Message ID | 20160719093242.GA23753@ircssh.c.rugged-nimbus-611.internal |
---|---|
State | Changes Requested, archived |
Delegated to: | David Miller |
Headers | show |
Hi Sargun, On 07/19/2016 11:32 AM, Sargun Dhillon wrote: > This allows user memory to be written to during the course of a kprobe. > It shouldn't be used to implement any kind of security mechanism > because of TOC-TOU attacks, but rather to debug, divert, and > manipulate execution of semi-cooperative processes. > > Although it uses probe_kernel_write, we limit the address space > the probe can write into by checking the space with access_ok. > This call shouldn't sleep on any architectures based on review. > > It was tested with the tracex7 program on x86-64. > > Signed-off-by: Sargun Dhillon <sargun@sargun.me> > Cc: Alexei Starovoitov <ast@kernel.org> [...] > diff --git a/kernel/trace/bpf_trace.c b/kernel/trace/bpf_trace.c > index ebfbb7d..45878f3 100644 > --- a/kernel/trace/bpf_trace.c > +++ b/kernel/trace/bpf_trace.c > @@ -81,6 +81,35 @@ static const struct bpf_func_proto bpf_probe_read_proto = { > .arg3_type = ARG_ANYTHING, > }; > > +static u64 bpf_copy_to_user(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5) > +{ > + void *to = (void *) (long) r1; > + void *from = (void *) (long) r2; > + int size = (int) r3; Nit: you have two spaces here? > + struct task_struct *task = current; > + > + /* check if we're in a user context */ > + if (unlikely(in_interrupt())) > + return -EINVAL; > + if (unlikely(!task || !task->pid)) Why !task->pid is needed? Can you point to the code where this is set up as such? I'd assume if that's the case, there's a generic helper for it to determine that the task_struct is a kernel task? > + return -EINVAL; > + > + /* Is this a user address, or a kernel address? */ > + if (!access_ok(VERIFY_WRITE, to, size)) > + return -EINVAL; > + > + return probe_kernel_write(to, from, size); I'm still worried that this can lead to all kind of hard to find bugs or races for user processes, if you make this writable to entire user address space (which is the only thing that access_ok() checks for). What if the BPF program has bugs and writes to wrong addresses for example, introducing bugs in some other, non-intended processes elsewhere? Can this be limited to syscalls only? And if so, to the passed memory only? Have you played around with ptrace() to check whether you could achieve similar functionality (was thinking about things like [1], PTRACE_PEEK{TEXT,DATA} / PTRACE_POKE{TEXT,DATA}). If not, why can't this be limited to a similar functionality for only the current task. ptrace() utilizes helpers like access_process_vm(), maybe this can similarly be adapted here, too (under the circumstances that sleeping is not allowed)? [1] https://code.google.com/archive/p/ptrace-parasite/ > +} > + > +static const struct bpf_func_proto bpf_copy_to_user_proto = { > + .func = bpf_copy_to_user, > + .gpl_only = false, > + .ret_type = RET_INTEGER, > + .arg1_type = ARG_ANYTHING, > + .arg2_type = ARG_PTR_TO_RAW_STACK, > + .arg3_type = ARG_CONST_STACK_SIZE, Nit: please tab-align '=' like we have in the other *_proto cases in bpf_trace. > +}; > + > /* > * limited trace_printk() > * only %d %u %x %ld %lu %lx %lld %llu %llx %p %s conversion specifiers allowed Thanks, Daniel
On Tue, Jul 19, 2016 at 01:17:53PM +0200, Daniel Borkmann wrote: > >+ return -EINVAL; > >+ > >+ /* Is this a user address, or a kernel address? */ > >+ if (!access_ok(VERIFY_WRITE, to, size)) > >+ return -EINVAL; > >+ > >+ return probe_kernel_write(to, from, size); > > I'm still worried that this can lead to all kind of hard to find > bugs or races for user processes, if you make this writable to entire > user address space (which is the only thing that access_ok() checks > for). What if the BPF program has bugs and writes to wrong addresses > for example, introducing bugs in some other, non-intended processes > elsewhere? Can this be limited to syscalls only? And if so, to the > passed memory only? my understanding that above code will write only to memory of current process, so impact is contained and in that sense buggy kprobe program is no different from buggy seccomp prorgram. Limiting this to syscalls will make it too limited. I'm in favor of this change, because it allows us to experiment with restartable sequences