user/sven/linux.git/include/linux, branch v4.17.9

bpf: undo prog rejection on read-only lock failure

2018-07-22T13:16:09Z

commit 85782e037f8aba8922dadb24a1523ca0b82ab8bc upstream. Partially undo commit 9facc336876f ("bpf: reject any prog that failed read-only lock") since it caused a regression, that is, syzkaller was able to manage to cause a panic via fault injection deep in set_memory_ro() path by letting an allocation fail: In x86's __change_page_attr_set_clr() it was able to change the attributes of the primary mapping but not in the alias mapping via cpa_process_alias(), so the second, inner call to the __change_page_attr() via __change_page_attr_set_clr() had to split a larger page and failed in the alloc_pages() with the artifically triggered allocation error which is then propagated down to the call site. Thus, for set_memory_ro() this means that it returned with an error, but from debugging a probe_kernel_write() revealed EFAULT on that memory since the primary mapping succeeded to get changed. Therefore the subsequent hdr->locked = 0 reset triggered the panic as it was performed on read-only memory, so call-site assumptions were infact wrong to assume that it would either succeed /or/ not succeed at all since there's no such rollback in set_memory_*() calls from partial change of mappings, in other words, we're left in a state that is "half done". A later undo via set_memory_rw() is succeeding though due to matching permissions on that part (aka due to the try_preserve_large_page() succeeding). While reproducing locally with explicitly triggering this error, the initial splitting only happens on rare occasions and in real world it would additionally need oom conditions, but that said, it could partially fail. Therefore, it is definitely wrong to bail out on set_memory_ro() error and reject the program with the set_memory_*() semantics we have today. Shouldn't have gone the extra mile since no other user in tree today infact checks for any set_memory_*() errors, e.g. neither module_enable_ro() / module_disable_ro() for module RO/NX handling which is mostly default these days nor kprobes core with alloc_insn_page() / free_insn_page() as examples that could be invoked long after bootup and original 314beb9bcabf ("x86: bpf_jit_comp: secure bpf jit against spraying attacks") did neither when it got first introduced to BPF so "improving" with bailing out was clearly not right when set_memory_*() cannot handle it today. Kees suggested that if set_memory_*() can fail, we should annotate it with __must_check, and all callers need to deal with it gracefully given those set_memory_*() markings aren't "advisory", but they're expected to actually do what they say. This might be an option worth to move forward in future but would at the same time require that set_memory_*() calls from supporting archs are guaranteed to be "atomic" in that they provide rollback if part of the range fails, once that happened, the transition from RW -> RO could be made more robust that way, while subsequent RO -> RW transition /must/ continue guaranteeing to always succeed the undo part. Reported-by: syzbot+a4eb8c7766952a1ca872@syzkaller.appspotmail.com Reported-by: syzbot+d866d1925855328eac3b@syzkaller.appspotmail.com Fixes: 9facc336876f ("bpf: reject any prog that failed read-only lock") Cc: Laura Abbott Cc: Kees Cook Signed-off-by: Daniel Borkmann Acked-by: Alexei Starovoitov Signed-off-by: Alexei Starovoitov Signed-off-by: Greg Kroah-Hartman

bpf: enforce correct alignment for instructions

2018-07-22T13:16:08Z

commit 9262478220eac908ae6e168c3df2c453c87e2da3 upstream. After commit 9facc336876f ("bpf: reject any prog that failed read-only lock") offsetof(struct bpf_binary_header, image) became 3 instead of 4, breaking powerpc BPF badly, since instructions need to be word aligned. Fixes: 9facc336876f ("bpf: reject any prog that failed read-only lock") Signed-off-by: Eric Dumazet Cc: Daniel Borkmann Cc: Martin KaFai Lau Cc: Alexei Starovoitov Signed-off-by: David S. Miller Signed-off-by: Greg Kroah-Hartman

arm64: Call ARCH_WORKAROUND_2 on transitions between EL0 and EL1

2018-07-22T13:16:07Z

commit 8e2906245f1e3b0d027169d9f2e55ce0548cb96e upstream. In order for the kernel to protect itself, let's call the SSBD mitigation implemented by the higher exception level (either hypervisor or firmware) on each transition between userspace and kernel. We must take the PSCI conduit into account in order to target the right exception level, hence the introduction of a runtime patching callback. Reviewed-by: Mark Rutland Reviewed-by: Julien Grall Acked-by: Will Deacon Signed-off-by: Marc Zyngier Signed-off-by: Catalin Marinas Signed-off-by: Greg Kroah-Hartman

