Skip to content
  1. Jun 18, 2020
    • notro's avatar
      BCM2708: Add core Device Tree support · 292ab07a
      notro authored 
      

Add the bare minimum needed to boot BCM2708 from a Device Tree.

Signed-off-by: default avatarNoralf Tronnes <notro@tronnes.org>

BCM2708: DT: change 'axi' nodename to 'soc'

Change DT node named 'axi' to 'soc' so it matches ARCH_BCM2835.
The VC4 bootloader fills in certain properties in the 'axi' subtree,
but since this is part of an upstreaming effort, the name is changed.

Signed-off-by: default avatarNoralf Tronnes <notro@tronnes.org>

BCM2708_DT: Correct length of the peripheral space

Use dts-dirs feature for overlays.

The kernel makefiles have a dts-dirs target that is for vendor subdirectories.

Using this fixes the install_dtbs target, which previously did not install the overlays.

BCM270X_DT: configure I2S DMA channels

Signed-off-by: default avatarMatthias Reichl <hias@horus.com>

BCM270X_DT: switch to bcm2835-i2s

I2S soundcard drivers with proper devicetree support (i.e. not linking
to the cpu_dai/platform via name but to cpu/platform via of_node)
will work out of the box without any modifications.

When the kernel is compiled without devicetree support the platform
code will instantiate the bcm2708-i2s driver and I2S soundcard drivers
will link to it via name, as before.

Signed-off-by: default avatarMatthias Reichl <hias@horus.com>

SDIO-overlay: add poll_once-boolean parameter

Add paramter to toggle sdio-device-polling
done every second or once at boot-time.

Signed-off-by: default avatarPatrick Boettcher <patrick.boettcher@posteo.de>

BCM270X_DT: Make mmc overlay compatible with current firmware

The original DT overlay logic followed a merge-then-patch procedure,
i.e. parameters are applied to the loaded overlay before the overlay
is merged into the base DTB. This sequence has been changed to
patch-then-merge, in order to support parameterised node names, and
to protect against bad overlays. As a result, overrides (parameters)
must only target labels in the overlay, but the overlay can obviously target nodes in the base DTB.

mmc-overlay.dts (that switches back to the original mmc sdcard
driver) is the only overlay violating that rule, and this patch
fixes it.

bcm270x_dt: Use the sdhost MMC controller by default

The "mmc" overlay reverts to using the other controller.

squash: Add cprman to dt

BCM270X_DT: Use clk_core for I2C interfaces

BCM270X_DT: Use bcm283x.dtsi, bcm2835.dtsi and bcm2836.dtsi

The mainline Device Tree files are quite close to downstream now.
Let's use bcm283x.dtsi, bcm2835.dtsi and bcm2836.dtsi as base files
for our dts files.

Mainline dts files are based on these files:

          bcm2835-rpi.dtsi
  bcm2835.dtsi    bcm2836.dtsi
          bcm283x.dtsi

Current downstream are based on these:

  bcm2708.dtsi    bcm2709.dtsi    bcm2710.dtsi
             bcm2708_common.dtsi

This patch introduces this dependency:

  bcm2708.dtsi    bcm2709.dtsi
          bcm2708-rpi.dtsi
          bcm270x.dtsi
  bcm2835.dtsi    bcm2836.dtsi
          bcm283x.dtsi

And:
          bcm2710.dtsi
          bcm2708-rpi.dtsi
          bcm270x.dtsi
          bcm283x.dtsi

bcm270x.dtsi contains the downstream bcm283x.dtsi diff.
bcm2708-rpi.dtsi is the downstream version of bcm2835-rpi.dtsi.

Other changes:
- The led node has moved from /soc/leds to /leds. This is not a problem
  since the label is used to reference it.
- The clk_osc reg property changes from 6 to 3.
- The gpu nodes has their interrupt property set in the base file.
- the clocks label does not point to the /clocks node anymore, but
  points to the cprman node. This is not a problem since the overlays
  that use the clock node refer to it directly: target-path = "/clocks";
- some nodes now have 2 labels since mainline and downstream differs in
  this respect: cprman/clocks, spi0/spi, gpu/vc4.
- some nodes doesn't have an explicit status = "okay" since they're not
  disabled in the base file: watchdog and random.
- gpiomem doesn't need an explicit status = "okay".
- bcm2708-rpi-cm.dts got the hpd-gpios property from bcm2708_common.dtsi,
  it's now set directly in that file.
- bcm2709-rpi-2-b.dts has the timer node moved from /soc/timer to /timer.
- Removed clock-frequency property on the bcm{2709,2710}.dtsi timer nodes.

Signed-off-by: default avatarNoralf Trønnes <noralf@tronnes.org>

BCM270X_DT: Use raspberrypi-power to turn on USB power

Use the raspberrypi-power driver to turn on USB power.

Signed-off-by: default avatarNoralf Trønnes <noralf@tronnes.org>

BCM270X_DT: Add a .dtbo target, use for overlays

Change the filenames and extensions to keep the pre-DDT style of
overlay (<name>-overlay.dtb) distinct from new ones that use a
different style of local fixups (<name>.dtbo), and to match other
platforms.

The RPi firmware uses the DDTK trailer atom to choose which type of
overlay to use for each kernel.

Signed-off-by: default avatarPhil Elwell <phil@raspberrypi.org>

BCM270X_DT: Don't generate "linux,phandle" props

The EPAPR standard says to use "phandle" properties to store phandles,
rather than the deprecated "linux,phandle" version. By default, dtc
generates both, but adding "-H epapr" causes it to only generate
"phandle"s, saving some space and clutter.

Signed-off-by: default avatarPhil Elwell <phil@raspberrypi.org>

BCM270X_DT: Add overlay for enc28j60 on SPI2

Works on SPI2 for compute module

BCM270X_DT: Add midi-uart0 overlay

MIDI requires 31.25kbaud, a baudrate unsupported by Linux. The
midi-uart0 overlay configures uart0 (ttyAMA0) to use a fake clock
so that requesting 38.4kbaud actually gets 31.25kbaud.

Signed-off-by: default avatarPhil Elwell <phil@raspberrypi.org>

BCM270X_DT: Add i2c-sensor overlay

The i2c-sensor overlay is a container for various pressure and
temperature sensors, currently bmp085 and bmp280. The standalone
bmp085_i2c-sensor overlay is now deprecated.

Signed-off-by: default avatarPhil Elwell <phil@raspberrypi.org>

BCM270X_DT: overlays/*-overlay.dtb -> overlays/*.dtbo (#1752)

We now create overlays as .dtbo files.

build: support for .dtbo files for dtb overlays

Kernel 4.4.6+ on RaspberryPi support .dtbo files for overlays, instead of .dtb.
Patch the kernel, which has faulty rules to generate .dtbo the way yocto does

Signed-off-by: default avatarHerve Jourdain <herve.jourdain@neuf.fr>
Signed-off-by: default avatarKhem Raj <raj.khem@gmail.com>

BCM270X: Drop position requirement for CMA in VC4 overlay.

