Gal Pressman [Thu, 19 Jan 2023 12:09:00 +0000 (14:09 +0200)]
net/mlx5e: Remove incorrect debugfs_create_dir NULL check in TLS
Remove the NULL check on debugfs_create_dir() return value as the
function returns an ERR pointer on failure, not NULL.
The check is not replaced with a IS_ERR_OR_NULL() as
debugfs_create_file(), and debugfs functions in general don't need error
checking.
Gal Pressman [Thu, 19 Jan 2023 12:09:00 +0000 (14:09 +0200)]
net/mlx5e: Remove incorrect debugfs_create_dir NULL check in hairpin
Remove the NULL check on debugfs_create_dir() return value as the
function returns an ERR pointer on failure, not NULL.
The check is not replaced with a IS_ERR_OR_NULL() as
debugfs_create_file(), and debugfs functions in general don't need error
checking.
Leon Romanovsky [Mon, 9 Jan 2023 12:11:01 +0000 (14:11 +0200)]
net/mlx5e: Don't listen to remove flows event
remove_flow_enable event isn't really needed as it will be
triggered once user and/or XFRM core explicitly asked to delete
state. In such situation, we won't be interested in any event
at all.
So don't trigger and listen to remove_flow_enable event.
Moshe Shemesh [Thu, 26 Jan 2023 06:58:33 +0000 (08:58 +0200)]
net/mlx5: fw reset: Skip device ID check if PCI link up failed
In case where after reset the PCI link is not ready within timeout, skip
reading device ID as if there is no PCI link up we can't have FW
response to pci config cycles either.
This also fixes err value not used and overwritten in such flow.
Shay Drory [Thu, 5 Jan 2023 08:03:46 +0000 (10:03 +0200)]
net/mlx5: Remove redundant health work lock
Commit 90e7cb78b815 ("net/mlx5: fix missing mutex_unlock in
mlx5_fw_fatal_reporter_err_work()") introduced another checking of
MLX5_DROP_HEALTH_NEW_WORK. At this point, the first check of
MLX5_DROP_HEALTH_NEW_WORK is redundant and so is the lock that
protects it.
Remove the lock and rename MLX5_DROP_HEALTH_NEW_WORK to reflect these
changes.
Arnd Bergmann [Tue, 17 Jan 2023 21:01:55 +0000 (22:01 +0100)]
mlx5: reduce stack usage in mlx5_setup_tc
Clang warns about excessive stack usage on 32-bit targets:
drivers/net/ethernet/mellanox/mlx5/core/en_main.c:3597:12: error: stack frame size (1184) exceeds limit (1024) in 'mlx5e_setup_tc' [-Werror,-Wframe-larger-than]
static int mlx5e_setup_tc(struct net_device *dev, enum tc_setup_type type,
It turns out that both the mlx5e_setup_tc_mqprio_dcb() function and
the mlx5e_safe_switch_params() function it calls have a copy of
'struct mlx5e_params' on the stack, and this structure is fairly
large.
David Howells [Tue, 31 Jan 2023 21:40:18 +0000 (21:40 +0000)]
rxrpc: Reduce unnecessary ack transmission
rxrpc_recvmsg_data() schedules an ACK to be transmitted every time at least
two packets have been consumed and any time it runs out of data and would
return -EAGAIN to the caller. Both events may occur within a single loop,
however, and if the I/O thread is quick enough it may send duplicate ACKs.
The ACKs are sent to inform the peer that more space has been made in the
local Rx window, but the I/O thread is going to send an ACK every couple of
DATA packets anyway, so we end up sending a lot more ACKs than we really
need to.
So reduce the rate at which recvmsg() schedules ACKs, such that if the I/O
thread sends ACKs at its normal faster rate, recvmsg() won't actually
schedule ACKs until the Rx flow stops (call->rx_consumed is cleared any
time we transmit an ACK for that call, resetting the counter used by
recvmsg).
David Howells [Tue, 29 Nov 2022 12:37:37 +0000 (12:37 +0000)]
rxrpc: Fix overwaking on call poking
If an rxrpc call is given a poke, it will get woken up unconditionally,
even if there's already a poke pending (for which there will have been a
wake) or if the call refcount has gone to 0.
Fix this by only waking the call if it is still referenced and if it
doesn't already have a poke pending.
Linus Torvalds [Tue, 7 Feb 2023 22:17:12 +0000 (14:17 -0800)]
Merge tag 'devicetree-fixes-for-6.2-2' of git://git.kernel.org/pub/scm/linux/kernel/git/robh/linux
Pull devicetree fixes from Rob Herring:
- Fix handling of multiple OF framebuffer devices
- Fix booting on Socionext Synquacer with bad 'dma-ranges' entries
- Add DT binding .yamllint to .gitignore
* tag 'devicetree-fixes-for-6.2-2' of git://git.kernel.org/pub/scm/linux/kernel/git/robh/linux:
dt-bindings: interrupt-controller: arm,gic-v3: Fix typo in description of msi-controller property
dt-bindings: Fix .gitignore
of/address: Return an error when no valid dma-ranges are found
of: Make OF framebuffer device names unique
- After Kees Cook patches and this one, we might
be able to revert commit dbae2b062824 ("net: skb: introduce and use a single page frag cache")
because GRO_MAX_HEAD is also small.
Linus Torvalds [Tue, 7 Feb 2023 15:54:40 +0000 (07:54 -0800)]
Merge tag 'trace-v6.2-rc6' of git://git.kernel.org/pub/scm/linux/kernel/git/trace/linux-trace
Pull tracing fix from Steven Rostedt:
"Fix regression in poll() and select()
With the fix that made poll() and select() block if read would block
caused a slight regression in rasdaemon, as it needed that kind of
behavior. Add a way to make that behavior come back by writing zero
into the 'buffer_percentage', which means to never block on read"
* tag 'trace-v6.2-rc6' of git://git.kernel.org/pub/scm/linux/kernel/git/trace/linux-trace:
tracing: Fix poll() and select() do not work on per_cpu trace_pipe and trace_pipe_raw
this is a pull request of 47 patches for net-next/master.
The first two patch is by Oliver Hartkopp. One adds missing error
checking to the CAN_GW protocol, the other adds a missing CAN address
family check to the CAN ISO TP protocol.
Thomas Kopp contributes a performance optimization to the mcp251xfd
driver.
The next 11 patches are by Geert Uytterhoeven and add support for
R-Car V4H systems to the rcar_canfd driver.
Stephane Grosjean and Lukas Magel contribute 8 patches to the peak_usb
driver, which add support for configurable CAN channel ID.
The last 17 patches are by me and target the CAN bit timing
configuration. The bit timing is cleaned up, error messages are
improved and forwarded to user space via NL_SET_ERR_MSG_FMT() instead
of netdev_err(), and the SJW handling is updated, including the
definition of a new default value that will benefit CAN-FD
controllers, by increasing their oscillator tolerance.