and lock-free algorithms that need ultrafast access to cpuid without burdening the kernel with stable abi. > Have you played around with ptrace() to check whether you could > achieve similar functionality (was thinking about things like [1], > PTRACE_PEEK{TEXT,DATA} / PTRACE_POKE{TEXT,DATA}). If not, why can't > this be limited to a similar functionality for only the current task. > ptrace() utilizes helpers like access_process_vm(), maybe this can > similarly be adapted here, too (under the circumstances that sleeping > is not allowed)? If we hack access_process_vm I think at the end it will look like probe_kernel_write. Walking mm requires semaphore, so we would only be able to do it in task_work and there we can do normal copy_to_user just as well, but it will complicate the programs quite a bit, since writes will be asynchronous and batched. Looks like with access_ok+probe_write we can achieve the same with a lot less code. As far as races between user and bpf program, yeah, if process is multithreaded, the other threads may access the same mem that bpf is writing to, but that's no different from reading. For tracing we walk complex datastructures and pointers. They can be changed by user space on the fly and bpf will see garbage. Like we have uprobe+bpf that walks clang c++ internal datastructures to figure out how long it takes clang to process .h headers. There is a lot of fragility in the bpf script, yet it's pretty useful to quickly analyze compile times. I see bpf_copy_to_user to be hugely valuable too, not as a stable interface, but rather as a tool to quickly debug and experiment.
On Tue, 19 Jul 2016, Daniel Borkmann wrote: > Hi Sargun, > > On 07/19/2016 11:32 AM, Sargun Dhillon wrote: >> This allows user memory to be written to during the course of a kprobe. >> It shouldn't be used to implement any kind of security mechanism >> because of TOC-TOU attacks, but rather to debug, divert, and >> manipulate execution of semi-cooperative processes. >> >> Although it uses probe_kernel_write, we limit the address space >> the probe can write into by checking the space with access_ok. >> This call shouldn't sleep on any architectures based on review. >> >> It was tested with the tracex7 program on x86-64. >> >> Signed-off-by: Sargun Dhillon <sargun@sargun.me> >> Cc: Alexei Starovoitov <ast@kernel.org> > [...] >> diff --git a/kernel/trace/bpf_trace.c b/kernel/trace/bpf_trace.c >> index ebfbb7d..45878f3 100644 >> --- a/kernel/trace/bpf_trace.c >> +++ b/kernel/trace/bpf_trace.c >> @@ -81,6 +81,35 @@ static const struct bpf_func_proto bpf_probe_read_proto >> = { >> .arg3_type = ARG_ANYTHING, >> }; >> >> +static u64 bpf_copy_to_user(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5) >> +{ >> + void *to = (void *) (long) r1; >> + void *from = (void *) (long) r2; >> + int size = (int) r3; > > Nit: you have two spaces here? > >> + struct task_struct *task = current; >> + >> + /* check if we're in a user context */ >> + if (unlikely(in_interrupt())) >> + return -EINVAL; >> + if (unlikely(!task || !task->pid)) > > Why !task->pid is needed? Can you point to the code where this is > set up as such? I'd assume if that's the case, there's a generic > helper for it to determine that the task_struct is a kernel task? > One example of precedent: http://lxr.free-electrons.com/source/arch/x86/kernel/process.c#L270 Also, while testing this out, I instrumented ip_route_output_flow, and I found that there were cases that were invoked from the kernel there current was not NULL, but task->pid was reliably 0. >> + return -EINVAL; >> + >> + /* Is this a user address, or a kernel address? */ >> + if (!access_ok(VERIFY_WRITE, to, size)) >> + return -EINVAL; >> + >> + return probe_kernel_write(to, from, size); > > I'm still worried that this can lead to all kind of hard to find > bugs or races for user processes, if you make this writable to entire > user address space (which is the only thing that access_ok() checks > for). What if the BPF program has bugs and writes to wrong addresses > for example, introducing bugs in some other, non-intended processes > elsewhere? Can this be limited to syscalls only? And if so, to the > passed memory only? > > Have you played around with ptrace() to check whether you could > achieve similar functionality (was thinking about things like [1], > PTRACE_PEEK{TEXT,DATA} / PTRACE_POKE{TEXT,DATA}). If not, why can't > this be limited to a similar functionality for only the current task. > ptrace() utilizes helpers like access_process_vm(), maybe this can > similarly be adapted here, too (under the circumstances that sleeping > is not allowed)? > > [1] https://code.google.com/archive/p/ptrace-parasite/ > I agree that this can get really complicated if it clashes with memory that a task is currently manipulating while we do the copy_to_user. I highlighted that in one of the commit messages not to rely on this behaviour for anything "real" because TOC-TOU. We could limit the use to be within syscalls and uprobes if you think that's the best idea. Is there an easy way to find out, but I think it's somewhat difficult to restrict it to passed memory only. In the example app there's only one level of indirection that occurs, and without some new rules in the verifier, I don't how to handle this. A simple, problematic example is dealing with iovecs, and process_vm_readv, and process_vm_writev, where to find the actual buffer content I need to dereference once to get the list of iovecs, and then again, to find the buffer the iovec points to. We looked at using ptrace, but our software is often working with blackbox code that sometimes does not like to ptrace'd. ptrace is kinda also a pain to work with, and has the same potential for data races you highlighted earlier. In fact, we have a demo of tracex7 that uses ptrace + process_vm_writev + seccomp, and it's much more complicated and fragile. Just casually poking around, it looks like access_process_vm sleeps, resulting in a whole new set of complexities that Alexei highlighted earlier. >> +} >> + >> +static const struct bpf_func_proto bpf_copy_to_user_proto = { >> + .func = bpf_copy_to_user, >> + .gpl_only = false, >> + .ret_type = RET_INTEGER, >> + .arg1_type = ARG_ANYTHING, >> + .arg2_type = ARG_PTR_TO_RAW_STACK, >> + .arg3_type = ARG_CONST_STACK_SIZE, > > Nit: please tab-align '=' like we have in the other *_proto cases in > bpf_trace. > >> +}; >> + >> /* >> * limited trace_printk() >> * only %d %u %x %ld %lu %lx %lld %llu %llx %p %s conversion specifiers >> allowed > > Thanks, > Daniel > Thanks for the review. Unless there's any other input, I'll go ahead and resubmit the patchset with the formatting nits fixed.
On 07/19/2016 06:34 PM, Alexei Starovoitov wrote: > On Tue, Jul 19, 2016 at 01:17:53PM +0200, Daniel Borkmann wrote: >>> + return -EINVAL; >>> + >>> + /* Is this a user address, or a kernel address? */ >>> + if (!access_ok(VERIFY_WRITE, to, size)) >>> + return -EINVAL; >>> + >>> + return probe_kernel_write(to, from, size); >> >> I'm still worried that this can lead to all kind of hard to find >> bugs or races for user processes, if you make this writable to entire >> user address space (which is the only thing that access_ok() checks >> for). What if the BPF program has bugs and writes to wrong addresses >> for example, introducing bugs in some other, non-intended processes >> elsewhere? Can this be limited to syscalls only? And if so, to the >> passed memory only? > > my understanding that above code will write only to memory of current process, > so impact is contained and in that sense buggy kprobe program is no different > from buggy seccomp prorgram. Compared to seccomp, you might not notice that a race has happened, in seccomp case you might have killed your process, which is visible. But ok, in ptrace() case it might be similar issue perhaps ... The asm-generic version does __access_ok(..) { return 1; } for nommu case, I haven't checked closely enough whether there's actually an arch that uses this, but for example arm nommu with only one addr space would certainly result in access_ok() as 1, and then you could also go into probe_kernel_write(), no? Don't know that code well enough, but I believe the check would only ensure in normal use-cases that user process doesn't fiddle with kernel address space, but not necessarily guarantee that this really only belongs to the process address space. x86 code comments this with "note that, depending on architecture, this function probably just checks that the pointer is in the user space range - after calling this function, memory access functions may still return -EFAULT". Also, what happens in case of