No longer necessary since 2aefcd57

,
and will probably let peeople that want to choose a larger CMA
allocation (particularly on pi0/1).

Signed-off-by: default avatarEric Anholt <eric@anholt.net>

BCM270X_DT: RPi Device Tree tidy

Use the upstream sdhost node, add thermal-zones, and factor out some
common elements.

Signed-off-by: default avatarPhil Elwell <phil@raspberrypi.org>

kbuild: Silence unhelpful DTC warnings

Signed-off-by: default avatarPhil Elwell <phil@raspberrypi.org>

BCM270X_DT: DT build rules no longer arch-specific

Signed-off-by: default avatarPhil Elwell <phil@raspberrypi.org>
      292ab07a
    • popcornmix's avatar
      Add dwc_otg driver · f5f6d397
      popcornmix authored 
      

Signed-off-by: default avatarpopcornmix <popcornmix@gmail.com>

usb: dwc: fix lockdep false positive

Signed-off-by: default avatarKari Suvanto <karis79@gmail.com>

usb: dwc: fix inconsistent lock state

Signed-off-by: default avatarKari Suvanto <karis79@gmail.com>

Add FIQ patch to dwc_otg driver. Enable with dwc_otg.fiq_fix_enable=1. Should give about 10% more ARM performance.
Thanks to Gordon and Costas

Avoid dynamic memory allocation for channel lock in USB driver. Thanks ddv2005.

Add NAK holdoff scheme. Enabled by default, disable with dwc_otg.nak_holdoff_enable=0. Thanks gsh

Make sure we wait for the reset to finish

dwc_otg: fix bug in dwc_otg_hcd.c resulting in silent kernel
	 memory corruption, escalating to OOPS under high USB load.

dwc_otg: Fix unsafe access of QTD during URB enqueue

In dwc_otg_hcd_urb_enqueue during qtd creation, it was possible that the
transaction could complete almost immediately after the qtd was assigned
to a host channel during URB enqueue, which meant the qtd pointer was no
longer valid having been completed and removed. Usually, this resulted in
an OOPS during URB submission. By predetermining whether transactions
need to be queued or not, this unsafe pointer access is avoided.

This bug was only evident on the Pi model A where a device was attached
that had no periodic endpoints (e.g. USB pendrive or some wlan devices).

dwc_otg: Fix incorrect URB allocation error handling

If the memory allocation for a dwc_otg_urb failed, the kernel would OOPS
because for some reason a member of the *unallocated* struct was set to
zero. Error handling changed to fail correctly.

dwc_otg: fix potential use-after-free case in interrupt handler

If a transaction had previously aborted, certain interrupts are
enabled to track error counts and reset where necessary. On IN
endpoints the host generates an ACK interrupt near-simultaneously
with completion of transfer. In the case where this transfer had
previously had an error, this results in a use-after-free on
the QTD memory space with a 1-byte length being overwritten to
0x00.

dwc_otg: add handling of SPLIT transaction data toggle errors

Previously a data toggle error on packets from a USB1.1 device behind
a TT would result in the Pi locking up as the driver never handled
the associated interrupt. Patch adds basic retry mechanism and
interrupt acknowledgement to cater for either a chance toggle error or
for devices that have a broken initial toggle state (FT8U232/FT232BM).

dwc_otg: implement tasklet for returning URBs to usbcore hcd layer

The dwc_otg driver interrupt handler for transfer completion will spend
a very long time with interrupts disabled when a URB is completed -
this is because usb_hcd_giveback_urb is called from within the handler
which for a USB device driver with complicated processing (e.g. webcam)
will take an exorbitant amount of time to complete. This results in
missed completion interrupts for other USB packets which lead to them
being dropped due to microframe overruns.

This patch splits returning the URB to the usb hcd layer into a
high-priority tasklet. This will have most benefit for isochronous IN
transfers but will also have incidental benefit where multiple periodic
devices are active at once.

dwc_otg: fix NAK holdoff and allow on split transactions only

This corrects a bug where if a single active non-periodic endpoint
had at least one transaction in its qh, on frnum == MAX_FRNUM the qh
would get skipped and never get queued again. This would result in
a silent device until error detection (automatic or otherwise) would
either reset the device or flush and requeue the URBs.

Additionally the NAK holdoff was enabled for all transactions - this
would potentially stall a HS endpoint for 1ms if a previous error state
enabled this interrupt and the next response was a NAK. Fix so that
only split transactions get held off.

dwc_otg: Call usb_hcd_unlink_urb_from_ep with lock held in completion handler

usb_hcd_unlink_urb_from_ep must be called with the HCD lock held.  Calling it
asynchronously in the tasklet was not safe (regression in
c4564d4a).

This change unlinks it from the endpoint prior to queueing it for handling in
the tasklet, and also adds a check to ensure the urb is OK to be unlinked
before doing so.

NULL pointer dereference kernel oopses had been observed in usb_hcd_giveback_urb
when a USB device was unplugged/replugged during data transfer.  This effect
was reproduced using automated USB port power control, hundreds of replug
events were performed during active transfers to confirm that the problem was
eliminated.

USB fix using a FIQ to implement split transactions

This commit adds a FIQ implementaion that schedules
the split transactions using a FIQ so we don't get
held off by the interrupt latency of Linux

dwc_otg: fix device attributes and avoid kernel warnings on boot

dcw_otg: avoid logging function that can cause panics

See: https://github.com/raspberrypi/firmware/issues/21
Thanks to cleverca22 for fix

dwc_otg: mask correct interrupts after transaction error recovery

The dwc_otg driver will unmask certain interrupts on a transaction
that previously halted in the error state in order to reset the
QTD error count. The various fine-grained interrupt handlers do not
consider that other interrupts besides themselves were unmasked.

By disabling the two other interrupts only ever enabled in DMA mode
for this purpose, we can avoid unnecessary function calls in the
IRQ handler. This will also prevent an unneccesary FIQ interrupt
from being generated if the FIQ is enabled.

dwc_otg: fiq: prevent FIQ thrash and incorrect state passing to IRQ

In the case of a transaction to a device that had previously aborted
due to an error, several interrupts are enabled to reset the error
count when a device responds. This has the side-effect of making the
FIQ thrash because the hardware will generate multiple instances of
a NAK on an IN bulk/interrupt endpoint and multiple instances of ACK
on an OUT bulk/interrupt endpoint. Make the FIQ mask and clear the
associated interrupts.

Additionally, on non-split transactions make sure that only unmasked
interrupts are cleared. This caused a hard-to-trigger but serious
race condition when you had the combination of an endpoint awaiting
error recovery and a transaction completed on an endpoint - due to
the sequencing and timing of interrupts generated by the dwc_otg core,
it was possible to confuse the IRQ handler.

