63681 2011-06-29 19:25:24 -0700 Chromium bug: Compositing on a lost context causes latch deadlocks 2011-06-30 16:47:10 -0700 1 1 1 Unclassified WebKit New Bugs 528+ (Nightly build) Unspecified Unspecified RESOLVED FIXED P2 Normal --- 1 jbates jbates jamesr kbr vangelis webkit.review.bot oldest_to_newest 430106 0 jbates 2011-06-29 19:25:24 -0700 Chromium bug: Compositing on a lost context causes latch deadlocks 430107 1 99211 jbates 2011-06-29 19:26:04 -0700 Created attachment 99211 Patch 430113 2 jbates 2011-06-29 19:28:52 -0700 This bug is triggered by opening two tabs in the same renderer that support lost context recovery. Killing the GPU process would recreate the contexts. Then when the background tab is clicked, the first composite occurs on the old compositor context, thereby latching on a lost context and leading to deadlock. 430121 3 99211 jamesr 2011-06-29 19:49:39 -0700 Comment on attachment 99211 Patch isCompositorContextLost() is not a synchronous operation, so it seems like this will still be racy if the context is lost and the renderer side hasn't noticed yet - no? 430482 4 jbates 2011-06-30 09:44:16 -0700 (In reply to comment #3) > (From update of attachment 99211 [details]) > isCompositorContextLost() is not a synchronous operation, so it seems like this will still be racy if the context is lost and the renderer side hasn't noticed yet - no? This change fixes a condition that was actually not a race. The background tab has already lost its contexts (at GPU process kill time), and later on, it is displayed. Since the first thing we do is composite, we haven't discovered yet that the context was already lost, and we go ahead latching on it. For the cases where the compositor context is lost after this point, I think it's handled already. Here are the cases: - compositor context is lost due to error (memory, etc): GPU process detects error and kills whole GpuChannel, bringing down all child contexts as well. No deadlock because IPCs don't block on closed channels. - GPU crashed/killed: Also no deadlock because IPCs don't block on closed channels. 430688 5 99211 kbr 2011-06-30 14:43:05 -0700 Comment on attachment 99211 Patch Looks fine as long as it's been tested. I am also concerned about what would happen if the GPU process died immediately after the call to isCompositorContextLost, and also about what happens if we skip doing a render at this point, but it seems this is an improvement. 