Closed Bug 848644 (CVE-2013-0787) Opened 12 years ago Closed 12 years ago

Use-after-free caused by us replacing the nsHTMLEditor's edit rules object while running scripts from the flush happening in nsEditor::IsPreformatted triggered by execCommand from web content

Categories

(Core :: DOM: Editor, defect)

defect
Not set
critical

Tracking

()

VERIFIED FIXED
mozilla22
blocking-b2g tef+
Tracking Status
firefox19 + verified
firefox20 + verified
firefox21 + verified
firefox22 + verified
firefox-esr17 19+ verified
b2g18 + fixed
b2g18-v1.0.0 --- wontfix
b2g18-v1.0.1 --- fixed

People

(Reporter: reed, Assigned: ehsan.akhgari)

References

Details

(6 keywords, Whiteboard: Pwn2Own 2013 bug)

Crash Data

Attachments

(6 files)

Placeholder for Pwn2Own bug found in Firefox by VUPEN at CanSecWest 2013.
document.designMode = "on"; ... document.execCommand("insertOrderedList",false,true); Looks like a designMode / contentEditable / editor bug at first glance? --> preemptive cc of ehsan & roc. Some stuff with form <inputs>, but it very very vaguely looks like that's related to the heap spraying and not the actual flaw?
Here's an ASan debug trace with symbols, taken from m-c rev 216ec69cc531.
Memory freed in nsHTMLInputElement::SetValueInternal then used in nsHTMLEditRules::GetPromotedPoint.
Hmm. So the PoC crashes on Firefox 18 for me but not in nightlies...
OK, given comment 3 that last is just luck presumably.
Assignee: nobody → ehsan
Does not crash: 20120101 Win32 nightly on Win7 Crashes: 20130101 Win32 nightly on Win7 Still narrowing this down...
Does not crash: 20121005 Win32 nightly on Win7 Crashes: 20121006 Win32 nightly on Win7 http://hg.mozilla.org/mozilla-central/pushloghtml?fromchange=fd724f194a1f&tochange=2da1f2bde40e (I *think* this is somewhat correct)
Taking a guess from the regression range - it might be Core: Editor bug 796839.
(In reply to Christian Holler (:decoder) from comment #4) > Memory freed in nsHTMLInputElement::SetValueInternal then used in > nsHTMLEditRules::GetPromotedPoint. In this changeset from bug 796839, https://hg.mozilla.org/mozilla-central/rev/d5abea9273a9 the file editor/libeditor/html/nsHTMLEditRules.cpp is changed. And nsHTMLEditRules::GetPromotedPoint from comment 4 is found in editor/libeditor/html/nsHTMLEditRules.cpp
(In reply to Christian Holler (:decoder) from comment #4) > Memory freed in nsHTMLInputElement::SetValueInternal then used in > nsHTMLEditRules::GetPromotedPoint. > Taking a guess from the regression range - it might be Core: Editor bug > 796839. Or it could be bug 795610, from that regression window. nsHTMLInputElement::SetValueInternal is defined in nsHTMLInputElement.cpp which was modified in bug 795610.
Setting Matt Wobensmith as QA contact for this bug. Please advise how we can help.
QA Contact: mwobensmith
(In reply to Gary Kwong [:gkw] [:nth10sd] from comment #12) > (In reply to Christian Holler (:decoder) from comment #4) > > Memory freed in nsHTMLInputElement::SetValueInternal then used in > > nsHTMLEditRules::GetPromotedPoint. > > > Taking a guess from the regression range - it might be Core: Editor bug > > 796839. > > Or it could be bug 795610, from that regression window. > > nsHTMLInputElement::SetValueInternal is defined in nsHTMLInputElement.cpp > which was modified in bug 795610. I don't think this bug is relevant...
(In reply to Anthony Hughes, Mozilla QA (:ashughes) from comment #13) > Setting Matt Wobensmith as QA contact for this bug. Please advise how we can > help. Please confirm that: http://ftp.mozilla.org/pub/mozilla.org/firefox/nightly/2012/10/2012-10-05-03-06-09-mozilla-central/firefox-18.0a1.en-US.win32.zip does not crash but: http://ftp.mozilla.org/pub/mozilla.org/firefox/nightly/2012/10/2012-10-06-03-05-34-mozilla-central/firefox-18.0a1.en-US.win32.zip does crash, on Win7. To get it to crash, try reloading, or closing the browser window and reloading the PoC again.
I crashed today's Win32 binary on Win7 at: bp-608c66ce-5f5a-4fbc-be88-ead642130307
Crash Signature: [@ nsHTMLEditor::GetPriorHTMLNode(nsINode*, int, bool) ]
This looks like ESR17 is unaffected based on the 10/05 date, which puts 17 on Aurora at that point.