Fix function tracing

dwc_otg: whitespace cleanup in dwc_otg_urb_enqueue

dwc_otg: prevent OOPSes during device disconnects

The dwc_otg_urb_enqueue function is thread-unsafe. In particular the
access of urb->hcpriv, usb_hcd_link_urb_to_ep, dwc_otg_urb->qtd and
friends does not occur within a critical section and so if a device
was unplugged during activity there was a high chance that the
usbcore hub_thread would try to disable the endpoint with partially-
formed entries in the URB queue. This would result in BUG() or null
pointer dereferences.

Fix so that access of urb->hcpriv, enqueuing to the hardware and
adding to usbcore endpoint URB lists is contained within a single
critical section.

dwc_otg: prevent BUG() in TT allocation if hub address is > 16

A fixed-size array is used to track TT allocation. This was
previously set to 16 which caused a crash because
dwc_otg_hcd_allocate_port would read past the end of the array.

This was hit if a hub was plugged in which enumerated as addr > 16,
due to previous device resets or unplugs.

Also add #ifdef FIQ_DEBUG around hcd->hub_port_alloc[], which grows
to a large size if 128 hub addresses are supported. This field is
for debug only for tracking which frame an allocate happened in.

dwc_otg: make channel halts with unknown state less damaging

If the IRQ received a channel halt interrupt through the FIQ
with no other bits set, the IRQ would not release the host
channel and never complete the URB.

Add catchall handling to treat as a transaction error and retry.

dwc_otg: fiq_split: use TTs with more granularity

This fixes certain issues with split transaction scheduling.

- Isochronous multi-packet OUT transactions now hog the TT until
  they are completed - this prevents hubs aborting transactions
  if they get a periodic start-split out-of-order
- Don't perform TT allocation on non-periodic endpoints - this
  allows simultaneous use of the TT's bulk/control and periodic
  transaction buffers

This commit will mainly affect USB audio playback.

dwc_otg: fix potential sleep while atomic during urb enqueue

Fixes a regression introduced with eb1b482a. Kmalloc called from
dwc_otg_hcd_qtd_add / dwc_otg_hcd_qtd_create did not always have
the GPF_ATOMIC flag set. Force this flag when inside the larger
critical section.

dwc_otg: make fiq_split_enable imply fiq_fix_enable

Failing to set up the FIQ correctly would result in
"IRQ 32: nobody cared" errors in dmesg.

dwc_otg: prevent crashes on host port disconnects

Fix several issues resulting in crashes or inconsistent state
if a Model A root port was disconnected.

- Clean up queue heads properly in kill_urbs_in_qh_list by
  removing the empty QHs from the schedule lists
- Set the halt status properly to prevent IRQ handlers from
  using freed memory
- Add fiq_split related cleanup for saved registers
- Make microframe scheduling reclaim host channels if
  active during a disconnect
- Abort URBs with -ESHUTDOWN status response, informing
  device drivers so they respond in a more correct fashion
  and don't try to resubmit URBs
- Prevent IRQ handlers from attempting to handle channel
  interrupts if the associated URB was dequeued (and the
  driver state was cleared)

dwc_otg: prevent leaking URBs during enqueue

A dwc_otg_urb would get leaked if the HCD enqueue function
failed for any reason. Free the URB at the appropriate points.

dwc_otg: Enable NAK holdoff for control split transactions

Certain low-speed devices take a very long time to complete a
data or status stage of a control transaction, producing NAK
responses until they complete internal processing - the USB2.0
spec limit is up to 500mS. This causes the same type of interrupt
storm as seen with USB-serial dongles prior to c8edb238.

In certain circumstances, usually while booting, this interrupt
storm could cause SD card timeouts.

dwc_otg: Fix for occasional lockup on boot when doing a USB reset

dwc_otg: Don't issue traffic to LS devices in FS mode

Issuing low-speed packets when the root port is in full-speed mode
causes the root port to stop responding. Explicitly fail when
enqueuing URBs to a LS endpoint on a FS bus.

Fix ARM architecture issue with local_irq_restore()

If local_fiq_enable() is called before a local_irq_restore(flags) where
the flags variable has the F bit set, the FIQ will be erroneously disabled.

Fixup arch_local_irq_restore to avoid trampling the F bit in CPSR.

Also fix some of the hacks previously implemented for previous dwc_otg
incarnations.

dwc_otg: fiq_fsm: Base commit for driver rewrite

This commit removes the previous FIQ fixes entirely and adds fiq_fsm.

This rewrite features much more complete support for split transactions
and takes into account several OTG hardware bugs. High-speed
isochronous transactions are also capable of being performed by fiq_fsm.

All driver options have been removed and replaced with:
  - dwc_otg.fiq_enable (bool)
  - dwc_otg.fiq_fsm_enable (bool)
  - dwc_otg.fiq_fsm_mask (bitmask)
  - dwc_otg.nak_holdoff (unsigned int)

Defaults are specified such that fiq_fsm behaves similarly to the
previously implemented FIQ fixes.

fiq_fsm: Push error recovery into the FIQ when fiq_fsm is used

If the transfer associated with a QTD failed due to a bus error, the HCD
would retry the transfer up to 3 times (implementing the USB2.0
three-strikes retry in software).

Due to the masking mechanism used by fiq_fsm, it is only possible to pass
a single interrupt through to the HCD per-transfer.

In this instance host channels would fall off the radar because the error
reset would function, but the subsequent channel halt would be lost.

Push the error count reset into the FIQ handler.

fiq_fsm: Implement timeout mechanism

For full-speed endpoints with a large packet size, interrupt latency
runs the risk of the FIQ starting a transaction too late in a full-speed
frame. If the device is still transmitting data when EOF2 for the
downstream frame occurs, the hub will disable the port. This change is
not reflected in the hub status endpoint and the device becomes
unresponsive.

Prevent high-bandwidth transactions from being started too late in a
frame. The mechanism is not guaranteed: a combination of bit stuffing
and hub latency may still result in a device overrunning.

fiq_fsm: fix bounce buffer utilisation for Isochronous OUT

Multi-packet isochronous OUT transactions were subject to a few bounday
bugs. Fix them.

Audio playback is now much more robust: however, an issue stands with
devices that have adaptive sinks - ALSA plays samples too fast.

dwc_otg: Return full-speed frame numbers in HS mode

The frame counter increments on every *microframe* in high-speed mode.
Most device drivers expect this number to be in full-speed frames - this
caused considerable confusion to e.g. snd_usb_audio which uses the
frame counter to estimate the number of samples played.

fiq_fsm: save PID on completion of interrupt OUT transfers

Also add edge case handling for interrupt transports.

