma_map+0xd7/0x160
 pqi_raid_submit_scsi_cmd_with_io_request+0x1b8/0x420 [smartpqi]

smartpqi 0000:23:00.0: AMD-Vi: Event logged [IO_PAGE_FAULT domain=0x0000 address=0xffffffffffec2800 flags=0x0010]
pqi_scsi_queue_command+0x8ab/0xe00 [smartpqi]
smartpqi 0000:23:00.0: AMD-Vi: Event logged [IO_PAGE_FAULT domain=0x0000 address=0xffffffffffec2c00 flags=0x0010]
scsi_queue_rq+0xd19/0x1360
smartpqi 0000:23:00.0: AMD-Vi: Event logged [IO_PAGE_FAULT domain=0x0000 address=0xffffffffffec3000 flags=0x0010]
__blk_mq_try_issue_directly+0x295/0x3f0
smartpqi 0000:23:00.0: AMD-Vi: Event logged [IO_PAGE_FAULT domain=0x0000 address=0xffffffffffec3400 flags=0x0010]
AMD-Vi: Event logged [IO_PAGE_FAULT device=23:00.0 domain=0x0000 address=0xffffffffffec3800 flags=0x0010]
blk_mq_request_issue_directly+0xb5/0x100
AMD-Vi: Event logged [IO_PAGE_FAULT device=23:00.0 domain=0x0000 address=0xffffffffffec3c00 flags=0x0010]

 blk_mq_try_issue_list_directly+0xa9/0x160
 blk_mq_sched_insert_requests+0x228/0x380
 blk_mq_flush_plug_list+0x448/0x7e0
 blk_flush_plug_list+0x1eb/0x230
 blk_finish_plug+0x43/0x5d
 shrink_node_memcg+0x9c5/0x1550

smartpqi 0000:23:00.0: controller is offline: status code 0x14803 smartpqi 0000:23:00.0: controller offline

Fixes 754265bcab78 ("iommu/amd: Fix race in increase_address_space()")

---

BTW, Joerg, this line from the commit "iommu/amd: Remove domain->updated" looks suspicious. Not sure what the purpose of it.

*updated = increase_address_space(domain, gfp) || *updated;

 drivers/iommu/amd_iommu.c | 10 +++++-----
 1 file changed, 5 insertions(+), 5 deletions(-)
drivers/iommu/amd_iommu.c +5 -5
index 2369b8af81f3..a5754068aa29 100644
@@ -1465,12 +1465,9 @@ static void free_pagetable(struct protection_domain *domain)
 static bool increase_address_space(struct protection_domain *domain,
 				   gfp_t gfp)
 {
-	unsigned long flags;
 	bool ret = false;
 	u64 *pte;
 
-	spin_lock_irqsave(&domain->lock, flags);
-
 	if (WARN_ON_ONCE(domain->mode == PAGE_MODE_6_LEVEL))
 		/* address space already 64 bit large */
 		goto out;
@@ -1487,8 +1484,6 @@ static bool increase_address_space(struct protection_domain *domain,
 	ret = true;
 
 out:
-	spin_unlock_irqrestore(&domain->lock, flags);
-
 	return ret;
 }
 
@@ -1499,14 +1494,19 @@ static u64 *alloc_pte(struct protection_domain *domain,
 		      gfp_t gfp,
 		      bool *updated)
 {
+	unsigned long flags;
 	int level, end_lvl;
 	u64 *pte, *page;
 
 	BUG_ON(!is_power_of_2(page_size));
 
+	spin_lock_irqsave(&domain->lock, flags);
+
 	while (address > PM_LEVEL_SIZE(domain->mode))
 		*updated = increase_address_space(domain, gfp) || *updated;
 
+	spin_unlock_irqrestore(&domain->lock, flags);
+
 	level   = domain->mode - 1;
 	pte     = &domain->pt_root[PM_LEVEL_INDEX(level, address)];
 	address = PAGE_SIZE_ALIGN(address, page_size);
Signature · 1 line
1.8.3.1
#1 lore raw
Quoted 82 lines · Qian Cai