Con(In reply to Gary Kwong [:gkw] [:nth10sd] from comment #15) > Please confirm that: > > /nightly/2012/10/2012-10-05-03-06-09-mozilla-central/firefox-18.0a1.en-US.win32.zip > > does not crash but: > > /nightly/2012/10/2012-10-06-03-05-34-mozilla-central/firefox-18.0a1.en-US.win32.zip > > does crash, on Win7. On Win7 that is what I see. 10-06 crashes but -05 does not.
So the execCommand does some stuff that ends up calling IsPreformatted, which flushes style, which flushes pending resize events (presumably there's a pending resize from the "I blocked your popup" notification), which runs the resize handler, which runs page script while editor code is on the stack. That page script calls document.open() and then document.write(), which reinitialized the editor and gives it a new editrules. But the old editrules is still on the stack, and the only strong ref to it was from the editor, so it dies and now we're working with a deleted object and things are all bad. If I hold a strong ref to the editrules on the stack in frame 49 of this callstack, I end up with a null-deref crash instead of a bogus-memory crash (because the mEditor of the editrules ends up null).
(In reply to Al Billings [:abillings] from comment #18) > This looks like ESR17 is unaffected based on the 10/05 date, which puts 17 > on Aurora at that point. Not true. I crashed ESR 17 nightly Win32 on Win7: bp-719b5a3c-122e-4754-86cc-62daf2130307
One problem here is that nsHTMLDocument::EditingStateChanged tries to reuse the nsHTMLEditor object if one is around previously, but we can't easily change that. Putting a script blocker inside IsPreformatted is also not an option.
> (In reply to Al Billings [:abillings] from comment #18) > > This looks like ESR17 is unaffected based on the 10/05 date, which puts 17 > > on Aurora at that point. > > Not true. I crashed ESR 17 nightly Win32 on Win7: > > bp-719b5a3c-122e-4754-86cc-62daf2130307 Bug 796839 was backported to Aurora. See bug 796839 comment 27. Thus, everything from 17 onwards is likely affected assuming bug 796839 is indeed the regressor.
Putting a script blocker inside ExecCommand might be an option though!
Nominating for all tracking/status flags on all current branches in existence.
> assuming bug 796839 is indeed the regressor. Backing out part 4 of that bug locally seems to fix this particular testcase at first glance, so the assumption that it's the regressor is probably safe.
(In reply to Boris Zbarsky (:bz) from comment #26) > > assuming bug 796839 is indeed the regressor. > > Backing out part 4 of that bug locally seems to fix this particular testcase > at first glance, so the assumption that it's the regressor is probably safe. Adding a bunch of keywords, and setting Trunk as the latest branch that this bug affects. Not setting this to block bug 796839 until this bug is public, but feel free to set it if it doesn't matter.
Version: unspecified → Trunk
Given the speed of this release it's not going to matter.
So a possible brute-force approach here is: 1) Make sure to always hold on-stack strong refs to the edit rules when calling into it. Most callers already do. I've just audited the rest and will post a patch to fix them all. That converts the crash from a "virtual call on deleted memory" to a "virtual call on null", which is a slight improvement. 2) Add null-checks on mEditor in the edit rules as needed.
Assignee: ehsan → bzbarsky
(In reply to :Ehsan Akhgari from comment #24) > Putting a script blocker inside ExecCommand might be an option though! Except that doing that breaks _tons_ of tests :(
(In reply to Boris Zbarsky (:bz) from comment #29) > So a possible brute-force approach here is: > > 1) Make sure to always hold on-stack strong refs to the edit rules when > calling into it. Most callers already do. I've just audited the rest and > will post a patch to fix them all. That converts the crash from a "virtual > call on deleted memory" to a "virtual call on null", which is a slight > improvement. > > 2) Add null-checks on mEditor in the edit rules as needed. ehsan and the stupid fix would not be sufficient ehsan because the edit rules object is not the only thing that can get messed up ehsan if we hold a weak ref to a content node and the script causes that node to get deleted, then we lose ehsan and I'm sure we do a ton of those types of weak refs in the editor code :(