kernel thread? As it stands, it does ... if (unlikely(in_interrupt())) return -EINVAL; if (unlikely(!task || !task->pid)) return -EINVAL; So up to here, irq/sirq, NULL current and that current is not the 'idle' process is being checked (still fail to see the point for the !task->pid, I believe the intend here is different). /* Is this a user address, or a kernel address? */ if (!access_ok(VERIFY_WRITE, to, size)) return -EINVAL; Now here. What if it's a kernel thread? You'll have KERNEL_DS segment, task->pid was non-zero as well for the kthread, so access_ok() will pass and you can still execute probe_kernel_write() ... > Limiting this to syscalls will make it too limited. > I'm in favor of this change, because it allows us to experiment > with restartable sequences and lock-free algorithms that need ultrafast > access to cpuid without burdening the kernel with stable abi. > >> Have you played around with ptrace() to check whether you could >> achieve similar functionality (was thinking about things like [1], >> PTRACE_PEEK{TEXT,DATA} / PTRACE_POKE{TEXT,DATA}). If not, why can't >> this be limited to a similar functionality for only the current task. >> ptrace() utilizes helpers like access_process_vm(), maybe this can >> similarly be adapted here, too (under the circumstances that sleeping >> is not allowed)? > > If we hack access_process_vm I think at the end it will look like > probe_kernel_write. Walking mm requires semaphore, so we would only > be able to do it in task_work and there we can do normal copy_to_user > just as well, but it will complicate the programs quite a bit, since > writes will be asynchronous and batched. > Looks like with access_ok+probe_write we can achieve the same > with a lot less code. I believe it may not quite be the same as it currently stands. No fundamental objection, just that this needs to be made "safe" to the limits you state above yourself. ;) > As far as races between user and bpf program, yeah, if process > is multithreaded, the other threads may access the same mem that > bpf is writing to, but that's no different from reading. > For tracing we walk complex datastructures and pointers. They > can be changed by user space on the fly and bpf will see garbage. > Like we have uprobe+bpf that walks clang c++ internal datastructures > to figure out how long it takes clang to process .h headers. > There is a lot of fragility in the bpf script, yet it's pretty > useful to quickly analyze compile times. > I see bpf_copy_to_user to be hugely valuable too, not as a stable > interface, but rather as a tool to quickly debug and experiment. Right, so maybe there should be a warn once to the dmesg log that this is just experimental? Thanks, Daniel
On Wed, Jul 20, 2016 at 01:19:51AM +0200, Daniel Borkmann wrote: > On 07/19/2016 06:34 PM, Alexei Starovoitov wrote: > >On Tue, Jul 19, 2016 at 01:17:53PM +0200, Daniel Borkmann wrote: > >>>+ return -EINVAL; > >>>+ > >>>+ /* Is this a user address, or a kernel address? */ > >>>+ if (!access_ok(VERIFY_WRITE, to, size)) > >>>+ return -EINVAL; > >>>+ > >>>+ return probe_kernel_write(to, from, size); > >> > >>I'm still worried that this can lead to all kind of hard to find > >>bugs or races for user processes, if you make this writable to entire > >>user address space (which is the only thing that access_ok() checks > >>for). What if the BPF program has bugs and writes to wrong addresses > >>for example, introducing bugs in some other, non-intended processes > >>elsewhere? Can this be limited to syscalls only? And if so, to the > >>passed memory only? > > > >my understanding that above code will write only to memory of current process, > >so impact is contained and in that sense buggy kprobe program is no different > >from buggy seccomp prorgram. > > Compared to seccomp, you might not notice that a race has happened, > in seccomp case you might have killed your process, which is visible. > But ok, in ptrace() case it might be similar issue perhaps ... > > The asm-generic version does __access_ok(..) { return 1; } for nommu > case, I haven't checked closely enough whether there's actually an arch > that uses this, but for example arm nommu with only one addr space would > certainly result in access_ok() as 1, and then you could also go into > probe_kernel_write(), no? good point. how arm