This patch fixes some performance bugs observed with some workloads
when unlinking EHCI queue header (QH) descriptors from the async ring
(control/bulk schedule).

The mechanism intended to defer unlinking an empty QH (so there is no
penalty in common cases where it's quickly reused) was not working as
intended.  Sometimes the unlink was scheduled:

 - too quickly ... which can be a *strong* negative effect, since
   that QH becomes unavailable for immediate re-use;

 - too slowly ... wasting DMA cycles, usually a minor issue except
   for increased bus contention and power usage;

Plus there was an extreme case of "too slowly":  a logical error in the
IAA watchdog-timer conversion meant that sometimes the unlink never
got scheduled.

The fix replaces a simple counter with a timestamp derived from the
controller's 8 KHz microframe counter, and adjusts the timer usage
for some issues associated with HZ being less than 8K.

(Based on a patch originally by Alan Stern, and good troubleshooting
from  Leonid.)
Signed-off-by: NDavid Brownell <dbrownell@users.sourceforge.net>
Cc: Alan Stern <stern@rowland.harvard.edu>
Cc: Leonid <leonidv11@gmail.com>
Signed-off-by: NGreg Kroah-Hartman <gregkh@suse.de>

This reverts commit fa38dfcc.

It wasn't really a regression and David and Alan are still working
through the issues reported.
Signed-off-by: NGreg Kroah-Hartman <gregkh@suse.de>

This patch (as1099) fixes a performance regression in ehci-hcd.  The
fundamental problem is that queue headers get removed from the
schedule too quickly, since the code checks for a counter advancing
rather than making an actual time-based check.  The latency involved
in removing the queue header and then relinking it can severely
degrade certain kinds of workloads.

The patch replaces a simple counter with a timestamp derived from the
controller's uframe value.  In addition, the delay for unlinking an
idle queue header is increased from 5 ms to 10 ms; since some
controllers (nVidia) have a latency of up to 1 ms for unlinking, this
reduces the relative impact from 20% to 10%.

Finally, a logical error left over from the IAA watchdog-timer
conversion is corrected.  Now the driver will always either unlink an
idle queue header or set up a timer to unlink it later.  The old code
would sometimes fail to do either.
Signed-off-by: NAlan Stern <stern@rowland.harvard.edu>
Cc: David Brownell <david-b@pacbell.net>
Cc: Leonid <leonidv11@gmail.com>
Signed-off-by: NGreg Kroah-Hartman <gregkh@suse.de>

* endianness annotations
* endianness fixes
* missing get_unaligned/put_unaligned

It's pretty much all over the place, changes to different files are independent.
Signed-off-by: NAl Viro <viro@zeniv.linux.org.uk>
Serial-parts-Acked-by: NAlan Cox <alan@redhat.com>
Signed-off-by: NLinus Torvalds <torvalds@linux-foundation.org>

__FUNCTION__ is gcc-specific, use __func__
Signed-off-by: NHarvey Harrison <harvey.harrison@gmail.com>
Signed-off-by: NGreg Kroah-Hartman <gregkh@suse.de>

Simplify processing of completed qtds, and correct handling of short
reads, by removing two state variables:

 - "qtd_status" wasn't needed.  The current URB's status is either
   OK (-EINPROGRESS) or some fault status.  Once a fault appears,
   the queue halts and any later QTDs are immediately removed, so
   no temporary status is needed.  (Or for typical short reads,
   it's not treated as a fault, so no queue halt is needed.)

 - "do_status" was erroneous.  Because of how the queue is set up,
   short control reads can (and should!) be treated like full size
   reads, and cleaned up the usual way.  The status stage will be
   executed transparently, and usbcore handles the choice of whether
   to report this status as unexected.
 
The "do_status" problem caused a rather perplexing timing-dependent
problem with usbtest case 10.  Sometimes it would make the controller
skip a dozen transactions while (wrongly) trying to clean up after a
short transfer.  Fortunately, removing a dcache contention issue made
this become trivial to reproduce (on one test rig), so enough clues
finally presented themselves ... I think this has been around for a
very long time, but was worsened by recent urb->status changes.
Signed-off-by: NDavid Brownell <dbrownell@users.sourceforge.net>
Signed-off-by: NGreg Kroah-Hartman <gregkh@suse.de>

Provide better comments about qh_completions() and QTD handling.
That code can be *VERY* confusing, since it's evolved over a few
years to cope with both hardware races and silicon quirks.