* tag 'linux-can-next-for-6.3-20230206' of git://git.kernel.org/pub/scm/linux/kernel/git/mkl/linux-can-next: (47 commits)
can: bittiming: can_validate_bitrate(): report error via netlink
can: bittiming: can_calc_bittiming(): convert from netdev_err() to NL_SET_ERR_MSG_FMT()
can: bittiming: can_calc_bittiming(): clean up SJW handling
can: bittiming: can_sjw_set_default(): use Phase Seg2 / 2 as default for SJW
can: bittiming: can_sjw_check(): check that SJW is not longer than either Phase Buffer Segment
can: bittiming: can_sjw_check(): report error via netlink and harmonize error value
can: bittiming: can_fixup_bittiming(): report error via netlink and harmonize error value
can: bittiming: factor out can_sjw_set_default() and can_sjw_check()
can: bittiming: can_changelink() pass extack down callstack
can: netlink: can_changelink(): convert from netdev_err() to NL_SET_ERR_MSG_FMT()
can: netlink: can_validate(): validate sample point for CAN and CAN-FD
can: dev: register_candev(): bail out if both fixed bit rates and bit timing constants are provided
can: dev: register_candev(): ensure that bittiming const are valid
can: bittiming: can_get_bittiming(): use direct return and remove unneeded else
can: bittiming: can_fixup_bittiming(): set effective tq
can: bittiming: can_fixup_bittiming(): use CAN_SYNC_SEG instead of 1
can: bittiming(): replace open coded variants of can_bit_time()
can: peak_usb: Reorder include directives alphabetically
can: peak_usb: align CAN channel ID format in log with sysfs attribute
can: peak_usb: export PCAN CAN channel ID as sysfs device attribute
...
====================
Vladimir Oltean [Mon, 6 Feb 2023 09:49:32 +0000 (11:49 +0200)]
ethtool: mm: fix get_mm() return code not propagating to user space
If ops->get_mm() returns a non-zero error code, we goto out_complete,
but there, we return 0. Fix that to propagate the "ret" variable to the
caller. If ops->get_mm() succeeds, it will always return 0.
can: j1939: do not wait 250 ms if the same addr was already claimed
The ISO 11783-5 standard, in "4.5.2 - Address claim requirements", states:
d) No CF shall begin, or resume, transmission on the network until 250
ms after it has successfully claimed an address except when
responding to a request for address-claimed.
But "Figure 6" and "Figure 7" in "4.5.4.2 - Address-claim
prioritization" show that the CF begins the transmission after 250 ms
from the first AC (address-claimed) message even if it sends another AC
message during that time window to resolve the address contention with
another CF.
As stated in "4.4.2.3 - Address-claimed message":
In order to successfully claim an address, the CF sending an address
claimed message shall not receive a contending claim from another CF
for at least 250 ms.
As stated in "4.4.3.2 - NAME management (NM) message":
1) A commanding CF can
d) request that a CF with a specified NAME transmit the address-
claimed message with its current NAME.
2) A target CF shall
d) send an address-claimed message in response to a request for a
matching NAME
Taking the above arguments into account, the 250 ms wait is requested
only during network initialization.
Do not restart the timer on AC message if both the NAME and the address
match and so if the address has already been claimed (timer has expired)
or the AC message has been sent to resolve the contention with another
CF (timer is still running).
Jiri Pirko [Mon, 6 Feb 2023 09:41:51 +0000 (10:41 +0100)]
devlink: change port event netdev notifier from per-net to global
Currently only the network namespace of devlink instance is monitored
for port events. If netdev is moved to a different namespace and then
unregistered, NETDEV_PRE_UNINIT is missed which leads to trigger
following WARN_ON in devl_port_unregister().
WARN_ON(devlink_port->type != DEVLINK_PORT_TYPE_NOTSET);
Fix this by changing the netdev notifier from per-net to global so no
event is missed.
Vladimir Oltean [Sun, 5 Feb 2023 19:24:09 +0000 (21:24 +0200)]
selftests: ocelot: tc_flower_chains: make test_vlan_ingress_modify() more comprehensive
We have two IS1 filters of the OCELOT_VCAP_KEY_ANY key type (the one with
"action vlan pop" and the one with "action vlan modify") and one of the
OCELOT_VCAP_KEY_IPV4 key type (the one with "action skbedit priority").
But we have no IS1 filter with the OCELOT_VCAP_KEY_ETYPE key type, and
there was an uncaught breakage there.
To increase test coverage, convert one of the OCELOT_VCAP_KEY_ANY
filters to OCELOT_VCAP_KEY_ETYPE, by making the filter also match on the
MAC SA of the traffic sent by mausezahn, $h1_mac.
Vladimir Oltean [Sun, 5 Feb 2023 19:24:08 +0000 (21:24 +0200)]
net: mscc: ocelot: fix VCAP filters not matching on MAC with "protocol 802.1Q"
Alternative short title: don't instruct the hardware to match on
EtherType with "protocol 802.1Q" flower filters. It doesn't work for the
reasons detailed below.
the created filter is set by ocelot_flower_parse_key() to be of type
OCELOT_VCAP_KEY_ETYPE, and etype is set to {value=0x8100, mask=0xffff}.
This gets propagated all the way to is1_entry_set() which commits it to
hardware (the VCAP_IS1_HK_ETYPE field of the key). Compare this to the
case where src_mac isn't specified - the key type is OCELOT_VCAP_KEY_ANY,
and is1_entry_set() doesn't populate VCAP_IS1_HK_ETYPE.
The problem is that for VLAN-tagged frames, the hardware interprets the
ETYPE field as holding the encapsulated VLAN protocol. So the above
filter will only match those packets which have an encapsulated protocol
of 0x8100, rather than all packets with VLAN ID 200 and the given src_mac.
The reason why this is allowed to occur is because, although we have a
block of code in ocelot_flower_parse_key() which sets "match_protocol"
to false when VLAN keys are present, that code executes too late.
There is another block of code, which executes for Ethernet addresses,
and has a "goto finished_key_parsing" and skips the VLAN header parsing.
By skipping it, "match_protocol" remains with the value it was
initialized with, i.e. "true", and "proto" is set to f->common.protocol,
or 0x8100.
The concept of ignoring some keys rather than erroring out when they are
present but can't be offloaded is dubious in itself, but is present
since the initial commit fe3490e6107e ("net: mscc: ocelot: Hardware
ofload for tc flower filter"), and it's outside of the scope of this
patch to change that.
The problem was introduced when the driver started to interpret the
flower filter's protocol, and populate the VCAP filter's ETYPE field
based on it.
To fix this, it is sufficient to move the code that parses the VLAN keys
earlier than the "goto finished_key_parsing" instruction. This will
ensure that if we have a flower filter with both VLAN and Ethernet
address keys, it won't match on ETYPE 0x8100, because the VLAN key
parsing sets "match_protocol = false".
Vladimir Oltean [Sun, 5 Feb 2023 14:07:13 +0000 (16:07 +0200)]
net: dsa: mt7530: don't change PVC_EG_TAG when CPU port becomes VLAN-aware
Frank reports that in a mt7530 setup where some ports are standalone and
some are in a VLAN-aware bridge, 8021q uppers of the standalone ports
lose their VLAN tag on xmit, as seen by the link partner.
This seems to occur because once the other ports join the VLAN-aware
bridge, mt7530_port_vlan_filtering() also calls
mt7530_port_set_vlan_aware(ds, cpu_dp->index), and this affects the way
that the switch processes the traffic of the standalone port.
Relevant is the PVC_EG_TAG bit. The MT7530 documentation says about it:
EG_TAG: Incoming Port Egress Tag VLAN Attribution
0: disabled (system default)
1: consistent (keep the original ingress tag attribute)
My interpretation is that this setting applies on the ingress port, and
"disabled" is basically the normal behavior, where the egress tag format
of the packet (tagged or untagged) is decided by the VLAN table
(MT7530_VLAN_EGRESS_UNTAG or MT7530_VLAN_EGRESS_TAG).