After the commit 754265bcab78 ("iommu/amd: Fix race in increase_address_space()"), it could still possible trigger a race condition under some heavy memory pressure below. The race to trigger a warning is,

CPU0: CPU1:
in alloc_pte(): in increase_address_space():
while (address > PM_LEVEL_SIZE(domain->mode)) [1]

				spin_lock_irqsave(&domain->lock
				domain->mode    += 1;
				spin_unlock_irqrestore(&domain->lock

in increase_address_space():
spin_lock_irqsave(&domain->lock
if (WARN_ON_ONCE(domain->mode == PAGE_MODE_6_LEVEL))

[1] domain->mode = 5

It is unclear the triggering of the warning is the root cause of the smartpqi offline yet, but let's fix it first by lifting the locking.

WARNING: CPU: 57 PID: 124314 at drivers/iommu/amd_iommu.c:1474
iommu_map_page+0x718/0x7e0

smartpqi 0000:23:00.0: AMD-Vi: Event logged [IO_PAGE_FAULT domain=0x0000 address=0xffffffffffec0000 flags=0x0010]
smartpqi 0000:23:00.0: AMD-Vi: Event logged [IO_PAGE_FAULT domain=0x0000 address=0xffffffffffec1000 flags=0x0010]

CPU: 57 PID: 124314 Comm: oom01 Tainted: G           O
Hardware name: HPE ProLiant DL385 Gen10/ProLiant DL385 Gen10, BIOS A40
07/10/2019
RIP: 0010:iommu_map_page+0x718/0x7e0
Code: 88 a5 70 ff ff ff e9 5d fa ff ff 48 8b b5 70 ff ff ff 4c 89 ef e8

08 32 2f 00 41 80 fc 01 0f 87 b7 3d 00 00 41 83 e4 01 eb be <0f> 0b 48 8b b5 70 ff ff ff 4c 89 ef e8 e7 31 2f 00 eb dd 0f 0b 48
RSP: 0018:ffff888da4816cb8 EFLAGS: 00010046
RAX: 0000000000000000 RBX: ffff8885fe689000 RCX: ffffffff96f4a6c4
RDX: 0000000000000007 RSI: dffffc0000000000 RDI: ffff8885fe689124
RBP: ffff888da4816da8 R08: ffffed10bfcd120e R09: ffffed10bfcd120e
R10: ffffed10bfcd120d R11: ffff8885fe68906b R12: 0000000000000000 smartpqi 0000:23:00.0: AMD-Vi: Event logged [IO_PAGE_FAULT domain=0x0000 address=0xffffffffffec1a00 flags=0x0010]
R13: ffff8885fe689068 R14: ffff8885fe689124 R15: 0000000000000000 smartpqi 0000:23:00.0: AMD-Vi: Event logged [IO_PAGE_FAULT domain=0x0000 address=0xffffffffffec1e00 flags=0x0010]
FS: 00007f29722ba700(0000) GS:ffff88902f880000(0000) knlGS:0000000000000000

CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
CR2: 00007f27f82d8000 CR3: 000000102ed9c000 CR4: 00000000003406e0
Call Trace:

smartpqi 0000:23:00.0: AMD-Vi: Event logged [IO_PAGE_FAULT domain=0x0000 address=0xffffffffffec2000 flags=0x0010]
map_sg+0x1ce/0x2f0
smartpqi 0000:23:00.0: AMD-Vi: Event logged [IO_PAGE_FAULT domain=0x0000 address=0xffffffffffec2400 flags=0x0010]

scsi_dma_map+0xd7/0x160
pqi_raid_submit_scsi_cmd_with_io_request+0x1b8/0x420 [smartpqi]

