Chuck Lever [Mon, 4 Dec 2017 19:04:04 +0000 (14:04 -0500)]
xprtrdma: Spread reply processing over more CPUs
Commit d8f532d20ee4 ("xprtrdma: Invoke rpcrdma_reply_handler
directly from RECV completion") introduced a performance regression
for NFS I/O small enough to not need memory registration. In multi-
threaded benchmarks that generate primarily small I/O requests,
IOPS throughput is reduced by nearly a third. This patch restores
the previous level of throughput.
Because workqueues are typically BOUND (in particular ib_comp_wq,
nfsiod_workqueue, and rpciod_workqueue), NFS/RDMA workloads tend
to aggregate on the CPU that is handling Receive completions.
The usual approach to addressing this problem is to create a QP
and CQ for each CPU, and then schedule transactions on the QP
for the CPU where you want the transaction to complete. The
transaction then does not require an extra context switch during
completion to end up on the same CPU where the transaction was
started.
This approach doesn't work for the Linux NFS/RDMA client because
currently the Linux NFS client does not support multiple connections
per client-server pair, and the RDMA core API does not make it
straightforward for ULPs to determine which CPU is responsible for
handling Receive completions for a CQ.
So for the moment, record the CPU number in the rpcrdma_req before
the transport sends each RPC Call. Then during Receive completion,
queue the RPC completion on that same CPU.
Additionally, move all RPC completion processing to the deferred
handler so that even RPCs with simple small replies complete on
the CPU that sent the corresponding RPC Call.
Fixes: d8f532d20ee4 ("xprtrdma: Invoke rpcrdma_reply_handler ...") Signed-off-by: Chuck Lever <[email protected]> Signed-off-by: Anna Schumaker <[email protected]>
Scott Mayhew [Tue, 5 Dec 2017 18:55:44 +0000 (13:55 -0500)]
nfs: fix a deadlock in nfs client initialization
The following deadlock can occur between a process waiting for a client
to initialize in while walking the client list during nfsv4 server trunking
detection and another process waiting for the nfs_clid_init_mutex so it
can initialize that client:
Process 1 Process 2
--------- ---------
spin_lock(&nn->nfs_client_lock);
list_add_tail(&CLIENTA->cl_share_link,
&nn->nfs_client_list);
spin_unlock(&nn->nfs_client_lock);
spin_lock(&nn->nfs_client_lock);
list_add_tail(&CLIENTB->cl_share_link,
&nn->nfs_client_list);
spin_unlock(&nn->nfs_client_lock);
mutex_lock(&nfs_clid_init_mutex);
nfs41_walk_client_list(clp, result, cred);
nfs_wait_client_init_complete(CLIENTA);
(waiting for nfs_clid_init_mutex)
Make sure nfs_match_client() only evaluates clients that have completed
initialization in order to prevent that deadlock.
This patch also fixes v4.0 trunking behavior by not marking the client
NFS_CS_READY until the clientid has been confirmed.
====================
net: sched: Make qdisc offload uapi uniform
Several qdiscs can already be offloaded to hardware, but there's an
inconsistecy in regard to the uapi through which they indicate such
an offload is taking place - indication is passed to the user via
TCA_OPTIONS where each qdisc retains private logic for setting it.
The recent addition of offloading to RED in 602f3baf2218 ("net_sch: red: Add offload ability to RED qdisc") caused
the addition of yet another uapi field for this purpose -
TC_RED_OFFLOADED.
For clarity and prevention of bloat in the uapi we want to eliminate
said added uapi, replacing it with a common mechanism that can be used
to reflect offload status of the various qdiscs.
The first patch introduces TCA_HW_OFFLOAD as the generic message meant
for this purpose. The second changes the current RED implementation into
setting the internal bits necessary for passing it, and the third removes
TC_RED_OFFLOADED as its no longer needed.
====================
Yuval Mintz [Thu, 14 Dec 2017 13:54:31 +0000 (15:54 +0200)]
pkt_sched: Remove TC_RED_OFFLOADED from uapi
Following the previous patch, RED is now using the new uniform uapi
for indicating it's offloaded. As a result, TC_RED_OFFLOADED is no
longer utilized by kernel and can be removed [as it's still not
part of any stable release].
Fixes: 602f3baf2218 ("net_sch: red: Add offload ability to RED qdisc") Signed-off-by: Yuval Mintz <[email protected]> Acked-by: Jiri Pirko <[email protected]> Signed-off-by: David S. Miller <[email protected]>
Yuval Mintz [Thu, 14 Dec 2017 13:54:29 +0000 (15:54 +0200)]
net: sched: Add TCA_HW_OFFLOAD
Qdiscs can be offloaded to HW, but current implementation isn't uniform.
Instead, qdiscs either pass information about offload status via their
TCA_OPTIONS or omit it altogether.
Introduce a new attribute - TCA_HW_OFFLOAD that would form a uniform
uAPI for the offloading status of qdiscs.
The patchset contains important hardware fix for machines with large MRRS
and couple of improvement in stats and capabilities reporting
patch v3:
- Fixed patch #7 after Andrew's finding. NIC level stats actually
have to be cleaned only on hw struct creation (and this is done
in kzalloc). On each hwinit we only have to reset link state
to make sure hw stats update will not increment nic stats during init.
patch v2:
- split into more detailed commits
Comment from David on wrong defines case will be submitted separately later
====================
Igor Russkikh [Thu, 14 Dec 2017 09:34:46 +0000 (12:34 +0300)]
net: aquantia: Update hw counters on hw init
On very first start we should read out current HW counter values
to make diff based calculations later.
This also should be done each time NIC gets down/up or wakes up
after sleep state. We reset link state explicitly to prevent diffs
from being summed this first time.
Igor Russkikh [Thu, 14 Dec 2017 09:34:43 +0000 (12:34 +0300)]
net: aquantia: Fill ndev stat couters from hardware
Originally they were filled from ring sw counters.
These sometimes incorrectly calculate byte and packet amounts
when using LRO/LSO and jumboframes. Filling ndev counters from
hardware makes them precise.
Igor Russkikh [Thu, 14 Dec 2017 09:34:42 +0000 (12:34 +0300)]
net: aquantia: Extend stat counters to 64bit values
Device hardware provides only 32bit counters. Using these directly
causes byte counters to overflow soon. A separate nic level structure
with 64 bit counters is now used to collect incrementally all the stats
and report these counters to ethtool stats and ndev stats.
