Fix tcp_rcv_synrecv_state_fastopen() to not zero retrans_stamp
if retransmits are outstanding.
tcp_fastopen_synack_timer() sets retrans_stamp, so typically we'll
need to zero retrans_stamp here to prevent spurious
retransmits_timed_out(). The logic to zero retrans_stamp is from this
2019 commit:
commit cd736d8b67 ("tcp: fix retrans timestamp on passive Fast Open")
However, in the corner case where the ACK of our TFO SYNACK carried
some SACK blocks that caused us to enter TCP_CA_Recovery then that
non-zero retrans_stamp corresponds to the active fast recovery, and we
need to leave retrans_stamp with its current non-zero value, for
correct ETIMEDOUT and undo behavior.
Fixes: cd736d8b67 ("tcp: fix retrans timestamp on passive Fast Open")
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Link: https://patch.msgid.link/20241001200517.2756803-4-ncardwell.sw@gmail.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Fix tcp_enter_recovery() so that if there are no retransmits out then
we zero retrans_stamp when entering fast recovery. This is necessary
to fix two buggy behaviors.
Currently a non-zero retrans_stamp value can persist across multiple
back-to-back loss recovery episodes. This is because we generally only
clears retrans_stamp if we are completely done with loss recoveries,
and get to tcp_try_to_open() and find !tcp_any_retrans_done(sk). This
behavior causes two bugs:
(1) When a loss recovery episode (CA_Loss or CA_Recovery) is followed
immediately by a new CA_Recovery, the retrans_stamp value can persist
and can be a time before this new CA_Recovery episode starts. That
means that timestamp-based undo will be using the wrong retrans_stamp
(a value that is too old) when comparing incoming TS ecr values to
retrans_stamp to see if the current fast recovery episode can be
undone.
(2) If there is a roughly minutes-long sequence of back-to-back fast
recovery episodes, one after another (e.g. in a shallow-buffered or
policed bottleneck), where each fast recovery successfully makes
forward progress and recovers one window of sequence space (but leaves
at least one retransmit in flight at the end of the recovery),
followed by several RTOs, then the ETIMEDOUT check may be using the
wrong retrans_stamp (a value set at the start of the first fast
recovery in the sequence). This can cause a very premature ETIMEDOUT,
killing the connection prematurely.
This commit changes the code to zero retrans_stamp when entering fast
recovery, when this is known to be safe (no retransmits are out in the
network). That ensures that when starting a fast recovery episode, and
it is safe to do so, retrans_stamp is set when we send the fast
retransmit packet. That addresses both bug (1) and bug (2) by ensuring
that (if no retransmits are out when we start a fast recovery) we use
the initial fast retransmit of this fast recovery as the time value
for undo and ETIMEDOUT calculations.
This makes intuitive sense, since the start of a new fast recovery
episode (in a scenario where no lost packets are out in the network)
means that the connection has made forward progress since the last RTO
or fast recovery, and we should thus "restart the clock" used for both
undo and ETIMEDOUT logic.
Note that if when we start fast recovery there *are* retransmits out
in the network, there can still be undesirable (1)/(2) issues. For
example, after this patch we can still have the (1) and (2) problems
in cases like this:
+ round 1: sender sends flight 1
+ round 2: sender receives SACKs and enters fast recovery 1,
retransmits some packets in flight 1 and then sends some new data as
flight 2
+ round 3: sender receives some SACKs for flight 2, notes losses, and
retransmits some packets to fill the holes in flight 2
+ fast recovery has some lost retransmits in flight 1 and continues
for one or more rounds sending retransmits for flight 1 and flight 2
+ fast recovery 1 completes when snd_una reaches high_seq at end of
flight 1
+ there are still holes in the SACK scoreboard in flight 2, so we
enter fast recovery 2, but some retransmits in the flight 2 sequence
range are still in flight (retrans_out > 0), so we can't execute the
new retrans_stamp=0 added here to clear retrans_stamp
It's not yet clear how to fix these remaining (1)/(2) issues in an
efficient way without breaking undo behavior, given that retrans_stamp
is currently used for undo and ETIMEDOUT. Perhaps the optimal (but
expensive) strategy would be to set retrans_stamp to the timestamp of
the earliest outstanding retransmit when entering fast recovery. But
at least this commit makes things better.
Note that this does not change the semantics of retrans_stamp; it
simply makes retrans_stamp accurate in some cases where it was not
before:
(1) Some loss recovery, followed by an immediate entry into a fast
recovery, where there are no retransmits out when entering the fast
recovery.
(2) When a TFO server has a SYNACK retransmit that sets retrans_stamp,
and then the ACK that completes the 3-way handshake has SACK blocks
that trigger a fast recovery. In this case when entering fast recovery
we want to zero out the retrans_stamp from the TFO SYNACK retransmit,
and set the retrans_stamp based on the timestamp of the fast recovery.
We introduce a tcp_retrans_stamp_cleanup() helper, because this
two-line sequence already appears in 3 places and is about to appear
in 2 more as a result of this bug fix patch series. Once this bug fix
patches series in the net branch makes it into the net-next branch
we'll update the 3 other call sites to use the new helper.
This is a long-standing issue. The Fixes tag below is chosen to be the
oldest commit at which the patch will apply cleanly, which is from
Linux v3.5 in 2012.
Fixes: 1fbc340514 ("tcp: early retransmit: tcp_enter_recovery()")
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Link: https://patch.msgid.link/20241001200517.2756803-3-ncardwell.sw@gmail.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Fix the TCP loss recovery undo logic in tcp_packet_delayed() so that
it can trigger undo even if TSQ prevents a fast recovery episode from
reaching tcp_retransmit_skb().
Geumhwan Yu <geumhwan.yu@samsung.com> recently reported that after
this commit from 2019:
commit bc9f38c832 ("tcp: avoid unconditional congestion window undo
on SYN retransmit")
...and before this fix we could have buggy scenarios like the
following:
+ Due to reordering, a TCP connection receives some SACKs and enters a
spurious fast recovery.
+ TSQ prevents all invocations of tcp_retransmit_skb(), because many
skbs are queued in lower layers of the sending machine's network
stack; thus tp->retrans_stamp remains 0.
