Messages in this thread | | | Date | Tue, 15 Mar 2011 23:34:23 +0530 | From | Srikar Dronamraju <> | Subject | Re: [PATCH v2 2.6.38-rc8-tip 5/20] 5: Uprobes: register/unregister probes. |
| |
* Peter Zijlstra <peterz@infradead.org> [2011-03-15 18:50:11]:
> On Tue, 2011-03-15 at 13:47 -0400, Steven Rostedt wrote: > > On Tue, 2011-03-15 at 22:45 +0530, Srikar Dronamraju wrote: > > > > > + } > > > > > + list_for_each_entry_safe(mm, tmpmm, &tmp_list, uprobes_list) { > > > > > + down_read(&mm->mmap_sem); > > > > > + if (!install_uprobe(mm, uprobe)) > > > > > + ret = 0; > > > > > > > > Installing it once is success ? > > > > > > This is a little tricky. My intention was to return success even if one > > > install is successful. If we return error, then the caller can go > > > ahead and free the consumer. Since we return success if there are > > > currently no processes that have mapped this inode, I was tempted to > > > return success on atleast one successful install. > > > > What about an all or nothing approach. If one fails, remove all that > > were installed, and give up. > > That sounds like a much saner semantic and is what we generally do in > the kernel.
One of the install_uprobe could be failing because the process was almost exiting, something like there was no mm->owner. Also lets assume that the first few install_uprobes go thro and the last install_uprobe fails. There could be breakpoint hits corresponding to the already installed uprobes that get displayed. i.e all breakpoint hits from the first install_uprobe to the time we detect a failed a install_uprobe and revert all inserted breakpoints will be shown as being captured.
Also register_uprobe will only install probes for processes that are actively and have mapped the inode. However install_uprobe called from uprobe_mmap which also registers the probe can fail. Now should we take the same yardstick and remove all the probes corresponding to the successful register_uprobe?
| |