Remove two unlikely() annotations that match the GCC defaults
(and are thus pointless); add an "else" to highlight code flow.

This patch doesn't change driver behavior.
Signed-off-by: NDavid Brownell <dbrownell@users.sourceforge.net>
Signed-off-by: NGreg Kroah-Hartman <gregkh@suse.de>

This teaches EHCI how to to work around bugs in certain high speed
devices, by accomodating "bulk" packets that exceed the 512 byte
constant value required by the USB 2.0 specification.  (Have a
look at section 5.8.3, paragraphs 1 and 3.)

It also makes the descriptor parsing code warn when it encounters
such bugs.  (We've had reports of maybe two or three such devices,
all pretty recent.)

Such devices are nonconformant.  The proper fix is have the vendors
of those devices do the simple, obvious, and correct thing ... which
will let them be used with USB hosts that don't have workarounds for
this particular vendor bug.  But unless/until they do, we can at least
have one of the high speed HCDs work with such buggy devices.
Signed-off-by: NDavid Brownell <dbrownell@users.sourceforge.net>
Signed-off-by: NGreg Kroah-Hartman <gregkh@suse.de>

USB: ehci: Fixes completion for multi-qtd URB the short read case

When use of urb->status in the EHCI driver was reworked last August
(commit 14c04c0f), a bug was inserted
in the handling of early completion for bulk transactions that need
more than one qTD (e.g. more than 20KB in one URB).

This patch resolves that problem by ensuring that the early completion
status is preserved until the URB is handed back to its submitter,
instead of resetting it after each qTD.
Signed-off-by: NMisha Zhilin <misha@epiphan.com>
Signed-off-by: NDavid Brownell <dbrownell@users.sourceforge.net>
Acked-by: NAlan Stern <stern@rowland.harvard.edu>
Cc: stable@kernel.org
Signed-off-by: NGreg Kroah-Hartman <gregkh@suse.de>

Kernel bugzilla entry #9569 reports a potential OOPS in some code
supporting the integrated root hub TT support used on ARC/TDI
derived cores.  (This seems to have been a longstanding issue.)

This patch cleans up usage of urb->dev->tt to avoid that potential
oops and also fixes some overly long lines.
Signed-off-by: NDavid Brownell <dbrownell@users.sourceforge.net>
Signed-off-by: NGreg Kroah-Hartman <gregkh@suse.de>

This patch (as1028) was mostly written by David Brownell; I made only
a few changes (extra log info and a small bug fix -- which might
account for why David's version had to be reverted).  It adds a new
watchdog timer to the ehci-hcd driver to be used exclusively for
detecting lost or missing IAA notifications.

Previously a shared timer had been used, which may have led to some
problems as reported by Christian Hoffmann.
Signed-off-by: NAlan Stern <stern@rowland.harvard.edu>
Cc: David Brownell <david-b@pacbell.net>
Signed-off-by: NGreg Kroah-Hartman <gregkh@suse.de>

This patch (as979) removes the last vestiges of urb->status from the
host controller drivers and the root-hub emulator.  Now the field
doesn't get set until just before the URB's completion routine is
called.
Signed-off-by: NAlan Stern <stern@rowland.harvard.edu>
CC: David Brownell <david-b@pacbell.net>
CC: Olav Kongas <ok@artecdesign.ee>
CC: Yoshihiro Shimoda <shimoda.yoshihiro@renesas.com>
CC: Tony Olech <tony.olech@elandigitalsystems.com>
Signed-off-by: NGreg Kroah-Hartman <gregkh@suse.de>

This patch (as974) reorganizes the way ehci-hcd sets urb->status.  It
now keeps the information in a local variable until the last moment.

