mbox series

[00/12] UBS: Cleanup in_interupt/in_irq/in_atomic() usage

Message ID 20201014145215.518912759@linutronix.de
Headers show
Series UBS: Cleanup in_interupt/in_irq/in_atomic() usage | expand

Message

Thomas Gleixner Oct. 14, 2020, 2:52 p.m. UTC
Folks,

in the discussion about preempt count consistency accross kernel configurations:

  https://lore.kernel.org/r/20200914204209.256266093@linutronix.de/

Linus clearly requested that code in drivers and libraries which changes
behaviour based on execution context should either be split up so that
e.g. task context invocations and BH invocations have different interfaces
or if that's not possible the context information has to be provided by the
caller which knows in which context it is executing.

This includes conditional locking, allocation mode (GFP_*) decisions and
avoidance of code paths which might sleep.

In the long run, usage of 'preemptible, in_*irq etc.' should be banned from
driver code completely.

The usage of such constructs in USB is rather limited. Most of it is in
debug code and (misleading) comments. But of course there are also a few
few bugs including one unfixable.

With the following series applied, USB is clean.

Thanks,

	tglx
---
 atm/usbatm.c             |    2 
 core/buffer.c            |    6 +-
 core/hcd-pci.c           |    6 +-
 core/hcd.c               |   21 ++++----
 core/hub.c               |    3 -
 core/message.c           |   35 +++++++++-----
 core/usb.c               |    4 -
 gadget/udc/core.c        |    2 
 gadget/udc/dummy_hcd.c   |    5 +-
 gadget/udc/pxa27x_udc.c  |   19 ++++---
 host/ehci-fsl.c          |    9 +--
 host/ehci-pmcmsp.c       |   15 +++---
 host/isp1362.h           |    5 +-
 host/ohci-at91.c         |   11 +++-
 host/ohci-omap.c         |    7 +-
 host/ohci-pxa27x.c       |   11 ++--
 host/ohci-s3c2410.c      |   12 ++---
 host/xhci-mem.c          |    2 
 host/xhci.c              |    6 --
 misc/sisusbvga/Kconfig   |    2 
 serial/digi_acceleport.c |    7 +-
 serial/keyspan_pda.c     |  112 ++++++++++++++++++++---------------------------
 usbip/usbip_common.c     |    5 --
 23 files changed, 156 insertions(+), 151 deletions(-)

Comments

Alan Stern Oct. 14, 2020, 4:14 p.m. UTC | #1
On Wed, Oct 14, 2020 at 04:52:18PM +0200, Thomas Gleixner wrote:
> From: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
> 
> Having two copies of the same code doesn't make the code more readable and
> allocating a buffer of 1 byte for a synchronous operation is a pointless
> exercise.

Not so.  In fact, it is required, because a portion of a structure 
cannot be mapped for DMA unless it is aligned at a cache line boundary.

> Add a byte buffer to struct keyspan_pda_private which can be used
> instead. The buffer is only used in open() and tty->write().

This won't work.

>  Console writes
> are not calling into the query. open() obviously happens before write() and
> the writes are serialized by bit 0 of port->write_urbs_free which protects
> also the transaction itself.
> 
> Move the actual query into a helper function and cleanup the usage sites in
> keyspan_pda_write() and keyspan_pda_open().
> 
> Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
> Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
> Cc: Johan Hovold <johan@kernel.org>
> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
> Cc: linux-usb@vger.kernel.org
> ---
>  drivers/usb/serial/keyspan_pda.c |  102 ++++++++++++++++-----------------------
>  1 file changed, 43 insertions(+), 59 deletions(-)
> 
> --- a/drivers/usb/serial/keyspan_pda.c
> +++ b/drivers/usb/serial/keyspan_pda.c
> @@ -47,6 +47,7 @@ struct keyspan_pda_private {
>  	struct work_struct			unthrottle_work;
>  	struct usb_serial	*serial;
>  	struct usb_serial_port	*port;
> +	u8			query_buf;
>  };
>  
>  
> @@ -436,6 +437,31 @@ static int keyspan_pda_tiocmset(struct t
>  	return rc;
>  }
>  
> +/*
> + * Using priv->query_buf is safe here because this is only called for TTY
> + * operations open() and write(). write() comes post open() obviously and
> + * write() itself is serialized via bit 0 of port->write_urbs_free. Console
> + * writes are never calling into this.
> + */
> +static int keyspan_pda_query_room(struct usb_serial *serial,
> +				  struct keyspan_pda_private *priv)
> +{
> +	int res;
> +
> +	res = usb_control_msg(serial->dev, usb_rcvctrlpipe(serial->dev, 0),
> +			      6, /* write_room */
> +			      USB_TYPE_VENDOR | USB_RECIP_INTERFACE | USB_DIR_IN,
> +			      0, /* value */
> +			      0, /* index */
> +			      &priv->query_buf,
> +			      1,
> +			      2000);

Instead, consider using the new usb_control_msg_recv() API.  But it 
might be better to allocate the buffer once and for all.