smartpqi 0000:23:00.0: AMD-Vi: Event logged [IO_PAGE_FAULT domain=0x0000 address=0xffffffffffec2800 flags=0x0010] pqi_scsi_queue_command+0x8ab/0xe00 [smartpqi]
smartpqi 0000:23:00.0: AMD-Vi: Event logged [IO_PAGE_FAULT domain=0x0000 address=0xffffffffffec2c00 flags=0x0010] scsi_queue_rq+0xd19/0x1360
smartpqi 0000:23:00.0: AMD-Vi: Event logged [IO_PAGE_FAULT domain=0x0000 address=0xffffffffffec3000 flags=0x0010] __blk_mq_try_issue_directly+0x295/0x3f0
smartpqi 0000:23:00.0: AMD-Vi: Event logged [IO_PAGE_FAULT domain=0x0000 address=0xffffffffffec3400 flags=0x0010]
AMD-Vi: Event logged [IO_PAGE_FAULT device=23:00.0 domain=0x0000 address=0xffffffffffec3800 flags=0x0010] blk_mq_request_issue_directly+0xb5/0x100
AMD-Vi: Event logged [IO_PAGE_FAULT device=23:00.0 domain=0x0000 address=0xffffffffffec3c00 flags=0x0010]

blk_mq_try_issue_list_directly+0xa9/0x160
blk_mq_sched_insert_requests+0x228/0x380
blk_mq_flush_plug_list+0x448/0x7e0
blk_flush_plug_list+0x1eb/0x230
blk_finish_plug+0x43/0x5d
shrink_node_memcg+0x9c5/0x1550

smartpqi 0000:23:00.0: controller is offline: status code 0x14803 smartpqi 0000:23:00.0: controller offline

Quoted 9 lines
Fixes 754265bcab78 ("iommu/amd: Fix race in increase_address_space()")

---

BTW, Joerg, this line from the commit "iommu/amd: Remove domain->updated" looks suspicious. Not sure what the purpose of it.

*updated = increase_address_space(domain, gfp) || *updated;

Looking at it again I think that isn't an issue really, it would just not lose updated being set in a previous loop iteration, but now I'm wondering about the loop itself. In the cases where it would return false, how does the evaluation of the condition for the while loop change?

Quoted 56 lines

drivers/iommu/amd_iommu.c | 10 +++++-----

1 file changed, 5 insertions(+), 5 deletions(-)
drivers/iommu/amd_iommu.c +0 -3
index 2369b8af81f3..a5754068aa29 100644
@@ -1465,12 +1465,9 @@ static void free_pagetable(struct protection_domain *domain)
 static bool increase_address_space(struct protection_domain *domain,
 				   gfp_t gfp)
 {
-	unsigned long flags;
 	bool ret = false;
 	u64 *pte;
 
-	spin_lock_irqsave(&domain->lock, flags);
-
 	if (WARN_ON_ONCE(domain->mode == PAGE_MODE_6_LEVEL))
 		/* address space already 64 bit large */
 		goto out;
@@ -1487,8 +1484,6 @@ static bool increase_address_space(struct protection_domain *domain,
 	ret = true;
 

out:

drivers/iommu/amd_iommu.c +0 -2
-	spin_unlock_irqrestore(&domain->lock, flags);
-
	return ret;
}
drivers/iommu/amd_iommu.c +5 -0
@@ -1499,14 +1494,19 @@ static u64 *alloc_pte(struct protection_domain *domain,
 		      gfp_t gfp,
 		      bool *updated)
 {
+	unsigned long flags;
 	int level, end_lvl;
 	u64 *pte, *page;
 
 	BUG_ON(!is_power_of_2(page_size));
 
+	spin_lock_irqsave(&domain->lock, flags);
+
 	while (address > PM_LEVEL_SIZE(domain->mode))
 		*updated = increase_address_space(domain, gfp) || *updated;
 
+	spin_unlock_irqrestore(&domain->lock, flags);
+
 	level   = domain->mode - 1;
 	pte     = &domain->pt_root[PM_LEVEL_INDEX(level, address)];
 	address = PAGE_SIZE_ALIGN(address, page_size);
Signature · 6 lines
1.8.3.1