The patch also simplifies the handling of -EREMOTEIO, since the only
use of that code is to set the do_status flag.
Signed-off-by: NAlan Stern <stern@rowland.harvard.edu>
CC: David Brownell <david-b@pacbell.net>
Signed-off-by: NGreg Kroah-Hartman <gregkh@suse.de>

This patch (as970) adds a new urb->unlinked field, which is used to
store the status of unlinked URBs since we can't use urb->status for
that purpose any more.  To help simplify the HCDs, usbcore will check
urb->unlinked before calling the completion handler; if the value is
set it will automatically override the status reported by the HCD.
Signed-off-by: NAlan Stern <stern@rowland.harvard.edu>
CC: David Brownell <david-b@pacbell.net>
CC: Olav Kongas <ok@artecdesign.ee>
CC: Yoshihiro Shimoda <shimoda.yoshihiro@renesas.com>
CC: Tony Olech <tony.olech@elandigitalsystems.com>
Signed-off-by: NGreg Kroah-Hartman <gregkh@suse.de>

This patch (as969) continues the ongoing changes to the way HCDs
report URB statuses.  The programming interface has been simplified by
making usbcore responsible for clearing urb->hcpriv and for setting
-EREMOTEIO status when an URB with the URB_SHORT_NOT_OK flag ends up
as a short transfer.

By moving the work out of the HCDs, this removes a fair amount of
repeated code.
Signed-off-by: NAlan Stern <stern@rowland.harvard.edu>
CC: David Brownell <david-b@pacbell.net>
CC: Olav Kongas <ok@artecdesign.ee>
CC: Yoshihiro Shimoda <shimoda.yoshihiro@renesas.com>
CC: Tony Olech <tony.olech@elandigitalsystems.com>
Signed-off-by: NGreg Kroah-Hartman <gregkh@suse.de>

This patch (as954) implements a suggestion of David Brownell's.  Now
the host controller drivers are responsible for linking and unlinking
URBs to/from their endpoint queues.  This eliminates the possiblity of
strange situations where usbcore thinks an URB is linked but the HCD
thinks it isn't.  It also means HCDs no longer have to check for URBs
being dequeued before they were fully enqueued.

In addition to the core changes, this requires changing every host
controller driver and the root-hub URB handler.  For the most part the
required changes are fairly small; drivers have to call
usb_hcd_link_urb_to_ep() in their urb_enqueue method,
usb_hcd_check_unlink_urb() in their urb_dequeue method, and
usb_hcd_unlink_urb_from_ep() before giving URBs back.  A few HCDs make
matters more complicated by the way they split up the flow of control.

In addition some method interfaces get changed.  The endpoint argument
for urb_enqueue is now redundant so it is removed.  The unlink status
is required by usb_hcd_check_unlink_urb(), so it has been added to
urb_dequeue.
Signed-off-by: NAlan Stern <stern@rowland.harvard.edu>
CC: David Brownell <david-b@pacbell.net>
CC: Olav Kongas <ok@artecdesign.ee>
CC: Tony Olech <tony.olech@elandigitalsystems.com>
CC: Yoshihiro Shimoda <shimoda.yoshihiro@renesas.com>
Signed-off-by: NGreg Kroah-Hartman <gregkh@suse.de>

This reverts commit 196705c9.  It was
reported to cause a regression by Daniel Exner, and Arjan van de Ven
points out that we actually already have infrastructure in place for
setting limits on acceptable DMA latency that would be the much more
correct fix for the problem with some Broadcom EHCI controllers.

Fixed up trivial conflicts due to the changes to support big-endian host
controller descriptors in drivers/usb/host/{ehci-sched.c,ehci.h}.
Signed-off-by: NLinus Torvalds <torvalds@linux-foundation.org>

This patch implements supports for EHCI controllers whose in-memory
data structures are represented in big-endian format. This is needed
(unfortunately) for the AMCC PPC440EPx SoC EHCI controller; the EHCI
spec doesn't specify little-endian format, although that's what most
other implementations use.

The guts of the patch are to introduce the hc32 type and change all
references from le32 to hc32.  All access routines are converted from
cpu_to_le32(...) to cpu_to_hc32(ehci, ...) and similar for the other
"direction".  (This is the same approach used with OHCI.)