But there is also an option of overriding the system default behavior,
and for the egress tagging format of packets to be decided not by the
VLAN table, but simply by copying the ingress tag format (if ingress was
tagged, egress is tagged; if ingress was untagged, egress is untagged;
aka "consistent). This is useful in 2 scenarios:
- VLAN-unaware bridge ports will always encounter a miss in the VLAN
table. They should forward a packet as-is, though. So we use
"consistent" there. See commit e045124e9399 ("net: dsa: mt7530: fix
tagged frames pass-through in VLAN-unaware mode").
- Traffic injected from the CPU port. The operating system is in god
mode; if it wants a packet to exit as VLAN-tagged, it sends it as
VLAN-tagged. Otherwise it sends it as VLAN-untagged*.
*This is true only if we don't consider the bridge TX forwarding offload
feature, which mt7530 doesn't support.
So for now, make the CPU port always stay in "consistent" mode to allow
software VLANs to be forwarded to their egress ports with the VLAN tag
intact, and not stripped.
Eddy Tao [Sun, 5 Feb 2023 01:35:37 +0000 (09:35 +0800)]
net: openvswitch: reduce cpu_used_mask memory
Use actual CPU number instead of hardcoded value to decide the size
of 'cpu_used_mask' in 'struct sw_flow'. Below is the reason.
'struct cpumask cpu_used_mask' is embedded in struct sw_flow.
Its size is hardcoded to CONFIG_NR_CPUS bits, which can be
8192 by default, it costs memory and slows down ovs_flow_alloc.
To address this:
Redefine cpu_used_mask to pointer.
Append cpumask_size() bytes after 'stat' to hold cpumask.
Initialization cpu_used_mask right after stats_last_writer.
APIs like cpumask_next and cpumask_set_cpu never access bits
beyond cpu count, cpumask_size() bytes of memory is enough.
Arnd Bergmann [Fri, 3 Feb 2023 12:15:36 +0000 (13:15 +0100)]
amd-xgbe: fix mismatched prototype
The forward declaration was introduced with a prototype that does
not match the function definition:
drivers/net/ethernet/amd/xgbe/xgbe-phy-v2.c:2166:13: error: conflicting types for 'xgbe_phy_perform_ratechange' due to enum/integer mismatch; have 'void(struct xgbe_prv_data *, enum xgbe_mb_cmd, enum xgbe_mb_subcmd)' [-Werror=enum-int-mismatch]
2166 | static void xgbe_phy_perform_ratechange(struct xgbe_prv_data *pdata,
| ^~~~~~~~~~~~~~~~~~~~~~~~~~~
drivers/net/ethernet/amd/xgbe/xgbe-phy-v2.c:391:13: note: previous declaration of 'xgbe_phy_perform_ratechange' with type 'void(struct xgbe_prv_data *, unsigned int, unsigned int)'
391 | static void xgbe_phy_perform_ratechange(struct xgbe_prv_data *pdata,
| ^~~~~~~~~~~~~~~~~~~~~~~~~~~
Ideally there should not be any forward declarations here, which
would make it easier to show that there is no unbounded recursion.
I tried fixing this but could not figure out how to avoid the
recursive call.
As a hotfix, address only the broken prototype to fix the build
problem instead.
There are no external users of the vsc7514_*_regmap[] symbols or
vsc7514_vcap_* functions. They were exported in commit 32ecd22ba60b ("net:
mscc: ocelot: split register definitions to a separate file") with the
intention of being used, but the actual structure used in commit 2efaca411c96 ("net: mscc: ocelot: expose vsc7514_regmap definition") ended
up being all that was needed.
Jakub Kicinski [Tue, 7 Feb 2023 06:24:33 +0000 (22:24 -0800)]
Merge branch 'aux-bus-v11' of https://github.com/ajitkhaparde1/linux
Ajit Khaparde says:
====================
bnxt: Add Auxiliary driver support
Add auxiliary device driver for Broadcom devices.
The bnxt_en driver will register and initialize an aux device
if RDMA is enabled in the underlying device.
The bnxt_re driver will then probe and initialize the
RoCE interfaces with the infiniband stack.
We got rid of the bnxt_en_ops which the bnxt_re driver used to
communicate with bnxt_en.
Similarly We have tried to clean up most of the bnxt_ulp_ops.
In most of the cases we used the functions and entry points provided
by the auxiliary bus driver framework.
And now these are the minimal functions needed to support the functionality.
We will try to work on getting rid of the remaining if we find any
other viable option in future.
* 'aux-bus-v11' of https://github.com/ajitkhaparde1/linux:
bnxt_en: Remove runtime interrupt vector allocation
RDMA/bnxt_re: Remove the sriov config callback
bnxt_en: Remove struct bnxt access from RoCE driver
bnxt_en: Use auxiliary bus calls over proprietary calls
bnxt_en: Use direct API instead of indirection
bnxt_en: Remove usage of ulp_id
RDMA/bnxt_re: Use auxiliary driver interface
bnxt_en: Add auxiliary driver support
====================
Dan Carpenter [Fri, 28 Oct 2022 15:02:24 +0000 (18:02 +0300)]
ice: Fix off by one in ice_tc_forward_to_queue()
The > comparison should be >= to prevent reading one element beyond
the end of the array.
The "vsi->num_rxq" is not strictly speaking the number of elements in
the vsi->rxq_map[] array. The array has "vsi->alloc_rxq" elements and
"vsi->num_rxq" is less than or equal to the number of elements in the
array. The array is allocated in ice_vsi_alloc_arrays(). It's still
an off by one but it might not access outside the end of the array.
Brett Creeley [Mon, 16 Jan 2023 12:34:58 +0000 (13:34 +0100)]
ice: Fix disabling Rx VLAN filtering with port VLAN enabled
If the user turns on the vf-true-promiscuous-support flag, then Rx VLAN
filtering will be disabled if the VF requests to enable promiscuous
mode. When the VF is in a port VLAN, this is the incorrect behavior
because it will allow the VF to receive traffic outside of its port VLAN
domain. Fortunately this only resulted in the VF(s) receiving broadcast
traffic outside of the VLAN domain because all of the VLAN promiscuous
rules are based on the port VLAN ID. Fix this by setting the
.disable_rx_filtering VLAN op to a no-op when a port VLAN is enabled on
the VF.
Also, make sure to make this fix for both Single VLAN Mode and Double
VLAN Mode enabled devices.
Fixes: c31af68a1b94 ("ice: Add outer_vlan_ops and VSI specific VLAN ops implementations") Signed-off-by: Brett Creeley <[email protected]> Signed-off-by: Karen Ostrowska <[email protected]> Tested-by: Marek Szlosek <[email protected]> Signed-off-by: Tony Nguyen <[email protected]>
The ICE_AQ_LINK_SPEED_UNKNOWN define is BIT(15). The value is bigger
than both legacy and normal link speed tables. Add one element (0 -
unknown) to both tables. There is no need to explicitly set table size,
leave it empty.
Fixes: 1d0e28a9be1f ("ice: Remove and replace ice speed defines with ethtool.h versions") Signed-off-by: Michal Swiatkowski <[email protected]> Reviewed-by: Alexander Lobakin <[email protected]> Tested-by: Gurucharan G <[email protected]> (A Contingent worker at Intel) Signed-off-by: Tony Nguyen <[email protected]> Reviewed-by: Leon Romanovsky <[email protected]>
When both ice and the irdma driver are loaded, a warning in
check_flush_dependency is being triggered. This is due to ice driver
workqueue being allocated with the WQ_MEM_RECLAIM flag and the irdma one
is not.