nommu handles copy_to_user? if there is nommu then there is no user/kernel mm ? Crazy archs. I guess we have to disable this helper on all such archs. > Don't know that code well enough, but I believe the check would only > ensure in normal use-cases that user process doesn't fiddle with kernel > address space, but not necessarily guarantee that this really only > belongs to the process address space. why? on x86 that exactly what it does. access_ok=true means it's user space address and since we're in _this user context_ probe_kernel_write can only affect this user. > x86 code comments this with "note that, depending on architecture, > this function probably just checks that the pointer is in the user > space range - after calling this function, memory access functions may > still return -EFAULT". Yes. I've read that comment to :) Certainly not an expert, but the archs I've looked at, access_ok has the same meaning as on x86. They check the space range to make sure address doesn't belong to kernel. Could I have missed something? Certainly. Please double check :) > Also, what happens in case of kernel thread? my understanding if access_ok(addr)=true the addr will never point to memory of kernel thread. We need expert opinion. Whom should we ping? > As it stands, it does ... > > if (unlikely(in_interrupt())) > return -EINVAL; > if (unlikely(!task || !task->pid)) > return -EINVAL; > > So up to here, irq/sirq, NULL current and that current is not the 'idle' > process is being checked (still fail to see the point for the !task->pid, > I believe the intend here is different). > > /* Is this a user address, or a kernel address? */ > if (!access_ok(VERIFY_WRITE, to, size)) > return -EINVAL; > > Now here. What if it's a kernel thread? You'll have KERNEL_DS segment, > task->pid was non-zero as well for the kthread, so access_ok() will > pass and you can still execute probe_kernel_write() ... I think user_addr_max() should be zero for kthread, but worth checking for sure. > >Limiting this to syscalls will make it too limited. > >I'm in favor of this change, because it allows us to experiment > >with restartable sequences and lock-free algorithms that need ultrafast > >access to cpuid without burdening the kernel with stable abi. > > > >>Have you played around with ptrace() to check whether you could > >>achieve similar functionality (was thinking about things like [1], > >>PTRACE_PEEK{TEXT,DATA} / PTRACE_POKE{TEXT,DATA}). If not, why can't > >>this be limited to a similar functionality for only the current task. > >>ptrace() utilizes helpers like access_process_vm(), maybe this can > >>similarly be adapted here, too (under the circumstances that sleeping > >>is not allowed)? > > > >If we hack access_process_vm I think at the end it will look like > >probe_kernel_write. Walking mm requires semaphore, so we would only > >be able to do it in task_work and there we can do normal copy_to_user > >just as well, but it will complicate the programs quite a bit, since > >writes will be asynchronous and batched. > >Looks like with access_ok+probe_write we can achieve the same > >with a lot less code. > > I believe it may not quite be the same as it currently stands. No > fundamental objection, just that this needs to be made "safe" to the > limits you state above yourself. ;) completely agree :) the more eyes on the patch the better. > >As far as races between user and bpf program, yeah, if process > >is multithreaded, the other threads may access the same mem that > >bpf is writing to, but that's no different from reading. > >For tracing we walk complex datastructures and pointers. They > >can be changed by user space on the fly and bpf will see garbage. > >Like we have uprobe+bpf that walks clang c++ internal datastructures > >to figure out how long it takes clang to process .h headers. > >There is a lot of fragility in the bpf script, yet it's pretty > >useful to quickly analyze compile times. > >I see bpf_copy_to_user to be hugely valuable too, not as a stable > >interface, but rather as a tool to quickly debug and experiment. > > Right, so maybe there should be a warn once to the dmesg log that this > is just experimental? good point. The tracing experience showed that trace_printk warning was quite effective on steering user space assumptions. Let's do something here as well.