Note that for periodic split IN, data toggles are unimplemented in the
OTG host hardware - it unconditionally accepts any PID.

fiq_fsm: add missing case for fiq_fsm_tt_in_use()

Certain combinations of bitrate and endpoint activity could
result in a periodic transaction erroneously getting started
while the previous Isochronous OUT was still active.

fiq_fsm: clear hcintmsk for aborted transactions

Prevents the FIQ from erroneously handling interrupts
on a timed out channel.

fiq_fsm: enable by default

fiq_fsm: fix dequeues for non-periodic split transactions

If a dequeue happened between the SSPLIT and CSPLIT phases of the
transaction, the HCD would never receive an interrupt.

fiq_fsm: Disable by default

fiq_fsm: Handle HC babble errors

The HCTSIZ transfer size field raises a babble interrupt if
the counter wraps. Handle the resulting interrupt in this case.

dwc_otg: fix interrupt registration for fiq_enable=0

Additionally make the module parameter conditional for wherever
hcd->fiq_state is touched.

fiq_fsm: Enable by default

dwc_otg: Fix various issues with root port and transaction errors

Process the host port interrupts correctly (and don't trample them).
Root port hotplug now functional again.

Fix a few thinkos with the transaction error passthrough for fiq_fsm.

fiq_fsm: Implement hack for Split Interrupt transactions

Hubs aren't too picky about which endpoint we send Control type split
transactions to. By treating Interrupt transfers as Control, it is
possible to use the non-periodic queue in the OTG core as well as the
non-periodic FIFOs in the hub itself. This massively reduces the
microframe exclusivity/contention that periodic split transactions
otherwise have to enforce.

It goes without saying that this is a fairly egregious USB specification
violation, but it works.

Original idea by Hans Petter Selasky @ FreeBSD.org.

dwc_otg: FIQ support on SMP. Set up FIQ stack and handler on Core 0 only.

dwc_otg: introduce fiq_fsm_spin(un|)lock()

SMP safety for the FIQ relies on register read-modify write cycles being
completed in the correct order. Several places in the DWC code modify
registers also touched by the FIQ. Protect these by a bare-bones lock
mechanism.

This also makes it possible to run the FIQ and IRQ handlers on different
cores.

fiq_fsm: fix build on bcm2708 and bcm2709 platforms

dwc_otg: put some barriers back where they should be for UP

bcm2709/dwc_otg: Setup FIQ on core 1 if >1 core active

dwc_otg: fixup read-modify-write in critical paths

Be more careful about read-modify-write on registers that the FIQ
also touches.

Guard fiq_fsm_spin_lock with fiq_enable check

fiq_fsm: Falling out of the state machine isn't fatal

This edge case can be hit if the port is disabled while the FIQ is
in the middle of a transaction. Make the effects less severe.

Also get rid of the useless return value.

squash: dwc_otg: Allow to build without SMP

usb: core: make overcurrent messages more prominent

Hub overcurrent messages are more serious than "debug". Increase loglevel.

usb: dwc_otg: Don't use dma_to_virt()

Commit 6ce0d200

 changes dma_to_virt() which breaks this driver.
Open code the old dma_to_virt() implementation to work around this.

Limit the use of __bus_to_virt() to cases where transfer_buffer_length
is set and transfer_buffer is not set. This is done to increase the
chance that this driver will also work on ARCH_BCM2835.

transfer_buffer should not be NULL if the length is set, but the
comment in the code indicates that there are situations where this
might happen. drivers/usb/isp1760/isp1760-hcd.c also has a similar
comment pointing to a possible: 'usb storage / SCSI bug'.

Signed-off-by: default avatarNoralf Trønnes <noralf@tronnes.org>

dwc_otg: Fix crash when fiq_enable=0

dwc_otg: fiq_fsm: Make high-speed isochronous strided transfers work properly

Certain low-bandwidth high-speed USB devices (specialist audio devices,
compressed-frame webcams) have packet intervals > 1 microframe.

Stride these transfers in the FIQ by using the start-of-frame interrupt
to restart the channel at the right time.

dwc_otg: Force host mode to fix incorrect compute module boards

dwc_otg: Add ARCH_BCM2835 support

Signed-off-by: default avatarNoralf Trønnes <noralf@tronnes.org>

dwc_otg: Simplify FIQ irq number code

Dropping ATAGS means we can simplify the FIQ irq number code.
Also add error checking on the returned irq number.

Signed-off-by: default avatarNoralf Trønnes <noralf@tronnes.org>

dwc_otg: Remove duplicate gadget probe/unregister function

dwc_otg: Properly set the HFIR

Douglas Anderson reported:

According to the most up to date version of the dwc2 databook, the FRINT
field of the HFIR register should be programmed to:
* 125 us * (PHY clock freq for HS) - 1
* 1000 us * (PHY clock freq for FS/LS) - 1

This is opposed to older versions of the doc that claimed it should be:
* 125 us * (PHY clock freq for HS)
* 1000 us * (PHY clock freq for FS/LS)

and reported lower timing jitter on a USB analyser

dcw_otg: trim xfer length when buffer larger than allocated size is received

dwc_otg: Don't free qh align buffers in atomic context

dwc_otg: Enable the hack for Split Interrupt transactions by default

dwc_otg.fiq_fsm_mask=0xF has long been a suggestion for users with audio stutters or other USB bandwidth issues.
So far we are aware of many success stories but no failure caused by this setting.
Make it a default to learn more.

See: https://www.raspberrypi.org/forums/viewtopic.php?f=28&t=70437



Signed-off-by: default avatarpopcornmix <popcornmix@gmail.com>

dwc_otg: Use kzalloc when suitable

dwc_otg: Pass struct device to dma_alloc*()

This makes it possible to get the bus address from Device Tree.

Signed-off-by: default avatarNoralf Trønnes <noralf@tronnes.org>

dwc_otg: fix summarize urb->actual_length for isochronous transfers

Kernel does not copy input data of ISO transfers to userspace
if actual_length is set only in ISO transfers and not summarized
in urb->actual_length. Fixes raspberrypi/linux#903

fiq_fsm: Use correct states when starting isoc OUT transfers

In fiq_fsm_start_next_periodic() if an isochronous OUT transfer
was selected, no regard was given as to whether this was a single-packet
transfer or a multi-packet staged transfer.

For single-packet transfers, this had the effect of repeatedly sending
OUT packets with bogus data and lengths.

Eventually if the channel was repeatedly enabled enough times, this
would lock up the OTG core and no further bus transfers would happen.

Set the FSM state up properly if we select a single-packet transfer.

Fixes https://github.com/raspberrypi/linux/issues/1842

dwc_otg: make nak_holdoff work as intended with empty queues

If URBs reading from non-periodic split endpoints were dequeued and
the last transfer from the endpoint was a NAK handshake, the resulting
qh->nak_frame value was stale which would result in unnecessarily long
polling intervals for the first subsequent transfer with a fresh URB.

Fixup qh->nak_frame in dwc_otg_hcd_urb_dequeue and also guard against
a case where a single URB is submitted to the endpoint, a NAK was
received on the transfer immediately prior to receiving data and the
device subsequently resubmits another URB past the qh->nak_frame interval.