Igor Russkikh [Thu, 14 Dec 2017 09:34:41 +0000 (12:34 +0300)]
net: aquantia: Fix hardware DMA stream overload on large MRRS
Systems with large MRRS on device (2K, 4K) with high data rates and/or
large MTU, atlantic observes DMA packet buffer overflow. On some systems
that causes PCIe transaction errors, hardware NMIs or datapath freeze.
This patch
1) Limits MRRS from device side to 2K (thats maximum our hardware supports)
2) Limit maximum size of outstanding TX DMA data read requests. This makes
hardware buffers running fine.
Igor Russkikh [Thu, 14 Dec 2017 09:34:40 +0000 (12:34 +0300)]
net: aquantia: Fix actual speed capabilities reporting
Different hardware device Ids correspond to different maximum speed
available. Extra checks were added for devices D108 and D109 to
remove unsupported speeds from these device capabilities list.
====================
Two fixes that deal with buggy usage of bpf_helper_changes_pkt_data()
in the sense that they also reload cached skb data when there's no
skb context but xdp one, for example. A fix where skb meta data is
reloaded out of the wrong register on helper call, rest is test cases
and making sure on verifier side that there's always the guarantee
that ctx sits in r1. Thanks!
====================
Daniel Borkmann [Thu, 14 Dec 2017 20:07:27 +0000 (21:07 +0100)]
bpf: add test case for ld_abs and helper changing pkt data
Add a test that i) uses LD_ABS, ii) zeroing R6 before call, iii) calls
a helper that triggers reload of cached skb data, iv) uses LD_ABS again.
It's added for test_bpf in order to do runtime testing after JITing as
well as test_verifier to test that the sequence is allowed.
Daniel Borkmann [Thu, 14 Dec 2017 20:07:26 +0000 (21:07 +0100)]
bpf, sparc: fix usage of wrong reg for load_skb_regs after call
When LD_ABS/IND is used in the program, and we have a BPF helper
call that changes packet data (bpf_helper_changes_pkt_data() returns
true), then in case of sparc JIT, we try to reload cached skb data
from bpf2sparc[BPF_REG_6]. However, there is no such guarantee or
assumption that skb sits in R6 at this point, all helpers changing
skb data only have a guarantee that skb sits in R1. Therefore,
store BPF R1 in L7 temporarily and after procedure call use L7 to
reload cached skb data. skb sitting in R6 is only true at the time
when LD_ABS/IND is executed.
Daniel Borkmann [Thu, 14 Dec 2017 20:07:25 +0000 (21:07 +0100)]
bpf: guarantee r1 to be ctx in case of bpf_helper_changes_pkt_data
Some JITs don't cache skb context on stack in prologue, so when
LD_ABS/IND is used and helper calls yield bpf_helper_changes_pkt_data()
as true, then they temporarily save/restore skb pointer. However,
the assumption that skb always has to be in r1 is a bit of a
gamble. Right now it turned out to be true for all helpers listed
in bpf_helper_changes_pkt_data(), but lets enforce that from verifier
side, so that we make this a guarantee and bail out if the func
proto is misconfigured in future helpers.
In case of BPF helper calls from cBPF, bpf_helper_changes_pkt_data()
is completely unrelevant here (since cBPF is context read-only) and
therefore always false.
Daniel Borkmann [Thu, 14 Dec 2017 20:07:24 +0000 (21:07 +0100)]
bpf, ppc64: do not reload skb pointers in non-skb context
The assumption of unconditionally reloading skb pointers on
BPF helper calls where bpf_helper_changes_pkt_data() holds
true is wrong. There can be different contexts where the helper
would enforce a reload such as in case of XDP. Here, we do
have a struct xdp_buff instead of struct sk_buff as context,
thus this will access garbage.
JITs only ever need to deal with cached skb pointer reload
when ld_abs/ind was seen, therefore guard the reload behind
SEEN_SKB.
Daniel Borkmann [Thu, 14 Dec 2017 20:07:23 +0000 (21:07 +0100)]
bpf, s390x: do not reload skb pointers in non-skb context
The assumption of unconditionally reloading skb pointers on
BPF helper calls where bpf_helper_changes_pkt_data() holds
true is wrong. There can be different contexts where the
BPF helper would enforce a reload such as in case of XDP.
Here, we do have a struct xdp_buff instead of struct sk_buff
as context, thus this will access garbage.
JITs only ever need to deal with cached skb pointer reload
when ld_abs/ind was seen, therefore guard the reload behind
SEEN_SKB only. Tested on s390x.
Fixes: 9db7f2b81880 ("s390/bpf: recache skb->data/hlen for skb_vlan_push/pop") Signed-off-by: Daniel Borkmann <[email protected]> Acked-by: Alexei Starovoitov <[email protected]> Cc: Michael Holzheu <[email protected]> Signed-off-by: Alexei Starovoitov <[email protected]>
Willem de Bruijn [Wed, 13 Dec 2017 19:41:06 +0000 (14:41 -0500)]
sock: free skb in skb_complete_tx_timestamp on error
skb_complete_tx_timestamp must ingest the skb it is passed. Call
kfree_skb if the skb cannot be enqueued.
Fixes: b245be1f4db1 ("net-timestamp: no-payload only sysctl") Fixes: 9ac25fc06375 ("net: fix socket refcounting in skb_complete_tx_timestamp()") Reported-by: Richard Cochran <[email protected]> Signed-off-by: Willem de Bruijn <[email protected]> Reviewed-by: Eric Dumazet <[email protected]> Signed-off-by: David S. Miller <[email protected]>
David S. Miller [Fri, 15 Dec 2017 16:29:44 +0000 (11:29 -0500)]
Merge branch 's390-fixes'
Julian Wiedmann says:
====================
s390/qeth: fixes 2017-12-13
some more patches for 4.15, that fix multiple issues with IP Takeover
configuration in qeth.
Please queue them up for stable kernels as well (4.9 and newer).
====================
Julian Wiedmann [Wed, 13 Dec 2017 17:56:32 +0000 (18:56 +0100)]
s390/qeth: update takeover IPs after configuration change
Any modification to the takeover IP-ranges requires that we re-evaluate
which IP addresses are takeover-eligible. Otherwise we might do takeover
for some addresses when we no longer should, or vice-versa.
