344000 - dom cellmap crash

Status: RESOLVED FIXED

(Core :: Layout: Tables, defect)

Description

[Bernd]

the testcase tries to access foo[-1] or foo[0xffffffff] marking sensitive.

Comment 1

[Bernd]

Attachment: testcase

Comment 2

[Bernd]

Attachment: patch

Comment 3

[Daniel Veditz [:dveditz]]

Affects the 1.8 branches also, nominating.

Why do these routines use signed indexes? Could they be changed to use unsigned ints instead?

Comment 4

[Daniel Veditz [:dveditz]]

Comment on attachment 228602 [details] [diff] [review]
patch

>-
>+  NS_ASSERTION(aFirstRowIndex >= 0, "nsCellMap::InsertRows called with negative rowIndex");

If you don't change the API's to use unsigned should we do more than assert? Add the assertion to other routines as well?

>   PRInt32 numRows = mRows.Count();
>   PRInt32 numCols = aMap.GetColCount();
> 
>   if (aFirstRowIndex >= numRows) {
>+    mRowCount -= aNumRowsToRemove;
>     return;

seems dodgy here, can mRowCount be made negative this way? Or has something else already incremented mRowCount by exactly the amount aFirstRowIndex is larger than numRows.

What if aFirstRowIndex is ever negative here? We didn't expect it could happen in the other place either.

>     if (aEndRowIndex < mRowCount - 1) {
>       cellData = GetDataAt(aMap, aEndRowIndex + 1, colX, PR_TRUE);
>       if ((cellData) && (cellData->IsRowSpan())) {
>         return PR_TRUE; // there is a row span out of the region
>       }
>     }
>+    else {
>+      // this cell might be the cause of a dead row which should be removed
>+      return PR_TRUE;
>+    }

"might be"? what happens if it isn't?

Comment 5

[Bernd]

>Why do these routines use signed indexes?
Because Chris wrote them. -1 is very common as a hidden boolean parameter to toggle function behaviour.

I haven't really analyzed the code what would happen with unsigned indices. I am currently more in panic mode trying to get the whole thing survive bug 339128. The upcoming patch does exactly this at least on my machine.

>seems dodgy here, can mRowCount be made negative this way? Or has something
>else already incremented mRowCount by exactly the amount aFirstRowIndex is
>larger than numRows.
Every row that we have increases mRowCount. Every row that we remove has to reduce it. We have a scenario where the latter does not happen without the patch.

>What if aFirstRowIndex is ever negative here? We didn't expect it could happen
>in the other place either.

If the routine returns PR_TRUE we will call a more expensive routine thats not a issue. 

Comment 6

[Bernd]

Attachment: patch

Comment 7

[Boris Zbarsky [:bzbarsky]]

I'd really like an explanation of what the various parts of the patch do to be able to review this...  For example, if aEndRowIndex >= mRowCount - 1, why is it ok to call GetDataAt(aMap, aEndRowIndex, colX, PR_TRUE)?  Why is the numOrigRows check needed?   That sort of thing.

Comment 8

[Bernd]

Attachment: revised patch

Comment 9

[Boris Zbarsky [:bzbarsky]]

Comment on attachment 231583 [details] [diff] [review]
revised patch

>--- nsCellMap.cpp	2006-08-01 16:26:40.296875000 +0200
>+++ nsCellMap.local	2006-08-01 16:33:42.140625000 +0200
>@@ -1498,30 +1502,51 @@ PRBool nsCellMap::CellsSpanOut(nsVoidArr
>+                                   // current cellmap extension.

s/extension/extent/ ?

>@@ -2023,18 +2048,23 @@ nsCellMap::RebuildConsideringRows(nsTabl

>+  PRInt32 copyEndRowIndex = PR_MIN(numOrigRows, aStartRowIndex);
>+
>   // put back the rows before the affected ones just as before
>-  for (rowX = 0; rowX < aStartRowIndex; rowX++) {
>+  // aStartRowIndex might be after all existing rows so we should limit the
>+  // copy to the amount of exisiting rows

Wouldn't this comment make more sense before you declare copyEndRowIndex and do the PR_MIN thing?

With that, looks ok.  Thanks for the comments!

Comment 10

[Bernd]

fixed on trunk

Comment 11

[Mike Schroepfer]

Is this patch suitable for 1.8 branch?

Comment 12

[Bernd]

fixed on branch by the patch in bug 346980

Comment 13

[Bernd]

the only bug that should block 1.8.0.7 is bug 346980 which is blocked by the
regression 347725

asking for 1.8.0.7- on this

Comment 14

[Bob Clary [:bc] (inactive)]

https://bugzilla.mozilla.org/attachment.cgi?id=228578&action=view
ff2b2 debug windows/linux no crash, no assert, nightly windows/linux no crash
verified fixed 1.8

Comment 15

[Daniel Veditz [:dveditz]]

Moving to 1.8.0.8 at Bernd's request (regression worries)

Comment 16

[Bernd]

fixed on branches by the cellmap branch patch

Comment 17

[alice nodelman [:alice] [:anode]]

Can I get some clarification on this bug's status?  It seems to have been minus-ed for 1807, but it is marked 'fixed1807'.  Has the patch been checked in?  Has this issue been fixed by another patch?  Or are we planning on checking in this patch to the 1.8.0 branch later, even though this particular crash has been fixed?

Comment 18

[Daniel Veditz [:dveditz]]

This did land in 1.8.0.7 after all. It wasn't "blocking", but was still approved.

Comment 19

[alice nodelman [:alice] [:anode]]

https://bugzilla.mozilla.org/attachment.cgi?id=228578&action=view should not crash

Mozilla/5.0 (Macintosh; U; Intel Mac OS X; en-US; rv:1.8.0.7) Gecko/20060906 Firefox/1.5.0.7

verified 1.8.0.7

Comment 20

[Bob Clary [:bc] (inactive)]

crash test landed
http://hg.mozilla.org/mozilla-central/rev/5561696a146d