Messages in this thread | | | Date | Mon, 16 Apr 2012 23:47:07 +0200 | From | Oleg Nesterov <> | Subject | Re: [RFC 0/6] uprobes: kill uprobes_srcu/uprobe_srcu_id |
| |
On 04/16, Peter Zijlstra wrote: > > On Mon, 2012-04-16 at 01:44 +0200, Oleg Nesterov wrote: > > > And. I have another reason for down_write() in register/unregister. > > I am still not sure this is possible (I had no time to try to > > implement), but it seems to me we can kill the uprobe counter in > > mm_struct. > > You mean by making register/unregister down_write, you're exclusive with > munmap()
.. and with register/unregister.
Why do we need mm->uprobes_state.count? It is writeonly, except we check it in the DIE_INT3 notifier before anything else to avoid the unnecessary uprobes overhead.
Suppose we kill it, and add the new MMF_HAS_UPROBE flag instead. install_breakpoint() sets it unconditionally, uprobe_pre_sstep_notifier() checks it.
(And perhaps we can stop right here? I mean how often this can slow down the debugger which installs int3 in the same mm?)
Now we need to clear MMF_HAS_UPROBE somehowe, when the last uprobe goes away. Lets ignore uprobe_map/unmap for simplicity.
- We add another flag, MMF_UPROBE_RECALC, it is set by remove_breakpoint().
- We change handle_swbp(). Ignoring all details it does:
if (find_uprobe(vaddr)) process_uprobe(); else if (test_bit(MMF_HAS_UPROBE) && test_bit(MMF_UPROBE_RECALC)) recalc_mmf_uprobe_flag();
where recalc_mmf_uprobe_flag() checks all vmas and either clears both flags or MMF_UPROBE_RECALC only.
This is the really slow O(n) path, but it can only happen after unregister, and only if we hit another non-uprobe breakpoint in the same mm.
Something like this. What do you think?
Oleg.
| |