Julian Wiedmann [Wed, 13 Dec 2017 17:56:31 +0000 (18:56 +0100)]
s390/qeth: lock IP table while applying takeover changes
Modifying the flags of an IP addr object needs to be protected against
eg. concurrent removal of the same object from the IP table.
Fixes: 5f78e29ceebf ("qeth: optimize IP handling in rx_mode callback") Signed-off-by: Julian Wiedmann <[email protected]> Signed-off-by: David S. Miller <[email protected]>
Julian Wiedmann [Wed, 13 Dec 2017 17:56:30 +0000 (18:56 +0100)]
s390/qeth: don't apply takeover changes to RXIP
When takeover is switched off, current code clears the 'TAKEOVER' flag on
all IPs. But the flag is also used for RXIP addresses, and those should
not be affected by the takeover mode.
Fix the behaviour by consistenly applying takover logic to NORMAL
addresses only.
Julian Wiedmann [Wed, 13 Dec 2017 17:56:29 +0000 (18:56 +0100)]
s390/qeth: apply takeover changes when mode is toggled
Just as for an explicit enable/disable, toggling the takeover mode also
requires that the IP addresses get updated. Otherwise all IPs that were
added to the table before the mode-toggle, get registered with the old
settings.
Will Deacon [Fri, 15 Dec 2017 16:07:22 +0000 (16:07 +0000)]
arm64: fpsimd: Fix copying of FP state from signal frame into task struct
Commit 9de52a755cfb6da5 ("arm64: fpsimd: Fix failure to restore FPSIMD
state after signals") fixed an issue reported in our FPSIMD signal
restore code but inadvertently introduced another issue which tends to
manifest as random SEGVs in userspace.
The problem is that when we copy the struct fpsimd_state from the kernel
stack (populated from the signal frame) into the struct held in the
current thread_struct, we blindly copy uninitialised stack into the
"cpu" field, which means that context-switching of the FP registers is
no longer reliable.
This patch fixes the problem by copying only the user_fpsimd member of
struct fpsimd_state. We should really rework the function prototypes
to take struct user_fpsimd_state * instead, but let's just get this
fixed for now.
Cc: Dave Martin <[email protected]> Fixes: 9de52a755cfb6da5 ("arm64: fpsimd: Fix failure to restore FPSIMD state after signals") Reported-by: Geert Uytterhoeven <[email protected]> Signed-off-by: Will Deacon <[email protected]>
Yuval Mintz [Fri, 15 Dec 2017 07:44:21 +0000 (08:44 +0100)]
mlxsw: spectrum: Disable MAC learning for ovs port
Learning is currently enabled for ports which are OVS slaves -
even though OVS doesn't need this indication.
Since we're not associating a fid with the port, HW would continuously
notify driver of learned [& aged] MACs which would be logged as errors.
Steven Rostedt [Sat, 2 Dec 2017 18:04:54 +0000 (13:04 -0500)]
sched/rt: Do not pull from current CPU if only one CPU to pull
Daniel Wagner reported a crash on the BeagleBone Black SoC.
This is a single CPU architecture, and does not have a functional
arch_send_call_function_single_ipi() implementation which can crash
the kernel if that is called.
As it only has one CPU, it shouldn't be called, but if the kernel is
compiled for SMP, the push/pull RT scheduling logic now calls it for
irq_work if the one CPU is overloaded, it can use that function to call
itself and crash the kernel.
Ideally, we should disable the SCHED_FEAT(RT_PUSH_IPI) if the system
only has a single CPU. But SCHED_FEAT is a constant if sched debugging
is turned off. Another fix can also be used, and this should also help
with normal SMP machines. That is, do not initiate the pull code if
there's only one RT overloaded CPU, and that CPU happens to be the
current CPU that is scheduling in a lower priority task.
Even on a system with many CPUs, if there's many RT tasks waiting to
run on a single CPU, and that CPU schedules in another RT task of lower
priority, it will initiate the PULL logic in case there's a higher
priority RT task on another CPU that is waiting to run. But if there is
no other CPU with waiting RT tasks, it will initiate the RT pull logic
on itself (as it still has RT tasks waiting to run). This is a wasted
effort.
Not only does this help with SMP code where the current CPU is the only
one with RT overloaded tasks, it should also solve the issue that
Daniel encountered, because it will prevent the PULL logic from
executing, as there's only one CPU on the system, and the check added
here will cause it to exit the RT pull code.
Song Liu [Fri, 15 Dec 2017 01:17:56 +0000 (17:17 -0800)]
xdp: linearize skb in netif_receive_generic_xdp()
In netif_receive_generic_xdp(), it is necessary to linearize all
nonlinear skb. However, in current implementation, skb with
troom <= 0 are not linearized. This patch fixes this by calling
skb_linearize() for all nonlinear skb.
Andy Lutomirski [Thu, 14 Dec 2017 21:19:07 +0000 (13:19 -0800)]
x86/power: Make restore_processor_context() sane
My previous attempt to fix a couple of bugs in __restore_processor_context():
5b06bbcfc2c6 ("x86/power: Fix some ordering bugs in __restore_processor_context()")
... introduced yet another bug, breaking suspend-resume.
Rather than trying to come up with a minimal fix, let's try to clean it up
for real. This patch fixes quite a few things:
- The old code saved a nonsensical subset of segment registers.
The only registers that need to be saved are those that contain
userspace state or those that can't be trivially restored without
percpu access working. (On x86_32, we can restore percpu access
by writing __KERNEL_PERCPU to %fs. On x86_64, it's easier to
save and restore the kernel's GSBASE.) With this patch, we
restore hardcoded values to the kernel state where applicable and
explicitly restore the user state after fixing all the descriptor
tables.
- We used to use an unholy mix of inline asm and C helpers for
segment register access. Let's get rid of the inline asm.
This fixes the reported s2ram hangs and make the code all around
more logical.
Andy Lutomirski [Thu, 14 Dec 2017 21:19:06 +0000 (13:19 -0800)]
x86/power/32: Move SYSENTER MSR restoration to fix_processor_context()
x86_64 restores system call MSRs in fix_processor_context(), and
x86_32 restored them along with segment registers. The 64-bit
variant makes more sense, so move the 32-bit code to match the
64-bit code.