David fixed:
	Whitespace fixes; refresh against ehci cpufreq patch; move glue
	for that PPC driver to the patch adding it; fix free symbol
	capture bugs in modified "constant" macros; and make "hc32" etc
	be "le32" unless we really need the BE options, so "sparse" can
	do some real good.
Signed-off-by: NStefan Roese <sr@denx.de>
Signed-off-by: NDavid Brownell <dbrownell@users.sourceforge.net>
Signed-off-by: NGreg Kroah-Hartman <gregkh@suse.de>

EHCI controllers that don't cache enough microframes can get MMF errors
when CPU frequency changes occur between the start and completion of
split interrupt transactions, due to delays in reading main memory
(caused by CPU cache snoop delays).

This patch adds a cpufreq notifier to the EHCI driver that will
inactivate split interrupt transactions during frequency transitions.
It was tested on Intel ICH7 and Serverworks/Broadcom HT1000 EHCI
controllers.
Signed-off-by: NStuart Hayes <stuart_hayes@dell.com>
Signed-off-by: NDavid Brownell <dbrownell@users.sourceforge.net>
Signed-off-by: NGreg Kroah-Hartman <gregkh@suse.de>

This patch implements supports for EHCI controllers whose MMIO
registers are big endian and enables that functionality for
the Toshiba SCC chip. It does _not_ add support for big endian
in-memory data structures as this is not needed for that chip
and I hope it will never be.

The guts of the patch are to convert readl(...) to
ehci_readl(ehci, ...) and similarly for register writes.
Signed-off-by: NKou Ishizaki <kou.ishizaki@toshiba.co.jp>
Signed-off-by: NBenjamin Herrenschmidt <benh@kernel.crashing.org>
Acked-by: NGeoff Levand <geoffrey.levand@am.sony.com>
Acked-by: NDavid Brownell <dbrownell@users.sourceforge.net>
Signed-off-by: NGreg Kroah-Hartman <gregkh@suse.de>

This reverts 26f953fd which caused
resume problems on the mac mini.

Cc: David Brownell <david-b@pacbell.net>
Signed-off-by: NGreg Kroah-Hartman <gregkh@suse.de>

Maintain a per-CPU global "struct pt_regs *" variable which can be used instead
of passing regs around manually through all ~1800 interrupt handlers in the
Linux kernel.

The regs pointer is used in few places, but it potentially costs both stack
space and code to pass it around.  On the FRV arch, removing the regs parameter
from all the genirq function results in a 20% speed up of the IRQ exit path
(ie: from leaving timer_interrupt() to leaving do_IRQ()).

Where appropriate, an arch may override the generic storage facility and do
something different with the variable.  On FRV, for instance, the address is
maintained in GR28 at all times inside the kernel as part of general exception
handling.

Having looked over the code, it appears that the parameter may be handed down
through up to twenty or so layers of functions.  Consider a USB character
device attached to a USB hub, attached to a USB controller that posts its
interrupts through a cascaded auxiliary interrupt controller.  A character
device driver may want to pass regs to the sysrq handler through the input
layer which adds another few layers of parameter passing.

I've build this code with allyesconfig for x86_64 and i386.  I've runtested the
main part of the code on FRV and i386, though I can't test most of the drivers.
I've also done partial conversion for powerpc and MIPS - these at least compile
with minimal configurations.

This will affect all archs.  Mostly the changes should be relatively easy.
Take do_IRQ(), store the regs pointer at the beginning, saving the old one:

	struct pt_regs *old_regs = set_irq_regs(regs);

And put the old one back at the end:

	set_irq_regs(old_regs);

Don't pass regs through to generic_handle_irq() or __do_IRQ().

In timer_interrupt(), this sort of change will be necessary:

	-	update_process_times(user_mode(regs));
	-	profile_tick(CPU_PROFILING, regs);
	+	update_process_times(user_mode(get_irq_regs()));
	+	profile_tick(CPU_PROFILING);

I'd like to move update_process_times()'s use of get_irq_regs() into itself,
except that i386, alone of the archs, uses something other than user_mode().

Some notes on the interrupt handling in the drivers:

 (*) input_dev() is now gone entirely.  The regs pointer is no longer stored in
     the input_dev struct.

 (*) finish_unlinks() in drivers/usb/host/ohci-q.c needs checking.  It does
     something different depending on whether it's been supplied with a regs
     pointer or not.