_______________________________________________
iommu mailing list
[email protected]
https://lists.linuxfoundation.org/mailman/listinfo/iommu

_______________________________________________
iommu mailing list
[email protected]
https://lists.linuxfoundation.org/mailman/listinfo/iommu

#2 lore raw
Quoted 5 lines · Jerry Snitselaar
Quoted 4 lines · Qian Cai

BTW, Joerg, this line from the commit "iommu/amd: Remove domain->updated" looks suspicious. Not sure what the purpose of it.

Quoted 9 lines
Quoted 2 lines

*updated = increase_address_space(domain, gfp) || *updated;

Looking at it again I think that isn't an issue really, it would just not lose updated being set in a previous loop iteration, but now I'm wondering about the loop itself. In the cases where it would return false, how does the evaluation of the condition for the while loop change?

I guess the mode level 6 check is really for other potential callers increase_address_space, none exist at the moment, and the condition of the while loop in alloc_pte should fail if the mode level is 6.

author
#3 lore raw
Quoted 7 lines · Jerry Snitselaar

I guess the mode level 6 check is really for other potential callers increase_address_space, none exist at the moment, and the condition of the while loop in alloc_pte should fail if the mode level is 6.

Because there is no locking around iommu_map_page(), if there are several concurrent callers of it for the same domain, could it be that it silently corrupt data due to invalid access?
_______________________________________________
iommu mailing list
[email protected]
https://lists.linuxfoundation.org/mailman/listinfo/iommu

#4 lore raw

PATCH 1/1 iommu/amd: Check PM_LEVEL_SIZE() condition in locked section

Quoted 11 lines · Qian Cai
Quoted 7 lines · Jerry Snitselaar

I guess the mode level 6 check is really for other potential callers increase_address_space, none exist at the moment, and the condition of the while loop in alloc_pte should fail if the mode level is 6.

Because there is no locking around iommu_map_page(), if there are several concurrent callers of it for the same domain, could it be that it silently corrupt data due to invalid access?

No, that can't happen because increase_address_space locks the domain before actually doing anything. So the address space can't grow above domain->mode == 6. But what can happen is that the WARN_ON_ONCE triggers in there and that the address space is increased multiple times when only one increase would be sufficient.

To fix this we just need to check the PM_LEVEL_SIZE() condition again when we hold the lock:

From e930e792a998e89dfd4feef15fbbf289c45124dc Mon Sep 17 00:00:00 2001
From: Joerg Roedel <[email protected]>
Date: Fri, 18 Oct 2019 11:34:22 +0200
Subject: [PATCH] iommu/amd: Check PM_LEVEL_SIZE() condition in locked section

The increase_address_space() function has to check the PM_LEVEL_SIZE() condition again under the domain->lock to avoid a false trigger of the WARN_ON_ONCE() and to avoid that the address space is increase more often than necessary.

Fixes 754265bcab78 ("iommu/amd: Fix race in increase_address_space()")
---
 drivers/iommu/amd_iommu.c | 7 ++++---
 1 file changed, 4 insertions(+), 3 deletions(-)
drivers/iommu/amd_iommu.c +4 -3
index 2369b8af81f3..a0639e511ffe 100644
@@ -1463,6 +1463,7 @@ static void free_pagetable(struct protection_domain *domain)
  * to 64 bits.
  */
 static bool increase_address_space(struct protection_domain *domain,
+				   unsigned long address,
 				   gfp_t gfp)
 {
 	unsigned long flags;
@@ -1471,8 +1472,8 @@ static bool increase_address_space(struct protection_domain *domain,
 
 	spin_lock_irqsave(&domain->lock, flags);
 