430749 6 99382 jbates 2011-06-30 15:51:09 -0700 Created attachment 99382 Patch 430774 7 99382 jamesr 2011-06-30 16:06:01 -0700 Comment on attachment 99382 Patch View in context: https://bugs.webkit.org/attachment.cgi?id=99382&action=review > Source/WebKit/chromium/ChangeLog:9 > + The cause of deadlock was that a setLatch command is added to the lost > + compositor GL context, and a waitLatch would never complete on a child context. > + This change checks whether the compositor context is in error state before > + doing the compositing. nit: in the future, put this after the bug link and a newline. pattern is: Reviewed by ... <newline> bug title bug URL <newline> long desc long desc long desc <newline> per-file stuff 430844 8 99382 webkit.review.bot 2011-06-30 16:47:06 -0700 Comment on attachment 99382 Patch Clearing flags on attachment: 99382 Committed r90185: <http://trac.webkit.org/changeset/90185> 430845 9 webkit.review.bot 2011-06-30 16:47:10 -0700 All reviewed patches have been landed. Closing bug. 99211 2011-06-29 19:26:04 -0700 2011-06-30 15:51:06 -0700 Patch bug-63681-20110629192603.patch text/plain 1565 jbates U3VidmVyc2lvbiBSZXZpc2lvbjogOTAwNTYKZGlmZiAtLWdpdCBhL1NvdXJjZS9XZWJLaXQvY2hy b21pdW0vQ2hhbmdlTG9nIGIvU291cmNlL1dlYktpdC9jaHJvbWl1bS9DaGFuZ2VMb2cKaW5kZXgg Zjg3MmEyYjgyYmRmZGUwOTg5OTg4YzVlMjA5MmM2NmJkMTgyZWFiZi4uZTU0ODJlNDM0YTRlNGUy MTZiMzRlODdjM2QxZmU4ZTQ5Nzk4NTdiYSAxMDA2NDQKLS0tIGEvU291cmNlL1dlYktpdC9jaHJv bWl1bS9DaGFuZ2VMb2cKKysrIGIvU291cmNlL1dlYktpdC9jaHJvbWl1bS9DaGFuZ2VMb2cKQEAg LTEsMyArMSwxMyBAQAorMjAxMS0wNi0yOSAgSm9obiBCYXRlcyAgPGpiYXRlc0Bnb29nbGUuY29t PgorCisgICAgICAgIFJldmlld2VkIGJ5IE5PQk9EWSAoT09QUyEpLgorCisgICAgICAgIENocm9t aXVtIGJ1ZzogQ29tcG9zaXRpbmcgb24gYSBsb3N0IGNvbnRleHQgY2F1c2VzIGxhdGNoIGRlYWRs b2NrcworICAgICAgICBodHRwczovL2J1Z3Mud2Via2l0Lm9yZy9zaG93X2J1Zy5jZ2k/aWQ9NjM2 ODEKKworICAgICAgICAqIHNyYy9XZWJWaWV3SW1wbC5jcHA6CisgICAgICAgIChXZWJLaXQ6Oldl YlZpZXdJbXBsOjpjb21wb3NpdGUpOgorCiAyMDExLTA2LTI5ICBKYW1lcyBIYXdraW5zICA8amhh d2tpbnNAY2hyb21pdW0ub3JnPgogCiAgICAgICAgIFJldmlld2VkIGJ5IFBhdmVsIEZlbGRtYW4u CmRpZmYgLS1naXQgYS9Tb3VyY2UvV2ViS2l0L2Nocm9taXVtL3NyYy9XZWJWaWV3SW1wbC5jcHAg Yi9Tb3VyY2UvV2ViS2l0L2Nocm9taXVtL3NyYy9XZWJWaWV3SW1wbC5jcHAKaW5kZXggYjg1MDI3 MDEyYWNkMmM3ZmM1MDE2MmViZmRjYjU1OTZjMzg1Nzc0Ni4uMjRhOGU4NDI3ZWI2NjI4MDM0NzQ0 NGU1OWM4M2NhMmE3ZDY1Mzg1NCAxMDA2NDQKLS0tIGEvU291cmNlL1dlYktpdC9jaHJvbWl1bS9z cmMvV2ViVmlld0ltcGwuY3BwCisrKyBiL1NvdXJjZS9XZWJLaXQvY2hyb21pdW0vc3JjL1dlYlZp ZXdJbXBsLmNwcApAQCAtMTE2NywxMCArMTE2NywxNCBAQCB2b2lkIFdlYlZpZXdJbXBsOjpjb21w b3NpdGUoYm9vbCBmaW5pc2gpCiAgICAgICAgIG1fcmVjcmVhdGluZ0dyYXBoaWNzQ29udGV4dCA9 IGZhbHNlOwogICAgICAgICByZXR1cm47CiAgICAgfQotICAgIGRvQ29tcG9zaXRlKCk7CiAKLSAg ICAvLyBQdXQgcmVzdWx0IG9uc2NyZWVuLgotICAgIG1fbGF5ZXJSZW5kZXJlci0+cHJlc2VudCgp OworICAgIC8vIERvIG5vdCBjb21wb3NpdGUgaWYgdGhlIGNvbXBvc2l0b3IgY29udGV4dCBpcyBh bHJlYWR5IGxvc3QuCisgICAgaWYgKCFtX2xheWVyUmVuZGVyZXItPmlzQ29tcG9zaXRvckNvbnRl eHRMb3N0KCkpIHsKKyAgICAgICAgZG9Db21wb3NpdGUoKTsKKworICAgICAgICAvLyBQdXQgcmVz dWx0IG9uc2NyZWVuLgorICAgICAgICBtX2xheWVyUmVuZGVyZXItPnByZXNlbnQoKTsKKyAgICB9 