arm/arm64: smccc: Add SMCCC-specific return codes

2018-07-22T13:16:07Z

commit eff0e9e1078ea7dc1d794dc50e31baef984c46d7 upstream. We've so far used the PSCI return codes for SMCCC because they were extremely similar. But with the new ARM DEN 0070A specification, "NOT_REQUIRED" (-2) is clashing with PSCI's "PSCI_RET_INVALID_PARAMS". Let's bite the bullet and add SMCCC specific return codes. Users can be repainted as and when required. Acked-by: Will Deacon Reviewed-by: Mark Rutland Signed-off-by: Marc Zyngier Signed-off-by: Catalin Marinas Signed-off-by: Greg Kroah-Hartman

block: don't use blocking queue entered for recursive bio submits

2018-07-22T13:16:06Z

commit cd4a4ae4683dc2e09380118e205e057896dcda2b upstream. If we end up splitting a bio and the queue goes away between the initial submission and the later split submission, then we can block forever in blk_queue_enter() waiting for the reference to drop to zero. This will never happen, since we already hold a reference. Mark a split bio as already having entered the queue, so we can just use the live non-blocking queue enter variant. Thanks to Tetsuo Handa for the analysis. Reported-by: syzbot+c4f9cebf9d651f6e54de@syzkaller.appspotmail.com Signed-off-by: Jens Axboe Signed-off-by: Greg Kroah-Hartman

bpf: reject any prog that failed read-only lock

2018-07-22T13:16:06Z

commit 9facc336876f7ecf9edba4c67b90426fde4ec898 upstream. We currently lock any JITed image as read-only via bpf_jit_binary_lock_ro() as well as the BPF image as read-only through bpf_prog_lock_ro(). In the case any of these would fail we throw a WARN_ON_ONCE() in order to yell loudly to the log. Perhaps, to some extend, this may be comparable to an allocation where __GFP_NOWARN is explicitly not set. Added via 65869a47f348 ("bpf: improve read-only handling"), this behavior is slightly different compared to any of the other in-kernel set_memory_ro() users who do not check the return code of set_memory_ro() and friends /at all/ (e.g. in the case of module_enable_ro() / module_disable_ro()). Given in BPF this is mandatory hardening step, we want to know whether there are any issues that would leave both BPF data writable. So it happens that syzkaller enabled fault injection and it triggered memory allocation failure deep inside x86's change_page_attr_set_clr() which was triggered from set_memory_ro(). Now, there are two options: i) leaving everything as is, and ii) reworking the image locking code in order to have a final checkpoint out of the central bpf_prog_select_runtime() which probes whether any of the calls during prog setup weren't successful, and then bailing out with an error. Option ii) is a better approach since this additional paranoia avoids altogether leaving any potential W+X pages from BPF side in the system. Therefore, lets be strict about it, and reject programs in such unlikely occasion. While testing I noticed also that one bpf_prog_lock_ro() call was missing on the outer dummy prog in case of calls, e.g. in the destructor we call bpf_prog_free_deferred() on the main prog where we try to bpf_prog_unlock_free() the program, and since we go via bpf_prog_select_runtime() do that as well. Reported-by: syzbot+3b889862e65a98317058@syzkaller.appspotmail.com Reported-by: syzbot+9e762b52dd17e616a7a5@syzkaller.appspotmail.com Signed-off-by: Daniel Borkmann Acked-by: Martin KaFai Lau Signed-off-by: Alexei Starovoitov Signed-off-by: Greg Kroah-Hartman

bdi: Fix another oops in wb_workfn()

2018-07-22T13:16:06Z

commit 3ee7e8697d5860b173132606d80a9cd35e7113ee upstream. syzbot is reporting NULL pointer dereference at wb_workfn() [1] due to wb->bdi->dev being NULL. And Dmitry confirmed that wb->state was WB_shutting_down after wb->bdi->dev became NULL. This indicates that unregister_bdi() failed to call wb_shutdown() on one of wb objects. The problem is in cgwb_bdi_unregister() which does cgwb_kill() and thus drops bdi's reference to wb structures before going through the list of wbs again and calling wb_shutdown() on each of them. This way the loop iterating through all wbs can easily miss a wb if that wb has already passed through cgwb_remove_from_bdi_list() called from wb_shutdown() from cgwb_release_workfn() and as a result fully shutdown bdi although wb_workfn() for this wb structure is still running. In fact there are also other ways cgwb_bdi_unregister() can race with cgwb_release_workfn() leading e.g. to use-after-free issues: CPU1 CPU2 cgwb_bdi_unregister() cgwb_kill(*slot); cgwb_release() queue_work(cgwb_release_wq, &wb->release_work); cgwb_release_workfn() wb = list_first_entry(&bdi->wb_list, ...) spin_unlock_irq(&cgwb_lock); wb_shutdown(wb); ... kfree_rcu(wb, rcu); wb_shutdown(wb); -> oops use-after-free We solve these issues by synchronizing writeback structure shutdown from cgwb_bdi_unregister() with cgwb_release_workfn() using a new mutex. That way we also no longer need synchronization using WB_shutting_down as the mutex provides it for CONFIG_CGROUP_WRITEBACK case and without CONFIG_CGROUP_WRITEBACK wb_shutdown() can be called only once from bdi_unregister(). Reported-by: syzbot Acked-by: Tejun Heo Signed-off-by: Jan Kara Signed-off-by: Jens Axboe Signed-off-by: Greg Kroah-Hartman

