-
Bruce Dawson authored
For the last few months Chrome has been seeing many "impossible" crashes on Intel Gemini Lake, family 6 model 122 stepping 1 CPUs. These crashes only happen with 64-bit Chrome and only happen in the prologue of two functions. The crashes come and go across different Chrome versions. Analysis of most of the crashes shows that the address of the crashing instruction follows some patterns: When crashing in GetFieldIndex() the last byte of the address is always 1c, 5c, 9c, or dc. When crashing in UpdateCaches (fewer unique samples) the last byte of the address is always 5d or 9d. The address of the function is 0xc or 0xd bytes earlier so the crashing functions always start with an address that ends in 10, 50, 90, or d0. Those addresses are for the crashes on a load of the __security_cookie. The crashes also occasionally happen on the two instructions that follow the __security_cookie load in which case the crashing instruction's address has been seen to end with 23 or a3. This corresponds to a function start address of 10 or 90. Since the crash involves reading incorrect instruction bytes when crossing a 16-byte boundary and since the crash appears to only happen with particular 16-byte alignments it seems reasonable to force the function's alignments to a multiple of 32 to see if this reliably avoids the crashes. This change uses the gcc/clang __attribute__ directive to force 32-byte alignment. I have tested this change enough to verify that it triggers the desired alignment (with up to 31 "int 3" instructions added for padding) but since I have never reproduced this crash I have no way of testing its efficacy. Bug: chromium:968683, chromium:964273 Change-Id: Ia6e1c6d1e044b84d274817374b25523303e78b51 Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/1803775Reviewed-by: Toon Verwaest <verwaest@chromium.org> Commit-Queue: Bruce Dawson <brucedawson@chromium.org> Cr-Commit-Position: refs/heads/master@{#63804}
10360127