+ The connection receives a TCP timestamp ECR value echoing a
timestamp before the fast recovery, indicating that the fast
recovery was spurious.
+ The connection fails to undo the spurious fast recovery because
tp->retrans_stamp is 0, and thus tcp_packet_delayed() returns false,
due to the new logic in the 2019 commit: commit bc9f38c832 ("tcp:
avoid unconditional congestion window undo on SYN retransmit")
This fix tweaks the logic to be more similar to the
tcp_packet_delayed() logic before bc9f38c832, except that we take
care not to be fooled by the FLAG_SYN_ACKED code path zeroing out
tp->retrans_stamp (the bug noted and fixed by Yuchung in
bc9f38c832).
Note that this returns the high-level behavior of tcp_packet_delayed()
to again match the comment for the function, which says: "Nothing was
retransmitted or returned timestamp is less than timestamp of the
first retransmission." Note that this comment is in the original
2005-04-16 Linux git commit, so this is evidently long-standing
behavior.
Fixes: bc9f38c832 ("tcp: avoid unconditional congestion window undo on SYN retransmit")
Reported-by: Geumhwan Yu <geumhwan.yu@samsung.com>
Diagnosed-by: Geumhwan Yu <geumhwan.yu@samsung.com>
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Link: https://patch.msgid.link/20241001200517.2756803-2-ncardwell.sw@gmail.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Abhishek Chauhan says:
====================
Fix AQR PMA capabilities
Patch 1:-
AQR115c reports incorrect PMA capabilities which includes
10G/5G and also incorrectly disables capabilities like autoneg
and 10Mbps support.
AQR115c as per the Marvell databook supports speeds up to 2.5Gbps
with autonegotiation.
Patch 2:-
Remove the use of phy_set_max_speed in phy driver as the
function is mainly used in MAC driver to set the max
speed.
Instead use get_features to fix up Phy PMA capabilities for
AQR111, AQR111B0, AQR114C and AQCS109
====================
Link: https://patch.msgid.link/20241001224626.2400222-1-quic_abchauha@quicinc.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Remove the use of phy_set_max_speed in phy driver as the
function is mainly used in MAC driver to set the max
speed.
Instead use get_features to fix up Phy PMA capabilities for
AQR111, AQR111B0, AQR114C and AQCS109
Fixes: 038ba1dc4e ("net: phy: aquantia: add AQR111 and AQR111B0 PHY ID")
Fixes: 0974f1f03b ("net: phy: aquantia: remove false 5G and 10G speed ability for AQCS109")
Fixes: c278ec6443 ("net: phy: aquantia: add support for AQR114C PHY ID")
Link: https://lore.kernel.org/all/20240913011635.1286027-1-quic_abchauha@quicinc.com/T/
Signed-off-by: Abhishek Chauhan <quic_abchauha@quicinc.com>
Reviewed-by: Russell King (Oracle) <rmk+kernel@armlinux.org.uk>
Link: https://patch.msgid.link/20241001224626.2400222-3-quic_abchauha@quicinc.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Mix netns into all IPv4 FIB hashes to avoid massive collision when
inserting the same address in many netns.
Signed-off-by: Alexandre Ferrieux <alexandre.ferrieux@orange.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Link: https://patch.msgid.link/20241001231438.3855035-1-alexandre.ferrieux@orange.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Joe Damato says:
====================
ena: Link IRQs, queues, and NAPI instances
This series uses the netdev-genl API to link IRQs and queues to NAPI IDs
so that this information is queryable by user apps. This is particularly
useful for epoll-based busy polling apps which rely on having access to
the NAPI ID.
I've tested these commits on an EC2 instance with an ENA NIC configured
and have included test output in the commit messages for each patch
showing how to query the information.
I noted in the implementation that the driver requests an IRQ for
management purposes which does not have an associated NAPI. I tried
to take this into account in patch 1, but would appreciate if ENA
maintainers can verify I did this correctly.
v1: https://lore.kernel.org/all/20240930195617.37369-1-jdamato@fastly.com/
====================
Link: https://patch.msgid.link/20241002001331.65444-1-jdamato@fastly.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Link IRQs to NAPI instances with netif_napi_set_irq. This information
can be queried with the netdev-genl API. Note that the ENA device
appears to allocate an IRQ for management purposes which does not have a
NAPI associated with it; this commit takes this into consideration to
accurately construct a map between IRQs and NAPI instances.
Compare the output of /proc/interrupts for my ena device with the output of
netdev-genl after applying this patch:
$ cat /proc/interrupts | grep enp55s0 | cut -f1 --delimiter=':'
94
95
96
97
98
99
100
101
$ ./tools/net/ynl/cli.py --spec Documentation/netlink/specs/netdev.yaml \
--dump napi-get --json='{"ifindex": 2}'
[{'id': 8208, 'ifindex': 2, 'irq': 101},
{'id': 8207, 'ifindex': 2, 'irq': 100},
{'id': 8206, 'ifindex': 2, 'irq': 99},
{'id': 8205, 'ifindex': 2, 'irq': 98},
{'id': 8204, 'ifindex': 2, 'irq': 97},
{'id': 8203, 'ifindex': 2, 'irq': 96},
{'id': 8202, 'ifindex': 2, 'irq': 95},
{'id': 8201, 'ifindex': 2, 'irq': 94}]
Signed-off-by: Joe Damato <jdamato@fastly.com>
Reviewed-by: David Arinzon <darinzon@amazon.com>
Link: https://patch.msgid.link/20241002001331.65444-2-jdamato@fastly.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Brandon reports sporadic, non-sensical spikes in cumulative pressure
time (total=) when reading cpu.pressure at a high rate. This is due to
a race condition between reader aggregation and tasks changing states.
While it affects all states and all resources captured by PSI, in
practice it most likely triggers with CPU pressure, since scheduling
events are so frequent compared to other resource events.