Alan Stern
Thomas Gleixner Oct. 14, 2020, 4:17 p.m. UTC | #2
On Wed, Oct 14 2020 at 12:14, Alan Stern wrote:
> On Wed, Oct 14, 2020 at 04:52:18PM +0200, Thomas Gleixner wrote:
>> From: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
>> 
>> Having two copies of the same code doesn't make the code more readable and
>> allocating a buffer of 1 byte for a synchronous operation is a pointless
>> exercise.
>
> Not so.  In fact, it is required, because a portion of a structure 
> cannot be mapped for DMA unless it is aligned at a cache line boundary.
>
>> Add a byte buffer to struct keyspan_pda_private which can be used
>> instead. The buffer is only used in open() and tty->write().
>
> This won't work.

Ok.

>> +	res = usb_control_msg(serial->dev, usb_rcvctrlpipe(serial->dev, 0),
>> +			      6, /* write_room */
>> +			      USB_TYPE_VENDOR | USB_RECIP_INTERFACE | USB_DIR_IN,
>> +			      0, /* value */
>> +			      0, /* index */
>> +			      &priv->query_buf,
>> +			      1,
>> +			      2000);
>
> Instead, consider using the new usb_control_msg_recv() API.  But it 
> might be better to allocate the buffer once and for all.

Let me have a look.

Thanks,

        tglx
Sebastian Andrzej Siewior Oct. 14, 2020, 4:27 p.m. UTC | #3
On 2020-10-14 12:14:33 [-0400], Alan Stern wrote:
> Instead, consider using the new usb_control_msg_recv() API.  But it 
> might be better to allocate the buffer once and for all.

This will still allocate and free buffer on each invocation. What about
moving the query_buf to the begin of the struct / align it?

> Alan Stern

Sebastian
Alan Stern Oct. 14, 2020, 4:34 p.m. UTC | #4
On Wed, Oct 14, 2020 at 06:27:14PM +0200, Sebastian Andrzej Siewior wrote:
> On 2020-10-14 12:14:33 [-0400], Alan Stern wrote:
> > Instead, consider using the new usb_control_msg_recv() API.  But it 
> > might be better to allocate the buffer once and for all.
> 
> This will still allocate and free buffer on each invocation. What about

Yes.  That's why I suggesting doing a single buffer allocation at the 
start and using it for each I/O transfer.  (But I'm not familiar with 
this code, and I don't know if there might be multiple transfers going 
on concurrently.)

> moving the query_buf to the begin of the struct / align it?

No, thank won't work either.  The key to the issue is that while some 
memory is mapped for DMA, the CPU must not touch it or anything else in 
the same cache line.  If a field is a member of a data structure, the 
CPU might very well access a neighboring member while this one is 
mapped, thereby messing up the cache line.

Alan Stern
Sebastian Andrzej Siewior Oct. 14, 2020, 4:44 p.m. UTC | #5
On 2020-10-14 12:34:25 [-0400], Alan Stern wrote:
> On Wed, Oct 14, 2020 at 06:27:14PM +0200, Sebastian Andrzej Siewior wrote:
> > On 2020-10-14 12:14:33 [-0400], Alan Stern wrote:
> > > Instead, consider using the new usb_control_msg_recv() API.  But it 
> > > might be better to allocate the buffer once and for all.
> > 
> > This will still allocate and free buffer on each invocation. What about
> 
> Yes.  That's why I suggesting doing a single buffer allocation at the 
> start and using it for each I/O transfer.  (But I'm not familiar with 
> this code, and I don't know if there might be multiple transfers going 
> on concurrently.)

There are no concurrent transfer. There is a bit used as a lock. The
first one does the transfer, the other wait.

> > moving the query_buf to the begin of the struct / align it?
> 
> No, thank won't work either.  The key to the issue is that while some 
> memory is mapped for DMA, the CPU must not touch it or anything else in 
> the same cache line.  If a field is a member of a data structure, the 
> CPU might very well access a neighboring member while this one is 
> mapped, thereby messing up the cache line.

that is unfortunately true. Let me do the single buffer.

> Alan Stern

Sebastian
Pavel Machek Oct. 23, 2020, 8:01 a.m. UTC | #6
> The usage of such constructs in USB is rather limited. Most of it is in
> debug code and (misleading) comments. But of course there are also a few
> few bugs including one unfixable.

The "UBS" typo in the subject got me confused for a while.

Best regards,
									Pavel