-	if (WARN_ON_ONCE(domain->mode == PAGE_MODE_6_LEVEL))
-		/* address space already 64 bit large */
+	if (address <= PM_LEVEL_SIZE(domain->mode) ||
+	    WARN_ON_ONCE(domain->mode == PAGE_MODE_6_LEVEL))
 		goto out;
 
 	pte = (void *)get_zeroed_page(gfp);
@@ -1505,7 +1506,7 @@ static u64 *alloc_pte(struct protection_domain *domain,
 	BUG_ON(!is_power_of_2(page_size));
 
 	while (address > PM_LEVEL_SIZE(domain->mode))
-		*updated = increase_address_space(domain, gfp) || *updated;
+		*updated = increase_address_space(domain, address, gfp) || *updated;
 
 	level   = domain->mode - 1;
 	pte     = &domain->pt_root[PM_LEVEL_INDEX(level, address)];
Signature · 1 line
2.16.4
Reviewed-by on v1 1/1
#5 lore raw

Re: [PATCH] iommu/amd: Check PM_LEVEL_SIZE() condition in locked section

Quoted 74 lines · Joerg Roedel
Quoted 11 lines · Qian Cai
Quoted 7 lines · Jerry Snitselaar

I guess the mode level 6 check is really for other potential callers increase_address_space, none exist at the moment, and the condition of the while loop in alloc_pte should fail if the mode level is 6.

Because there is no locking around iommu_map_page(), if there are several concurrent callers of it for the same domain, could it be that it silently corrupt data due to invalid access?

No, that can't happen because increase_address_space locks the domain before actually doing anything. So the address space can't grow above domain->mode == 6. But what can happen is that the WARN_ON_ONCE triggers in there and that the address space is increased multiple times when only one increase would be sufficient.

To fix this we just need to check the PM_LEVEL_SIZE() condition again when we hold the lock:

From e930e792a998e89dfd4feef15fbbf289c45124dc Mon Sep 17 00:00:00 2001
From: Joerg Roedel <[email protected]>
Date: Fri, 18 Oct 2019 11:34:22 +0200
Subject: [PATCH] iommu/amd: Check PM_LEVEL_SIZE() condition in locked section

The increase_address_space() function has to check the PM_LEVEL_SIZE() condition again under the domain->lock to avoid a false trigger of the WARN_ON_ONCE() and to avoid that the address space is increase more often than necessary.

Fixes 754265bcab78 ("iommu/amd: Fix race in increase_address_space()")

---
drivers/iommu/amd_iommu.c | 7 ++++---

1 file changed, 4 insertions(+), 3 deletions(-)
drivers/iommu/amd_iommu.c +4 -3
index 2369b8af81f3..a0639e511ffe 100644
@@ -1463,6 +1463,7 @@ static void free_pagetable(struct protection_domain *domain)
 * to 64 bits.
 */
 static bool increase_address_space(struct protection_domain *domain,
+				   unsigned long address,
 				   gfp_t gfp)
 {
 	unsigned long flags;
@@ -1471,8 +1472,8 @@ static bool increase_address_space(struct protection_domain *domain,
 
 	spin_lock_irqsave(&domain->lock, flags);
 
-	if (WARN_ON_ONCE(domain->mode == PAGE_MODE_6_LEVEL))
-		/* address space already 64 bit large */
+	if (address <= PM_LEVEL_SIZE(domain->mode) ||
+	    WARN_ON_ONCE(domain->mode == PAGE_MODE_6_LEVEL))
 		goto out;
 
 	pte = (void *)get_zeroed_page(gfp);
@@ -1505,7 +1506,7 @@ static u64 *alloc_pte(struct protection_domain *domain,
 	BUG_ON(!is_power_of_2(page_size));
 
 	while (address > PM_LEVEL_SIZE(domain->mode))
-		*updated = increase_address_space(domain, gfp) || *updated;
+		*updated = increase_address_space(domain, address, gfp) || *updated;
 
 	level   = domain->mode - 1;
 	pte     = &domain->pt_root[PM_LEVEL_INDEX(level, address)];
Signature · 1 line
2.16.4