Fixes https://github.com/raspberrypi/linux/issues/1709

dwc_otg: fix split transaction data toggle handling around dequeues

See https://github.com/raspberrypi/linux/issues/1709



Fix several issues regarding endpoint state when URBs are dequeued
- If the HCD is disconnected, flush FIQ-enabled channels properly
- Save the data toggle state for bulk endpoints if the last transfer
  from an endpoint where URBs were dequeued returned a data packet
- Reset hc->start_pkt_count properly in assign_and_init_hc()

dwc_otg: fix several potential crash sources

On root port disconnect events, the host driver state is cleared and
in-progress host channels are forcibly stopped. This doesn't play
well with the FIQ running in the background, so:
- Guard the disconnect callback with both the host spinlock and FIQ
  spinlock
- Move qtd dereference in dwc_otg_handle_hc_fsm() after the early-out
  so we don't dereference a qtd that has gone away
- Turn catch-all BUG()s in dwc_otg_handle_hc_fsm() into warnings.

dwc_otg: delete hcd->channel_lock

The lock serves no purpose as it is only held while the HCD spinlock
is already being held.

dwc_otg: remove unnecessary dma-mode channel halts on disconnect interrupt

Host channels are already halted in kill_urbs_in_qh_list() with the
subsequent interrupt processing behaving as if the URB was dequeued
via HCD callback.

There's no need to clobber the host channel registers a second time
as this exposes races between the driver and host channel resulting
in hcd->free_hc_list becoming corrupted.

dwcotg: Allow to build without FIQ on ARM64

Signed-off-by: default avatarpopcornmix <popcornmix@gmail.com>

dwc_otg: make periodic scheduling behave properly for FS buses

If the root port is in full-speed mode, transfer times at 12mbit/s
would be calculated but matched against high-speed quotas.

Reinitialise hcd->frame_usecs[i] on each port enable event so that
full-speed bandwidth can be tracked sensibly.

Also, don't bother using the FIQ for transfers when in full-speed
mode - at the slower bus speed, interrupt frequency is reduced by
an order of magnitude.

Related issue: https://github.com/raspberrypi/linux/issues/2020

dwc_otg: fiq_fsm: Make isochronous compatibility checks work properly

Get rid of the spammy printk and local pointer mangling.
Also, there is a nominal benefit for using fiq_fsm for isochronous
transfers in FS mode (~1.1k IRQs per second vs 2.1k IRQs per second)
so remove the root port speed check.

dwc_otg: add module parameter int_ep_interval_min

Add a module parameter (defaulting to ignored) that clamps the polling rate
of high-speed Interrupt endpoints to a minimum microframe interval.

The parameter is modifiable at runtime as it is used when activating new
endpoints (such as on device connect).

dwc_otg: fiq_fsm: Add non-periodic TT exclusivity constraints

Certain hub types do not discriminate between pipe direction (IN or OUT)
when considering non-periodic transfers. Therefore these hubs get confused
if multiple transfers are issued in different directions with the same
device address and endpoint number.

Constrain queuing non-periodic split transactions so they are performed
serially in such cases.

Related: https://github.com/raspberrypi/linux/issues/2024



dwc_otg: Fixup change to DRIVER_ATTR interface

dwc_otg: Fix compilation warnings

Signed-off-by: default avatarPhil Elwell <phil@raspberrypi.org>

USB_DWCOTG: Disable building dwc_otg as a module (#2265)

When dwc_otg is built as a module, build will fail with the following
error:

ERROR: "DWC_TASK_HI_SCHEDULE" [drivers/usb/host/dwc_otg/dwc_otg.ko] undefined!
scripts/Makefile.modpost:91: recipe for target '__modpost' failed
make[1]: *** [__modpost] Error 1
Makefile:1199: recipe for target 'modules' failed
make: *** [modules] Error 2

Even if the error is solved by including the missing
DWC_TASK_HI_SCHEDULE function, the kernel will panic when loading
dwc_otg.

As a workaround, simply prevent user from building dwc_otg as a module
as the current kernel does not support it.

See: https://github.com/raspberrypi/linux/issues/2258



Signed-off-by: default avatarMalik Olivier Boussejra <malik@boussejra.com>

dwc_otg: New timer API

dwc_otg: Fix removed ACCESS_ONCE->READ_ONCE

dwc_otg: don't unconditionally force host mode in dwc_otg_cil_init()

Add the ability to disable force_host_mode for those that want to use
dwc_otg in both device and host modes.

dwc_otg: Fix a regression when dequeueing isochronous transfers

In 282bed95 (dwc_otg: make nak_holdoff work as intended with empty queues)
the dequeue mechanism was changed to leave FIQ-enabled transfers to run
to completion - to avoid leaving hub TT buffers with stale packets lying
around.

This broke FIQ-accelerated isochronous transfers, as this then meant that
dozens of transfers were performed after the dequeue function returned.

Restore the state machine fence for isochronous transfers.

fiq_fsm: rewind DMA pointer for OUT transactions that fail (#2288)

See: https://github.com/raspberrypi/linux/issues/2140

dwc_otg: add smp_mb() to prevent driver state corruption on boot

Occasional crashes have been seen where the FIQ code dereferences
invalid/random pointers immediately after being set up, leading to
panic on boot.

The crash occurs as the FIQ code races against hcd_init_fiq() and
the hcd_init_fiq() code races against the outstanding memory stores
from dwc_otg_hcd_init(). Use explicit barriers after touching
driver state.

usb: dwc_otg: fix memory corruption in dwc_otg driver

[Upstream commit 51b1b649

]

The move from the staging tree to the main tree exposed a
longstanding memory corruption bug in the dwc2 driver. The
reordering of the driver initialization caused the dwc2 driver
to corrupt the initialization data of the sdhci driver on the
Raspberry Pi platform, which made the bug show up.

The error is in calling to_usb_device(hsotg->dev), since ->dev
is not a member of struct usb_device. The easiest fix is to
just remove the offending code, since it is not really needed.

Thanks to Stephen Warren for tracking down the cause of this.

Reported-by: default avatarAndre Heider <a.heider@gmail.com>
Tested-by: default avatarStephen Warren <swarren@wwwdotorg.org>
Signed-off-by: default avatarPaul Zimmerman <paulz@synopsys.com>
Signed-off-by: default avatarGreg Kroah-Hartman <gregkh@linuxfoundation.org>
[lukas: port from upstream dwc2 to out-of-tree dwc_otg driver]
Signed-off-by: default avatarLukas Wunner <lukas@wunner.de>

usb: dwb_otg: Fix unreachable switch statement warning

This warning appears with GCC 7.3.0 from toolchains.bootlin.com:

../drivers/usb/host/dwc_otg/dwc_otg_fiq_fsm.c: In function ‘fiq_fsm_update_hs_isoc’:
../drivers/usb/host/dwc_otg/dwc_otg_fiq_fsm.c:595:61: warning: statement will never be executed [-Wswitch-unreachable]
   st->hctsiz_copy.b.xfersize = nrpackets * st->hcchar_copy.b.mps;
                                            ~~~~~~~~~~~~~~~~~^~~~

Signed-off-by: default avatarNathan Chancellor <natechancellor@gmail.com>

dwc_otg: fiq_fsm: fix incorrect DMA register offset calculation

Rationalise the offset and update all call sites.