According to kernel documentation, this flag should be set if the
workqueue will be involved in the kernel's memory reclamation flow.
Since it is not, there is no need for the ice driver's WQ to have this
flag set so remove it.
The /proc/version_signature file doesn't present in at least the
older versions of Debian distributions (eg, Debian 9, 10). The Debian
kernel has a similar issue where the release information from uname()
syscall doesn't give the kernel code version that matches what the
kernel actually expects. Below is an example content from Debian 10.
Debian reports incorrect kernel version in utsname::release returned
by uname() syscall, which in older kernels (Debian 9, 10) leads to
kprobe BPF programs failing to load due to the version check mismatch.
Fortunately, the correct kernel code version presents in the
utsname::version returned by uname() syscall in Debian kernels. This
change adds another get kernel version function to handle Debian in
addition to the previously added get kernel version function to handle
Ubuntu. Some minor refactoring work is also done to make the code more
readable.
Linus Torvalds [Mon, 6 Feb 2023 22:07:38 +0000 (14:07 -0800)]
Merge tag 'cgroup-for-6.2-rc7-fixes' of git://git.kernel.org/pub/scm/linux/kernel/git/tj/cgroup
Pull cgroup fixes from Tejun Heo:
"During the v6.2 cycle, there were a series of changes to task cpu
affinity handling which fixed cpuset inadvertently clobbering
user-configured affinity masks. Unfortunately, they broke the affinity
handling on hybrid heterogeneous CPUs which have cores that can
execute both 64 and 32bit along with cores that can only execute 32bit
code.
This contains two fix patches for the above issue. While reverting the
changes that caused the regression is definitely an option, the
origial patches do improve how cpuset behave signficantly in some
cases and the fixes seem fairly safe, so I think it'd be better to try
to fix them first"
* tag 'cgroup-for-6.2-rc7-fixes' of git://git.kernel.org/pub/scm/linux/kernel/git/tj/cgroup:
cpuset: Call set_cpus_allowed_ptr() with appropriate mask for task
cgroup/cpuset: Don't filter offline CPUs in cpuset_cpus_allowed() for top cpuset tasks
Linus Torvalds [Mon, 6 Feb 2023 22:05:16 +0000 (14:05 -0800)]
Merge tag 'for-6.2-rc7-tag' of git://git.kernel.org/pub/scm/linux/kernel/git/kdave/linux
Pull btrfs fixes from David Sterba:
- explicitly initialize zlib work memory to fix a KCSAN warning
- limit number of send clones by maximum memory allocated
- limit device size extent in case it device shrink races with chunk
allocation
- raid56 fixes:
- fix copy&paste error in RAID6 stripe recovery
- make error bitmap update atomic
* tag 'for-6.2-rc7-tag' of git://git.kernel.org/pub/scm/linux/kernel/git/kdave/linux:
btrfs: raid56: make error_bitmap update atomic
btrfs: send: limit number of clones and allocated memory size
btrfs: zlib: zero-initialize zlib workspace
btrfs: limit device extents to the device size
btrfs: raid56: fix stripes if vertical errors are found
Will Deacon [Tue, 31 Jan 2023 22:17:19 +0000 (22:17 +0000)]
cpuset: Call set_cpus_allowed_ptr() with appropriate mask for task
set_cpus_allowed_ptr() will fail with -EINVAL if the requested
affinity mask is not a subset of the task_cpu_possible_mask() for the
task being updated. Consequently, on a heterogeneous system with cpusets
spanning the different CPU types, updates to the cgroup hierarchy can
silently fail to update task affinities when the effective affinity
mask for the cpuset is expanded.
For example, consider an arm64 system with 4 CPUs, where CPUs 2-3 are
the only cores capable of executing 32-bit tasks. Attaching a 32-bit
task to a cpuset containing CPUs 0-2 will correctly affine the task to
CPU 2. Extending the cpuset to CPUs 0-3, however, will fail to extend
the affinity mask of the 32-bit task because update_tasks_cpumask() will
pass the full 0-3 mask to set_cpus_allowed_ptr().
Extend update_tasks_cpumask() to take a temporary 'cpumask' paramater
and use it to mask the 'effective_cpus' mask with the possible mask for
each task being updated.
Fixes: 431c69fac05b ("cpuset: Honour task_cpu_possible_mask() in guarantee_online_cpus()") Signed-off-by: Will Deacon <[email protected]> Acked-by: Waiman Long <[email protected]> Signed-off-by: Tejun Heo <[email protected]>
Waiman Long [Mon, 6 Feb 2023 03:48:53 +0000 (22:48 -0500)]
cgroup/cpuset: Don't filter offline CPUs in cpuset_cpus_allowed() for top cpuset tasks
Since commit 8f9ea86fdf99 ("sched: Always preserve the user
requested cpumask"), relax_compatible_cpus_allowed_ptr() is calling
__sched_setaffinity() unconditionally. This helps to expose a bug in
the current cpuset hotplug code where the cpumasks of the tasks in
the top cpuset are not updated at all when some CPUs become online or
offline. It is likely caused by the fact that some of the tasks in the
top cpuset, like percpu kthreads, cannot have their cpu affinity changed.
One way to reproduce this as suggested by Peter is:
- boot machine
- offline all CPUs except one
- taskset -p ffffffff $$
- online all CPUs
Fix this by allowing cpuset_cpus_allowed() to return a wider mask that
includes offline CPUs for those tasks that are in the top cpuset. For
tasks not in the top cpuset, the old rule applies and only online CPUs
will be returned in the mask since hotplug events will update their
cpumasks accordingly.
Jacob Keller [Thu, 19 Jan 2023 01:16:53 +0000 (17:16 -0800)]
ice: remove unnecessary virtchnl_ether_addr struct use
The dev_lan_addr and hw_lan_addr members of ice_vf are used only to store
the MAC address for the VF. They are defined using virtchnl_ether_addr, but
only the .addr sub-member is actually used. Drop the use of
virtchnl_ether_addr and just use a u8 array of length [ETH_ALEN].
Jacob Keller [Thu, 19 Jan 2023 01:16:52 +0000 (17:16 -0800)]
ice: introduce .irq_close VF operation
The Scalable IOV implementation will require notifying the VDCM driver when
an IRQ must be closed. This allows the VDCM to handle releasing stale IRQ
context values and properly reconfigure.
To handle this, introduce a new optional .irq_close callback to the VF
operations structure. This will be implemented by Scalable IOV to handle
the shutdown of the IRQ context.
Since the SR-IOV implementation does not need this, we must check that its
non-NULL before calling it.
Jacob Keller [Thu, 19 Jan 2023 01:16:51 +0000 (17:16 -0800)]
ice: introduce clear_reset_state operation
When hardware is reset, the VF relies on the VFGEN_RSTAT register to detect
when the VF is finished resetting. This is a tri-state register where 0
indicates a reset is in progress, 1 indicates the hardware is done
resetting, and 2 indicates that the software is done resetting.
Currently the PF driver relies on the device hardware resetting VFGEN_RSTAT
when a global reset occurs. This works ok, but it does mean that the VF
might not immediately notice a reset when the driver first detects that the
global reset is occurring.
This is also problematic for Scalable IOV, because there is no read/write
equivalent VFGEN_RSTAT register for the Scalable VSI type. Instead, the
Scalable IOV VFs will need to emulate this register.