On Wed, 20 Jul 2016, Daniel Borkmann wrote: > On 07/19/2016 06:34 PM, Alexei Starovoitov wrote: >> On Tue, Jul 19, 2016 at 01:17:53PM +0200, Daniel Borkmann wrote: >>>> + return -EINVAL; >>>> + >>>> + /* Is this a user address, or a kernel address? */ >>>> + if (!access_ok(VERIFY_WRITE, to, size)) >>>> + return -EINVAL; >>>> + >>>> + return probe_kernel_write(to, from, size); >>> >>> I'm still worried that this can lead to all kind of hard to find >>> bugs or races for user processes, if you make this writable to entire >>> user address space (which is the only thing that access_ok() checks >>> for). What if the BPF program has bugs and writes to wrong addresses >>> for example, introducing bugs in some other, non-intended processes >>> elsewhere? Can this be limited to syscalls only? And if so, to the >>> passed memory only? >> >> my understanding that above code will write only to memory of current >> process, >> so impact is contained and in that sense buggy kprobe program is no >> different >> from buggy seccomp prorgram. > > Compared to seccomp, you might not notice that a race has happened, > in seccomp case you might have killed your process, which is visible. > But ok, in ptrace() case it might be similar issue perhaps ... > > The asm-generic version does __access_ok(..) { return 1; } for nommu > case, I haven't checked closely enough whether there's actually an arch > that uses this, but for example arm nommu with only one addr space would > certainly result in access_ok() as 1, and then you could also go into > probe_kernel_write(), no? > > Don't know that code well enough, but I believe the check would only > ensure in normal use-cases that user process doesn't fiddle with kernel > address space, but not necessarily guarantee that this really only > belongs to the process address space. > You're right, it doesn't gaurantee that the memory being written to is the user, but only the current process should be mapped into that chunk of memory that is being operated at in the point in time. I think it'd be fairly difficult to write to another process' memory to corrupt it. It also might make sense to me to call access_ok with a known kernel address, and either deny access to the bpf helper, or warn in dmesg saying that it is running in an environment without any protection. -- I'm not sure if there is a preferred kernel address to check? What do you think? > x86 code comments this with "note that, depending on architecture, > this function probably just checks that the pointer is in the user > space range - after calling this function, memory access functions may > still return -EFAULT". > > Also, what happens in case of kernel thread? It seems fairly reasonable to me to check task->flags & PF_KTHREAD in order to prevent this scenario since there is no memory to copy_to_user to. > > As it stands, it does ... > > if (unlikely(in_interrupt())) > return -EINVAL; > if (unlikely(!task || !task->pid)) > return -EINVAL; > > So up to here, irq/sirq, NULL current and that current is not the 'idle' > process is being checked (still fail to see the point for the !task->pid, > I believe the intend here is different). > > /* Is this a user address, or a kernel address? */ > if (!access_ok(VERIFY_WRITE, to, size)) > return -EINVAL; > > Now here. What if it's a kernel thread? You'll have KERNEL_DS segment, > task->pid was non-zero as well for the kthread, so access_ok() will > pass and you can still execute probe_kernel_write() ... > >> Limiting this to syscalls will make it too limited. >> I'm in favor of this change, because it allows us to experiment >> with restartable sequences and lock-free algorithms that need ultrafast >> access to cpuid without burdening the kernel with stable abi. >> >>> Have you played around with ptrace() to check whether you could >>> achieve similar functionality (was thinking about things like [1], >>> PTRACE_PEEK{TEXT,DATA} / PTRACE_POKE{TEXT,DATA}). If not, why can't >>> this be limited to a similar functionality for only the current task. >>> ptrace() utilizes helpers like access_process_vm(), maybe this can >>> similarly be adapted here, too (under the circumstances that sleeping >>> is not allowed)? >> >> If we hack access_process_vm I think at the end it will look like >> probe_kernel_write. Walking mm requires semaphore, so we would only >> be able to do it in task_work and there we can do normal copy_to_user >> just as well, but it will complicate the programs quite a bit, since >> writes will be asynchronous and batched. >> Looks like with access_ok+probe_write we can achieve the same >> with a lot less code. > > I believe it may not quite be the same as it currently stands. No > fundamental objection, just that this needs to be made "safe" to the > limits you state above yourself. ;) > At least for my current use cases, doing this asynchronously could prove to be unworkable. >> As far as races between user and bpf program, yeah, if process >> is multithreaded, the other threads may access the same mem that >> bpf is writing to, but that's no different from reading. >> For tracing we walk complex datastructures and pointers. They >> can be changed by user space on the fly and bpf will see garbage. >> Like we have uprobe+bpf that walks clang c++ internal datastructures >> to figure out how long it takes clang to process .h headers. >> There is a lot of fragility in the bpf script, yet it's pretty >> useful to quickly analyze compile times. >> I see bpf_copy_to_user to be hugely valuable too, not as a stable >> interface, but rather as a tool to quickly debug and experiment. > > Right, so maybe there should be a warn once to the dmesg log that this > is just experimental? > > Thanks, > Daniel > It seems reasonable to me to give people a heads up, that they have bpf programs that can modify user memory. I can add a warn once with along the lines of: pr_warn_once("bpf_copy_to_user in use: Experimental feature, may corrupt memory"); --- I think that should scare people away to the downsides. That warning along with verifying that the access_ok function is doing useful things on their platform, and warning if it isn't seems to be a strong enough signal to people that might accidentally abuse it.
