From mboxrd@z Thu Jan 1 00:00:00 1970 From: Petr Tesarik Date: Fri, 19 Oct 2007 19:42:52 +0000 Subject: Re: [PATCH] ptrace RSE bug Message-Id: <1192822972.2727.47.camel@elijah.suse.cz> List-Id: References: <1188357710.22637.7.camel@sli10-conroe.sh.intel.com> In-Reply-To: <1188357710.22637.7.camel@sli10-conroe.sh.intel.com> MIME-Version: 1.0 Content-Type: text/plain; charset="us-ascii" Content-Transfer-Encoding: 7bit To: linux-ia64@vger.kernel.org On Thu, 2007-10-18 at 12:59 +0200, Petr Tesarik wrote: > Shaohua Li wrote: > > On Wed, 2007-10-17 at 16:56 +0200, Petr Tesarik wrote: > >> Shaohua Li wrote: > >>> On Fri, 2007-09-07 at 09:11 -0600, David Mosberger-Tang wrote: > >>>> Anything that avoids complicating the kernel exit path is worth doing! > >>>> The exit path is complicated enough as it is. > >>>> > >>>> --david > >>>> > >>>> On 9/7/07, Petr Tesarik wrote: > >>>>> -----BEGIN PGP SIGNED MESSAGE----- > >>>>> Hash: SHA1 > >>>>> > >>>>> Shaohua Li wrote: > >>>>>> On Thu, 2007-09-06 at 15:59 +0200, Petr Tesarik wrote: > >>>>>>> [...] > >>>>>>> So, what happens if upon syscall entry notification the debugger > >>>>>>> modifies the part of the RBS (in user-space) which corresponds to the > >>>>>>> arguments of that syscall? Currently, the syscall takes the modified > >>>>>>> arguments, but with your change it would still take the stale data > >>>>>>> from > >>>>>>> the kernel RBS. > >>>>>> The patch does sync from user RBS to kernel RBS just after syscall trace > >>>>>> enter. this is an exception I said doing sync just before syscall > >>>>>> return. I thought this covers your case, no? > >>>>> Ah, I'm sorry, I missed that part of the patch. Well, if we have to do a > >>>>> sync on every syscall_trace_enter() and syscall_trace_leave(), then the > >>>>> only cases where introducing TIF_RESTORE_RSE saves us a duplicate sync > >>>>> seems to be in the clone/fork and exit paths. In other words, it's > >>>>> probably not worth the added complexity. But since you have written the > >>>>> whole complex thing already, I have no objections against it. > >>> Ok, this is a simplified patch. please review. > >> Well, it's been quite some time, but here we go. > >> > >> I'm generally fine with this patch, but pleas note that it can't be > >> included on its own: > >> > >> 1. There still is the race condition introduced by moving > >> set_current_state(TASK_TRACED) after the spin_unlock_irq > > I don't know the details, but Roland said if other parts are ok, he can help fix the issue. > > > >> 2. You must couple it with the (planned) changes to the ptrace, > >> because otherwise PTRACE_{PEEK,POKE}{TEXT,DATA} still access the kernel > >> RBS, but it gets later overwritten back from userspace when it is synced. > > > >> I have verified that failing to do so breaks "strace -f", because > >> strace > >> relies on intercepting the clone() system call and setting the > >> CLONE_PTRACE bit in the flags argument. Of course, if the bit is only > >> set in the kernel RBS, which is overwritten with the (old) value from > >> the user RBS on a PTRACE_CONT, the new process is not traced. > > The patch sync kernel RBS to user just before the task is suspended, so > > I think we should be fine here. I did test 'strace -f', and test is ok. > > Maybe you're right. I was porting this to 2.6.16 for SUSE Linux > Enterprise Server 10, so my patch was a bit different. I'll retest with > latest git. Nevertheless, I still think that ia64_poke() can't do the > right thing here, because the changes made by PTRACE_PEEKDATA should > also be visible in /proc//mem, for example. OK, I retested everything again with 2.6.23 and I can confirm that the kernel behaves consistently with this patch applied - modifying syscall arguments works (both for break and for fsyscalls), changes are refleced in /proc//mem and accessing the RNAT bits works too. I would still like to get rid of ia64_peek() and ia64_poke(), because it is no longer needed and is inefficient. For example, currently each PTRACE_POKE first non-trivially finds out the correct location within the kernel RBS and then immediately synchronizes the RBS to user space. Not to mention that for peeking/poking a process with more threads the kernel must first find the correct thread for a given address. Shaohua's patch allows us to greatly simplify the architecture-specific bits of ptrace. I'll send a patch soon. In short, you've got my ack (whatever it's worth). Cheers, Petr Tesarik