To support this, introduce a new VF operation, clear_reset_state, which is
called when the PF driver first detects a global reset. The Single Root IOV
implementation can just write to VFGEN_RSTAT to ensure it's cleared
immediately, without waiting for the actual hardware reset to begin. The
Scalable IOV implementation will use this as part of its tracking of the
reset status to allow properly reporting the emulated VFGEN_RSTAT to the VF
driver.
Jacob Keller [Thu, 19 Jan 2023 01:16:50 +0000 (17:16 -0800)]
ice: convert vf_ops .vsi_rebuild to .create_vsi
The .vsi_rebuild function exists for ice_reset_vf. It is used to release
and re-create the VSI during a single-VF reset.
This function is only called when we need to re-create the VSI, and not
when rebuilding an existing VSI. This makes the single-VF reset process
different from the process used to restore functionality after a
hardware reset such as the PF reset or EMP reset.
When we add support for Scalable IOV VFs, the implementation will be very
similar. The primary difference will be in the fact that each VF type uses
a different underlying VSI type in hardware.
Move the common functionality into a new ice_vf_recreate VSI function. This
will allow the two IOV paths to share this functionality. Rework the
.vsi_rebuild vf_op into .create_vsi, only performing the task of creating a
new VSI.
This creates a nice dichotomy between the ice_vf_rebuild_vsi and
ice_vf_recreate_vsi, and should make it more clear why the two flows atre
distinct.
Jacob Keller [Thu, 19 Jan 2023 01:16:49 +0000 (17:16 -0800)]
ice: introduce ice_vf_init_host_cfg function
Introduce a new generic helper ice_vf_init_host_cfg which performs common
host configuration initialization tasks that will need to be done for both
Single Root IOV and the new Scalable IOV implementation.
Jacob Keller [Thu, 19 Jan 2023 01:16:48 +0000 (17:16 -0800)]
ice: add a function to initialize vf entry
Some of the initialization code for Single Root IOV VFs will need to be
reused when we introduce Scalable IOV. Pull this code out into a new
ice_initialize_vf_entry helper function.
Jacob Keller [Thu, 19 Jan 2023 01:16:47 +0000 (17:16 -0800)]
ice: Pull common tasks into ice_vf_post_vsi_rebuild
The Single Root IOV implementation of .post_vsi_rebuild performs some tasks
that will ultimately need to be shared with the Scalable IOV implementation
such as rebuilding the host configuration.
Refactor by introducing a new wrapper function, ice_vf_post_vsi_rebuild
which performs the tasks that will be shared between SR-IOV and Scalable
IOV. Move the ice_vf_rebuild_host_cfg and ice_vf_set_initialized calls into
this wrapper. Then call the implementation specific post_vsi_rebuild
handler afterwards.
This ensures that we will properly re-initialize filters and expected
settings for both SR-IOV and Scalable IOV.
Jacob Keller [Thu, 19 Jan 2023 01:16:46 +0000 (17:16 -0800)]
ice: move ice_vf_vsi_release into ice_vf_lib.c
The ice_vf_vsi_release function will be used in a future change to
refactor the .vsi_rebuild function. Move this over to ice_vf_lib.c so
that it can be used there.
Jacob Keller [Thu, 19 Jan 2023 01:16:44 +0000 (17:16 -0800)]
ice: move vsi_type assignment from ice_vsi_alloc to ice_vsi_cfg
The ice_vsi_alloc and ice_vsi_cfg functions are used together to allocate
and configure a new VSI, called as part of the ice_vsi_setup function.
In the future with the addition of the subfunction code the ice driver
will want to be able to allocate a VSI while delaying the configuration to
a later point of the port activation.
Currently this requires that the port code know what type of VSI should
be allocated. This is required because ice_vsi_alloc assigns the VSI type.
Refactor the ice_vsi_alloc and ice_vsi_cfg functions so that VSI type
assignment isn't done until the configuration stage. This will allow the
devlink port addition logic to reserve a VSI as early as possible before
the type of the port is known. In this way, the port add can fail in the
event that all hardware VSI resources are exhausted.
Since the ice_vsi_cfg function already takes the ice_vsi_cfg_params
structure, this is relatively straight forward.
Jacob Keller [Thu, 19 Jan 2023 01:16:43 +0000 (17:16 -0800)]
ice: refactor VSI setup to use parameter structure
The ice_vsi_setup function, ice_vsi_alloc, and ice_vsi_cfg functions have
grown a large number of parameters. These parameters are used to initialize
a new VSI, as well as re-configure an existing VSI
Any time we want to add a new parameter to this function chain, even if it
will usually be unset, we have to change many call sites due to changing
the function signature.
A future change is going to refactor ice_vsi_alloc and ice_vsi_cfg to move
the VSI configuration and initialization all into ice_vsi_cfg.
Before this, refactor the VSI setup flow to use a new ice_vsi_cfg_params
structure. This will contain the configuration (mainly pointers) used to
initialize a VSI.
Pass this from ice_vsi_setup into the related functions such as
ice_vsi_alloc, ice_vsi_cfg, and ice_vsi_cfg_def.
Introduce a helper, ice_vsi_to_params to convert an existing VSI to the
parameters used to initialize it. This will aid in the flows where we
rebuild an existing VSI.
Since we also pass the ICE_VSI_FLAG_INIT to more functions which do not
need (or cannot yet have) the VSI parameters, lets make this clear by
renaming the function parameter to vsi_flags and using a u32 instead of a
signed integer. The name vsi_flags also makes it clear that we may extend
the flags in the future.
This change will make it easier to refactor the setup flow in the future,
and will reduce the complexity required to add a new parameter for
configuration in the future.
Jacob Keller [Thu, 19 Jan 2023 01:16:42 +0000 (17:16 -0800)]
ice: drop unnecessary VF parameter from several VSI functions
The vsi->vf pointer gets assigned early on during ice_vsi_alloc. Several
functions currently take a VF pointer, but they can just use the existing
vsi->vf pointer as needed. Modify these functions to drop the unnecessary
VF parameter.
Note that ice_vsi_cfg is not changed as a following change will refactor so
that the VF pointer is assigned during ice_vsi_cfg rather than
ice_vsi_alloc.
Jacob Keller [Thu, 19 Jan 2023 01:16:41 +0000 (17:16 -0800)]
ice: fix function comment referring to ice_vsi_alloc
Since commit 1d2e32275de7 ("ice: split ice_vsi_setup into smaller
functions") ice_vsi_alloc has not been responsible for all of the behavior
implied by the comment for ice_vsi_setup_vector_base.
Fix the comment to refer to the new function ice_vsi_alloc_def().
Brett Creeley [Thu, 1 Dec 2022 11:58:59 +0000 (12:58 +0100)]
ice: Add more usage of existing function ice_get_vf_vsi(vf)
Extend the usage of function ice_get_vf_vsi(vf) in multiple places
instead of VF's VSI by using a long string of dereferences
(i.e. vf->pf->vsi[vf->lan_vsi_idx]).
Luka Guzenko [Fri, 20 Jan 2023 22:37:41 +0000 (23:37 +0100)]
HID: Ignore battery for ELAN touchscreen 29DF on HP
The touchscreen reports a battery status of 0% and jumps to 1% when a
stylus is used. The device ID was added and the battery ignore quirk was
enabled for it.
several people noticed that on modern CAN controllers with wide bit
timing registers the default SJW of 1 can result in unstable or no
synchronization to the CAN network. See Patch 14/17 for details.
During review of v1 Vincent pointed out that the original code and the
series doesn't always check user provided bit timing parameters,
sometimes silently limits them and the return error values are not
consistent.