On 07/20/2016 05:02 AM, Alexei Starovoitov wrote: > On Wed, Jul 20, 2016 at 01:19:51AM +0200, Daniel Borkmann wrote: >> On 07/19/2016 06:34 PM, Alexei Starovoitov wrote: >>> On Tue, Jul 19, 2016 at 01:17:53PM +0200, Daniel Borkmann wrote: >>>>> + return -EINVAL; >>>>> + >>>>> + /* Is this a user address, or a kernel address? */ >>>>> + if (!access_ok(VERIFY_WRITE, to, size)) >>>>> + return -EINVAL; >>>>> + >>>>> + return probe_kernel_write(to, from, size); >>>> >>>> I'm still worried that this can lead to all kind of hard to find >>>> bugs or races for user processes, if you make this writable to entire >>>> user address space (which is the only thing that access_ok() checks >>>> for). What if the BPF program has bugs and writes to wrong addresses >>>> for example, introducing bugs in some other, non-intended processes >>>> elsewhere? Can this be limited to syscalls only? And if so, to the >>>> passed memory only? >>> >>> my understanding that above code will write only to memory of current process, >>> so impact is contained and in that sense buggy kprobe program is no different >> >from buggy seccomp prorgram. >> >> Compared to seccomp, you might not notice that a race has happened, >> in seccomp case you might have killed your process, which is visible. >> But ok, in ptrace() case it might be similar issue perhaps ... >> >> The asm-generic version does __access_ok(..) { return 1; } for nommu >> case, I haven't checked closely enough whether there's actually an arch >> that uses this, but for example arm nommu with only one addr space would >> certainly result in access_ok() as 1, and then you could also go into >> probe_kernel_write(), no? > > good point. how arm nommu handles copy_to_user? if there is nommu Should probably boil down to something similar as plain memcpy(). > then there is no user/kernel mm ? Crazy archs. > I guess we have to disable this helper on all such archs. > >> Don't know that code well enough, but I believe the check would only >> ensure in normal use-cases that user process doesn't fiddle with kernel >> address space, but not necessarily guarantee that this really only >> belongs to the process address space. > > why? on x86 that exactly what it does. access_ok=true means > it's user space address and since we're in _this user context_ > probe_kernel_write can only affect this user. > >> x86 code comments this with "note that, depending on architecture, >> this function probably just checks that the pointer is in the user >> space range - after calling this function, memory access functions may >> still return -EFAULT". > > Yes. I've read that comment to :) > Certainly not an expert, but the archs I've looked at, access_ok > has the same meaning as on x86. They check the space range to > make sure address doesn't belong to kernel. > Could I have missed something? Certainly. Please double check :) > >> Also, what happens in case of kernel thread? > > my understanding if access_ok(addr)=true the addr will never point > to memory of kernel thread. If you're coming from user context only, this should be true, it'll check whether it's some user space pointer. >> As it stands, it does ... >> >> if (unlikely(in_interrupt())) >> return -EINVAL; >> if (unlikely(!task || !task->pid)) >> return -EINVAL; >> >> So up to here, irq/sirq, NULL current and that current is not the 'idle' >> process is being checked (still fail to see the point for the !task->pid, >> I believe the intend here is different). >> >> /* Is this a user address, or a kernel address? */ >> if (!access_ok(VERIFY_WRITE, to, size)) >> return -EINVAL; >> >> Now here. What if it's a kernel thread? You'll have KERNEL_DS segment, >> task->pid was non-zero as well for the kthread, so access_ok() will >> pass and you can still execute probe_kernel_write() ... > > I think user_addr_max() should be zero for kthread, but > worth checking for sure. It's 0xffffffffffffffff, I did a quick test yesterday night with creating a kthread, so access_ok() should pass for such case.