CiAKICAgICBpZiAobV9sYXllclJlbmRlcmVyLT5pc0NvbXBvc2l0b3JDb250ZXh0TG9zdCgpKSB7 CiAgICAgICAgIC8vIFRyeWluZyB0byByZWNvdmVyIHRoZSBjb250ZXh0IHJpZ2h0IGhlcmUgd2ls bCBub3Qgd29yayBpZiBHUFUgcHJvY2Vzcwo= 99382 2011-06-30 15:51:09 -0700 2011-06-30 16:47:06 -0700 Patch bug-63681-20110630155108.patch text/plain 1851 jbates U3VidmVyc2lvbiBSZXZpc2lvbjogOTAxMzIKZGlmZiAtLWdpdCBhL1NvdXJjZS9XZWJLaXQvY2hy b21pdW0vQ2hhbmdlTG9nIGIvU291cmNlL1dlYktpdC9jaHJvbWl1bS9DaGFuZ2VMb2cKaW5kZXgg ZTcwZGUyNDRlNmU2ZmVhOGUyMzY0OGNkZTc3YjBjMjZjMWViNzQ2MS4uNzlmZWYwOGYxMTQwYjk0 YTAwMmFhNTc5ZDM0NmU0OTJiNTVlMTJiZiAxMDA2NDQKLS0tIGEvU291cmNlL1dlYktpdC9jaHJv bWl1bS9DaGFuZ2VMb2cKKysrIGIvU291cmNlL1dlYktpdC9jaHJvbWl1bS9DaGFuZ2VMb2cKQEAg LTEsMyArMSwxNyBAQAorMjAxMS0wNi0yOSAgSm9obiBCYXRlcyAgPGpiYXRlc0Bnb29nbGUuY29t PgorCisgICAgICAgIFJldmlld2VkIGJ5IE5PQk9EWSAoT09QUyEpLgorCisgICAgICAgIENocm9t aXVtIGJ1ZzogQ29tcG9zaXRpbmcgb24gYSBsb3N0IGNvbnRleHQgY2F1c2VzIGxhdGNoIGRlYWRs b2Nrcy4KKyAgICAgICAgVGhlIGNhdXNlIG9mIGRlYWRsb2NrIHdhcyB0aGF0IGEgc2V0TGF0Y2gg Y29tbWFuZCBpcyBhZGRlZCB0byB0aGUgbG9zdAorICAgICAgICBjb21wb3NpdG9yIEdMIGNvbnRl eHQsIGFuZCBhIHdhaXRMYXRjaCB3b3VsZCBuZXZlciBjb21wbGV0ZSBvbiBhIGNoaWxkIGNvbnRl eHQuCisgICAgICAgIFRoaXMgY2hhbmdlIGNoZWNrcyB3aGV0aGVyIHRoZSBjb21wb3NpdG9yIGNv bnRleHQgaXMgaW4gZXJyb3Igc3RhdGUgYmVmb3JlCisgICAgICAgIGRvaW5nIHRoZSBjb21wb3Np dGluZy4KKyAgICAgICAgaHR0cHM6Ly9idWdzLndlYmtpdC5vcmcvc2hvd19idWcuY2dpP2lkPTYz NjgxCisKKyAgICAgICAgKiBzcmMvV2ViVmlld0ltcGwuY3BwOgorICAgICAgICAoV2ViS2l0OjpX ZWJWaWV3SW1wbDo6Y29tcG9zaXRlKToKKwogMjAxMS0wNi0zMCAgQWxleGFuZGVyIFBhdmxvdiAg PGFwYXZsb3ZAY2hyb21pdW0ub3JnPgogCiAgICAgICAgIFJldmlld2VkIGJ5IEtlbnQgVGFtdXJh LgpkaWZmIC0tZ2l0IGEvU291cmNlL1dlYktpdC9jaHJvbWl1bS9zcmMvV2ViVmlld0ltcGwuY3Bw IGIvU291cmNlL1dlYktpdC9jaHJvbWl1bS9zcmMvV2ViVmlld0ltcGwuY3BwCmluZGV4IGVjN2Ri MTQ3MTU0OTAzMjk3YmU1MjFlODIxMDZmY2RmYmQ4MDBlZjEuLjllY2RlMjkxYmEzZTU4YTJmNDdi MzMyOWNhOWRkZDI1ZjZkYTgyOWIgMTAwNjQ0Ci0tLSBhL1NvdXJjZS9XZWJLaXQvY2hyb21pdW0v c3JjL1dlYlZpZXdJbXBsLmNwcAorKysgYi9Tb3VyY2UvV2ViS2l0L2Nocm9taXVtL3NyYy9XZWJW aWV3SW1wbC5jcHAKQEAgLTExNjcsMTAgKzExNjcsMTQgQEAgdm9pZCBXZWJWaWV3SW1wbDo6Y29t cG9zaXRlKGJvb2wgZmluaXNoKQogICAgICAgICBtX3JlY3JlYXRpbmdHcmFwaGljc0NvbnRleHQg PSBmYWxzZTsKICAgICAgICAgcmV0dXJuOwogICAgIH0KLSAgICBkb0NvbXBvc2l0ZSgpOwogCi0g ICAgLy8gUHV0IHJlc3VsdCBvbnNjcmVlbi4KLSAgICBtX2xheWVyUmVuZGVyZXItPnByZXNlbnQo KTsKKyAgICAvLyBEbyBub3QgY29tcG9zaXRlIGlmIHRoZSBjb21wb3NpdG9yIGNvbnRleHQgaXMg YWxyZWFkeSBsb3N0LgorICAgIGlmICghbV9sYXllclJlbmRlcmVyLT5pc0NvbXBvc2l0b3JDb250 ZXh0TG9zdCgpKSB7CisgICAgICAgIGRvQ29tcG9zaXRlKCk7CisKKyAgICAgICAgLy8gUHV0IHJl c3VsdCBvbnNjcmVlbi4KKyAgICAgICAgbV9sYXllclJlbmRlcmVyLT5wcmVzZW50KCk7CisgICAg fQogCiAgICAgaWYgKG1fbGF5ZXJSZW5kZXJlci0+aXNDb21wb3NpdG9yQ29udGV4dExvc3QoKSkg ewogICAgICAgICAvLyBUcnlpbmcgdG8gcmVjb3ZlciB0aGUgY29udGV4dCByaWdodCBoZXJlIHdp bGwgbm90IHdvcmsgaWYgR1BVIHByb2Nlc3MK