Andy Lutomirski [Thu, 14 Dec 2017 21:19:05 +0000 (13:19 -0800)]
x86/power/64: Use struct desc_ptr for the IDT in struct saved_context
x86_64's saved_context nonsensically used separate idt_limit and
idt_base fields and then cast &idt_limit to struct desc_ptr *.
This was correct (with -fno-strict-aliasing), but it's confusing,
served no purpose, and required #ifdeffery. Simplify this by
using struct desc_ptr directly.
Linus Torvalds [Fri, 15 Dec 2017 02:25:03 +0000 (18:25 -0800)]
Merge tag 'pm-4.15-rc4' of git://git.kernel.org/pub/scm/linux/kernel/git/rafael/linux-pm
Pull power management fix from Rafael Wysocki:
"This fixes an issue in two recent commits that may cause
pm_runtime_enable() to be called for too many times for some devices
during the "thaw" transition belonging to hibernation"
* tag 'pm-4.15-rc4' of git://git.kernel.org/pub/scm/linux/kernel/git/rafael/linux-pm:
PM / sleep: Avoid excess pm_runtime_enable() calls in device_resume()
Linus Torvalds [Fri, 15 Dec 2017 02:21:33 +0000 (18:21 -0800)]
Merge tag 'trace-v4.15-rc1' of git://git.kernel.org/pub/scm/linux/kernel/git/rostedt/linux-trace
Pull tracing fixes from Steven Rostedt:
"Various fix-ups:
- comment fixes
- build fix
- better memory alloction (don't use NR_CPUS)
- configuration fix
- build warning fix
- enhanced callback parameter (to simplify users of trace hooks)
- give up on stack tracing when RCU isn't watching (it's a lost
cause)"
* tag 'trace-v4.15-rc1' of git://git.kernel.org/pub/scm/linux/kernel/git/rostedt/linux-trace:
tracing: Have stack trace not record if RCU is not watching
tracing: Pass export pointer as argument to ->write()
ring-buffer: Remove unused function __rb_data_page_index()
tracing: make PREEMPTIRQ_EVENTS depend on TRACING
tracing: Allocate mask_str buffer dynamically
tracing: always define trace_{irq,preempt}_{enable_disable}
tracing: Fix code comments in trace.c
tracing: Have stack trace not record if RCU is not watching
The stack tracer records a stack dump whenever it sees a stack usage that is
more than what it ever saw before. This can happen at any function that is
being traced. If it happens when the CPU is going idle (or other strange
locations), RCU may not be watching, and in this case, the recording of the
stack trace will trigger a warning. There's been lots of efforts to make
hacks to allow stack tracing to proceed even if RCU is not watching, but
this only causes more issues to appear. Simply do not trace a stack if RCU
is not watching. It probably isn't a bad stack anyway.
Linus Torvalds [Fri, 15 Dec 2017 01:02:39 +0000 (17:02 -0800)]
Merge tag 'pci-v4.15-fixes-1' of git://git.kernel.org/pub/scm/linux/kernel/git/helgaas/pci
Pull PCI fixes from Bjorn Helgaas:
- add a pci_get_domain_bus_and_slot() stub for the CONFIG_PCI=n case to
avoid build breakage in the v4.16 merge window if a
pci_get_bus_and_slot() -> pci_get_domain_bus_and_slot() patch gets
merged before the PCI tree (Randy Dunlap)
- fix an AMD boot regression in the 64bit BAR support added in v4.15
(Christian König)
- fix an R-Car use-after-free that causes a crash if no PCIe card is
present (Geert Uytterhoeven)
* tag 'pci-v4.15-fixes-1' of git://git.kernel.org/pub/scm/linux/kernel/git/helgaas/pci:
PCI: rcar: Fix use-after-free in probe error path
x86/PCI: Only enable a 64bit BAR on single-socket AMD Family 15h
x86/PCI: Fix infinite loop in search for 64bit BAR placement
PCI: Add pci_get_domain_bus_and_slot() stub
Linus Torvalds [Fri, 15 Dec 2017 00:35:20 +0000 (16:35 -0800)]
Merge branch 'akpm' (patches from Andrew)
Merge misc fixes from Andrew Morton:
"17 fixes"
* emailed patches from Andrew Morton <[email protected]>:
arch: define weak abort()
mm, oom_reaper: fix memory corruption
kernel: make groups_sort calling a responsibility group_info allocators
mm/frame_vector.c: release a semaphore in 'get_vaddr_frames()'
tools/slabinfo-gnuplot: force to use bash shell
kcov: fix comparison callback signature
mm/slab.c: do not hash pointers when debugging slab
mm/page_alloc.c: avoid excessive IRQ disabled times in free_unref_page_list()
mm/memory.c: mark wp_huge_pmd() inline to prevent build failure
scripts/faddr2line: fix CROSS_COMPILE unset error
Documentation/vm/zswap.txt: update with same-value filled page feature
exec: avoid gcc-8 warning for get_task_comm
autofs: fix careless error in recent commit
string.h: workaround for increased stack usage
mm/kmemleak.c: make cond_resched() rate-limiting more efficient
lib/rbtree,drm/mm: add rbtree_replace_node_cached()
include/linux/idr.h: add #include <linux/bug.h>
Sudip Mukherjee [Thu, 14 Dec 2017 23:33:19 +0000 (15:33 -0800)]
arch: define weak abort()
gcc toggle -fisolate-erroneous-paths-dereference (default at -O2
onwards) isolates faulty code paths such as null pointer access, divide
by zero etc. If gcc port doesnt implement __builtin_trap, an abort() is
generated which causes kernel link error.
In this case, gcc is generating abort due to 'divide by zero' in
lib/mpi/mpih-div.c.
Currently 'frv' and 'arc' are failing. Previously other arch was also
broken like m32r was fixed by commit d22e3d69ee1a ("m32r: fix build
failure").
Let's define this weak function which is common for all arch and fix the
problem permanently. We can even remove the arch specific 'abort' after
this is done.
Tetsuo Handa has noticed that the synchronization inside exit_mmap is
insufficient. We only synchronize with the oom reaper if
tsk_is_oom_victim which is not true if the final __mmput is called from
a different context than the oom victim exit path. This can trivially
happen from context of any task which has grabbed mm reference (e.g. to
read /proc/<pid>/ file which requires mm etc.).