The race context is the live snooping of ongoing stalls during a
pressure read. The read aggregates per-cpu records for stalls that
have concluded, but will also incorporate ad-hoc the duration of any
active state that hasn't been recorded yet. This is important to get
timely measurements of ongoing stalls. Those ad-hoc samples are
calculated on-the-fly up to the current time on that CPU; since the
stall hasn't concluded, it's expected that this is the minimum amount
of stall time that will enter the per-cpu records once it does.
The problem is that the path that concludes the state uses a CPU clock
read that is not synchronized against aggregators; the clock is read
outside of the seqlock protection. This allows aggregators to race and
snoop a stall with a longer duration than will actually be recorded.
With the recorded stall time being less than the last snapshot
remembered by the aggregator, a subsequent sample will underflow and
observe a bogus delta value, resulting in an erratic jump in pressure.
Fix this by moving the clock read of the state change into the seqlock
protection. This ensures no aggregation can snoop live stalls past the
time that's recorded when the state concludes.
Reported-by: Brandon Duffany <brandon@buildbuddy.io>
Link: https://bugzilla.kernel.org/show_bug.cgi?id=219194
Link: https://lore.kernel.org/lkml/20240827121851.GB438928@cmpxchg.org/
Fixes: df77430639 ("psi: Reduce calls to sched_clock() in psi")
Cc: stable@vger.kernel.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Chengming Zhou <chengming.zhou@linux.dev>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
As was tried in commit 4e103134b8 ("KVM: x86/mmu: Zap only the relevant
pages when removing a memslot"), all shadow pages, i.e. non-leaf SPTEs,
need to be zapped. All of the accounting for a shadow page is tied to the
memslot, i.e. the shadow page holds a reference to the memslot, for all
intents and purposes. Deleting the memslot without removing all relevant
shadow pages, as is done when KVM_X86_QUIRK_SLOT_ZAP_ALL is disabled,
results in NULL pointer derefs when tearing down the VM.
Reintroduce from that commit the code that walks the whole memslot when
there are active shadow MMU pages.
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
With the patch 0b6c5371c0 "Add missing topdown metrics events" eight
topdown metric events with numbers ranging from 0x8000 to 0x8700 were
added to the test since they were added as 'perf stat' default events.
Later the patch 951efb9976 "Update no event/metric expectations" kept
only 4 of those events(0x8000-0x8300).
Currently, the topdown events with numbers 0x8400 to 0x8700 are missing
from the list of expected events resulting in a failure. Add back the
missing topdown events.
Fixes: 951efb9976 ("perf test attr: Update no event/metric expectations")
Signed-off-by: Veronika Molnarova <vmolnaro@redhat.com>
Tested-by: Ian Rogers <irogers@google.com>
Cc: mpetlan@redhat.com
Link: https://lore.kernel.org/r/20240311081611.7835-1-vmolnaro@redhat.com
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
Yury reported a crash in the sfc driver originated from
netpoll_send_udp(). The netconsole sends a message and then netpoll
invokes the driver's NAPI function with a budget of zero. It is
dedicated to allow driver to free TX resources, that it may have used
while sending the packet.
In the netpoll case the driver invokes xdp_do_flush() unconditionally,
leading to crash because bpf_net_context was never assigned.
Invoke xdp_do_flush() only if budget is not zero.
Fixes: 401cb7dae8 ("net: Reference bpf_redirect_info via task_struct on PREEMPT_RT.")
Reported-by: Yury Vostrikov <mon@unformed.ru>
Closes: https://lore.kernel.org/5627f6d1-5491-4462-9d75-bc0612c26a22@app.fastmail.com
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Reviewed-by: Edward Cree <ecree.xilinx@gmail.com>
Link: https://patch.msgid.link/20241002125837.utOcRo6Y@linutronix.de
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
When configuring the fiber port, the DP83869 PHY driver incorrectly
calls linkmode_set_bit() with a bit mask (1 << 10) rather than a bit
number (10). This corrupts some other memory location -- in case of
arm64 the priv pointer in the same structure.
Since the advertising flags are updated from supported at the end of the
function the incorrect line isn't needed at all and can be removed.
Fixes: a29de52ba2 ("net: dp83869: Add ability to advertise Fiber connection")
Signed-off-by: Ingo van Lil <inguin@gmx.de>
Reviewed-by: Alexander Sverdlin <alexander.sverdlin@siemens.com>
Reviewed-by: Andrew Lunn <andrew@lunn.ch>
Link: https://patch.msgid.link/20241002161807.440378-1-inguin@gmx.de
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Jacob Keller says:
====================
packing: various improvements and KUnit tests
This series contains a handful of improvements and fixes for the packing
library, including the addition of KUnit tests.
There are two major changes which might be considered bug fixes:
1) The library is updated to handle arbitrary buffer lengths, fixing
undefined behavior when operating on buffers which are not a multiple
of 4 bytes.
2) The behavior of QUIRK_MSB_ON_THE_RIGHT is fixed to match the intended
behavior when operating on packings that are not byte aligned.
These are not sent to net because no driver currently depends on this
behavior. For (1), the existing users of the packing API all operate on
buffers which are multiples of 4-bytes. For (2), no driver currently uses
QUIRK_MSB_ON_THE_RIGHT. The incorrect behavior was found while writing
KUnit tests.
This series also includes a handful of minor cleanups from Vladimir, as
well as a change to introduce a separated pack() and unpack() API. This API
is not (yet) used by a driver, but is the first step in implementing
pack_fields() and unpack_fields() which will be used in future changes for
the ice driver and changes Vladimir has in progress for other drivers using
the packing API.
This series is part 1 of a 2-part series for implementing use of
lib/packing in the ice driver. The 2nd part includes a new pack_fields()
and unpack_fields() implementation inspired by the ice driver's existing
bit packing code. It is built on top of the split pack() and unpack()
code. Additionally, the KUnit tests are built on top of pack() and
unpack(), based on original selftests written by Vladimir.
Fitting the entire library changes and drivers changes into a single series
exceeded the usual series limits.
v1: https://lore.kernel.org/r/20240930-packing-kunit-tests-and-split-pack-unpack-v1-0-94b1f04aca85@intel.com
====================
Link: https://patch.msgid.link/20241002-packing-kunit-tests-and-split-pack-unpack-v2-0-8373e551eae3@intel.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
The QUIRK_MSB_ON_THE_RIGHT quirk is intended to modify pack() and unpack()
so that the most significant bit of each byte in the packed layout is on
the right.