Fixes https://github.com/raspberrypi/linux/issues/2408

dwc_otg: fix bug with port_addr assignment for single-TT hubs

See https://github.com/raspberrypi/linux/issues/2734



The "Hub Port" field in the split transaction packet was always set
to 1 for single-TT hubs. The majority of single-TT hub products
apparently ignore this field and broadcast to all downstream enabled
ports, which masked the issue. A subset of hub devices apparently
need the port number to be exact or split transactions will fail.

usb: dwc_otg: Clean up build warnings on 64bit kernels

No functional changes. Almost all are changes to logging lines.

Signed-off-by: default avatarDave Stevenson <dave.stevenson@raspberrypi.org>

usb: dwc_otg: Use dma allocation for mphi dummy_send buffer

The FIQ driver used a kzalloc'ed buffer for dummy_send,
passing a kernel virtual address to the hardware block.
The buffer is only ever used for a dummy read, so it
should be harmless, but there is the chance that it will
cause exceptions.

Use a dma allocation so that we have a genuine bus address,
and read from that.
Free the allocation when done for good measure.

Signed-off-by: default avatarDave Stevenson <dave.stevenson@raspberrypi.org>

dwc_otg: only do_split when we actually need to do a split

The previous test would fail if the root port was in fullspeed mode
and there was a hub between the FS device and the root port. While
the transfer worked, the schedule mangling performed for high-speed
split transfers would break leading to an 8ms polling interval.

dwc_otg: fix locking around dequeueing and killing URBs

kill_urbs_in_qh_list() is practically only ever called with the fiq lock
already held, so don't spinlock twice in the case where we need to cancel
an isochronous transfer.

Also fix up a case where the global interrupt register could be read with
the fiq lock not held.

Fixes the deadlock seen in https://github.com/raspberrypi/linux/issues/2907



ARM64/DWC_OTG: Port dwc_otg driver to ARM64

In ARM64, the FIQ mechanism used by this driver is not current
implemented.   As a workaround, reqular IRQ is used instead
of FIQ.

In a separate change, the IRQ-CPU mapping is round robined
on ARM64 to increase concurrency and allow multiple interrupts
to be serviced at a time.  This reduces the need for FIQ.

Tests Run:

This mechanism is most likely to break when multiple USB devices
are attached at the same time.  So the system was tested under
stress.

Devices:

1. USB Speakers playing back a FLAC audio through VLC
   at 96KHz.(Higher then typically, but supported on my speakers).

2. sftp transferring large files through the buildin ethernet
   connection which is connected through USB.

3. Keyboard and mouse attached and being used.

Although I do occasionally hear some glitches, the music seems to
play quite well.

Signed-off-by: default avatarMichael Zoran <mzoran@crowfest.net>

usb: dwc_otg: Clean up interrupt claiming code

The FIQ/IRQ interrupt number identification code is scattered through
the dwc_otg driver. Rationalise it, simplifying the code and solving
an existing issue.

See: https://github.com/raspberrypi/linux/issues/2612



Signed-off-by: default avatarPhil Elwell <phil@raspberrypi.org>

dwc_otg: Choose appropriate IRQ handover strategy

2711 has no MPHI peripheral, but the ARM Control block can fake
interrupts. Use the size of the DTB "mphi" reg block to determine
which is required.

Signed-off-by: default avatarPhil Elwell <phil@raspberrypi.org>

usb: host: dwc_otg: fix compiling in separate directory

The dwc_otg Makefile does not respect the O=path argument correctly:
include paths in CFLAGS are given relatively to object path, not source
path. Compiling in a separate directory yields #include errors.

Signed-off-by: default avatarMarek Behún <marek.behun@nic.cz>

dwc_otg: use align_buf for small IN control transfers (#3150)

The hardware will do a 4-byte write to memory on any IN packet received
that is between 1 and 3 bytes long. This tramples memory in the uvcvideo
driver, as it uses a sequence of 1- and 2-byte control transfers to
query the min/max/range/step of each individual camera control and
gives us buffers that are offsets into a struct.

Catch small control transfers in the data phase and use the align_buf
to bounce the correct number of bytes into the URB's buffer.

In general, short packets on non-control endpoints should be OK as URBs
should have enough buffer space for a wMaxPacket size transfer.

See: https://github.com/raspberrypi/linux/issues/3148



Signed-off-by: default avatarJonathan Bell <jonathan@raspberrypi.org>

dwc_otg: Declare DMA capability with HCD_DMA flag

Following [1], USB controllers have to declare DMA capabilities in
order for them to be used by adding the HCD_DMA flag to their hc_driver
struct.

[1] 7b81cb6b

 ("usb: add a HCD_DMA flag instead of guestimating DMA capabilities")

Signed-off-by: default avatarPhil Elwell <phil@raspberrypi.org>

dwc_otg: checking the urb->transfer_buffer too early (#3332)

After enable the HIGHMEM and VMSPLIT_3G, the dwc_otg driver doesn't
work well on Pi2/3 boards with 1G physical ram. Users experience
the failure when copying a file of 600M size to the USB stick. And
at the same time, the dmesg shows:
usb 1-1.1.2: reset high-speed USB device number 8 using dwc_otg
sd 0:0:0:0: [sda] tag#0 FAILED Result: hostbyte=DID_ERROR driverbyte=DRIVER_OK
blk_update_request: I/O error, dev sda, sector 3024048 op 0x1:(WRITE) flags 0x4000 phys_seg 15 prio class 0

When this happens, the sg_buf sent to the driver is located in the
highmem region, the usb_sg_init() in the core/message.c will leave
transfer_buffer to NULL if the sg_buf is in highmem, but in the
dwc_otg driver, it returns -EINVAL unconditionally if transfer_buffer
is NULL.

The driver can handle the situation of buffer to be NULL, if it is in
DMA mode, it will convert an address from transfer_dma.

But if the conversion fails or it is in the PIO mode, we should check
buffer and return -EINVAL if it is NULL.

BugLink: https://bugs.launchpad.net/bugs/1852510


Signed-off-by: default avatarHui Wang <hui.wang@canonical.com>

dwc_otg: constrain endpoint max packet and transfer size on split IN

The hcd would unconditionally set the transfer length to the endpoint
packet size for non-isoc IN transfers. If the remaining buffer length
was less than the length of returned data, random memory would get
scribbled over, with bad effects if it crossed a page boundary.

Force a babble error if this happens by limiting the max transfer size
to the available buffer space. DMA will stop writing to memory on a
babble condition.