 (*) Various IRQ handler function pointers have been moved to type
     irq_handler_t.
Signed-Off-By: NDavid Howells <dhowells@redhat.com>
(cherry picked from 1b16e7ac850969f38b375e511e3fa2f474a33867 commit)

This revamps handling of the hardware "async advance" IRQ, and its watchdog
timer.  Basically it dis-entangles that important timeout from the others,
simplifying the associated state and code to make it more robust.

This reportedly improves behavior of EHCI on some systems with VIA chips,
and AFAIK won't affect non-VIA hardware.  VIA systems need this code to
recover from silcon bugs whereby the "async advance" IRQ isn't issued.
Signed-off-by: NDavid Brownell <dbrownell@users.sourceforge.net>
Signed-off-by: NGreg Kroah-Hartman <gregkh@suse.de>

[ ... when you have an editor set to remind you of whitespace bugs ... ]

Cosmetic EHCI changes: remove end-of-line whitespace, spaces before tabs.
Signed-off-by: NDavid Brownell <dbrownell@users.sourceforge.net>
Signed-off-by: NGreg Kroah-Hartman <gregkh@suse.de>

This patch (as640) removes several put_device and the corresponding
get_device calls from the USB core and HCDs.  Some of the puts were done
in atomic contexts, and none of them are needed since the core now
guarantees that every endpoint will be disabled and every URB completed
before a USB device is released.
Signed-off-by: NAlan Stern <stern@rowland.harvard.edu>
Acked-by: NDavid Brownell <david-b@pacbell.net>
Signed-off-by: NGreg Kroah-Hartman <gregkh@suse.de>

This patch modifies the behavior of the EHCI driver in an unlink path
that seems to be causing various issues on some systems.  Those problems
have included issues with disconnection, driver unbinding, and similar
cases where urb unlinking would just not work right.

This patch should help avoid those problems by not turning off the async
(control/bulk) schedule until it's not expecting an "async advance" IRQ,
which comes from the processing passing the schedule head.  Whether the
driver attempts to do such things is dependent on system timings, so
many folk would never have seen these problems.
Signed-off-by: NDavid Brownell <dbrownell@users.sourceforge.net>
Signed-off-by: NGreg Kroah-Hartman <gregkh@suse.de>

On the MPC834x processors the multiport host (MPH) EHCI controller has an
erratum in which the port number in the queue head expects to be 0..N-1
instead of 1..N.  If we are on one of these chips we subtract one from
the port number before putting it into the queue head.
Signed-off-by: NKumar Gala <galak@kernel.crashing.org>
Signed-off-by: NDavid Brownell <dbrownell@users.sourceforge.net>
Signed-off-by: NGreg Kroah-Hartman <gregkh@suse.de>

When the ehci-hcd driver prepares a control URB, it tests for a
zero-length data stage by looking at the transfer_dma value instead of
the transfer_buffer_length.  (In fact it does this even for non-control
URBs, which is an additional aspect of the same bug.)

However, under certain circumstances it's possible for transfer_dma to
be 0 while transfer_buffer_length is non-zero.  This can happen when a
freshly allocated page (mapped to address 0 and marked Copy-On-Write,
but never written to) is used as the source buffer for an OUT transfer.
This patch (as598) fixes the problem.
Signed-off-by: NAlan Stern <stern@rowland.harvard.edu>
Signed-off-by: NDavid Brownell <dbrownell@users.sourceforge.net>
Signed-off-by: NGreg Kroah-Hartman <gregkh@suse.de>

This patch closes the IRQ race and makes various other OHCI & EHCI code
path safer vs. suspend/resume.
I've been able to (finally !) successfully suspend and resume various
Mac models, with or without USB mouse plugged, or plugging while asleep,
or unplugging while asleep etc... all without a crash.

Alan, please verify the UHCI bit I did, I only verified that it builds.
It's very simple so I wouldn't expect any issue there. If you aren't
confident, then just drop the hunks that change uhci-hcd.c

I also made the patch a little bit more "safer" by making sure the store
to the interrupt register that disables interrupts is not posted before
I set the flag and drop the spinlock.

