This is the mail archive of the libc-alpha@sourceware.org mailing list for the glibc project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: Question about madvise(DONTNEED) in glibc malloc


On Sun, Apr 14, 2013 at 10:11:50PM -0700, KOSAKI Motohiro wrote:
> (4/14/13 9:12 PM), Siddhesh Poyarekar wrote:
> > On 14 April 2013 22:07, KOSAKI Motohiro <kosaki.motohiro@gmail.com> wrote:
> >> Now, we linux MM folks discuss are discussing about new memory discarding feature.
> >> (https://lkml.org/lkml/2013/3/12/105). The motivation is similar wtih MADV_FREE,
> >> but more efficient. (http://lwn.net/Articles/230799)
> > 
> > I just went through this and it looks interesting as a replacement for
> > MADV_DONTNEED in malloc.  Please let us know what becomes of this so
> > that we could explore the possibility of using it in malloc.
> 
> Firstly, this api is still under discusstion and any feedback has a chance to change 
> api design.
> 
> The idea is simple. vrange(VOLATILE) just mark the argument pages is discardable. but
> doesn't discard. No page table walk, No TLB shootdown, No anything more.
> Touching the VOLATILed range, makes SIGBUG. you need to call vrange(NOVOLATILE) because
> you begin to use again.

A SIGBUS makes using vrange out of question. We simply cannot ignore user generated SIGBUS.
I do not see how we could distinguish user/vrange SIGBUS.

> When kernel meet memory starvation, kerenl may discared VOLATILed pages instead of swap
> out. Of cource, jsut dropping is much faster than disk write. More importantly, as far as
> system has enough memory, this heaviy weight pte nuke never run.
> 
> Please feel free to ask everything. We add more document to the man page or change the design.
> 


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]