Component: Security → Editor
I think this is a better fix.
Assignee: bzbarsky → ehsan
Status: NEW → ASSIGNED
Attachment #722053 - Flags: review?(bzbarsky)
Comment on attachment 722053 [details] [diff] [review] Do not flush in IsPreformatted r=me
Attachment #722053 - Flags: review?(bzbarsky) → review+
Comment on attachment 722053 [details] [diff] [review] Do not flush in IsPreformatted I think this is the safest option that we currently have at hand.
Attachment #722053 - Flags: review?(bzbarsky)
Attachment #722053 - Flags: review+
Attachment #722053 - Flags: approval-mozilla-release?
Attachment #722053 - Flags: approval-mozilla-esr17?
Attachment #722053 - Flags: approval-mozilla-beta?
Attachment #722053 - Flags: approval-mozilla-b2g18?
Attachment #722053 - Flags: approval-mozilla-aurora?
Comment on attachment 722053 [details] [diff] [review] Do not flush in IsPreformatted That's r=me if it doesn't break anything, of course, as discussed. ;)
Attachment #722053 - Flags: review?(bzbarsky) → review+
Comment on attachment 722053 [details] [diff] [review] Do not flush in IsPreformatted a=dveditz
Attachment #722053 - Flags: sec-approval+
Attachment #722053 - Flags: approval-mozilla-release?
Attachment #722053 - Flags: approval-mozilla-release+
Attachment #722053 - Flags: approval-mozilla-esr17?
Attachment #722053 - Flags: approval-mozilla-esr17+
Attachment #722053 - Flags: approval-mozilla-beta?
Attachment #722053 - Flags: approval-mozilla-beta+
Attachment #722053 - Flags: approval-mozilla-b2g18?
Attachment #722053 - Flags: approval-mozilla-b2g18+
Attachment #722053 - Flags: approval-mozilla-aurora?
Attachment #722053 - Flags: approval-mozilla-aurora+
Can we statically analyze for functions and methods that must not flush, and annotate them to enforce the rule? /be
(In reply to :Ehsan Akhgari from comment #38) > https://hg.mozilla.org/integration/mozilla-inbound/rev/cddf5f75843f Please advise QA on where the earliest available build to test this patch is (try build, m-i on-change, etc). What areas should we test to ensure there are no regressions? This looks like a very simple fix, but I want to be sure we cover our bases.
It's been suggested in the past, just never happened. More precisely the suggestion was to have a way to tell whether a function can flush via static analysis...
(In reply to Boris Zbarsky (:bz) from comment #41) > It's been suggested in the past, just never happened. More precisely the > suggestion was to have a way to tell whether a function can flush via static > analysis... Perhaps this would be a(nother) reason for us to get serious about doing that! ;-)
Ryan, I don't have b2g18* checkouts locally. Would you mind doing the honors on those branches? Thanks!
Keywords: checkin-needed
(In reply to Alex Keybl [:akeybl] from comment #40) > (In reply to :Ehsan Akhgari from comment #38) > > https://hg.mozilla.org/integration/mozilla-inbound/rev/cddf5f75843f > > Please advise QA on where the earliest available build to test this patch is > (try build, m-i on-change, etc). This has landed on inbound <https://tbpl.mozilla.org/?tree=Mozilla-Inbound&rev=cddf5f75843f>, aurora, beta, release and esr17. I suggest focusing on testing this on the release branch. I expect builds to be available starting from 1-2 hours from now, and between 3-4 hours from now on Windows. Please watch <https://tbpl.mozilla.org/?tree=Mozilla-Release&rev=0c7a6bfdfd8c> and look for green B icons to know when the builds are done for each platform. I strongly suggest testing this on Windows with the exploit test case to make sure that we've covered all cases. The expected result on the exploit is no crash (and obviously, no exploit either -- not sure what the exploit shell code actually does; presumably it runs calc.exe or notepad.exe or something?) > What areas should we test to ensure there are no regressions? This looks > like a very simple