net/mlx5: E-Switch, Avoid setup attempt if not being e-switch manager

2018-07-22T13:16:00Z

[ Upstream commit 0efc8562491b7d36f6bbc4fbc8f3348cb6641e9c ] In smartnic env, the host (PF) driver might not be an e-switch manager, hence the FW will err on driver attempts to deal with setting/unsetting the eswitch and as a result the overall setup of sriov will fail. Fix that by avoiding the operation if e-switch management is not allowed for this driver instance. While here, move to use the correct name for the esw manager capability name. Fixes: 81848731ff40 ('net/mlx5: E-Switch, Add SR-IOV (FDB) support') Signed-off-by: Or Gerlitz Reported-by: Guy Kushnir Reviewed-by: Eli Cohen Tested-by: Eli Cohen Signed-off-by: Saeed Mahameed Signed-off-by: Greg Kroah-Hartman

net: fix use-after-free in GRO with ESP

2018-07-22T13:15:59Z

[ Upstream commit 603d4cf8fe095b1ee78f423d514427be507fb513 ] Since the addition of GRO for ESP, gro_receive can consume the skb and return -EINPROGRESS. In that case, the lower layer GRO handler cannot touch the skb anymore. Commit 5f114163f2f5 ("net: Add a skb_gro_flush_final helper.") converted some of the gro_receive handlers that can lead to ESP's gro_receive so that they wouldn't access the skb when -EINPROGRESS is returned, but missed other spots, mainly in tunneling protocols. This patch finishes the conversion to using skb_gro_flush_final(), and adds a new helper, skb_gro_flush_final_remcsum(), used in VXLAN and GUE. Fixes: 5f114163f2f5 ("net: Add a skb_gro_flush_final helper.") Signed-off-by: Sabrina Dubroca Reviewed-by: Stefano Brivio Signed-off-by: David S. Miller Signed-off-by: Greg Kroah-Hartman

atm: Preserve value of skb->truesize when accounting to vcc

2018-07-22T13:15:58Z

[ Upstream commit 9bbe60a67be5a1c6f79b3c9be5003481a50529ff ] ATM accounts for in-flight TX packets in sk_wmem_alloc of the VCC on which they are to be sent. But it doesn't take ownership of those packets from the sock (if any) which originally owned them. They should remain owned by their actual sender until they've left the box. There's a hack in pskb_expand_head() to avoid adjusting skb->truesize for certain skbs, precisely to avoid messing up sk_wmem_alloc accounting. Ideally that hack would cover the ATM use case too, but it doesn't — skbs which aren't owned by any sock, for example PPP control frames, still get their truesize adjusted when the low-level ATM driver adds headroom. This has always been an issue, it seems. The truesize of a packet increases, and sk_wmem_alloc on the VCC goes negative. But this wasn't for normal traffic, only for control frames. So I think we just got away with it, and we probably needed to send 2GiB of LCP echo frames before the misaccounting would ever have caused a problem and caused atm_may_send() to start refusing packets. Commit 14afee4b609 ("net: convert sock.sk_wmem_alloc from atomic_t to refcount_t") did exactly what it was intended to do, and turned this mostly-theoretical problem into a real one, causing PPPoATM to fail immediately as sk_wmem_alloc underflows and atm_may_send() *immediately* starts refusing to allow new packets. The least intrusive solution to this problem is to stash the value of skb->truesize that was accounted to the VCC, in a new member of the ATM_SKB(skb) structure. Then in atm_pop_raw() subtract precisely that value instead of the then-current value of skb->truesize. Fixes: 158f323b9868 ("net: adjust skb->truesize in pskb_expand_head()") Signed-off-by: David Woodhouse Tested-by: Kevin Darbyshire-Bryant Signed-off-by: David S. Miller Signed-off-by: Greg Kroah-Hartman