Sources Bugzilla – Bug 11447
default section layout should be improved to reduce pagefaults
Last modified: 2010-04-13 11:56:04 UTC
I graphed pagefaults on firefox startup which showed that there some bonus page faults caused by the default section order. The main problem is backwards io which always causes page faults. For reference click the second rectange in http://people.mozilla.com/~tglek/startup/systemtap_graphs/visualize.html?#ld_bug/log.txt.html x-axis is file offset, y-axis is time Log of the sections being read is in http://people.mozilla.com/~tglek/startup/systemtap_graphs/ld_bug/report.txt Fault #1: This likely can't fixed in ld, but mentioning it in case Going through the log chronologically there the problems are: .bss is being zeroed which causes copy-on-write io. Then .dynamic is being read. The fix should be in the dynamic linker to read .dynamic before memset. This would allow binaries(where distance between .dynamic and .bss is smaller than the kernel's readhead) to avoid the .bss pagefault. Perhaps .dynamic should be the last section before .bss? Alternatively ld could force .bss to point past the end of the file. Fault #2 .gnu.version* should be in front of .dyn(sym|str) Fault #3 .dynstr should be in front of .dynsym Would be nice if .eh_frame* could be in front of .fini, but as far as I can tell these have to be in different segments.
Subject: Re: New: default section layout should be improved to reduce pagefaults On Mon, Mar 29, 2010 at 08:29:31PM -0000, tglek at mozilla dot com wrote: > Fault #2 > .gnu.version* should be in front of .dyn(sym|str) > > Fault #3 > .dynstr should be in front of .dynsym Why? Both .gnu.version and .dynstr are referred to from .dynsym. ISTM that it would be as easy to improve the loader if it's reading them in a funny order.
Created attachment 4719 [details] Reorder .gnu.version, .dynstr and .dynsym
Hi Taras, This sounds like a loader problem to me, rather than a binutils bug. It is relatively easy to rearrange the order of the .gnu.version, .dynstr and .dynsym sections (see the uploaded patch), but I do not think that this is right way to solve the problem. If the loader is performing some backwards io to get at those sections then I would suggest that it ought to be recoded to keep them in memory after reading them for the first time. Cheers Nick