The hardware expects xfersize to be an integer multiple of maxpacket
size, so override hcchar.b.mps as well.

Signed-off-by: default avatarJonathan Bell <jonathan@raspberrypi.org>

dwc_otg: fiq_fsm: pause when cancelling split transactions

Non-periodic splits will DMA to/from the driver-provided transfer_buffer,
which may be freed immediately after the dequeue call returns. Block until
we know the transfer is complete.

A similar delay is needed when cleaning up disconnects, as the FIQ could
have started a periodic transfer in the previous microframe to the one
that triggered a disconnect.

Signed-off-by: default avatarJonathan Bell <jonathan@raspberrypi.org>

dwc_otg: fiq_fsm: add a barrier on entry into FIQ handler(s)

On BCM2835, there is no hardware guarantee that multiple outstanding
reads to different peripherals will complete in-order. The FIQ code
uses peripheral reads without barriers for performance, so in the case
where a read to a slow peripheral was issued immediately prior to FIQ
entry, the first peripheral read that the FIQ did could end up with
wrong read data returned.

Add dsb(sy) on entry so that all outstanding reads are retired.

The FIQ only issues reads to the dwc_otg core, so per-read barriers
in the handler itself are not required.

On BCM2836 and BCM2837 the barrier is not strictly required due to
differences in how the peripheral bus is implemented, but having
arch-specific handlers that introduce different latencies is risky.

Signed-off-by: default avatarJonathan Bell <jonathan@raspberrypi.org>
      f5f6d397
    • popcornmix's avatar
      Main bcm2708/bcm2709 linux port · a7457f0b
      popcornmix authored 
      

Signed-off-by: default avatarpopcornmix <popcornmix@gmail.com>
Signed-off-by: default avatarNoralf Trønnes <noralf@tronnes.org>

bcm2709: Drop platform smp and timer init code

irq-bcm2836 handles this through these functions:
bcm2835_init_local_timer_frequency()
bcm2836_arm_irqchip_smp_init()

Signed-off-by: default avatarNoralf Trønnes <noralf@tronnes.org>

bcm270x: Use watchdog for reboot/poweroff

The watchdog driver already has support for reboot/poweroff.
Make use of this and remove the code from the platform files.

Signed-off-by: default avatarNoralf Trønnes <noralf@tronnes.org>

board_bcm2835: Remove coherent dma pool increase - API has gone
      a7457f0b
    • Claggy3's avatar
      Update vfpmodule.c · e4652cc8
      Claggy3 authored 
      Christopher Alexander Tobias Schulze - May 2, 2015, 11:57 a.m.
This patch fixes a problem with VFP state save and restore related
to exception handling (panic with message "BUG: unsupported FP
instruction in kernel mode") present on VFP11 floating point units
(as used with ARM1176JZF-S CPUs, e.g. on first generation Raspberry
Pi boards). This patch was developed and discussed on

   https://github.com/raspberrypi/linux/issues/859



A precondition to see the crashes is that floating point exception
traps are enabled. In this case, the VFP11 might determine that a FPU
operation needs to trap at a point in time when it is not possible to
signal this to the ARM11 core any more. The VFP11 will then set the
FPEXC.EX bit and store the trapped opcode in FPINST. (In some cases,
a second opcode might have been accepted by the VFP11 before the
exception was detected and could be reported to the ARM11 - in this
case, the VFP11 also sets FPEXC.FP2V and stores the second opcode in
FPINST2.)

If FPEXC.EX is set, the VFP11 will "bounce" the next FPU opcode issued
by the ARM11 CPU, which will be seen by the ARM11 as an undefined opcode
trap. The VFP support code examines the FPEXC.EX and FPEXC.FP2V bits
to decide what actions to take, i.e., whether to emulate the opcodes
found in FPINST and FPINST2, and whether to retry the bounced instruction.

If a user space application has left the VFP11 in this "pending trap"
state, the next FPU opcode issued to the VFP11 might actually be the
VSTMIA operation vfp_save_state() uses to store the FPU registers
to memory (in our test cases, when building the signal stack frame).
In this case, the kernel crashes as described above.

This patch fixes the problem by making sure that vfp_save_state() is
always entered with FPEXC.EX cleared. (The current value of FPEXC has
already been saved, so this does not corrupt the context. Clearing
FPEXC.EX has no effects on FPINST or FPINST2. Also note that many
callers already modify FPEXC by setting FPEXC.EN before invoking
vfp_save_state().)

This patch also addresses a second problem related to FPEXC.EX: After
returning from signal handling, the kernel reloads the VFP context
from the user mode stack. However, the current code explicitly clears
both FPEXC.EX and FPEXC.FP2V during reload. As VFP11 requires these
bits to be preserved, this patch disables clearing them for VFP
implementations belonging to architecture 1. There should be no
negative side effects: the user can set both bits by executing FPU
opcodes anyway, and while user code may now place arbitrary values
into FPINST and FPINST2 (e.g., non-VFP ARM opcodes) the VFP support
code knows which instructions can be emulated, and rejects other
opcodes with "unhandled bounce" messages, so there should be no
security impact from allowing reloading FPEXC.EX and FPEXC.FP2V.

Signed-off-by: default avatarChristopher Alexander Tobias Schulze <cat.schulze@alice-dsl.net>
      e4652cc8
    • popcornmix's avatar
      reboot: Use power off rather than busy spinning when halt is requested · 0b93c40e
      popcornmix authored
      0b93c40e
    • Noralf Trønnes's avatar
      irqchip: bcm2835: Add FIQ support · 61255346
      Noralf Trønnes authored 
      

Add a duplicate irq range with an offset on the hwirq's so the
driver can detect that enable_fiq() is used.
Tested with downstream dwc_otg USB controller driver.

Signed-off-by: default avatarNoralf Trønnes <noralf@tronnes.org>
Reviewed-by: default avatarEric Anholt <eric@anholt.net>
Acked-by: default avatarStephen Warren <swarren@wwwdotorg.org>
      61255346
    • Dan Pasanen's avatar
      arm: partially revert 702b94bf · 3a61f83c
      Dan Pasanen authored 
      * Re-expose some dmi APIs for use in VCSM
      3a61f83c
  3. Jun 17, 2020
  5. Jun 03, 2020
  7. May 27, 2020
    • Daniel Borkmann's avatar
      bpf: Restrict bpf_probe_read{, str}() only to archs where they work · d21e07bc
      Daniel Borkmann authored 
      commit 0ebeea8c upstream.

Given the legacy bpf_probe_read{,str}() BPF helpers are broken on archs
with overlapping address ranges, we should really take the next step to
disable them from BPF use there.

To generally fix the situation, we've recently added new helper variants
bpf_probe_read_{user,kernel}() and bpf_probe_read_{user,kernel}_str().
For details on them, see 6ae08ae3

 ("bpf: Add probe_read_{user, kernel}
and probe_read_{user,kernel}_str helpers").