This series first cleans up some code in bittiming.c, replacing
open-coded variants by macros or functions (Patches 1, 2).
Patch 3 adds the missing assignment of the effective TQ if the
interface is configured with low level timing parameters.
Patch 4 is another code cleanup.
Patches 5, 6 check the bit timing parameter during interface
registration.
Patch 7 adds a validation of the sample point.
The patches 8-13 convert the error messages from netdev_err() to
NL_SET_ERR_MSG_FMT, factor out the SJW handling from
can_fixup_bittiming(), add checking and error messages for the
individual limits and harmonize the error return values.
Patch 14 changes the default SJW value from 1 to min(Phase Seg1, Phase
Seg2 / 2).
Patch 15 switches can_calc_bittiming() to use the new SJW handling.
Patch 16 converts can_calc_bittiming() to NL_SET_ERR_MSG_FMT().
And patch 16 adds a NL_SET_ERR_MSG_FMT() error message to
can_validate_bitrate().
can: bittiming: can_calc_bittiming(): convert from netdev_err() to NL_SET_ERR_MSG_FMT()
Replace the netdev_err() by NL_SET_ERR_MSG_FMT() to better inform the
user about the problem. While there, use %u to print unsigned values
and improve error message a bit.
In case of an error, return -EINVAL instead of -EDOM, this corresponds
better to the actual meaning of the error value.
can: bittiming: can_calc_bittiming(): clean up SJW handling
In the current code, if the user configures a bitrate, a default SJW
value of 1 is used. If the user configures both a bitrate and a SJW
value, can_calc_bittiming() silently limits the SJW value to SJW max
and TSEG2.
We came to the conclusion that if the user provided an invalid SJW
value, it's best to bail out and inform the user [1].
Further the ISO 11898-1:2015 standard mandates that "SJW shall be less
than or equal to the minimum of these two items: Phase_Seg1 and
Phase_Seg2." [2] The current code is missing that check.
The previous patches introduced
1) can_sjw_set_default() - sets a default value for SJW if unset
2) can_sjw_check() - implements a SJW check against SJW max, Phase
Seg1 and Phase Seg2. In the error case this function reports the error
to user space via netlink.
Replace both the open-coded SJW default setting and the open-coded and
insufficient checks of SJW with the helper functions
can_sjw_set_default() and can_sjw_check().
can: bittiming: can_sjw_set_default(): use Phase Seg2 / 2 as default for SJW
"The (Re-)Synchronization Jump Width (SJW) defines how far a
resynchronization may move the Sample Point inside the limits defined
by the Phase Buffer Segments to compensate for edge phase errors." [1]
In other words, this means that the SJW parameter controls the
tolerance of the CAN controller to frequency errors compared to other
CAN controllers.
If the user space does not provide an SJW parameter, the kernel
chooses a default value of 1. This has proven to be a good default
value for classic CAN controllers, but no longer for modern CAN-FD
controllers.
In the past there were CAN controllers like the sja1000 with a rather
limited range of bit timing parameters. For the standard bit rates
this results in the following bit timing parameters:
The attentive reader will notice that the SJW is 1 in most cases,
while the Seg2 phase is 2. Both values are given in TQ units, which in
turn is a duration in nanoseconds.
The TQ is 25ns, the Phase Seg 2 is "10" (== 250ns), the SJW is "1" (==
25ns).
Since the kernel chooses a default SJW of 1 regardless of the TQ, this
leads to a much smaller SJW and thus much smaller tolerances to
frequency errors.
To maintain the same oscillator tolerances on controllers with wide
bit timing registers, select a default SJW value of Phase Seg2 / 2
unless Phase Seg 1 is less. This results in the following bit timing
parameters:
can: bittiming: can_sjw_check(): report error via netlink and harmonize error value
If the user space has supplied an invalid SJW value (greater than the
maximum SJW value), report -EINVAL instead of -ERANGE, this better
matches the actual meaning of the error value.
Additionally report an error message via netlink to the user space.
can: bittiming: can_changelink() pass extack down callstack
This is a preparation patch.
In order to pass warning/error messages during netlink calls back to
user space, pass the extack struct down the callstack of
can_changelink(), the actual error messages will be added in the
following ptaches.
can: netlink: can_changelink(): convert from netdev_err() to NL_SET_ERR_MSG_FMT()
Since commit 51c352bdbcd2 ("netlink: add support for formatted extack
messages") formatted extack messages are supported to inform the user
space or warnings/errors during netlink calls.
Replace the netdev_err() by NL_SET_ERR_MSG_FMT() to better inform the
user about the problem. While there, use %u to print unsigned values
and improve error message a bit.
can: netlink: can_validate(): validate sample point for CAN and CAN-FD
The sample point is a value in tenths of a percent. Meaningful values
are between 0 and 1000. Invalid values are rejected and an error
message is returned to user space via netlink.
can: dev: register_candev(): ensure that bittiming const are valid
Implement the function can_bittiming_const_valid() to check the
validity of the specified bit timing constant. Call this function from
register_candev() to check the bit timing constants during the
registration of the CAN interface.
can: bittiming: can_fixup_bittiming(): set effective tq
The can_fixup_bittiming() function is used to validate the
user-supplied low-level bit timing parameters and calculate the
bitrate prescaler (brp) from the requested time quanta (tq) and the
CAN clock of the controller.
can_fixup_bittiming() selects the best matching integer bit rate
prescaler, which may result in a different time quantum than the value
specified by the user.
Calculate the resulting time quantum and assign it so that the user
sees the effective time quantum.
can: bittiming: can_fixup_bittiming(): use CAN_SYNC_SEG instead of 1
Commit 1c47fa6b31c2 ("can: dev: add a helper function to calculate the
duration of one bit") made the constant CAN_SYNC_SEG available in a
header file.
The magic number 1 in can_fixup_bittiming() represents the width of
the sync segment, replace it by CAN_SYNC_SEG to make the code more
readable.
sock_init_data() assumes that the `struct socket` passed in input is
contained in a `struct socket_alloc` allocated with sock_alloc().
However, tap_open() and tun_chr_open() pass a `struct socket` embedded
in a `struct tap_queue` and `struct tun_file` respectively, both
allocated with sk_alloc().
This causes a type confusion when issuing a container_of() with
SOCK_INODE() in sock_init_data() which results in assigning a wrong
sk_uid to the `struct sock` in input.
Due to the type confusion, both sockets happen to have their uid set
to 0, i.e. root.
While it will be often correct, as tuntap devices require
CAP_NET_ADMIN, it may not always be the case.
Not sure how widespread is the impact of this, it seems the socket uid
may be used for network filtering and routing, thus tuntap sockets may
be incorrectly managed.
Additionally, it seems a socket with an incorrect uid may be returned
to the vhost driver when issuing a get_socket() on a tuntap device in
vhost_net_set_backend().
Fix the bugs by adding and using sock_init_data_uid(), which
explicitly takes a uid as argument.
Signed-off-by: Pietro Borrello <[email protected]>
---
Changes in v3:
- Fix the bug by defining and using sock_init_data_uid()
- Link to v2: https://lore.kernel.org/r/20230131-tuntap-sk-uid-v2-0-29ec15592813@diag.uniroma1.it
Changes in v2:
- Shorten and format comments
- Link to v1: https://lore.kernel.org/r/20230131-tuntap-sk-uid-v1-0-af4f9f40979d@diag.uniroma1.it
====================
Pietro Borrello [Sat, 4 Feb 2023 17:39:22 +0000 (17:39 +0000)]
tap: tap_open(): correctly initialize socket uid
sock_init_data() assumes that the `struct socket` passed in input is
contained in a `struct socket_alloc` allocated with sock_alloc().