diff --git a/include/uapi/linux/bpf.h b/include/uapi/linux/bpf.h index c4d9224..d435f7c 100644 --- a/include/uapi/linux/bpf.h +++ b/include/uapi/linux/bpf.h @@ -364,6 +364,17 @@ enum bpf_func_id { */ BPF_FUNC_get_current_task, + /** + * copy_to_user(to, from, len) - Copy a block of data into user space. + * @to: destination address in userspace + * @from: source address on stack + * @len: number of bytes to copy + * Return: + * Returns number of bytes that could not be copied. + * On success, this will be zero + */ + BPF_FUNC_copy_to_user, + __BPF_FUNC_MAX_ID, }; diff --git a/kernel/trace/bpf_trace.c b/kernel/trace/bpf_trace.c index ebfbb7d..45878f3 100644 --- a/kernel/trace/bpf_trace.c +++ b/kernel/trace/bpf_trace.c @@ -81,6 +81,35 @@ static const struct bpf_func_proto bpf_probe_read_proto = { .arg3_type = ARG_ANYTHING, }; +static u64 bpf_copy_to_user(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5) +{ + void *to = (void *) (long) r1; + void *from = (void *) (long) r2; + int size = (int) r3; + struct task_struct *task = current; + + /* check if we're in a user context */ + if (unlikely(in_interrupt())) + return -EINVAL; + if (unlikely(!task || !task->pid)) + return -EINVAL; + + /* Is this a user address, or a kernel address? */ + if (!access_ok(VERIFY_WRITE, to, size)) + return -EINVAL; + + return probe_kernel_write(to, from, size); +} + +static const struct bpf_func_proto bpf_copy_to_user_proto = { + .func = bpf_copy_to_user, + .gpl_only = false, + .ret_type = RET_INTEGER, + .arg1_type = ARG_ANYTHING, + .arg2_type = ARG_PTR_TO_RAW_STACK, + .arg3_type = ARG_CONST_STACK_SIZE, +}; + /* * limited trace_printk() * only %d %u %x %ld %lu %lx %lld %llu %llx %p %s conversion specifiers allowed @@ -360,6 +389,8 @@ static const struct bpf_func_proto *tracing_func_proto(enum bpf_func_id func_id) return &bpf_get_smp_processor_id_proto; case BPF_FUNC_perf_event_read: return &bpf_perf_event_read_proto; + case BPF_FUNC_copy_to_user: + return &bpf_copy_to_user_proto; default: return NULL; } diff --git a/samples/bpf/bpf_helpers.h b/samples/bpf/bpf_helpers.h index 84e3fd9..a54a26c 100644 --- a/samples/bpf/bpf_helpers.h +++ b/samples/bpf/bpf_helpers.h @@ -41,6 +41,8 @@ static int (*bpf_perf_event_output)(void *ctx, void *map, int index, void *data, (void *) BPF_FUNC_perf_event_output; static int (*bpf_get_stackid)(void *ctx, void *map, int flags) = (void *) BPF_FUNC_get_stackid; +static int (*bpf_copy_to_user)(void *to, void *from, int size) = + (void *) BPF_FUNC_copy_to_user; /* llvm builtin functions that eBPF C program may use to * emit BPF_LD_ABS and BPF_LD_IND instructions
This allows user memory to be written to during the course of a kprobe. It shouldn't be used to implement any kind of security mechanism because of TOC-TOU attacks, but rather to debug, divert, and manipulate execution of semi-cooperative processes. Although it uses probe_kernel_write, we limit the address space the probe can write into by checking the space with access_ok. This call shouldn't sleep on any architectures based on review. It was tested with the tracex7 program on x86-64. Signed-off-by: Sargun Dhillon <sargun@sargun.me> Cc: Alexei Starovoitov <ast@kernel.org> --- include/uapi/linux/bpf.h | 11 +++++++++++ kernel/trace/bpf_trace.c | 31 +++++++++++++++++++++++++++++++ samples/bpf/bpf_helpers.h | 2 ++ 3 files changed, 44 insertions(+)