Re: [PATCH 01/12] uprobes: split THPs before trying replace them

[Johannes Weiner]

On Wed, Jan 25, 2017 at 08:55:22AM -0800, Srikar Dronamraju wrote:
> > > 
> > > > For THPs page_check_address() always fails. It's better to split them
> > > > first before trying to replace.
> > > 
> > > So what does this mean.  uprobes simply fails to work when trying to
> > > place a probe into a THP memory region?
> > 
> > Looks like we can end up with endless retry loop in uprobe_write_opcode().
> > 
> > > How come nobody noticed (and reported) this when using the feature?
> > 
> > I guess it's not often used for anon memory.
> > 
> 
> The first time the breakpoint is hit on a page, it replaces the text
> page with anon page.  Now lets assume we insert breakpoints in all the
> pages in a range. Here each page is individually replaced by a non THP
> anonpage. (since we dont have bulk breakpoint insertion support,
> breakpoint insertion happens one at a time). Now the only interesting
> case may be when each of these replaced pages happen to be physically
> contiguous so that THP kicks in to replace all of these pages with one
> THP page. Can happen in practice?
> 
> Are there any other cases that I have missed?

We use a hack in our applications where we open /proc/self/maps, copy
text segments to a staging area, then create overlay anon mappings on
top and copy the text back into them. Now we have THP-backed text and
very little iTLB pressure :-)

That said, we haven't run into the uprobes issue yet.