However, tap_open() passes a `struct socket` embedded in a `struct
tap_queue` allocated with sk_alloc().
This causes a type confusion when issuing a container_of() with
SOCK_INODE() in sock_init_data() which results in assigning a wrong
sk_uid to the `struct sock` in input.
On default configuration, the type confused field overlaps with
padding bytes between `int vnet_hdr_sz` and `struct tap_dev __rcu
*tap` in `struct tap_queue`, which makes the uid of all tap sockets 0,
i.e., the root one.
Fix the assignment by using sock_init_data_uid().
Fixes: 86741ec25462 ("net: core: Add a UID field to struct sock.") Signed-off-by: Pietro Borrello <[email protected]> Reviewed-by: Eric Dumazet <[email protected]> Signed-off-by: David S. Miller <[email protected]>
sock_init_data() assumes that the `struct socket` passed in input is
contained in a `struct socket_alloc` allocated with sock_alloc().
However, tun_chr_open() passes a `struct socket` embedded in a `struct
tun_file` allocated with sk_alloc().
This causes a type confusion when issuing a container_of() with
SOCK_INODE() in sock_init_data() which results in assigning a wrong
sk_uid to the `struct sock` in input.
On default configuration, the type confused field overlaps with the
high 4 bytes of `struct tun_struct __rcu *tun` of `struct tun_file`,
NULL at the time of call, which makes the uid of all tun sockets 0,
i.e., the root one.
Fix the assignment by using sock_init_data_uid().
Fixes: 86741ec25462 ("net: core: Add a UID field to struct sock.") Signed-off-by: Pietro Borrello <[email protected]> Reviewed-by: Eric Dumazet <[email protected]> Signed-off-by: David S. Miller <[email protected]>
Pietro Borrello [Sat, 4 Feb 2023 17:39:20 +0000 (17:39 +0000)]
net: add sock_init_data_uid()
Add sock_init_data_uid() to explicitly initialize the socket uid.
To initialise the socket uid, sock_init_data() assumes a the struct
socket* sock is always embedded in a struct socket_alloc, used to
access the corresponding inode uid. This may not be true.
Examples are sockets created in tun_chr_open() and tap_open().
Fixes: 86741ec25462 ("net: core: Add a UID field to struct sock.") Signed-off-by: Pietro Borrello <[email protected]> Reviewed-by: Eric Dumazet <[email protected]> Signed-off-by: David S. Miller <[email protected]>
Please excuse the increased patch set size compared to v4's 15 patches,
but Claudiu stirred up the pot :) when he pointed out that the mqprio
TXQ validation procedure is still incorrect, so I had to fix that, and
then do some consolidation work so that taprio doesn't duplicate
mqprio's bugs. Compared to v4, 3 patches are new and 1 was dropped for now
("net/sched: taprio: mask off bits in gate mask that exceed number of TCs"),
since there's not really much to gain from it. Since the previous patch
set has largely been reviewed, I hope that a delta overview will help
and make up for the large size.
v3->v4:
- adjusted patch 07/15 to not remove "#include <net/pkt_sched.h>" from
ti cpsw
https://patchwork.kernel.org/project/netdevbpf/cover/20230127001516[email protected]/
v2->v3:
- move min_num_stack_tx_queues definition so it doesn't conflict with
the ethtool mm patches I haven't submitted yet for enetc (and also to
make use of a 4 byte hole)
- warn and mask off excess TCs in gate mask instead of failing
- finally CC qdisc maintainers
v2 at:
https://patchwork.kernel.org/project/netdevbpf/patch/20230126125308.1199404[email protected]/
v1->v2:
- patches 1->4 are new
- update some header inclusions in drivers
- fix typo (said "taprio" instead of "mqprio")
- better enetc mqprio error handling
- dynamically reconstruct mqprio configuration in taprio offload
- also let stmmac and tsnep use per-TXQ gate_mask
v1 (RFC) at:
https://patchwork.kernel.org/project/netdevbpf/cover/20230120141537.1350744[email protected]/
The main goal of this patch set is to make taprio pass the mqprio queue
configuration structure down to ndo_setup_tc() - patch 13/17. But mqprio
itself is not in the best shape currently, so there are some
consolidation patches on that as well.
Next, there are some consolidation patches in the enetc driver's
handling of TX queues and their traffic class assignment. Then, there is
a consolidation between the TX queue configuration for mqprio and
taprio.
Finally, there is a change in the meaning of the gate_mask passed by
taprio through ndo_setup_tc(). We introduce a capability through which
drivers can request the gate mask to be per TXQ. The default is changed
so that it is per TC.
====================
Vladimir Oltean [Sat, 4 Feb 2023 13:53:07 +0000 (15:53 +0200)]
net: enetc: act upon mqprio queue config in taprio offload
We assume that the mqprio queue configuration from taprio has a simple
1:1 mapping between prio and traffic class, and one TX queue per TC.
That might not be the case. Actually parse and act upon the mqprio
config.
Vladimir Oltean [Sat, 4 Feb 2023 13:53:06 +0000 (15:53 +0200)]
net: enetc: act upon the requested mqprio queue configuration
Regardless of the requested queue count per traffic class, the enetc
driver allocates a number of TX rings equal to the number of TCs, and
hardcodes a queue configuration of "1@0 1@1 ... 1@max-tc". Other
configurations are silently ignored and treated the same.
Improve that by allowing what the user requests to be actually
fulfilled. This allows more than one TX ring per traffic class.
For example:
By requesting validation via the mqprio capability structure, this is no
longer allowed, and we bring what is accepted by hardware in line with
what is accepted by software.
The check that num_tc <= real_num_tx_queues also becomes superfluous and
can be dropped, because mqprio_validate_queue_counts() validates that no
TXQ range exceeds real_num_tx_queues. That is a stronger check, because
there is at least 1 TXQ per TC, so there are at least as many TXQs as TCs.
Vladimir Oltean [Sat, 4 Feb 2023 13:53:04 +0000 (15:53 +0200)]
net/sched: taprio: only pass gate mask per TXQ for igc, stmmac, tsnep, am65_cpsw
There are 2 classes of in-tree drivers currently:
- those who act upon struct tc_taprio_sched_entry :: gate_mask as if it
holds a bit mask of TXQs
- those who act upon the gate_mask as if it holds a bit mask of TCs
When it comes to the standard, IEEE 802.1Q-2018 does say this in the
second paragraph of section 8.6.8.4 Enhancements for scheduled traffic:
| A gate control list associated with each Port contains an ordered list
| of gate operations. Each gate operation changes the transmission gate
| state for the gate associated with each of the Port's traffic class
| queues and allows associated control operations to be scheduled.
In typically obtuse language, it refers to a "traffic class queue"
rather than a "traffic class" or a "queue". But careful reading of
802.1Q clarifies that "traffic class" and "queue" are in fact
synonymous (see 8.6.6 Queuing frames):
| A queue in this context is not necessarily a single FIFO data structure.
| A queue is a record of all frames of a given traffic class awaiting
| transmission on a given Bridge Port. The structure of this record is not
| specified.
i.o.w. their definition of "queue" isn't the Linux TX queue.
The gate_mask really is input into taprio via its UAPI as a mask of
traffic classes, but taprio_sched_to_offload() converts it into a TXQ
mask.