fix, but I want to be sure we cover our bases. Testing different richtext frameworks would be a good idea. ckeditor, aloha, richtext mediawiki editor, etherpad, gmail, wordpress, etc. That said, we do have a pretty good test coverage on the code path in question so I'm a bit less worried than I would be otherwise, but still we should look for regressions in richtext editors. A regression would show up as an editing command accessed through either a keyboard shortcut (such as Ctrl+B, etc.) or an editing toolbar icon to not work, or stop working after editing a bit. Please track the steps you're taking in testing carefully though, the editor is buggy as hell, so don't be surprised if you find bugs _not_ caused by this patch. Recording the screen while testing is probably a good idea...
To be perfectly clear, it _is_ possible that this patch breaks web content in ways that our automated and manual tests do not cover. This patch is the best of all evil options that we have for the release channel, and I will be on the hook to address any possible regressions quickly.
Summary: ZDI CanSecWest 2013 bug (use-after-free) → Use-after-free caused by us replacing the nsHTMLEditor's edit rules object while running scripts from the flush happening in nsEditor::IsPreformatted triggered by execCommand from web content
(In reply to :Ehsan Akhgari from comment #44) > Ryan, I don't have b2g18* checkouts locally. Would you mind doing the > honors on those branches? Thanks! No urgency here - our partners are still picking up changes. But thanks for calling this out :)
blocking-b2g: --- → tef+
There's going to be a couple of things we do before spinning up a 19.0.2 candidate: 1) Wait for green across all platforms for at least one build 2) Test to make sure that a build with the patch resolves the exploitable crash 3) Decide how much regression testing is required (QA, pre-release audiences, etc.), based upon the risk of regression - comment 46 directs QA testing very well, with that kind of targeted testing I think we can skip pre-release audience testing 4) Obviously land the patches to mozilla-release - this is done in comment 42 In parallel, we should figure out how much further investigation is necessary to ensure there are no similarly exploitable paths (especially obvious ones), so that we can still take further change for a second candidate as necessary. - Spoke with both bz/ehsan about this. They've done some code analysis tonight, but believe that Brendan's comment 39 would be the longer term investigation.
(In reply to Alex Keybl [:akeybl] from comment #49) > There's going to be a couple of things we do before spinning up a 19.0.2 > candidate: > > 1) Wait for green across all platforms for at least one build > 2) Test to make sure that a build with the patch resolves the exploitable > crash > 3) Decide how much regression testing is required (QA, pre-release > audiences, etc.), based upon the risk of regression > - comment 46 directs QA testing very well, with that kind of targeted > testing I think we can skip pre-release audience testing > 4) Obviously land the patches to mozilla-release > - this is done in comment 42 Yeah, FWIW I don't expect that the pre-release audience can find any regressions that our targeted QA cannot, so it makes sense to me not to block on that. > In parallel, we should figure out how much further investigation is > necessary to ensure there are no similarly exploitable paths (especially > obvious ones), so that we can still take further change for a second > candidate as necessary. > - Spoke with both bz/ehsan about this. They've done some code analysis > tonight, but believe that Brendan's comment 39 would be the longer term > investigation. Amen!
Alias: CVE-2013-0787
Boris, can you please file another bug and attach your patch there? I would still like to take that patch but I need to review it carefully...
(In reply to :Ehsan Akhgari from comment #46) > not sure what the exploit shell code actually does; presumably > it runs calc.exe or notepad.exe or something?) ZDI did not give us a working exploit, this PoC just crashes.