Fix this issue by providing a new mm_is_oom_victim() helper which
operates on the mm struct rather than a task. Any context which
operates on a remote mm struct should use this helper in place of
tsk_is_oom_victim. The flag is set in mark_oom_victim and never cleared
so it is stable in the exit_mmap path.
kernel: make groups_sort calling a responsibility group_info allocators
In testing, we found that nfsd threads may call set_groups in parallel
for the same entry cached in auth.unix.gid, racing in the call of
groups_sort, corrupting the groups for that entry and leading to
permission denials for the client.
This patch:
- Make groups_sort globally visible.
- Move the call to groups_sort to the modifiers of group_info
- Remove the call to groups_sort from set_groups
mm/slab.c: do not hash pointers when debugging slab
If CONFIG_DEBUG_SLAB/CONFIG_DEBUG_SLAB_LEAK are enabled, the slab code
prints extra debug information when e.g. corruption is detected. This
includes pointers, which are not very useful when hashed.
Fix this by using %px to print unhashed pointers instead where it makes
sense, and by removing the printing of a last user pointer referring to
code.
Lucas Stach [Thu, 14 Dec 2017 23:32:55 +0000 (15:32 -0800)]
mm/page_alloc.c: avoid excessive IRQ disabled times in free_unref_page_list()
Since commit 9cca35d42eb6 ("mm, page_alloc: enable/disable IRQs once
when freeing a list of pages") we see excessive IRQ disabled times of up
to 25ms on an embedded ARM system (tracing overhead included).
This is due to graphics buffers being freed back to the system via
release_pages(). Graphics buffers can be huge, so it's not hard to hit
cases where the list of pages to free has 2048 entries. Disabling IRQs
while freeing all those pages is clearly not a good idea.
Introduce a batch limit, which allows IRQ servicing once every few
pages. The batch count is the same as used in other parts of the MM
subsystem when dealing with IRQ disabled regions.
mm/memory.c: mark wp_huge_pmd() inline to prevent build failure
With gcc 4.1.2:
mm/memory.o: In function `wp_huge_pmd':
memory.c:(.text+0x9b4): undefined reference to `do_huge_pmd_wp_page'
Interestingly, wp_huge_pmd() is emitted in the assembler output, but
never called.
Apparently replacing the call to pmd_write() in __handle_mm_fault() by a
call to the more complex pmd_access_permitted() reduced the ability of
the compiler to remove unused code.
Fix this by marking wp_huge_pmd() inline, like was done in commit 91a90140f998 ("mm/memory.c: mark create_huge_pmd() inline to prevent
build failure") for a similar problem.
Documentation/vm/zswap.txt: update with same-value filled page feature
Update zswap document with details on same-value filled pages
identification feature. The usage of zswap.same_filled_pages_enabled
module parameter is explained.
Arnd Bergmann [Thu, 14 Dec 2017 23:32:41 +0000 (15:32 -0800)]
exec: avoid gcc-8 warning for get_task_comm
gcc-8 warns about using strncpy() with the source size as the limit:
fs/exec.c:1223:32: error: argument to 'sizeof' in 'strncpy' call is the same expression as the source; did you mean to use the size of the destination? [-Werror=sizeof-pointer-memaccess]
This is indeed slightly suspicious, as it protects us from source
arguments without NUL-termination, but does not guarantee that the
destination is terminated.
This keeps the strncpy() to ensure we have properly padded target
buffer, but ensures that we use the correct length, by passing the
actual length of the destination buffer as well as adding a build-time
check to ensure it is exactly TASK_COMM_LEN.
There are only 23 callsites which I all reviewed to ensure this is
currently the case. We could get away with doing only the check or
passing the right length, but it doesn't hurt to do both.
NeilBrown [Thu, 14 Dec 2017 23:32:38 +0000 (15:32 -0800)]
autofs: fix careless error in recent commit
Commit ecc0c469f277 ("autofs: don't fail mount for transient error") was
meant to replace an 'if' with a 'switch', but instead added the 'switch'
leaving the case in place.
Arnd Bergmann [Thu, 14 Dec 2017 23:32:34 +0000 (15:32 -0800)]
string.h: workaround for increased stack usage
The hardened strlen() function causes rather large stack usage in at
least one file in the kernel, in particular when CONFIG_KASAN is
enabled:
drivers/media/usb/em28xx/em28xx-dvb.c: In function 'em28xx_dvb_init':
drivers/media/usb/em28xx/em28xx-dvb.c:2062:1: error: the frame size of 3256 bytes is larger than 204 bytes [-Werror=frame-larger-than=]
Analyzing this problem led to the discovery that gcc fails to merge the
stack slots for the i2c_board_info[] structures after we strlcpy() into
them, due to the 'noreturn' attribute on the source string length check.
I reported this as a gcc bug, but it is unlikely to get fixed for gcc-8,
since it is relatively easy to work around, and it gets triggered
rarely. An earlier workaround I did added an empty inline assembly
statement before the call to fortify_panic(), which works surprisingly
well, but is really ugly and unintuitive.
This is a new approach to the same problem, this time addressing it by
not calling the 'extern __real_strnlen()' function for string constants
where __builtin_strlen() is a compile-time constant and therefore known
to be safe.
We do this by checking if the last character in the string is a
compile-time constant '\0'. If it is, we can assume that strlen() of
the string is also constant.
As a side-effect, this should also improve the object code output for
any other call of strlen() on a string constant.
Andrew Morton [Thu, 14 Dec 2017 23:32:31 +0000 (15:32 -0800)]
mm/kmemleak.c: make cond_resched() rate-limiting more efficient
Commit bde5f6bc68db ("kmemleak: add scheduling point to
kmemleak_scan()") tries to rate-limit the frequency of cond_resched()
calls, but does it in a way which might incur an expensive division
operation in the inner loop. Simplify this.
Add a variant of rbtree_replace_node() that maintains the leftmost cache
of struct rbtree_root_cached when replacing nodes within the rbtree.
As drm_mm is the only rb_replace_node() being used on an interval tree,
the mistake looks fairly self-contained. Furthermore the only user of
drm_mm_replace_node() is its testsuite...
Wei Wang [Thu, 14 Dec 2017 23:32:24 +0000 (15:32 -0800)]
include/linux/idr.h: add #include <linux/bug.h>
The <linux/bug.h> was removed from radix-tree.h by commit f5bba9d11a25
("include/linux/radix-tree.h: remove unneeded #include <linux/bug.h>").
Since that commit, tools/testing/radix-tree/ couldn't pass compilation
due to tools/testing/radix-tree/idr.c:17: undefined reference to
WARN_ON_ONCE. This patch adds the bug.h header to idr.h to solve the
issue.
Linus Torvalds [Thu, 14 Dec 2017 19:45:53 +0000 (11:45 -0800)]
Merge tag 'drm-misc-fixes-2017-12-14' of git://anongit.freedesktop.org/drm/drm-misc
Pull drm fixes from Daniel Vetter:
- two fixes for new core features
- a corner case fix for the connnector_iter fix from last week (this
one is cc: stable)
- one vc4 fix
* tag 'drm-misc-fixes-2017-12-14' of git://anongit.freedesktop.org/drm/drm-misc:
drm/drm_lease: Prevent deadlock in case drm_lease_create() fails
drm: rework delayed connector cleanup in connector_iter
drm: Update edid-derived drm_display_info fields at edid property set [v2]
drm/vc4: Release fence after signalling
Mark Rutland [Tue, 12 Dec 2017 13:45:50 +0000 (13:45 +0000)]
virtio_mmio: fix devm cleanup
Recent rework of the virtio_mmio probe/remove paths balanced a
devm_ioremap() with an iounmap() rather than its devm variant. This ends
up corrupting the devm datastructures, and results in the following
boot-time splat on arm64 under QEMU 2.9.0:
To fix this, we can simply rip out the explicit cleanup that the devm
infrastructure will do for us when our probe function returns an error
code, or when our remove function returns.
We only need to ensure that we call put_device() if a call to
register_virtio_device() fails in the probe path.
Dave Martin [Thu, 14 Dec 2017 14:03:44 +0000 (14:03 +0000)]
arm64/sve: Report SVE to userspace via CPUID only if supported
Currently, the SVE field in ID_AA64PFR0_EL1 is visible
unconditionally to userspace via the CPU ID register emulation,
irrespective of the kernel config. This means that if a kernel
configured with CONFIG_ARM64_SVE=n is run on SVE-capable hardware,
userspace will see SVE reported as present in the ID regs even
though the kernel forbids execution of SVE instructions.
This patch makes the exposure of the SVE field in ID_AA64PFR0_EL1
conditional on CONFIG_ARM64_SVE=y.
Since future architecture features are likely to encounter a
similar requirement, this patch adds a suitable helper macros for
use when declaring config-conditional ID register fields.
Mark Rutland [Wed, 13 Dec 2017 11:45:42 +0000 (11:45 +0000)]
arm64: fix CONFIG_DEBUG_WX address reporting
In ptdump_check_wx(), we pass walk_pgd() a start address of 0 (rather
than VA_START) for the init_mm. This means that any reported W&X
addresses are offset by VA_START, which is clearly wrong and can make
them appear like userspace addresses.
Fix this by telling the ptdump code that we're walking init_mm starting
at VA_START. We don't need to update the addr_markers, since these are
still valid bounds regardless.
drm_mode_create_lease_ioctl() calls drm_lease_create() which acquires a lock
on dev->mode_config.idr_mutex. In case of failure, drm_lease_create() calls
drm_master_put() which in turn tries to acquire the same lock when calling
drm_lease_destroy().
v2: - Reverse the order at exit in case of fail, so that unlocking takes place
before dropping the reference.
- Include detail information about deadlock (Daniel Vetter)
Linus Torvalds [Thu, 14 Dec 2017 04:15:49 +0000 (20:15 -0800)]
Merge tag 'xfs-4.15-fixes-5' of git://git.kernel.org/pub/scm/fs/xfs/xfs-linux
Pull xfs fixes from Darrick Wong:
"Here are a few more bug fixes & cleanups for 4.15-rc4:
- clean up duplicate includes
- remove ancient 'no-alloc' crap code that occasionally caused hard
fs shutdowns due to lack of proper space reservations
- fix regression in FIEMAP behavior when reporting xattr extents"
* tag 'xfs-4.15-fixes-5' of git://git.kernel.org/pub/scm/fs/xfs/xfs-linux:
xfs: make iomap_begin functions trim iomaps consistently
xfs: remove "no-allocation" reservations for file creations
fs: xfs: remove duplicate includes
Linus Torvalds [Thu, 14 Dec 2017 04:13:05 +0000 (20:13 -0800)]
Merge tag 'riscv-for-linus-4.15-rc4-riscv_fixes' of git://git.kernel.org/pub/scm/linux/kernel/git/palmer/linux
Pull RISC-V fixes from Palmer Dabbelt:
"This contains three small fixes:
- A fix to a typo in sys_riscv_flush_icache. This only effects error
handling, but I think it's a small and obvious enough change that
it's sane outside the merge window.
- The addition of smp_mb__after_spinlock(), which was recently
removed due to an incorrect comment. This is largly a comment
change (as there's a big one now), and while it's necessary for
complience with the RISC-V memory model the lack of this fence
shouldn't manifest as a bug on current implementations.
Nonetheless, it still seems saner to have the fence in 4.15.
- The removal of some of the HVC_RISCV_SBI driver that snuck into the
arch port. This is compile-time dead code in 4.15 (as the driver
isn't in yet), and during the review process we found a better way
to implement early printk on RISC-V. While this change doesn't do
anything, it will make staging our HVC driver easier: without this
change the HVC driver we hope to upstream won't build on 4.15
(because the 4.15 arch code would reference a function that no
longer exists).
I don't think this is the last patch set we'll want for 4.15: I think
I'll want to remove some of the first-level irqchip driver that snuck
in as well, which will look a lot like the HVC patch here. This is
pending some asm-generic cleanup I'm doing that I haven't quite gotten
clean enough to send out yet, though, but hopefully it'll be ready by
next week (and still OK for that late)"
* tag 'riscv-for-linus-4.15-rc4-riscv_fixes' of git://git.kernel.org/pub/scm/linux/kernel/git/palmer/linux:
RISC-V: Remove unused CONFIG_HVC_RISCV_SBI code
RISC-V: Resurrect smp_mb__after_spinlock()
RISC-V: Logical vs Bitwise typo
The following pull-request contains BPF updates for your *net* tree.
The main changes are:
1) Addition of explicit scheduling points to map alloc/free
in order to avoid having to hold the CPU for too long,
from Eric.
2) Fixing of a corruption in overlapping perf_event_output
calls from different BPF prog types on the same CPU out
of different contexts, from Daniel.
3) Fallout fixes for recent correction of broken uapi for
BPF_PROG_TYPE_PERF_EVENT. um had a missing asm header
that needed to be pulled in from asm-generic and for
BPF selftests the asm-generic include did not work,
so similar asm include scheme was adapted for that
problematic header that perf is having with other
header files under tools, from Daniel.
====================
Daniel Vetter [Wed, 13 Dec 2017 12:49:36 +0000 (13:49 +0100)]
drm: rework delayed connector cleanup in connector_iter
PROBE_DEFER also uses system_wq to reprobe drivers, which means when
that again fails, and we try to flush the overall system_wq (to get
all the delayed connectore cleanup work_struct completed), we
deadlock.
Fix this by using just a single cleanup work, so that we can only
flush that one and don't block on anything else. That means a free
list plus locking, a standard pattern.
v2:
- Correctly free connectors only on last ref. Oops (Chris).
- use llist_head/node (Chris).
David S. Miller [Wed, 13 Dec 2017 21:38:37 +0000 (16:38 -0500)]
Merge branch 'mlx4-misc-fixes'
Tariq Toukan says:
====================
mlx4 misc fixes
This patchset contains misc bug fixes from the team
to the mlx4 Core and Eth drivers.
Patch 1 by Eugenia fixes an MTU issue in selftest.
Patch 2 by Eran fixes an accounting issue in the resource tracker.
Patch 3 by Eran fixes a race condition that causes counter inconsistency.
Series generated against net commit: 200809716aed fou: fix some member types in guehdr
Eran Ben Elisha [Wed, 13 Dec 2017 16:12:11 +0000 (18:12 +0200)]
net/mlx4_en: Fill all counters under one call of stats lock
Before this patch, the stats_lock was acquired twice. In between the
locks Driver sent command to gather some more statistics (per priority
and counter statistics). If the stats lock was acquired by get
statistics NDO in between we would have report out of sync counters.
Fix this by collecting all stats from Firmware in advance and then
fill the Software structs under one lock.
Fixes: 0b131561a7d6 ("net/mlx4_en: Add Flow control statistics display via ethtool") Signed-off-by: Eran Ben Elisha <[email protected]> Signed-off-by: Tariq Toukan <[email protected]> Signed-off-by: David S. Miller <[email protected]>
Eran Ben Elisha [Wed, 13 Dec 2017 16:12:10 +0000 (18:12 +0200)]
net/mlx4_core: Fix wrong calculation of free counters
The field res_free indicates the total number of counters which are
available for allocation (reserved and unreserved). Fixed a bug where
the reserved counters were subtracted from res_free before any
allocation was performed.
Before this fix, free counters which were not reserved could not be
allocated.
Fixes: 9de92c60beaa ("net/mlx4_core: Adjust counter grant policy in the resource tracker") Signed-off-by: Eran Ben Elisha <[email protected]> Reviewed-by: Jack Morgenstein <[email protected]> Signed-off-by: Tariq Toukan <[email protected]> Signed-off-by: David S. Miller <[email protected]>
Set the minimal MTU threshold for running loopback selftest.
MTU should be big enough to include packet payload, NET_IP_ALIGN,
Ethernet headers and preamble length.
Fixes: e7c1c2c46201 ("mlx4_en: Added self diagnostics test implementation") Signed-off-by: Eugenia Emantayev <[email protected]> Signed-off-by: Tariq Toukan <[email protected]> Signed-off-by: David S. Miller <[email protected]>
Russell King [Wed, 13 Dec 2017 09:22:03 +0000 (09:22 +0000)]
net: phy: marvell: avoid configuring fiber page for SGMII-to-Copper
When in SGMII-to-Copper mode, the fiber page is used for the MAC facing
link, and does not require configuration of the fiber auto-negotiation
settings. Avoid trying.
Wei Wang [Wed, 13 Dec 2017 00:28:58 +0000 (16:28 -0800)]
tcp: fix potential underestimation on rcv_rtt
When ms timestamp is used, current logic uses 1us in
tcp_rcv_rtt_update() when the real rcv_rtt is within 1 - 999us.
This could cause rcv_rtt underestimation.
Fix it by always using a min value of 1ms if ms timestamp is used.
Fixes: 645f4c6f2ebd ("tcp: switch rcv_rtt_est and rcvq_space to high resolution timestamps") Signed-off-by: Wei Wang <[email protected]> Signed-off-by: Eric Dumazet <[email protected]> Acked-by: Neal Cardwell <[email protected]> Signed-off-by: David S. Miller <[email protected]>
Russell King [Tue, 12 Dec 2017 10:49:15 +0000 (10:49 +0000)]
of_mdio / mdiobus: ensure mdio devices have fwnode correctly populated
Ensure that all mdio devices populate the struct device fwnode pointer
as well as the of_node pointer to allow drivers that wish to use
fwnode APIs to work.
Russell King [Tue, 12 Dec 2017 10:45:36 +0000 (10:45 +0000)]
net: phy: fix resume handling
When a PHY has the BMCR_PDOWN bit set, it may decide to ignore writes
to other registers, or reset the registers to power-on defaults.
Micrel PHYs do this for their interrupt registers.
The current structure of phylib tries to enable interrupts before
resuming (and releasing) the BMCR_PDOWN bit. This fails, causing
Micrel PHYs to stop working after a suspend/resume sequence if they
are using interrupts.
Fix this by ensuring that the PHY driver resume methods do not take
the phydev->lock mutex themselves, but the callers of phy_resume()
take that lock. This then allows us to move the call to phy_resume()
before we enable interrupts in phy_start().
Russell King [Tue, 12 Dec 2017 09:29:51 +0000 (09:29 +0000)]
ARM: dts: vf610-zii-dev: use XAUI for DSA link ports
Use XAUI rather than XGMII for DSA link ports, as this is the interface
mode that the switches actually use. XAUI is the 4 lane bus with clock
per direction, whereas XGMII is a 32 bit bus with clock.
Russell King [Tue, 12 Dec 2017 09:29:46 +0000 (09:29 +0000)]
net: dsa: allow XAUI phy interface mode
XGMII is a 32-bit bus plus two clock signals per direction. XAUI is
four serial lanes per direction. The 88e6190 supports XAUI but not
XGMII as it doesn't have enough pins. The same is true of 88e6176.
Match on PHY_INTERFACE_MODE_XAUI for the XAUI port type, but keep
accepting XGMII for backwards compatibility.
The follow patchset contains Netfilter fixes for your net tree,
they are:
1) Fix compilation warning in x_tables with clang due to useless
redundant reassignment, from Colin Ian King.
2) Add bugtrap to net_exit to catch uninitialized lists, patch
from Vasily Averin.
3) Fix out of bounds memory reads in H323 conntrack helper, this
comes with an initial patch to remove replace the obscure
CHECK_BOUND macro as a dependency. From Eric Sesterhenn.
4) Reduce retransmission timeout when window is 0 in TCP conntrack,
from Florian Westphal.
6) ctnetlink clamp timeout to INT_MAX if timeout is too large,
otherwise timeout wraps around and it results in killing the
entry that is being added immediately.
7) Missing CAP_NET_ADMIN checks in cthelper and xt_osf, due to
no netns support. From Kevin Cernekee.
8) Missing maximum number of instructions checks in xt_bpf, patch
from Jann Horn.
9) With no CONFIG_PROC_FS ipt_CLUSTERIP compilation breaks,
patch from Arnd Bergmann.
10) Missing netlink attribute policy in nftables exthdr, from
Florian Westphal.
11) Enable conntrack with IPv6 MASQUERADE rules, as a357b3f80bc8
should have done in first place, from Konstantin Khlebnikov.
====================
net: ethernet: arc: fix error handling in emac_rockchip_probe
If clk_set_rate() fails, we should disable clk before return.
Found by Linux Driver Verification project (linuxtesting.org).
Changes since v2 [1]:
* Merged with latest code changes
Changes since v1:
Update made thanks to David's review, much appreciated David.
* Improved inconsistent failure handling of clock rate setting
* For completeness of usecase, added arc_emac_probe error handling
Kevin Cernekee [Mon, 11 Dec 2017 19:13:45 +0000 (11:13 -0800)]
net: igmp: Use correct source address on IGMPv3 reports
Closing a multicast socket after the final IPv4 address is deleted
from an interface can generate a membership report that uses the
source IP from a different interface. The following test script, run
from an isolated netns, reproduces the issue:
#!/bin/bash
ip link add dummy0 type dummy
ip link add dummy1 type dummy
ip link set dummy0 up
ip link set dummy1 up
ip addr add 10.1.1.1/24 dev dummy0
ip addr add 192.168.99.99/24 dev dummy1
sleep 1
ip addr del 10.1.1.1/24 dev dummy0
sleep 5
kill %tcpdump
RFC 3376 specifies that the report must be sent with a valid IP source
address from the destination subnet, or from address 0.0.0.0. Add an
extra check to make sure this is the case.
Jon Maloy [Mon, 11 Dec 2017 18:11:55 +0000 (19:11 +0100)]
tipc: eliminate potential memory leak
In the function tipc_sk_mcast_rcv() we call refcount_dec(&skb->users)
on received sk_buffers. Since the reference counter might hit zero at
this point, we have a potential memory leak.
We fix this by replacing refcount_dec() with kfree_skb().
Eric Dumazet [Mon, 11 Dec 2017 15:17:39 +0000 (07:17 -0800)]
ipv4: igmp: guard against silly MTU values
IPv4 stack reacts to changes to small MTU, by disabling itself under
RTNL.
But there is a window where threads not using RTNL can see a wrong
device mtu. This can lead to surprises, in igmp code where it is
assumed the mtu is suitable.
Fix this by reading device mtu once and checking IPv4 minimal MTU.
This patch adds missing IPV4_MIN_MTU define, to not abuse
ETH_MIN_MTU anymore.
Eric Dumazet [Mon, 11 Dec 2017 15:03:38 +0000 (07:03 -0800)]
ipv6: mcast: better catch silly mtu values
syzkaller reported crashes in IPv6 stack [1]
Xin Long found that lo MTU was set to silly values.
IPv6 stack reacts to changes to small MTU, by disabling itself under
RTNL.
But there is a window where threads not using RTNL can see a wrong
device mtu. This can lead to surprises, in mld code where it is assumed
the mtu is suitable.
Fix this by reading device mtu once and checking IPv6 minimal MTU.
Keith Packard [Wed, 13 Dec 2017 08:44:26 +0000 (00:44 -0800)]
drm: Update edid-derived drm_display_info fields at edid property set [v2]
There are a set of values in the drm_display_info structure for each
connector which hold information derived from EDID. These are computed
in drm_add_display_info. Before this patch, that was only called in
drm_add_edid_modes. This meant that they were only set when EDID was
present and never reset when EDID was not, as happened when the
display was disconnected.
One of these fields, non_desktop, is used from
drm_mode_connector_update_edid_property, the function responsible for
assigning the new edid value to the application-visible property.
Various drivers call these two functions (drm_add_edid_modes and
drm_mode_connector_update_edid_property) in different orders. This
means that even when EDID is present, the drm_display_info fields may
not have been computed at the time that
drm_mode_connector_update_edid_property used the non_desktop value to
set the non_desktop property.
I've added a public function (drm_reset_display_info) that resets the
drm_display_info field values to default values and then made the
drm_add_display_info function public. These two functions are now
called directly from drm_mode_connector_update_edid_property so that
the drm_display_info fields are always computed from the current EDID
information before being used in that function.
This means that the drm_display_info values are often computed twice,
once when the EDID property it set and a second time when EDID is used
to compute modes for the device. The alternative would be to uniformly
ensure that the values were computed once before being used, which
would require that all drivers reliably invoke the two paths in the
same order. The computation is inexpensive enough that it seems more
maintainable in the long term to simply compute them in both paths.
The API to drm_add_display_info has been changed so that it no longer
takes the set of edid-based quirks as a parameter. Rather, it now
computes those quirks itself and returns them for further use by
drm_add_edid_modes.
This patch also includes a number of 'const' additions caused by
drm_mode_connector_update_edid_property taking a 'const struct edid *'
parameter and wanting to pass that along to drm_add_display_info.
projects / linux.git / log