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

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [day] [month] [year] [list]

Message-ID: <20191018144805.ici3ewsvonlgketl@cantor>
Date:   Fri, 18 Oct 2019 07:48:05 -0700
From:   Jerry Snitselaar <jsnitsel@...hat.com>
To:     Joerg Roedel <jroedel@...e.de>
Cc:     Qian Cai <cai@....pw>, don.brace@...rosemi.com,
        martin.petersen@...cle.com, linux-scsi@...r.kernel.org,
        jejb@...ux.ibm.com, esc.storagedev@...rosemi.com,
        linux-kernel@...r.kernel.org, iommu@...ts.linux-foundation.org
Subject: Re: [PATCH] iommu/amd: Check PM_LEVEL_SIZE() condition in locked
 section

On Fri Oct 18 19, Joerg Roedel wrote:
>On Thu, Oct 17, 2019 at 07:36:51AM -0400, Qian Cai wrote:
>>
>>
>> > On Oct 16, 2019, at 6:59 PM, Jerry Snitselaar <jsnitsel@...hat.com> wrote:
>> >
>> > 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 <jroedel@...e.de>
>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.
>
>Reported-by: Qian Cai <cai@....pw>
>Fixes: 754265bcab78 ("iommu/amd: Fix race in increase_address_space()")
>Signed-off-by: Joerg Roedel <jroedel@...e.de>
>---
> drivers/iommu/amd_iommu.c | 7 ++++---
> 1 file changed, 4 insertions(+), 3 deletions(-)
>
>diff --git a/drivers/iommu/amd_iommu.c b/drivers/iommu/amd_iommu.c
>index 2369b8af81f3..a0639e511ffe 100644
>--- a/drivers/iommu/amd_iommu.c
>+++ b/drivers/iommu/amd_iommu.c
>@@ -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)];
>-- 
>2.16.4
>

Reviewed-by: Jerry Snitselaar <jsnitsel@...hat.com>