I forgot to mention here that I ran all of the editor tests locally and they all passed with this patch.
(In reply to Brendan Eich [:brendan] from comment #39) > Can we statically analyze for functions and methods that must not flush, and > annotate them to enforce the rule? Something like bug 477432 or bug 542364? If a static analysis is hard, I'd rather have assertions than nothing.
We're going to end up spinning a 19.0.2 for both desktop and mobile, but Lukas and I weren't able to repro with the PoC on Android. Is there any reason to believe this Fennec will be any less impacted? Our decision to actually release the mobile build candidate will depend on that answer.
Tested this on https://ftp.mozilla.org/pub/mozilla.org/firefox/tinderbox-builds/mozilla-release-macosx64/ and was not able to reproduce the crash so will go ahead with kicking off the 19.0.2 builds from the changset with Ehsan's patch applied. Fwiw, I tested on the current Aurora* and was not able to crash there either and this PoC was not showing the same thing as it does in Firefox 19.0.1 (or the patched tinderbox build) and it also does not crash - not sure what this means, but thought it worth mentioning. * Built from http://hg.mozilla.org/releases/mozilla-aurora/rev/357d21b20bf0
(In reply to Alex Keybl [:akeybl] from comment #55) > We're going to end up spinning a 19.0.2 for both desktop and mobile, but > Lukas and I weren't able to repro with the PoC on Android. > Confirming neither here; builds tested release → central (ARMv7/6)
I forgot to mention that one trick to reproduce this with the attached PoC is to have a single tab open when running the test. Basically, the popup block infobar that changes the size of the viewport as the test is running is crucial to be able to reproduce this.
Comment on attachment 722053 [details] [diff] [review] Do not flush in IsPreformatted Verified the fix on Linux using two ASan builds. The version with the patch shows no use-after-free occurring anymore.
Attachment #722053 - Flags: feedback+
> Is there any reason to believe this Fennec will be any less impacted? For the underlying bug, no. As for PoC testcase, it depends on window.open() triggering a resize of the content area of the web page it's called from. This is true in desktop firefox but I'd be it's not on Android. So a different vector would be needed to get the flush to trigger the script execution under editor code there.
(In reply to :Ehsan Akhgari from comment #42) > https://hg.mozilla.org/releases/mozilla-esr17/rev/fb56ba35db3b Just realized - this needs to land to GECKO1703_2013021512_RELBRANCH as well, so that 17.0.4 doesn't pick up any of the Firefox 20 security fixes.
> Just realized - this needs to land to GECKO1703_2013021512_RELBRANCH as well https://hg.mozilla.org/releases/mozilla-esr17/rev/19c96497b96a
(In reply to :Ehsan Akhgari from comment #44) > Ryan, I don't have b2g18* checkouts locally. Would you mind doing the > honors on those branches? Thanks! Not a problem, of course. I'll get it with my other uplifts today. :)
(In reply to Jesse Ruderman from comment #54) > If a static analysis is hard, I'd rather have assertions than nothing. Assertions will not be a very strong defense, since they require us to hit bad code paths like this to trigger them.
And for Thunderbird, the GECKO1703_2013021513_RELBRANCH branch: https://hg.mozilla.org/releases/mozilla-esr17/rev/2c3766cfba66
This crashes my Firefox on Android.
Attachment #722231 - Attachment mime type: text/html → text/plain
Mozilla/5.0 (X11; Linux i686; rv:19.0) Gecko/20100101 Firefox/19.0 Build ID: 20130307023931 Firefox 19.0.2/Ubuntu 12.10 doesn't crash when loading PoC test file.
Verified the crash for Firefox 19.0 (20130215130331) and 19.0.1 (20130226172142) by using the attachment from comment 1. No crashes for Firefox 19.0.2 (20130307023931) after loading PoC test file. Verified on: Mac OS X 10.8 Mozilla/5.0 (Macintosh; Intel Mac OS X 10.8; rv:19.0) Gecko/20100101 Firefox/19.0 Windows 7 x64 Mozilla/5.0 (Windows NT 6.1; WOW64; rv:19.0) Gecko/20100101 Firefox/19.0 Windows 8 x86 Mozilla/5.0 (Windows NT 6.2; rv:19.0) Gecko/20100101 Firefox/19.0