The way the quirk is currently implemented is broken whenever the packing
code packs or unpacks any value that is not exactly a full byte.
The broken behavior can occur when packing any values smaller than one
byte, when packing any value that is not exactly a whole number of bytes,
or when the packing is not aligned to a byte boundary.
This quirk is documented in the following way:
1. Normally (no quirks), we would do it like this:
::
63 62 61 60 59 58 57 56 55 54 53 52 51 50 49 48 47 46 45 44 43 42 41 40 39 38 37 36 35 34 33 32
7 6 5 4
31 30 29 28 27 26 25 24 23 22 21 20 19 18 17 16 15 14 13 12 11 10 9 8 7 6 5 4 3 2 1 0
3 2 1 0
<snip>
2. If QUIRK_MSB_ON_THE_RIGHT is set, we do it like this:
::
56 57 58 59 60 61 62 63 48 49 50 51 52 53 54 55 40 41 42 43 44 45 46 47 32 33 34 35 36 37 38 39
7 6 5 4
24 25 26 27 28 29 30 31 16 17 18 19 20 21 22 23 8 9 10 11 12 13 14 15 0 1 2 3 4 5 6 7
3 2 1 0
That is, QUIRK_MSB_ON_THE_RIGHT does not affect byte positioning, but
inverts bit offsets inside a byte.
Essentially, the mapping for physical bit offsets should be reserved for a
given byte within the payload. This reversal should be fixed to the bytes
in the packing layout.
The logic to implement this quirk is handled within the
adjust_for_msb_right_quirk() function. This function does not work properly
when dealing with the bytes that contain only a partial amount of data.
In particular, consider trying to pack or unpack the range 53-44. We should
always be mapping the bits from the logical ordering to their physical
ordering in the same way, regardless of what sequence of bits we are
unpacking.
This, we should grab the following logical bits:
Logical: 55 54 53 52 51 50 49 48 47 45 44 43 42 41 40 39
^ ^ ^ ^ ^ ^ ^ ^ ^
And pack them into the physical bits:
Physical: 48 49 50 51 52 53 54 55 40 41 42 43 44 45 46 47
Logical: 48 49 50 51 52 53 44 45 46 47
^ ^ ^ ^ ^ ^ ^ ^ ^ ^
The current logic in adjust_for_msb_right_quirk is broken. I believe it is
intending to map according to the following:
Physical: 48 49 50 51 52 53 54 55 40 41 42 43 44 45 46 47
Logical: 48 49 50 51 52 53 44 45 46 47
^ ^ ^ ^ ^ ^ ^ ^ ^ ^
That is, it tries to keep the bits at the start and end of a packing
together. This is wrong, as it makes the packing change what bit is being
mapped to what based on which bits you're currently packing or unpacking.
Worse, the actual calculations within adjust_for_msb_right_quirk don't make
sense.
Consider the case when packing the last byte of an unaligned packing. It
might have a start bit of 7 and an end bit of 5. This would have a width of
3 bits. The new_start_bit will be calculated as the width - the box_end_bit
- 1. This will underflow and produce a negative value, which will
ultimate result in generating a new box_mask of all 0s.
For any other values, the result of the calculations of the
new_box_end_bit, new_box_start_bit, and the new box_mask will result in the
exact same values for the box_end_bit, box_start_bit, and box_mask. This
makes the calculations completely irrelevant.
If box_end_bit is 0, and box_start_bit is 7, then the entire function of
adjust_for_msb_right_quirk will boil down to just:
*to_write = bitrev8(*to_write)
The other adjustments are attempting (incorrectly) to keep the bits in the
same place but just reversed. This is not the right behavior even if
implemented correctly, as it leaves the mapping dependent on the bit values
being packed or unpacked.
Remove adjust_for_msb_right_quirk() and just use bitrev8 to reverse the
byte order when interacting with the packed data.
In particular, for packing, we need to reverse both the box_mask and the
physical value being packed. This is done after shifting the value by
box_end_bit so that the reversed mapping is always aligned to the physical
buffer byte boundary. The box_mask is reversed as we're about to use it to
clear any stale bits in the physical buffer at this block.
For unpacking, we need to reverse the contents of the physical buffer
*before* masking with the box_mask. This is critical, as the box_mask is a
logical mask of the bit layout before handling the QUIRK_MSB_ON_THE_RIGHT.
Add several new tests which cover this behavior. These tests will fail
without the fix and pass afterwards. Note that no current drivers make use
of QUIRK_MSB_ON_THE_RIGHT. I suspect this is why there have been no reports
of this inconsistency before.
Signed-off-by: Jacob Keller <jacob.e.keller@intel.com>
Reviewed-by: Przemek Kitszel <przemyslaw.kitszel@intel.com>
Link: https://patch.msgid.link/20241002-packing-kunit-tests-and-split-pack-unpack-v2-8-8373e551eae3@intel.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
While reviewing the initial KUnit tests for lib/packing, Przemek pointed
out that the test values have duplicate bytes in the input sequence.
In addition, I noticed that the unit tests pack and unpack on a byte
boundary, instead of crossing bytes. Thus, we lack good coverage of the
corner cases of the API.
Add additional unit tests to cover packing and unpacking byte buffers which
do not have duplicate bytes in the unpacked value, and which pack and
unpack to an unaligned offset.
A careful reviewer may note the lack tests for QUIRK_MSB_ON_THE_RIGHT. This
is because I found issues with that quirk during test implementation. This
quirk will be fixed and the tests will be included in a future change.
Signed-off-by: Jacob Keller <jacob.e.keller@intel.com>
Reviewed-by: Przemek Kitszel <przemyslaw.kitszel@intel.com>
Link: https://patch.msgid.link/20241002-packing-kunit-tests-and-split-pack-unpack-v2-7-8373e551eae3@intel.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Add 24 simple KUnit tests for the lib/packing.c pack() and unpack() APIs.