Given bpf_probe_read{,str}() have been around for ~5 years by now, there
are plenty of users at least on x86 still relying on them today, so we
cannot remove them entirely w/o breaking the BPF tracing ecosystem.

However, their use should be restricted to archs with non-overlapping
address ranges where they are working in their current form. Therefore,
move this behind a CONFIG_ARCH_HAS_NON_OVERLAPPING_ADDRESS_SPACE and
have x86, arm64, arm select it (other archs supporting it can follow-up
on it as well).

For the remaining archs, they can workaround easily by relying on the
feature probe from bpftool which spills out defines that can be used out
of BPF C code to implement the drop-in replacement for old/new kernels
via: bpftool feature probe macro

Suggested-by: default avatarLinus Torvalds <torvalds@linux-foundation.org>
Signed-off-by: default avatarDaniel Borkmann <daniel@iogearbox.net>
Signed-off-by: default avatarAlexei Starovoitov <ast@kernel.org>
Reviewed-by: default avatarMasami Hiramatsu <mhiramat@kernel.org>
Acked-by: default avatarLinus Torvalds <torvalds@linux-foundation.org>
Cc: Brendan Gregg <brendan.d.gregg@gmail.com>
Cc: Christoph Hellwig <hch@lst.de>
Link: https://lore.kernel.org/bpf/20200515101118.6508-2-daniel@iogearbox.net


Signed-off-by: default avatarGreg Kroah-Hartman <gregkh@linuxfoundation.org>
      d21e07bc
    • Thomas Gleixner's avatar
      ARM: futex: Address build warning · f5a51dc9
      Thomas Gleixner authored 
      [ Upstream commit 8101b5a1 ]

Stephen reported the following build warning on a ARM multi_v7_defconfig
build with GCC 9.2.1:

kernel/futex.c: In function 'do_futex':
kernel/futex.c:1676:17: warning: 'oldval' may be used uninitialized in this function [-Wmaybe-uninitialized]
 1676 |   return oldval == cmparg;
      |          ~~~~~~~^~~~~~~~~
kernel/futex.c:1652:6: note: 'oldval' was declared here
 1652 |  int oldval, ret;
      |      ^~~~~~

introduced by commit a08971e9

 ("futex: arch_futex_atomic_op_inuser()
calling conventions change").

While that change should not make any difference it confuses GCC which
fails to work out that oldval is not referenced when the return value is
not zero.

GCC fails to properly analyze arch_futex_atomic_op_inuser(). It's not the
early return, the issue is with the assembly macros. GCC fails to detect
that those either set 'ret' to 0 and set oldval or set 'ret' to -EFAULT
which makes oldval uninteresting. The store to the callsite supplied oldval
pointer is conditional on ret == 0.

The straight forward way to solve this is to make the store unconditional.

Aside of addressing the build warning this makes sense anyway because it
removes the conditional from the fastpath. In the error case the stored
value is uninteresting and the extra store does not matter at all.

Reported-by: default avatarStephen Rothwell <sfr@canb.auug.org.au>
Signed-off-by: default avatarThomas Gleixner <tglx@linutronix.de>
Link: https://lkml.kernel.org/r/87pncao2ph.fsf@nanos.tec.linutronix.de


Signed-off-by: default avatarSasha Levin <sashal@kernel.org>
      f5a51dc9
  9. May 20, 2020
  11. May 14, 2020
    • Jason A. Donenfeld's avatar
      crypto: arch/lib - limit simd usage to 4k chunks · e7b8447e
      Jason A. Donenfeld authored 
      commit 706024a5 upstream.

The initial Zinc patchset, after some mailing list discussion, contained
code to ensure that kernel_fpu_enable would not be kept on for more than
a 4k chunk, since it disables preemption. The choice of 4k isn't totally
scientific, but it's not a bad guess either, and it's what's used in
both the x86 poly1305, blake2s, and nhpoly1305 code already (in the form
of PAGE_SIZE, which this commit corrects to be explicitly 4k for the
former two).

Ard did some back of the envelope calculations and found that
at 5 cycles/byte (overestimate) on a 1ghz processor (pretty slow), 4k
means we have a maximum preemption disabling of 20us, which Sebastian
confirmed was probably a good limit.

Unfortunately the chunking appears to have been left out of the final
patchset that added the glue code. So, this commit adds it back in.

Fixes: 84e03fa3 ("crypto: x86/chacha - expose SIMD ChaCha routine as library function")
Fixes: b3aad5ba ("crypto: arm64/chacha - expose arm64 ChaCha routine as library function")
Fixes: a44a3430 ("crypto: arm/chacha - expose ARM ChaCha routine as library function")
Fixes: d7d7b853 ("crypto: x86/poly1305 - wire up faster implementations for kernel")
Fixes: f569ca16 ("crypto: arm64/poly1305 - incorporate OpenSSL/CRYPTOGAMS NEON implementation")
Fixes: a6b803b3 ("crypto: arm/poly1305 - incorporate OpenSSL/CRYPTOGAMS NEON implementation")
Fixes: ed0356ed

 ("crypto: blake2s - x86_64 SIMD implementation")
Cc: Eric Biggers <ebiggers@google.com>
Cc: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Cc: stable@vger.kernel.org
Signed-off-by: default avatarJason A. Donenfeld <Jason@zx2c4.com>
Reviewed-by: default avatarArd Biesheuvel <ardb@kernel.org>
Signed-off-by: default avatarHerbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: default avatarGreg Kroah-Hartman <gregkh@linuxfoundation.org>
      e7b8447e
    • Jason A. Donenfeld's avatar
      crypto: arch/nhpoly1305 - process in explicit 4k chunks · d4487b63
      Jason A. Donenfeld authored 
      commit a9a8ba90 upstream.

Rather than chunking via PAGE_SIZE, this commit changes the arch
implementations to chunk in explicit 4k parts, so that calculations on
maximum acceptable latency don't suddenly become invalid on platforms
where PAGE_SIZE isn't 4k, such as arm64.

Fixes: 0f961f9f ("crypto: x86/nhpoly1305 - add AVX2 accelerated NHPoly1305")
Fixes: 012c8238 ("crypto: x86/nhpoly1305 - add SSE2 accelerated NHPoly1305")
Fixes: a00fa0c8 ("crypto: arm64/nhpoly1305 - add NEON-accelerated NHPoly1305")
Fixes: 16aae359

 ("crypto: arm/nhpoly1305 - add NEON-accelerated NHPoly1305")
Cc: stable@vger.kernel.org
Signed-off-by: default avatarJason A. Donenfeld <Jason@zx2c4.com>
Reviewed-by: default avatarEric Biggers <ebiggers@google.com>
Signed-off-by: default avatarHerbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: default avatarGreg Kroah-Hartman <gregkh@linuxfoundation.org>
      d4487b63
  13. May 06, 2020
  15. May 02, 2020