Without this patch, you cannot reliably sleep/wakeup any recent Mac, and
I suspect PCs have some more sneaky issues too (they don't frankly crash
with machine checks because x86 tend to silently swallow PCI errors but
that won't last afaik, at least PCI Express will blow up in those
situations, but the USB code may still misbehave).
Signed-off-by: NBenjamin Herrenschmidt <benh@kernel.crashing.org>
Signed-off-by: NGreg Kroah-Hartman <gregkh@suse.de>
Signed-off-by: NGreg Kroah-Hartman <gregkh@suse.de>

Signed-off-by: NAl Viro <viro@zeniv.linux.org.uk>
Signed-off-by: NLinus Torvalds <torvalds@osdl.org>

This adds the field tt_usecs to ehci_qh and ehci_iso_stream, and sets it
appropriately when setting them up as periodic endpoints.  It records
the transation translator's think_time (added in last patch) plus the
downstream (i.e. low or full speed) bustime of the transfer associated
with each interrupt or iso frame, as calculated by usb_calc_bus_time.
Signed-off-by: NDan Streetman <ddstreet@ieee.org>
Signed-off-by: NDavid Brownell <dbrownell@users.sourceforge.net>
Signed-off-by: NGreg Kroah-Hartman <gregkh@suse.de>

This patch has a one line oops fix, plus related cleanups.

 - The bugfix uses microframe scheduling data given to the hardware to
   test "is this a periodic QH", rather than testing for nonzero period.
   (Prevents an oops by providing the correct answer.)

 - The cleanup going along with the patch should make it clearer what's
   going on whenever those bitfields are accessed.

The bug came about when, around January, two new kinds of EHCI interrupt
scheduling operation were added, involving both the high speed (24 KBytes
per millisec) and low/full speed (1-64 bytes per millisec) microframe
scheduling.  A driver for the Edirol UA-1000 Audio Capture Unit ran into
the oops; it used one of the newly supported high speed modes.
Signed-off-by: NDavid Brownell <dbrownell@users.sourceforge.net>
Signed-off-by: NAndrew Morton <akpm@osdl.org>
Signed-off-by: NGreg Kroah-Hartman <gregkh@suse.de>
Signed-off-by: NLinus Torvalds <torvalds@osdl.org>

This patch does the same swap, i.e. use the ISO macro if (isoc).
Additionally, it fixes the return value - the usb_calc_bus_time function
returns the time in nanoseconds (I didn't notice that before) while the
HS_USECS and HS_USECS_ISO are microseconds.  This fixes the function to
return nanoseconds always, and adjusts ehci-q.c (the only high-speed
caller of the function) to wrap the call in NS_TO_US().
Signed-off-by: NDan Streetman <ddstreet@ieee.org>
Signed-off-by: NGreg Kroah-Hartman <gregkh@suse.de>
Signed-off-by: NLinus Torvalds <torvalds@osdl.org>

Greg,

This patch fixes the kmalloc() flags argument type in USB
subsystem; hopefully all of its occurences. The patch was
made against patch-2.6.12-git2 from Jun 20.

Cleanup of flags for kmalloc() in USB subsystem.
Signed-off-by: NOlav Kongas <ok@artecdesign.ee>
Signed-off-by: NGreg Kroah-Hartman <gregkh@suse.de>

Various minor EHCI updates

   * Dump some more info in the debug dumps, notably the product
     description (e.g. chip vendor), BIOS handhake flags, and
     debug port status (when it's not managed by the HCD).

   * Minor updates to the BIOS handoff code:  always flag the HCD
     as owned by Linux (in case BIOS doesn't grab it "early"),
     and on the buggy-BIOS path always match the "early handoff"
     code and forcibly disable SMI IRQs.

   * For the disabled 64bit DMA support, there's now a constant
     to use for the mask; use it.
Signed-off-by: NDavid Brownell <dbrownell@users.sourceforge.net>
Signed-off-by: NGreg Kroah-Hartman <gregkh@suse.de>

Initial git repository build. I'm not bothering with the full history,
even though we have it. We can create a separate "historical" git
archive of that later if we want to, and in the meantime it's about
3.2GB when imported into git - space that would just make the early
git days unnecessarily complicated, when we don't have a lot of good
infrastructure for it.

Let it rip!