The first 16 tests exercise all combinations of quirks with a simple magic
number value on a 16-byte buffer. The remaining 8 tests cover
non-multiple-of-4 buffer sizes.
These tests were originally written by Vladimir as simple selftest
functions. I adapted them to KUnit, refactoring them into a table driven
approach. This will aid in adding additional tests in the future.
Co-developed-by: Vladimir Oltean <vladimir.oltean@nxp.com>
Signed-off-by: Vladimir Oltean <vladimir.oltean@nxp.com>
Signed-off-by: Jacob Keller <jacob.e.keller@intel.com>
Reviewed-by: Przemek Kitszel <przemyslaw.kitszel@intel.com>
Reviewed-by: Vladimir Oltean <vladimir.oltean@nxp.com>
Tested-by: Vladimir Oltean <vladimir.oltean@nxp.com>
Link: https://patch.msgid.link/20241002-packing-kunit-tests-and-split-pack-unpack-v2-6-8373e551eae3@intel.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
packing() is now used in some hot paths, and it would be good to get rid
of some ifs and buts that depend on "op", to speed things up a little bit.
With the main implementations now taking size_t endbit, we no longer
have to check for negative values. Update the local integer variables to
also be size_t to match.
Signed-off-by: Vladimir Oltean <vladimir.oltean@nxp.com>
Signed-off-by: Jacob Keller <jacob.e.keller@intel.com>
Reviewed-by: Przemek Kitszel <przemyslaw.kitszel@intel.com>
Reviewed-by: Vladimir Oltean <vladimir.oltean@nxp.com>
Link: https://patch.msgid.link/20241002-packing-kunit-tests-and-split-pack-unpack-v2-5-8373e551eae3@intel.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Geert Uytterhoeven described packing() as "really bad API" because of
not being able to enforce const correctness. The same function is used
both when "pbuf" is input and "uval" is output, as in the other way
around.
Create 2 wrapper functions where const correctness can be ensured.
Do ugly type casts inside, to be able to reuse packing() as currently
implemented - which will _not_ modify the input argument.
Also, take the opportunity to change the type of startbit and endbit to
size_t - an unsigned type - in these new function prototypes. When int,
an extra check for negative values is necessary. Hopefully, when
packing() goes away completely, that check can be dropped.
My concern is that code which does rely on the conditional directionality
of packing() is harder to refactor without blowing up in size. So it may
take a while to completely eliminate packing(). But let's make alternatives
available for those who do not need that.
Link: https://lore.kernel.org/netdev/20210223112003.2223332-1-geert+renesas@glider.be/
Signed-off-by: Vladimir Oltean <vladimir.oltean@nxp.com>
Signed-off-by: Jacob Keller <jacob.e.keller@intel.com>
Reviewed-by: Przemek Kitszel <przemyslaw.kitszel@intel.com>
Reviewed-by: Vladimir Oltean <vladimir.oltean@nxp.com>
Link: https://patch.msgid.link/20241002-packing-kunit-tests-and-split-pack-unpack-v2-4-8373e551eae3@intel.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
It is not necessary to have the kernel-doc duplicated both in the
header and in the implementation. It is better to have it near the
implementation of the function, since in C, a function can have N
declarations, but only one definition.
Signed-off-by: Vladimir Oltean <vladimir.oltean@nxp.com>
Signed-off-by: Jacob Keller <jacob.e.keller@intel.com>
Reviewed-by: Przemek Kitszel <przemyslaw.kitszel@intel.com>
Reviewed-by: Vladimir Oltean <vladimir.oltean@nxp.com>
Link: https://patch.msgid.link/20241002-packing-kunit-tests-and-split-pack-unpack-v2-3-8373e551eae3@intel.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Jacob Keller has a use case for packing() in the intel/ice networking
driver, but it cannot be used as-is.
Simply put, the API quirks for LSW32_IS_FIRST and LITTLE_ENDIAN are
naively implemented with the undocumented assumption that the buffer
length must be a multiple of 4. All calculations of group offsets and
offsets of bytes within groups assume that this is the case. But in the
ice case, this does not hold true. For example, packing into a buffer
of 22 bytes would yield wrong results, but pretending it was a 24 byte
buffer would work.
Rather than requiring such hacks, and leaving a big question mark when
it comes to discontinuities in the accessible bit fields of such buffer,
we should extend the packing API to support this use case.
It turns out that we can keep the design in terms of groups of 4 bytes,
but also make it work if the total length is not a multiple of 4.
Just like before, imagine the buffer as a big number, and its most
significant bytes (the ones that would make up to a multiple of 4) are
missing. Thus, with a big endian (no quirks) interpretation of the
buffer, those most significant bytes would be absent from the beginning
of the buffer, and with a LSW32_IS_FIRST interpretation, they would be
absent from the end of the buffer. The LITTLE_ENDIAN quirk, in the
packing() API world, only affects byte ordering within groups of 4.
Thus, it does not change which bytes are missing. Only the significance
of the remaining bytes within the (smaller) group.
No change intended for buffer sizes which are multiples of 4. Tested
with the sja1105 driver and with downstream unit tests.
Link: https://lore.kernel.org/netdev/a0338310-e66c-497c-bc1f-a597e50aa3ff@intel.com/
Signed-off-by: Vladimir Oltean <vladimir.oltean@nxp.com>
Tested-by: Jacob Keller <jacob.e.keller@intel.com>
Signed-off-by: Jacob Keller <jacob.e.keller@intel.com>
Reviewed-by: Przemek Kitszel <przemyslaw.kitszel@intel.com>
Reviewed-by: Vladimir Oltean <vladimir.oltean@nxp.com>
Link: https://patch.msgid.link/20241002-packing-kunit-tests-and-split-pack-unpack-v2-2-8373e551eae3@intel.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
While reworking the implementation, it became apparent that this check
does not exist.
There is no functional issue yet, because at call sites, "startbit" and
"endbit" are always hardcoded to correct values, and never come from the
user.
Even with the upcoming support of arbitrary buffer lengths, the
"startbit >= 8 * pbuflen" check will remain correct. This is because
we intend to always interpret the packed buffer in a way that avoids
discontinuities in the available bit indices.
Signed-off-by: Vladimir Oltean <vladimir.oltean@nxp.com>
Tested-by: Jacob Keller <jacob.e.keller@intel.com>
Signed-off-by: Jacob Keller <jacob.e.keller@intel.com>
Reviewed-by: Przemek Kitszel <przemyslaw.kitszel@intel.com>
Reviewed-by: Vladimir Oltean <vladimir.oltean@nxp.com>
Link: https://patch.msgid.link/20241002-packing-kunit-tests-and-split-pack-unpack-v2-1-8373e551eae3@intel.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
This commit dumps metadata with version 2. It dumps metadata for header
and per CPU data respectively in the arm_spe_print_info() function to
support metadata version 2 format.
After:
0 0 0x3c0 [0x1b0]: PERF_RECORD_AUXTRACE_INFO type: 4
Header version :2
Header size :4
PMU type v2 :13
CPU number :8
Magic :0x1010101010101010
CPU # :0
Num of params :3
MIDR :0x410fd801
PMU Type :-1
Min Interval :0
Magic :0x1010101010101010
CPU # :1
Num of params :3
MIDR :0x410fd801
PMU Type :-1
Min Interval :0
Magic :0x1010101010101010
CPU # :2
Num of params :3
MIDR :0x410fd870
PMU Type :13
Min Interval :1024
Magic :0x1010101010101010
CPU # :3
Num of params :3
MIDR :0x410fd870
PMU Type :13
Min Interval :1024
Magic :0x1010101010101010
CPU # :4
Num of params :3
MIDR :0x410fd870
PMU Type :13
Min Interval :1024
Magic :0x1010101010101010
CPU # :5
Num of params :3
MIDR :0x410fd870
PMU Type :13
Min Interval :1024
Magic :0x1010101010101010
CPU # :6
Num of params :3
MIDR :0x410fd850
PMU Type :-1
Min Interval :0
Magic :0x1010101010101010
CPU # :7
Num of params :3
MIDR :0x410fd850
PMU Type :-1
Min Interval :0
Signed-off-by: Leo Yan <leo.yan@arm.com>
Reviewed-by: James Clark <james.clark@linaro.org>
Cc: Will Deacon <will@kernel.org>
Cc: Mike Leach <mike.leach@linaro.org>
Cc: linux-arm-kernel@lists.infradead.org
Cc: Besar Wicaksono <bwicaksono@nvidia.com>
Cc: John Garry <john.g.garry@oracle.com>
Link: https://lore.kernel.org/r/20241003184302.190806-6-leo.yan@arm.com
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
This commit is to support metadata version 2 and at the meantime it is
backward compatible for version 1's format.
The metadata version 1 doesn't include the ARM_SPE_HEADER_VERSION field.
As version 1 is fixed with two u64 fields, by checking the metadata
size, it distinguishes the metadata is version 1 or version 2 (and any
new versions if later will have). For version 2, it reads out CPU number
and retrieves the metadata info for every CPU.
Signed-off-by: Leo Yan <leo.yan@arm.com>
Reviewed-by: James Clark <james.clark@linaro.org>
Cc: Will Deacon <will@kernel.org>
Cc: Mike Leach <mike.leach@linaro.org>
Cc: linux-arm-kernel@lists.infradead.org
Cc: Besar Wicaksono <bwicaksono@nvidia.com>
Cc: John Garry <john.g.garry@oracle.com>
Link: https://lore.kernel.org/r/20241003184302.190806-5-leo.yan@arm.com
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
Save the Arm SPE information on a per-CPU basis. This approach is easier
in the decoding phase for retrieving metadata based on the CPU number of
every Arm SPE record.
Signed-off-by: Leo Yan <leo.yan@arm.com>
Reviewed-by: James Clark <james.clark@linaro.org>
Cc: Will Deacon <will@kernel.org>
Cc: Mike Leach <mike.leach@linaro.org>
Cc: linux-arm-kernel@lists.infradead.org
Cc: Besar Wicaksono <bwicaksono@nvidia.com>
Cc: John Garry <john.g.garry@oracle.com>
Link: https://lore.kernel.org/r/20241003184302.190806-4-leo.yan@arm.com
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
The metadata is designed to contain a header and per CPU information.
The arm_spe_find_cpus() function is introduced to identify how many CPUs
support ARM SPE. Based on the CPU number, calculates the metadata size.
Signed-off-by: Leo Yan <leo.yan@arm.com>
Reviewed-by: James Clark <james.clark@linaro.org>
Cc: Will Deacon <will@kernel.org>
Cc: Mike Leach <mike.leach@linaro.org>
Cc: linux-arm-kernel@lists.infradead.org
Cc: Besar Wicaksono <bwicaksono@nvidia.com>
Cc: John Garry <john.g.garry@oracle.com>
Link: https://lore.kernel.org/r/20241003184302.190806-3-leo.yan@arm.com
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
The first version's metadata header structure doesn't include a field to
indicate a header version, which is not friendly for extension.
Define the metadata version 2 format with a new header structure and
extend per CPU's metadata. In the meantime, the old metadata header will
still be supported for backward compatibility.
Signed-off-by: Leo Yan <leo.yan@arm.com>
Reviewed-by: James Clark <james.clark@linaro.org>
Cc: Will Deacon <will@kernel.org>
Cc: Mike Leach <mike.leach@linaro.org>
Cc: linux-arm-kernel@lists.infradead.org
Cc: Besar Wicaksono <bwicaksono@nvidia.com>
Cc: John Garry <john.g.garry@oracle.com>
Link: https://lore.kernel.org/r/20241003184302.190806-2-leo.yan@arm.com
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
After commit 0edb555a65 ("platform: Make platform_driver::remove() return
void") .remove() is (again) the right callback to implement for platform
drivers.
Convert all PCI controller drivers to use .remove(), with the eventual goal
to drop struct platform_driver::remove_new(). As .remove() and
.remove_new() have the same prototypes, conversion is done by just changing
the structure member name in the driver initializer.
Link: https://lore.kernel.org/r/20240927092449.44628-1-sergio.paracuellos@gmail.com
Signed-off-by: Sergio Paracuellos <sergio.paracuellos@gmail.com>
Signed-off-by: Bjorn Helgaas <bhelgaas@google.com>
After commit 0edb555a65 ("platform: Make platform_driver::remove() return
void") .remove() is (again) the right callback to implement for platform
drivers.
Convert all PCI controller drivers to use .remove(), with the eventual goal
to drop struct platform_driver::remove_new(). As .remove() and .remove_new()
have the same prototypes, conversion is done by just changing the structure
member name in the driver initializer.
Link: https://lore.kernel.org/r/20240923065706.728769-1-sergio.paracuellos@gmail.com
Signed-off-by: Sergio Paracuellos <sergio.paracuellos@gmail.com>
[bhelgaas: add pcie-xilinx-nwl.c and tidy whitespace per Uwe Kleine-König:
https://lore.kernel.org/r/tdxrmmqyzcufupnwkdbg7lwgadizm7v3lxjirykijbml7x54ze@upbdzycdsilm]
Signed-off-by: Bjorn Helgaas <bhelgaas@google.com>
Push regular plane/color management updates to the DSB,
if other constraints allow it.
The first part of the sequence will go as follows:
- CPU will kick off DSB0 immediately
- DSB0 writes double bufferd non-arming registers
- DSB0 evades the vblank
- DSB0 writes double buffered arming registers
If no color management updates is needed we follow that up with:
- DSB0 waits for the undelayed vblank
- DSB0 waits for the delayed vblank (usec wait)
- DSB0 emits an interrupt which will cause the CPU to complete the commit
If color management update is needed:
- DSB0 will start DSB1 with wait for undelayed vblank
- DSB0 will in parallel perform the force DEwake tricks
- DSB1 writes single buffered LUT registers
- DSB1 waits for the delayed vblank (usec wait)
- DSB1 emits an interrupt which will cause the CPU to complete the commit
With this sequence we don't need to increase the vblank delay
to make room for register programming during vblank, which is
a good thing for high refresh rate display. But I'll need to
still think of some way to eliminate VRR commit completion
related races under this scheme.
Stuff that isn't ready for DSB yet:
- modesets (potentially we could do
at least the plane enabling via DSB)
- fastsets
- VRR
- PSR
- scalers
- async flips
Reviewed-by: Animesh Manna <animesh.manna@intel.com>
Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20240930170415.23841-14-ville.syrjala@linux.intel.com
We need to be able to do both MMIO and DSB based pipe/plane
programming. To that end plumb the 'dsb' all way from the top
into the plane commit hooks.
The compiler appears smart enough to combine the branches from
all the back-to-back register writes into a single branch.
So the generated asm ends up looking more or less like this:
plane_hook()
{
if (dsb) {
intel_dsb_reg_write();
intel_dsb_reg_write();
...
} else {
intel_de_write_fw();
intel_de_write_fw();
...
}
}
which seems like a reasonably efficient way to do this.
An alternative I was also considering is some kind of closure
(register write function + display vs. dsb pointer passed to it).
That does result is smaller code as there are no branches anymore,
but having each register access go via function pointer sounds
less efficient.
Not that I actually measured the overhead of either approach yet.
Also the reg_rw tracepoint seems to be making a huge mess of the
generated code for the mmio path. And additionally there's some
kind of IS_GSI_REG() hack in __raw_uncore_read() which ends up
generating a pointless branch for every mmio register access.
So looks like there might be quite a bit of room for improvement
in the mmio path still.
Reviewed-by: Animesh Manna <animesh.manna@intel.com>
Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20240930170415.23841-12-ville.syrjala@linux.intel.com
Add intel_dsb_wait_vblank_delay() which instructs the DSB
to wait for duration between the undelayed and delayed vblanks.
We'll need this as the DSB can only directly wait for the
undelayed vblank, but we'll need to wait until the delayed
vblank has elapsed as well.
Reviewed-by: Animesh Manna <animesh.manna@intel.com>
Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20240930170415.23841-10-ville.syrjala@linux.intel.com
Introduce intel_scanlines_to_usecs() as a counterpart to
intel_usecs_to_scanlines().
We'll have some use for this in DSB code as we want to do
relative scanline waits to evade the delayed vblank, but
unfortunately DSB can't do relative scanline waits (only
absolute). So we'll instead convert the relative scanline
count to usec and do a relative usec wait instead (which the
DSB knows how to do).
Reviewed-by: Animesh Manna <animesh.manna@intel.com>
Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20240930170415.23841-9-ville.syrjala@linux.intel.com
Add a function to emit a DSB wait for vblank instruction. This
just waits until the specified number of vblanks.
Note that this triggers on the transcoder's undelayed vblank,
as opposed to the pipe's delayed vblank.
Reviewed-by: Animesh Manna <animesh.manna@intel.com>
Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20240930170415.23841-8-ville.syrjala@linux.intel.com
Add a function to emit the DSB "wait usecs" instruction.
This is just a usleep() for the DSB.
As a lower bound it seems pretty accurate, but the upper bound
seemed oddly relaxed (ie. sometimes I've seen waits that are
quite a bit longer than specified, not sure why).
Reviewed-by: Animesh Manna <animesh.manna@intel.com>
Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20240930170415.23841-7-ville.syrjala@linux.intel.com
Add a helper for performing vblank evasion on the DSB. DSB based
plane updates will need this to guarantee all the double buffered
arming registers will get programmed atomically within the same
frame.
With VRR we more or less have two vblanks to worry about:
- vmax vblank start in case no push was sent
- vmin vblank start in case a push was already sent during
the vertical active. Only a concern for mailbox updates,
which I suppose could happen if the legacy cursor updates
take the non-fastpath without setting
state->legacy_cursor_update to false.
Since we don't know which case is relevant we'll just evade
both.
We must also make sure to evade both the delayed vblank
(for pipe/plane registers) and the undelayed vblank
(for transcoder registers and chained DSBs w/
DSB_WAIT_FOR_VBLANK).
TODO: come up with a sensible usec number for the evasion...
Reviewed-by: Animesh Manna <animesh.manna@intel.com>
Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20240930170415.23841-6-ville.syrjala@linux.intel.com
The DSB can signal a programmable interrupt in response to
a specific DSB command getting executed. Hook that up.
For now we'll just use this to signal the completion of the
commit via a vblank event. If, in the future, we'll need to
do other things in response to DSB interrupts we may need to
come up with some kind of fancier DSB interrupt framework where
the caller can specify a custom handler...
Reviewed-by: Animesh Manna <animesh.manna@intel.com>
Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20240930170415.23841-5-ville.syrjala@linux.intel.com
Once we start using DSB for plane updates we'll need to defer
generating the DSB buffer until the clear color has been
read out. So we need to move at some of the DSB stuff into
commit_tail(). That is perhaps a better place for it anyway
as the ioctl thread can move on immediately without spending
time building the DSB commands.
We always have the MMIO fallback (in case the DSB buffer
allocation fails), so there's no real reason to keep any
of this in the synchronous part of the ioctl.
Because the DSB LUT programming doesn't depend on the plane
clear color we can still do that part before waiting for
fences/etc. which should help paralleize things a bit more.
The DSB plane programming will need to happen after those
however as that depends on the clear color.
Reviewed-by: Animesh Manna <animesh.manna@intel.com>
Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20240930170415.23841-4-ville.syrjala@linux.intel.com
Read out the clear color as soon as fences and the transient
data flush have finished. There is no need to wait for
all the display specific operations that might still be
going on. This could parallelize things a bit more effectively.
Reviewed-by: Animesh Manna <animesh.manna@intel.com>
Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20240930170415.23841-3-ville.syrjala@linux.intel.com
Reading from the DSB command buffer might be somewhat expensive on
discrete GPUs because the buffer resides in GPU local memory. Avoid
such reads in the indexed register write handling by tracking the
previous instruction in intel_dsb.
TODO: actually measure this
Reviewed-by: Animesh Manna <animesh.manna@intel.com>
Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20240930170415.23841-2-ville.syrjala@linux.intel.com
There is another found exception that the "timerlat/1" thread was
scheduled on CPU0, and lead to timer corruption finally:
```
ODEBUG: init active (active state 0) object: ffff888237c2e108 object type: hrtimer hint: timerlat_irq+0x0/0x220
WARNING: CPU: 0 PID: 426 at lib/debugobjects.c:518 debug_print_object+0x7d/0xb0
Modules linked in:
CPU: 0 UID: 0 PID: 426 Comm: timerlat/1 Not tainted 6.11.0-rc7+ #45
Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.13.0-1ubuntu1.1 04/01/2014
RIP: 0010:debug_print_object+0x7d/0xb0
...
Call Trace:
<TASK>
? __warn+0x7c/0x110
? debug_print_object+0x7d/0xb0
? report_bug+0xf1/0x1d0
? prb_read_valid+0x17/0x20
? handle_bug+0x3f/0x70
? exc_invalid_op+0x13/0x60
? asm_exc_invalid_op+0x16/0x20
? debug_print_object+0x7d/0xb0
? debug_print_object+0x7d/0xb0
? __pfx_timerlat_irq+0x10/0x10
__debug_object_init+0x110/0x150
hrtimer_init+0x1d/0x60
timerlat_main+0xab/0x2d0
? __pfx_timerlat_main+0x10/0x10
kthread+0xb7/0xe0
? __pfx_kthread+0x10/0x10
ret_from_fork+0x2d/0x40
? __pfx_kthread+0x10/0x10
ret_from_fork_asm+0x1a/0x30
</TASK>
```
After tracing the scheduling event, it was discovered that the migration
of the "timerlat/1" thread was performed during thread creation. Further
analysis confirmed that it is because the CPU online processing for
osnoise is implemented through workers, which is asynchronous with the
offline processing. When the worker was scheduled to create a thread, the
CPU may has already been removed from the cpu_online_mask during the offline
process, resulting in the inability to select the right CPU:
T1 | T2
[CPUHP_ONLINE] | cpu_device_down()
osnoise_hotplug_workfn() |
| cpus_write_lock()
| takedown_cpu(1)
| cpus_write_unlock()
[CPUHP_OFFLINE] |
cpus_read_lock() |
start_kthread(1) |
cpus_read_unlock() |
To fix this, skip online processing if the CPU is already offline.
Cc: stable@vger.kernel.org
Cc: Masami Hiramatsu <mhiramat@kernel.org>
Cc: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Link: https://lore.kernel.org/20240924094515.3561410-4-liwei391@huawei.com
Fixes: c8895e271f ("trace/osnoise: Support hotplug operations")
Signed-off-by: Wei Li <liwei391@huawei.com>
Signed-off-by: Steven Rostedt (Google) <rostedt@goodmis.org>
stop_kthread() is the offline callback for "trace/osnoise:online", since
commit 5bfbcd1ee5 ("tracing/timerlat: Add interface_lock around clearing
of kthread in stop_kthread()"), the following ABBA deadlock scenario is
introduced:
T1 | T2 [BP] | T3 [AP]
osnoise_hotplug_workfn() | work_for_cpu_fn() | cpuhp_thread_fun()
| _cpu_down() | osnoise_cpu_die()
mutex_lock(&interface_lock) | | stop_kthread()
| cpus_write_lock() | mutex_lock(&interface_lock)
cpus_read_lock() | cpuhp_kick_ap() |
As the interface_lock here in just for protecting the "kthread" field of
the osn_var, use xchg() instead to fix this issue. Also use
for_each_online_cpu() back in stop_per_cpu_kthreads() as it can take
cpu_read_lock() again.
Cc: stable@vger.kernel.org
Cc: Masami Hiramatsu <mhiramat@kernel.org>
Cc: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Link: https://lore.kernel.org/20240924094515.3561410-3-liwei391@huawei.com
Fixes: 5bfbcd1ee5 ("tracing/timerlat: Add interface_lock around clearing of kthread in stop_kthread()")
Signed-off-by: Wei Li <liwei391@huawei.com>
Signed-off-by: Steven Rostedt (Google) <rostedt@goodmis.org>