The breakdown of drivers which handle TC_SETUP_QDISC_TAPRIO is:
- hellcreek, felix, sja1105: these are DSA switches, it's not even very
clear what TXQs correspond to, other than purely software constructs.
Only the mqprio configuration with 8 TCs and 1 TXQ per TC makes sense.
So it's fine to convert these to a gate mask per TC.
- enetc: I have the hardware and can confirm that the gate mask is per
TC, and affects all TXQs (BD rings) configured for that priority.
- igc: in igc_save_qbv_schedule(), the gate_mask is clearly interpreted
to be per-TXQ.
- tsnep: Gerhard Engleder clarifies that even though this hardware
supports at most 1 TXQ per TC, the TXQ indices may be different from
the TC values themselves, and it is the TXQ indices that matter to
this hardware. So keep it per-TXQ as well.
- stmmac: I have a GMAC datasheet, and in the EST section it does
specify that the gate events are per TXQ rather than per TC.
- lan966x: again, this is a switch, and while not a DSA one, the way in
which it implements lan966x_mqprio_add() - by only allowing num_tc ==
NUM_PRIO_QUEUES (8) - makes it clear to me that TXQs are a purely
software construct here as well. They seem to map 1:1 with TCs.
- am65_cpsw: from looking at am65_cpsw_est_set_sched_cmds(), I get the
impression that the fetch_allow variable is treated like a prio_mask.
This definitely sounds closer to a per-TC gate mask rather than a
per-TXQ one, and TI documentation does seem to recomment an identity
mapping between TCs and TXQs. However, Roger Quadros would like to do
some testing before making changes, so I'm leaving this driver to
operate as it did before, for now. Link with more details at the end.
Based on this breakdown, we have 5 drivers with a gate mask per TC and
4 with a gate mask per TXQ. So let's make the gate mask per TXQ the
opt-in and the gate mask per TC the default.
Benefit from the TC_QUERY_CAPS feature that Jakub suggested we add, and
query the device driver before calling the proper ndo_setup_tc(), and
figure out if it expects one or the other format.
Vladimir Oltean [Sat, 4 Feb 2023 13:53:03 +0000 (15:53 +0200)]
net/sched: taprio: pass mqprio queue configuration to ndo_setup_tc()
The taprio qdisc does not currently pass the mqprio queue configuration
down to the offloading device driver. So the driver cannot act upon the
TXQ counts/offsets per TC, or upon the prio->tc map. It was probably
assumed that the driver only wants to offload num_tc (see
TC_MQPRIO_HW_OFFLOAD_TCS), which it can get from netdev_get_num_tc(),
but there's clearly more to the mqprio configuration than that.
I've considered 2 mechanisms to remedy that. First is to pass a struct
tc_mqprio_qopt_offload as part of the tc_taprio_qopt_offload. The second
is to make taprio actually call TC_SETUP_QDISC_MQPRIO, *in addition to*
TC_SETUP_QDISC_TAPRIO.
The difference is that in the first case, existing drivers (offloading
or not) all ignore taprio's mqprio portion currently, whereas in the
second case, we could control whether to call TC_SETUP_QDISC_MQPRIO,
based on a new capability. The question is which approach would be
better.
I'm afraid that calling TC_SETUP_QDISC_MQPRIO unconditionally (not based
on a taprio capability bit) would risk introducing regressions. For
example, taprio doesn't populate (or validate) qopt->hw, as well as
mqprio.flags, mqprio.shaper, mqprio.min_rate, mqprio.max_rate.
In comparison, adding a capability is functionally equivalent to just
passing the mqprio in a way that drivers can ignore it, except it's
slightly more complicated to use it (need to set the capability).
Ultimately, what made me go for the "mqprio in taprio" variant was that
it's easier for offloading drivers to interpret the mqprio qopt slightly
differently when it comes from taprio vs when it comes from mqprio,
should that ever become necessary.
Vladimir Oltean [Sat, 4 Feb 2023 13:53:02 +0000 (15:53 +0200)]
net/sched: refactor mqprio qopt reconstruction to a library function
The taprio qdisc will need to reconstruct a struct tc_mqprio_qopt from
netdev settings once more in a future patch, but this code was already
written twice, once in taprio and once in mqprio.
Refactor the code to a helper in the common mqprio library.
There is a lot of code in taprio which is "borrowed" from mqprio.
It makes sense to put a stop to the "borrowing" and start actually
reusing code.
Because taprio and mqprio are built as part of different kernel modules,
code reuse can only take place either by writing it as static inline
(limiting), putting it in sch_generic.o (not generic enough), or
creating a third auto-selectable kernel module which only holds library
code. I opted for the third variant.
In a previous change, mqprio gained support for reverse TC:TXQ mappings,
something which taprio still denies. Make taprio use the same validation
logic so that it supports this configuration as well.
The taprio code didn't enforce TXQ overlaps in txtime-assist mode and
that looks intentional, even if I've no idea why that might be. Preserve
that, but add a comment.
There isn't any dedicated MAINTAINERS entry for mqprio, so nothing to
update there.
Vladimir Oltean [Sat, 4 Feb 2023 13:53:00 +0000 (15:53 +0200)]
net/sched: mqprio: add extack messages for queue count validation
To make mqprio more user-friendly, create netlink extended ack messages
which say exactly what is wrong about the queue counts. This uses the
new support for printf-formatted extack messages.
Vladimir Oltean [Sat, 4 Feb 2023 13:52:59 +0000 (15:52 +0200)]
net/sched: mqprio: allow offloading drivers to request queue count validation
mqprio_parse_opt() proudly has a comment:
/* If hardware offload is requested we will leave it to the device
* to either populate the queue counts itself or to validate the
* provided queue counts.
*/
Unfortunately some device drivers did not get this memo, and don't
validate the queue counts, or populate them.
In case drivers don't want to populate the queue counts themselves, just
act upon the requested configuration, it makes sense to introduce a tc
capability, and make mqprio query it, so they don't have to do the
validation themselves.
Vladimir Oltean [Sat, 4 Feb 2023 13:52:58 +0000 (15:52 +0200)]
net/sched: mqprio: allow reverse TC:TXQ mappings
By imposing that the last TXQ of TC i is smaller than the first TXQ of
any TC j (j := i+1 .. n), mqprio imposes a strict ordering condition for
the TXQ indices (they must increase as TCs increase).
Claudiu points out that the complexity of the TXQ count validation is
too high for this logic, i.e. instead of iterating over j, it is
sufficient that the TXQ indices of TC i and i + 1 are ordered, and that
will eventually ensure global ordering.
This is true, however it doesn't appear to me that is what the code
really intended to do. Instead, based on the comments, it just wanted to
check for overlaps (and this isn't how one does that).
So the following mqprio configuration, which I had recommended to
Vinicius more than once for igb/igc (to account for the fact that on
this hardware, lower numbered TXQs have higher dequeue priority than
higher ones):
num_tc 4 map 0 1 2 3 queues 1@3 1@2 1@1 1@0
is in fact denied today by mqprio.
The full story is that in fact, it's only denied with "hw 0"; if
hardware offloading is requested, mqprio defers TXQ range overlap
validation to the device driver (a strange decision in itself).
This is most certainly a bug, but it's not one that has any merit for
being fixed on "stable" as far as I can tell. This is because mqprio
always rejected a configuration which was in fact valid, and this has
shaped the way in which mqprio configuration scripts got built for
various hardware (see igb/igc in the link below). Therefore, one could
consider it to be merely an improvement for mqprio to allow reverse
TC:TXQ mappings.
projects / linux.git / log