I filed bug 848788 for a mechanism to only block script execution. If we had such an infrastructure, fixing this and similar bugs would be a lot easier and safer.
Attachment #722231 - Attachment mime type: text/plain → text/html
Based on comments in this bug and the results of testing seen here[1] I'm calling this verified fixed for Firefox 19. Please advise if there is something more we need to be checking before signing off Firefox 19.0.2 for release. 1. https://etherpad.mozilla.org/19-0-2
Patch attachment #722053 [details] [diff] [review] in mozilla-release 19.0.2 works for me on Android (ARMv7/ARMv6); likewise verified fixed on firefox19.
Status: ASSIGNED → RESOLVED
Closed: 12 years ago
Flags: in-testsuite?
Resolution: --- → FIXED
Target Milestone: --- → mozilla22
Confirmed reproducible with Firefox 20.0b3 on Linux. Verified fixed with Firefox 20.0b4#1 on Linux.
Why was status-b2g18-v1.0.0 marked as wontfix?
(In reply to :Ehsan Akhgari from comment #80) > Why was status-b2g18-v1.0.0 marked as wontfix? IIUC, that branch isn't being used any more.
(In reply to Andrew McCreight [:mccr8] from comment #81) > (In reply to :Ehsan Akhgari from comment #80) > > Why was status-b2g18-v1.0.0 marked as wontfix? > > IIUC, that branch isn't being used any more. That's correct.
(In reply to Anthony Hughes, Mozilla QA (:ashughes) from comment #75) > Based on comments in this bug and the results of testing seen here[1] I'm > calling this verified fixed for Firefox 19. Please advise if there is > something more we need to be checking before signing off Firefox 19.0.2 for > release. > > 1. https://etherpad.mozilla.org/19-0-2 Spoke with Anthony about this on IRC. Given the verification testing at this etherpad, we can consider this bug verified and continue testing for regressions as we find the time.
Comment on attachment 722039 [details] [diff] possible fix part 1. Always hold an on-stack strong reference to mRules when calling into it. > Boris, can you please file another bug and attach your patch there? Bug 848895.
Confirmed reproducible with Firefox Aurora 21.0a2 2013-03-06. Verified fixed with Firefox Aurora 21.0a2 2013-03-07.
I can still reproduce this crash with Firefox Nightly 22.0a2 2013-03-07. Did this land in time for today's Nightlies? Is there another 22.0a2 build I can use for verification?
> Did this land in time for today's Nightlies? No. > Is there another 22.0a2 build I can use for verification? The builds at http://ftp.mozilla.org/pub/mozilla.org/firefox/tinderbox-builds/mozilla-central-win32/1362676792/ should have this patch, for Windows.
(In reply to Anthony Hughes, Mozilla QA (:ashughes) from comment #86) > I can still reproduce this crash with Firefox Nightly 22.0a2 2013-03-07. Did > this land in time for today's Nightlies? Is there another 22.0a2 build I can > use for verification? The patch was merged to central just this morning, so that is expected. You should not be able to reproduce on recent central/inbound tinderbox builds. This will be in tonight's Nightly.
Verified for 20.b4 candidates. Tried earlier betas, which crashed with the testcase. These didn't: Mozilla/5.0 (Windows NT 6.1; WOW64; rv:20.0) Gecko/20100101 Firefox/20.0 Mozilla/5.0 (X11; Linux i686; rv:20.0) Gecko/20100101 Firefox/20.0 Mozilla/5.0 (Macintosh; Intel Mac OS X 10.7; rv:20.0) Gecko/20100101 Firefox/20.0
Confirming reproducible with Firefox Beta for Android (mozilla-beta) 20.0 Beta 3 (ARMv7/ARMv6) Verified Fixed with Firefox Beta for Android (mozilla-beta) 20.0 Beta 4 (ARMv7/ARMv6)
FF17 ESR verified fixed using today's nightly build on the following platforms: MacOS 10.8 Win7 32 Win8 64 Ubuntu 12.04 64 Attachment no longer crashes, and a spot check of ~20 top URLs was performed per platform to verify no broken functionality.
Status: RESOLVED → VERIFIED
Group: core-security
Group: core-security
Group: core-security
Flags: in-testsuite? → in-testsuite+
Pushed by eakhgari@mozilla.com: https://hg.mozilla.org/integration/autoland/rev/6cf11160e686 Add a crashtest based on the test case for the bug
You need to log in before you can comment on or make changes to this bug.

Attachment

General

Created:
Updated:
Size: