This reverts commit da56a1bfba.
Bjorn Andersson, Fabio Estevam, Xiaolei Wang, and Jon Hunter reported that
da56a1bfba ("PCI: dwc: Wait for link up only if link is started") broke
controller probing by returning an error in case the link does not come up
during host initialisation, for example when the slot is empty.
As explained in commit 886a9c1347 ("PCI: dwc: Move link handling into
common code") and as indicated by the comment "Ignore errors, the link may
come up later" in the code, waiting for link up and ignoring errors is the
intended behaviour:
Let's standardize this to succeed as there are usecases where devices
(and the link) appear later even without hotplug. For example, a
reconfigured FPGA device.
Reverting the offending commit specifically fixes a regression on Qualcomm
platforms like the Lenovo ThinkPad X13s which no longer reach the
interconnect sync state if a slot does not have a device populated (e.g. an
optional modem).
Note that enabling asynchronous probing by default as was done for Qualcomm
platforms by commit c0e1eb441b ("PCI: qcom: Enable async probe by
default"), should take care of any related boot time concerns.
Finally, note that the intel-gw driver is the only driver currently not
providing a .start_link() callback and instead starts the link in its
.host_init() callback, which may avoid an additional one-second timeout
during probe by making the link-up wait conditional. If anyone cares, that
can be done in a follow-up patch with a proper motivation.
[bhelgaas: add Fabio Estevam, Xiaolei Wang, Jon Hunter reports]
Fixes: da56a1bfba ("PCI: dwc: Wait for link up only if link is started")
Link: https://lore.kernel.org/r/20230706082610.26584-1-johan+linaro@kernel.org
Reported-by: Bjorn Andersson <quic_bjorande@quicinc.com>
Signed-off-by: Johan Hovold <johan+linaro@kernel.org>
Reported-by: Fabio Estevam <festevam@gmail.com>
Link: https://lore.kernel.org/r/20230704122635.1362156-1-festevam@gmail.com/
Reported-by: Xiaolei Wang <xiaolei.wang@windriver.com>
Link: https://lore.kernel.org/r/20230705010624.3912934-1-xiaolei.wang@windriver.com/
Reported-by: Jon Hunter <jonathanh@nvidia.com>
Link: https://lore.kernel.org/r/6ca287a1-6c7c-7b90-9022-9e73fb82b564@nvidia.com
Signed-off-by: Bjorn Helgaas <bhelgaas@google.com>
Reviewed-by: Manivannan Sadhasivam <manivannan.sadhasivam@linaro.org>
Cc: Sajid Dalvi <sdalvi@google.com>
Cc: Ajay Agarwal <ajayagarwal@google.com>
Bail out with EOPNOTSUPP when adding rule to bound chain via
NFTA_RULE_CHAIN_ID. The following warning splat is shown when
adding a rule to a deleted bound chain:
WARNING: CPU: 2 PID: 13692 at net/netfilter/nf_tables_api.c:2013 nf_tables_chain_destroy+0x1f7/0x210 [nf_tables]
CPU: 2 PID: 13692 Comm: chain-bound-rul Not tainted 6.1.39 #1
RIP: 0010:nf_tables_chain_destroy+0x1f7/0x210 [nf_tables]
Fixes: d0e2c7de92 ("netfilter: nf_tables: add NFT_CHAIN_BINDING")
Reported-by: Kevin Rich <kevinrich1337@gmail.com>
Signed-off-by: Pablo Neira Ayuso <pablo@netfilter.org>
Signed-off-by: Florian Westphal <fw@strlen.de>
On error when building the rule, the immediate expression unbinds the
chain, hence objects can be deactivated by the transaction records.
Otherwise, it is possible to trigger the following warning:
WARNING: CPU: 3 PID: 915 at net/netfilter/nf_tables_api.c:2013 nf_tables_chain_destroy+0x1f7/0x210 [nf_tables]
CPU: 3 PID: 915 Comm: chain-bind-err- Not tainted 6.1.39 #1
RIP: 0010:nf_tables_chain_destroy+0x1f7/0x210 [nf_tables]
Fixes: 4bedf9eee0 ("netfilter: nf_tables: fix chain binding transaction logic")
Reported-by: Kevin Rich <kevinrich1337@gmail.com>
Signed-off-by: Pablo Neira Ayuso <pablo@netfilter.org>
Signed-off-by: Florian Westphal <fw@strlen.de>
The lazy gc on insert that should remove timed-out entries fails to release
the other half of the interval, if any.
Can be reproduced with tests/shell/testcases/sets/0044interval_overlap_0
in nftables.git and kmemleak enabled kernel.
Second bug is the use of rbe_prev vs. prev pointer.
If rbe_prev() returns NULL after at least one iteration, rbe_prev points
to element that is not an end interval, hence it should not be removed.
Lastly, check the genmask of the end interval if this is active in the
current generation.
Fixes: c9e6978e27 ("netfilter: nft_set_rbtree: Switch to node list walk for overlap detection")
Signed-off-by: Florian Westphal <fw@strlen.de>
An attempt to acquire exclusive lock can race with the current lock
owner closing the image:
1. lock is held by client123, rbd_lock() returns -EBUSY
2. get_lock_owner_info() returns client123 instance details
3. client123 closes the image, lock is released
4. find_watcher() returns 0 as there is no matching watcher anymore
5. client123 instance gets erroneously blocklisted
Particularly impacted is mirror snapshot scheduler in snapshot-based
mirroring since it happens to open and close images a lot (images are
opened only for as long as it takes to take the next mirror snapshot,
the same client instance is used for all images).
To reduce the potential for erroneous blocklisting, retrieve the lock
owner again after find_watcher() returns 0. If it's still there, make
sure it matches the previously detected lock owner.
Cc: stable@vger.kernel.org # f38cb9d9c2: rbd: make get_lock_owner_info() return a single locker or NULL
Cc: stable@vger.kernel.org # 8ff2c64c97: rbd: harden get_lock_owner_info() a bit
Cc: stable@vger.kernel.org
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
Reviewed-by: Dongsheng Yang <dongsheng.yang@easystack.cn>
- we want the exclusive lock type, so test for it directly
- use sscanf() to actually parse the lock cookie and avoid admitting
invalid handles
- bail if locker has a blank address
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
Reviewed-by: Dongsheng Yang <dongsheng.yang@easystack.cn>
Make the "num_lockers can be only 0 or 1" assumption explicit and
simplify the API by getting rid of output parameters in preparation
for calling get_lock_owner_info() twice before blocklisting.
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
Reviewed-by: Dongsheng Yang <dongsheng.yang@easystack.cn>
btrfs_attach_transaction_barrier() is used to get a handle pointing to the
current running transaction if the transaction has not started its commit
yet (its state is < TRANS_STATE_COMMIT_START). If the transaction commit
has started, then we wait for the transaction to commit and finish before
returning - however we completely ignore if the transaction was aborted
due to some error during its commit, we simply return ERR_PT(-ENOENT),
which makes the caller assume everything is fine and no errors happened.
This could make an fsync return success (0) to user space when in fact we
had a transaction abort and the target inode changes were therefore not
persisted.
Fix this by checking for the return value from btrfs_wait_for_commit(),
and if it returned an error, return it back to the caller.
Fixes: d4edf39bd5 ("Btrfs: fix uncompleted transaction")
CC: stable@vger.kernel.org # 4.19+
Reviewed-by: Qu Wenruo <wqu@suse.com>
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
In the patch ("spi: spi-qcom-qspi: Fallback to PIO for xfers that
aren't multiples of 4 bytes") we detect reads that we can't handle
properly and fallback to PIO mode. While that's correct behavior, we
can do better by adding "spi_controller_mem_ops" for our
controller. Once we do this then the caller will give us a transfer
that's a multiple of 4-bytes so we can DMA.
Fixes: b5762d9560 ("spi: spi-qcom-qspi: Add DMA mode support")
Signed-off-by: Douglas Anderson <dianders@chromium.org>
Reviewed-by: Vijaya Krishna Nivarthi <quic_vnivarth@quicinc.com>
Link: https://lore.kernel.org/r/20230725110226.2.Id4a39804e01e4a06dae9b73fd2a5194c4c7ea453@changeid
Signed-off-by: Mark Brown <broonie@kernel.org>
The Qualcomm QSPI driver appears to require that any reads using DMA
are a mutliple of 4 bytes. If this isn't true then the controller will
clobber any extra bytes in memory following the last word. Let's
detect this and falback to PIO.
This fixes problems reported by slub_debug=FZPUA, which would complain
about "kmalloc Redzone overwritten". One such instance said:
0xffffff80c29d541a-0xffffff80c29d541b @offset=21530. First byte 0x0 instead of 0xcc
Allocated in mtd_kmalloc_up_to+0x98/0xac age=36 cpu=3 pid=6658
Tracing through what was happening I saw that, while we often did DMA
tranfers of 0x1000 bytes, sometimes we'd end up doing ones of 0x41a
bytes. Those 0x41a byte transfers were the problem.
NOTE: a future change will enable the SPI "mem ops" to help avoid this
case, but it still seems good to add the extra check in the transfer.
Fixes: b5762d9560 ("spi: spi-qcom-qspi: Add DMA mode support")
Signed-off-by: Douglas Anderson <dianders@chromium.org>
Reviewed-by: Vijaya Krishna Nivarthi <quic_vnivarth@quicinc.com>
Reviewed-by: Bjorn Andersson <andersson@kernel.org>
Link: https://lore.kernel.org/r/20230725110226.1.Ia2f980fc7cd0b831e633391f0bb1272914d8f381@changeid
Signed-off-by: Mark Brown <broonie@kernel.org>
The qca8k switch doesn't support using 0 as VID and require a default
VID to be always set. MDB add/del function doesn't currently handle
this and are currently setting the default VID.
Fix this by correctly handling this corner case and internally use the
default VID for VID 0 case.
Fixes: ba8f870dfa ("net: dsa: qca8k: add support for mdb_add/del")
Signed-off-by: Christian Marangi <ansuelsmth@gmail.com>
Cc: stable@vger.kernel.org
Signed-off-by: David S. Miller <davem@davemloft.net>
On deleting an MDB entry for a port, fdb_search_and_del is used.
An FDB entry can't be modified so it needs to be deleted and readded
again with the new portmap (and the port deleted as requested)
We use the SEARCH operator to search the entry to edit by vid and mac
address and then we check the aging if we actually found an entry.
Currently the code suffer from a bug where the searched fdb entry is
never read again with the found values (if found) resulting in the code
always returning -EINVAL as aging was always 0.
Fix this by correctly read the fdb entry after it was searched.
Fixes: ba8f870dfa ("net: dsa: qca8k: add support for mdb_add/del")
Signed-off-by: Christian Marangi <ansuelsmth@gmail.com>
Cc: stable@vger.kernel.org
Signed-off-by: David S. Miller <davem@davemloft.net>
On inserting a mdb entry, fdb_search_and_insert is used to add a port to
the qca8k target entry in the FDB db.
A FDB entry can't be modified so it needs to be removed and insert again
with the new values.
To detect if an entry already exist, the SEARCH operation is used and we
check the aging of the entry. If the entry is not 0, the entry exist and
we proceed to delete it.
Current code have 2 main problem:
- The condition to check if the FDB entry exist is wrong and should be
the opposite.
- When a FDB entry doesn't exist, aging was never actually set to the
STATIC value resulting in allocating an invalid entry.
Fix both problem by adding aging support to the function, calling the
function with STATIC as aging by default and finally by correct the
condition to check if the entry actually exist.
Fixes: ba8f870dfa ("net: dsa: qca8k: add support for mdb_add/del")
Signed-off-by: Christian Marangi <ansuelsmth@gmail.com>
Cc: stable@vger.kernel.org
Signed-off-by: David S. Miller <davem@davemloft.net>
The qca8xxx switch supports 2 way to write reg values, a slow way using
mdio and a fast way by sending specially crafted mgmt packet to
read/write reg.
The fast way can support up to 32 bytes of data as eth packet are used
to send/receive.
This correctly works for almost the entire regmap of the switch but with
the use of some kernel selftests for dsa drivers it was found a funny
and interesting hw defect/limitation.
For some specific reg, bulk write won't work and will result in writing
only part of the requested regs resulting in half data written. This was
especially hard to track and discover due to the total strangeness of
the problem and also by the specific regs where this occurs.
This occurs in the specific regs of the ATU table, where multiple entry
needs to be written to compose the entire entry.
It was discovered that with a bulk write of 12 bytes on
QCA8K_REG_ATU_DATA0 only QCA8K_REG_ATU_DATA0 and QCA8K_REG_ATU_DATA2
were written, but QCA8K_REG_ATU_DATA1 was always zero.
Tcpdump was used to make sure the specially crafted packet was correct
and this was confirmed.
The problem was hard to track as the lack of QCA8K_REG_ATU_DATA1
resulted in an entry somehow possible as the first bytes of the mac
address are set in QCA8K_REG_ATU_DATA0 and the entry type is set in
QCA8K_REG_ATU_DATA2.
Funlly enough writing QCA8K_REG_ATU_DATA1 results in the same problem
with QCA8K_REG_ATU_DATA2 empty and QCA8K_REG_ATU_DATA1 and
QCA8K_REG_ATU_FUNC correctly written.
A speculation on the problem might be that there are some kind of
indirection internally when accessing these regs and they can't be
accessed all together, due to the fact that it's really a table mapped
somewhere in the switch SRAM.
Even more funny is the fact that every other reg was tested with all
kind of combination and they are not affected by this problem. Read
operation was also tested and always worked so it's not affected by this
problem.
The problem is not present if we limit writing a single reg at times.
To handle this hardware defect, enable use_single_write so that bulk
api can correctly split the write in multiple different operation
effectively reverting to a non-bulk write.
Cc: Mark Brown <broonie@kernel.org>
Fixes: c766e077d9 ("net: dsa: qca8k: convert to regmap read/write API")
Signed-off-by: Christian Marangi <ansuelsmth@gmail.com>
Cc: stable@vger.kernel.org
Signed-off-by: David S. Miller <davem@davemloft.net>
A previous commit tried to come up with more generic subpool
names, but this isn't quite working: the node name was used
elsewhere to match pools to consumers which regressed the
nVidia Tegra 2/3 video decoder.
Revert back to an earlier approach using of_node_full_name()
instead of just the name to make sure the pool name is more
unique, and change both sites using this in the kernel.
It is not perfect since two SRAM nodes could have the same
subpool name but it makes the situation better than before.
Reported-by: Dmitry Osipenko <digetx@gmail.com>
Fixes: 21e5a2d10c ("misc: sram: Generate unique names for subpools")
Signed-off-by: Linus Walleij <linus.walleij@linaro.org>
Tested-by: Dmitry Osipenko <digetx@gmail.com>
Link: https://lore.kernel.org/r/20230622074520.3058027-1-linus.walleij@linaro.org
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Yan-Hsuan has been away since 2021 and Ping has been the de facto maintainer
the past year, actively reviewing patches and doing all other maintainer
duties. So fix the MAINTAINERS file to show the current situation.
Signed-off-by: Kalle Valo <kvalo@kernel.org>
Link: https://lore.kernel.org/r/20230724104547.3061709-2-kvalo@kernel.org
Xen 4.17 supports the creation of static evtchns. To allow user space
application to bind static evtchns introduce new ioctl
"IOCTL_EVTCHN_BIND_STATIC". Existing IOCTL doing more than binding
that’s why we need to introduce the new IOCTL to only bind the static
event channels.
Static evtchns to be available for use during the lifetime of the
guest. When the application exits, __unbind_from_irq() ends up being
called from release() file operations because of that static evtchns
are getting closed. To avoid closing the static event channel, add the
new bool variable "is_static" in "struct irq_info" to mark the event
channel static when creating the event channel to avoid closing the
static evtchn.
Also, take this opportunity to remove the open-coded version of the
evtchn close in drivers/xen/evtchn.c file and use xen_evtchn_close().
Signed-off-by: Rahul Singh <rahul.singh@arm.com>
Reviewed-by: Oleksandr Tyshchenko <oleksandr_tyshchenko@epam.com>
Acked-by: Stefano Stabellini <sstabellini@kernel.org>
Link: https://lore.kernel.org/r/ae7329bf1713f83e4aad4f3fa0f316258c40a3e9.1689677042.git.rahul.singh@arm.com
Signed-off-by: Juergen Gross <jgross@suse.com>
Last year, the code that manages GSI channel transactions switched
from using spinlock-protected linked lists to using indexes into the
ring buffer used for a channel. Recently, Google reported seeing
transaction reference count underflows occasionally during shutdown.
Doug Anderson found a way to reproduce the issue reliably, and
bisected the issue to the commit that eliminated the linked lists
and the lock. The root cause was ultimately determined to be
related to unused transactions being committed as part of the modem
shutdown cleanup activity. Unused transactions are not normally
expected (except in error cases).
The modem uses some ranges of IPA-resident memory, and whenever it
shuts down we zero those ranges. In ipa_filter_reset_table() a
transaction is allocated to zero modem filter table entries. If
hashing is not supported, hashed table memory should not be zeroed.
But currently nothing prevents that, and the result is an unused
transaction. Something similar occurs when we zero routing table
entries for the modem.
By preventing any attempt to clear hashed tables when hashing is not
supported, the reference count underflow is avoided in this case.
Note that there likely remains an issue with properly freeing unused
transactions (if they occur due to errors). This patch addresses
only the underflows that Google originally reported.
Cc: <stable@vger.kernel.org> # 6.1.x
Fixes: d338ae28d8 ("net: ipa: kill all other transaction lists")
Tested-by: Douglas Anderson <dianders@chromium.org>
Signed-off-by: Alex Elder <elder@linaro.org>
Link: https://lore.kernel.org/r/20230724224055.1688854-1-elder@linaro.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
syzkaller found a bug in unix_bind_bsd() [0]. We can reproduce it
by bind()ing a socket on a path with length 108.
108 is the size of sun_addr of struct sockaddr_un and is the maximum
valid length for the pathname socket. When calling bind(), we use
struct sockaddr_storage as the actual buffer size, so terminating
sun_addr[108] with null is legitimate as done in unix_mkname_bsd().
However, strlen(sunaddr) for such a case causes fortify_panic() if
CONFIG_FORTIFY_SOURCE=y. __fortify_strlen() has no idea about the
actual buffer size and see the string as unterminated.
Let's use strnlen() to allow sun_addr to be unterminated at 107.
[0]:
detected buffer overflow in __fortify_strlen
kernel BUG at lib/string_helpers.c:1031!
Internal error: Oops - BUG: 00000000f2000800 [#1] PREEMPT SMP
Modules linked in:
CPU: 0 PID: 255 Comm: syz-executor296 Not tainted 6.5.0-rc1-00330-g60cc1f7d0605 #4
Hardware name: linux,dummy-virt (DT)
pstate: 60400005 (nZCv daif +PAN -UAO -TCO -DIT -SSBS BTYPE=--)
pc : fortify_panic+0x1c/0x20 lib/string_helpers.c:1030
lr : fortify_panic+0x1c/0x20 lib/string_helpers.c:1030
sp : ffff800089817af0
x29: ffff800089817af0 x28: ffff800089817b40 x27: 1ffff00011302f68
x26: 000000000000006e x25: 0000000000000012 x24: ffff800087e60140
x23: dfff800000000000 x22: ffff800089817c20 x21: ffff800089817c8e
x20: 000000000000006c x19: ffff00000c323900 x18: ffff800086ab1630
x17: 0000000000000000 x16: 0000000000000000 x15: 0000000000000001
x14: 1ffff00011302eb8 x13: 0000000000000000 x12: 0000000000000000
x11: 0000000000000000 x10: 0000000000000000 x9 : 64a26b65474d2a00
x8 : 64a26b65474d2a00 x7 : 0000000000000001 x6 : 0000000000000001
x5 : ffff800089817438 x4 : ffff800086ac99e0 x3 : ffff800080f19e8c
x2 : 0000000000000001 x1 : 0000000100000000 x0 : 000000000000002c
Call trace:
fortify_panic+0x1c/0x20 lib/string_helpers.c:1030
_Z16__fortify_strlenPKcU25pass_dynamic_object_size1 include/linux/fortify-string.h:217 [inline]
unix_bind_bsd net/unix/af_unix.c:1212 [inline]
unix_bind+0xba8/0xc58 net/unix/af_unix.c:1326
__sys_bind+0x1ac/0x248 net/socket.c:1792
__do_sys_bind net/socket.c:1803 [inline]
__se_sys_bind net/socket.c:1801 [inline]
__arm64_sys_bind+0x7c/0x94 net/socket.c:1801
__invoke_syscall arch/arm64/kernel/syscall.c:38 [inline]
invoke_syscall+0x98/0x2c0 arch/arm64/kernel/syscall.c:52
el0_svc_common+0x134/0x240 arch/arm64/kernel/syscall.c:139
do_el0_svc+0x64/0x198 arch/arm64/kernel/syscall.c:188
el0_svc+0x2c/0x7c arch/arm64/kernel/entry-common.c:647
el0t_64_sync_handler+0x84/0xfc arch/arm64/kernel/entry-common.c:665
el0t_64_sync+0x190/0x194 arch/arm64/kernel/entry.S:591
Code: aa0003e1 d0000e80 91030000 97ffc91a (d4210000)
Fixes: df8fc4e934 ("kbuild: Enable -fstrict-flex-arrays=3")
Reported-by: syzkaller <syzkaller@googlegroups.com>
Suggested-by: Kees Cook <keescook@chromium.org>
Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Link: https://lore.kernel.org/r/20230724213425.22920-2-kuniyu@amazon.com
Reviewed-by: Simon Horman <simon.horman@corigine.com>
Reviewed-by: Kees Cook <keescook@chromium.org>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
The previous commit 954d1fa1ac ("macvlan: Add netlink attribute for
broadcast cutoff") added one additional attribute named
IFLA_MACVLAN_BC_CUTOFF to allow broadcast cutfoff.
However, it forgot to describe the nla_policy at macvlan_policy
(drivers/net/macvlan.c). Hence, this suppose NLA_S32 (4 bytes) integer
can be faked as empty (0 bytes) by a malicious user, which could leads
to OOB in heap just like CVE-2023-3773.
To fix it, this commit just completes the nla_policy description for
IFLA_MACVLAN_BC_CUTOFF. This enforces the length check and avoids the
potential OOB read.
Fixes: 954d1fa1ac ("macvlan: Add netlink attribute for broadcast cutoff")
Signed-off-by: Lin Ma <linma@zju.edu.cn>
Reviewed-by: Simon Horman <simon.horman@corigine.com>
Link: https://lore.kernel.org/r/20230723080205.3715164-1-linma@zju.edu.cn
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Storage devices are free to send RSCNs, e.g. for internal state changes. If
this happens on all connected paths, zfcp risks temporarily losing all
paths at the same time. This has strong requirements on multipath
configuration such as "no_path_retry queue".
Avoid such situations by deferring fc_rport blocking until after the ADISC
response, when any actual state change of the remote port became clear.
The already existing port recovery triggers explicitly block the fc_rport.
The triggers are: on ADISC reject or timeout (typical cable pull case), and
on ADISC indicating that the remote port has changed its WWPN or
the port is meanwhile no longer open.
As a side effect, this also removes a confusing direct function call to
another work item function zfcp_scsi_rport_work() instead of scheduling
that other work item. It was probably done that way to have the rport block
side effect immediate and synchronous to the caller.
Fixes: a2fa0aede0 ("[SCSI] zfcp: Block FC transport rports early on errors")
Cc: stable@vger.kernel.org #v2.6.30+
Reviewed-by: Benjamin Block <bblock@linux.ibm.com>
Reviewed-by: Fedor Loshakov <loshakov@linux.ibm.com>
Signed-off-by: Steffen Maier <maier@linux.ibm.com>
Link: https://lore.kernel.org/r/20230724145156.3920244-1-maier@linux.ibm.com
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
Jiri Olsa says:
====================
bpf: Disable preemption in perf_event_output helpers code
hi,
we got report of kernel crash [1][3] within bpf_event_output helper.
The reason is the nesting protection code in bpf_event_output that expects
disabled preemption, which is not guaranteed for programs executed by
bpf_prog_run_array_cg.
I managed to reproduce on tracing side where we have the same problem
in bpf_perf_event_output. The reproducer [2] just creates busy uprobe
and call bpf_perf_event_output helper a lot.
v3 changes:
- added acks and fixed 'Fixes' tag style [Hou Tao]
- added Closes tag to patch 2
v2 changes:
- I changed 'Fixes' commits to where I saw we switched from preempt_disable
to migrate_disable, but I'm not completely sure about the patch 2, because
it was tricky to find, would be nice if somebody could check on that
thanks,
jirka
[1] https://github.com/cilium/cilium/issues/26756
[2] https://git.kernel.org/pub/scm/linux/kernel/git/jolsa/perf.git/commit/?h=bpf_output_fix_reproducer&id=8054dcc634121b884c7c331329d61d93351d03b5
[3] slack:
[66194.378161] BUG: kernel NULL pointer dereference, address: 0000000000000001
[66194.378324] #PF: supervisor instruction fetch in kernel mode
[66194.378447] #PF: error_code(0x0010) - not-present page
...
[66194.378692] Oops: 0010 [#1] PREEMPT SMP NOPTI
...
[66194.380666] <TASK>
[66194.380775] ? perf_output_sample+0x12a/0x9a0
[66194.380902] ? finish_task_switch.isra.0+0x81/0x280
[66194.381024] ? perf_event_output+0x66/0xa0
[66194.381148] ? bpf_event_output+0x13a/0x190
[66194.381270] ? bpf_event_output_data+0x22/0x40
[66194.381391] ? bpf_prog_dfc84bbde731b257_cil_sock4_connect+0x40a/0xacb
[66194.381519] ? xa_load+0x87/0xe0
[66194.381635] ? __cgroup_bpf_run_filter_sock_addr+0xc1/0x1a0
[66194.381759] ? release_sock+0x3e/0x90
[66194.381876] ? sk_setsockopt+0x1a1/0x12f0
[66194.381996] ? udp_pre_connect+0x36/0x50
[66194.382114] ? inet_dgram_connect+0x93/0xa0
[66194.382233] ? __sys_connect+0xb4/0xe0
[66194.382353] ? udp_setsockopt+0x27/0x40
[66194.382470] ? __pfx_udp_push_pending_frames+0x10/0x10
[66194.382593] ? __sys_setsockopt+0xdf/0x1a0
[66194.382713] ? __x64_sys_connect+0xf/0x20
[66194.382832] ? do_syscall_64+0x3a/0x90
[66194.382949] ? entry_SYSCALL_64_after_hwframe+0x72/0xdc
[66194.383077] </TASK>
---
====================
Link: https://lore.kernel.org/r/20230725084206.580930-1-jolsa@kernel.org
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
We received report [1] of kernel crash, which is caused by
using nesting protection without disabled preemption.
The bpf_event_output can be called by programs executed by
bpf_prog_run_array_cg function that disabled migration but
keeps preemption enabled.
This can cause task to be preempted by another one inside the
nesting protection and lead eventually to two tasks using same
perf_sample_data buffer and cause crashes like:
BUG: kernel NULL pointer dereference, address: 0000000000000001
#PF: supervisor instruction fetch in kernel mode
#PF: error_code(0x0010) - not-present page
...
? perf_output_sample+0x12a/0x9a0
? finish_task_switch.isra.0+0x81/0x280
? perf_event_output+0x66/0xa0
? bpf_event_output+0x13a/0x190
? bpf_event_output_data+0x22/0x40
? bpf_prog_dfc84bbde731b257_cil_sock4_connect+0x40a/0xacb
? xa_load+0x87/0xe0
? __cgroup_bpf_run_filter_sock_addr+0xc1/0x1a0
? release_sock+0x3e/0x90
? sk_setsockopt+0x1a1/0x12f0
? udp_pre_connect+0x36/0x50
? inet_dgram_connect+0x93/0xa0
? __sys_connect+0xb4/0xe0
? udp_setsockopt+0x27/0x40
? __pfx_udp_push_pending_frames+0x10/0x10
? __sys_setsockopt+0xdf/0x1a0
? __x64_sys_connect+0xf/0x20
? do_syscall_64+0x3a/0x90
? entry_SYSCALL_64_after_hwframe+0x72/0xdc
Fixing this by disabling preemption in bpf_event_output.
[1] https://github.com/cilium/cilium/issues/26756
Cc: stable@vger.kernel.org
Reported-by: Oleg "livelace" Popov <o.popov@livelace.ru>
Closes: https://github.com/cilium/cilium/issues/26756
Fixes: 2a916f2f54 ("bpf: Use migrate_disable/enable in array macros and cgroup/lirc code.")
Acked-by: Hou Tao <houtao1@huawei.com>
Signed-off-by: Jiri Olsa <jolsa@kernel.org>
Link: https://lore.kernel.org/r/20230725084206.580930-3-jolsa@kernel.org
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
The nesting protection in bpf_perf_event_output relies on disabled
preemption, which is guaranteed for kprobes and tracepoints.
However bpf_perf_event_output can be also called from uprobes context
through bpf_prog_run_array_sleepable function which disables migration,
but keeps preemption enabled.
This can cause task to be preempted by another one inside the nesting
protection and lead eventually to two tasks using same perf_sample_data
buffer and cause crashes like:
kernel tried to execute NX-protected page - exploit attempt? (uid: 0)
BUG: unable to handle page fault for address: ffffffff82be3eea
...
Call Trace:
? __die+0x1f/0x70
? page_fault_oops+0x176/0x4d0
? exc_page_fault+0x132/0x230
? asm_exc_page_fault+0x22/0x30
? perf_output_sample+0x12b/0x910
? perf_event_output+0xd0/0x1d0
? bpf_perf_event_output+0x162/0x1d0
? bpf_prog_c6271286d9a4c938_krava1+0x76/0x87
? __uprobe_perf_func+0x12b/0x540
? uprobe_dispatcher+0x2c4/0x430
? uprobe_notify_resume+0x2da/0xce0
? atomic_notifier_call_chain+0x7b/0x110
? exit_to_user_mode_prepare+0x13e/0x290
? irqentry_exit_to_user_mode+0x5/0x30
? asm_exc_int3+0x35/0x40
Fixing this by disabling preemption in bpf_perf_event_output.
Cc: stable@vger.kernel.org
Fixes: 8c7dcb84e3 ("bpf: implement sleepable uprobes by chaining gps")
Acked-by: Hou Tao <houtao1@huawei.com>
Signed-off-by: Jiri Olsa <jolsa@kernel.org>
Link: https://lore.kernel.org/r/20230725084206.580930-2-jolsa@kernel.org
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
