Some Fresco Logic hosts, including those found in the AUAU N533V laptop,
advertise MSI, but fail to actually generate MSI interrupts. Add a new
xHCI quirk to skip MSI enabling for the Fresco Logic host controllers.
Fresco Logic confirms that all chips with PCI vendor ID 0x1b73 and device
ID 0x1000, regardless of PCI revision ID, do not support MSI.
This should be backported to stable kernels as far back as 2.6.36, which
was the first kernel to support MSI on xHCI hosts.
Signed-off-by: Sarah Sharp <sarah.a.sharp@linux.intel.com>
Reported-by: Sergey Galanov <sergey.e.galanov@gmail.com>
Cc: stable@kernel.org
xHCI controllers respond to a Reset Device command when the Slot is in the
Enabled/Disabled state by returning an error. This is fine on other host
controllers, but the Etron xHCI host controller returns a vendor-specific
error code that the xHCI driver doesn't understand. The xHCI driver then
gives up on device enumeration.
Instead of issuing a command that will fail, just return. This fixes the
issue with the xhci driver not working on ASRock P67 Pro/Extreme boards.
This should be backported to stable kernels as far back as 2.6.34.
Signed-off-by: Maarten Lankhorst <m.b.lankhorst@gmail.com>
Signed-off-by: Sarah Sharp <sarah.a.sharp@linux.intel.com>
Cc: stable@kernel.org
Some of the recently-added cpu_to_leXX and leXX_to_cpu made things somewhat
messy; this patch neatens some of these areas, removing unnecessary casts
in those parts also. In some places (where Y & Z are constants) a
comparison of (leXX_to_cpu(X) & Y) == Z has been replaced with
(X & cpu_to_leXX(Y)) == cpu_to_leXX(Z). The endian reversal of the
constants should wash out at compile time.
Signed-off-by: Matt Evans <matt@ozlabs.org>
Signed-off-by: Sarah Sharp <sarah.a.sharp@linux.intel.com>
Commit 834cb0fc47 "xhci: Fix memory leak
bug when dropping endpoints" added a small endian bug. This patch fixes
xhci_check_bandwidth() to read add/drop_flags LE.
Signed-off-by: Matt Evans <matt@ozlabs.org>
Signed-off-by: Sarah Sharp <sarah.a.sharp@linux.intel.com>
The Panther Point chipset has an xHCI host controller that has a limit to
the number of active endpoints it can handle. Ideally, it would signal
that it can't handle anymore endpoints by returning a Resource Error for
the Configure Endpoint command, but they don't. Instead it needs software
to keep track of the number of active endpoints, across configure endpoint
commands, reset device commands, disable slot commands, and address device
commands.
Add a new endpoint context counter, xhci_hcd->num_active_eps, and use it
to track the number of endpoints the xHC has active. This gets a little
tricky, because commands to change the number of active endpoints can
fail. This patch adds a new xHCI quirk for these Intel hosts, and the new
code should not have any effect on other xHCI host controllers.
Fail a new device allocation if we don't have room for the new default
control endpoint. Use the endpoint ring pointers to determine what
endpoints were active before a Reset Device command or a Disable Slot
command, and drop those once the command completes.
Fail a configure endpoint command if it would add too many new endpoints.
We have to be a bit over zealous here, and only count the number of new
endpoints to be added, without subtracting the number of dropped
endpoints. That's because a second configure endpoint command for a
different device could sneak in before we know if the first command is
completed. If the first command dropped resources, the host controller
fails the command for some reason, and we're nearing the limit of
endpoints, we could end up oversubscribing the host.
To fix this race condition, when evaluating whether a configure endpoint
command will fix in our bandwidth budget, only add the new endpoints to
xhci->num_active_eps, and don't subtract the dropped endpoints. Ignore
changed endpoints (ones that are dropped and then re-added), as that
shouldn't effect the host's endpoint resources. When the configure
endpoint command completes, subtract off the dropped endpoints.
This may mean some configuration changes may temporarily fail, but it's
always better to under-subscribe than over-subscribe resources.
(Originally my plan had been to push the resource allocation down into the
ring allocation functions. However, that would cause us to allocate
unnecessary resources when endpoints were changed, because the xHCI driver
allocates a new ring for the changed endpoint, and only deletes the old
ring once the Configure Endpoint command succeeds. A further complication
would have been dealing with the per-device endpoint ring cache.)
Signed-off-by: Sarah Sharp <sarah.a.sharp@linux.intel.com>
When the xHCI host controller dies, the USB core may attempt to reset the
devices to their default configuration before disconnecting them. This
causes calls into the xHCI bandwidth allocation functions. Don't allow
those functions to submit commands or work on xHCI structures if the host
controller is marked as dying.
Signed-off-by: Sarah Sharp <sarah.a.sharp@linux.intel.com>
When the USB core wants to change to an alternate interface setting that
doesn't include an active endpoint, or de-configuring the device, the xHCI
driver needs to issue a Configure Endpoint command to tell the host to
drop some endpoints from the schedule. After the command completes, the
xHCI driver needs to free rings for any endpoints that were dropped.
Unfortunately, the xHCI driver wasn't actually freeing the endpoint rings
for dropped endpoints. The rings would be freed if the endpoint's
information was simply changed (and a new ring was installed), but dropped
endpoints never had their rings freed. This caused errors when the ring
segment DMA pool was freed when the xHCI driver was unloaded:
[ 5582.883995] xhci_hcd 0000:06:00.0: dma_pool_destroy xHCI ring segments, ffff88003371d000 busy
[ 5582.884002] xhci_hcd 0000:06:00.0: dma_pool_destroy xHCI ring segments, ffff880033716000 busy
[ 5582.884011] xhci_hcd 0000:06:00.0: dma_pool_destroy xHCI ring segments, ffff880033455000 busy
[ 5582.884018] xhci_hcd 0000:06:00.0: Freed segment pool
[ 5582.884026] xhci_hcd 0000:06:00.0: Freed device context pool
[ 5582.884033] xhci_hcd 0000:06:00.0: Freed small stream array pool
[ 5582.884038] xhci_hcd 0000:06:00.0: Freed medium stream array pool
[ 5582.884048] xhci_hcd 0000:06:00.0: xhci_stop completed - status = 1
[ 5582.884061] xhci_hcd 0000:06:00.0: USB bus 3 deregistered
[ 5582.884193] xhci_hcd 0000:06:00.0: PCI INT A disabled
Fix this issue and free endpoint rings when their endpoints are
successfully dropped.
This patch should be backported to kernels as old as 2.6.31.
Signed-off-by: Sarah Sharp <sarah.a.sharp@linux.intel.com>
Cc: stable@kernel.org
This is a new TRB Completion Code of the xHCI spec 1.0.
Asserted by the Evalute Context Command if the proposed Max Exit Latency would
not allow the periodic endpoints of the Device Slot to be scheduled.
Signed-off-by: Alex He <alex.he@amd.com>
Signed-off-by: Andiry Xu <andiry.xu@amd.com>
Signed-off-by: Sarah Sharp <sarah.a.sharp@linux.intel.com>
xHCI 1.0 spec says the TT Think Time field shall be set to zero if the device
is not a High-speed hub.
Signed-off-by: Andiry Xu <andiry.xu@amd.com>
Signed-off-by: Sarah Sharp <sarah.a.sharp@linux.intel.com>
Sparse complains about the arguments to xhci_evaluate_context_result() and
xhci_configure_endpoint_result():
CHECK drivers/usb/host/xhci.c
drivers/usb/host/xhci.c:1647:53: warning: incorrect type in argument 3 (different signedness)
drivers/usb/host/xhci.c:1647:53: expected int *cmd_status
drivers/usb/host/xhci.c:1647:53: got unsigned int [usertype] *[assigned] cmd_status
drivers/usb/host/xhci.c:1648:50: warning: incorrect type in argument 3 (different signedness)
drivers/usb/host/xhci.c:1648:50: expected int *cmd_status
drivers/usb/host/xhci.c:1648:50: got unsigned int [usertype] *[assigned] cmd_status
The command status is taken from the command completion event TRB, and
will always be a positive number. Change the signature of
xhci_evaluate_context_result() and xhci_configure_endpoint_result() to
take a u32 for cmd_status.
Signed-off-by: Sarah Sharp <sarah.a.sharp@linux.intel.com>
During a "plug-unplug" stress test on an NEC xHCI card, a null pointer
dereference was observed. xhci_address_device() dereferenced a null
virt_dev (possibly an erroneous udev->slot_id?); this patch adds a WARN_ON &
message to aid debug if it can be recreated.
Signed-off-by: Matt Evans <matt@ozlabs.org>
Signed-off-by: Sarah Sharp <sarah.a.sharp@linux.intel.com>
This patch changes the struct members defining access to xHCI device-visible
memory to use __le32/__le64 where appropriate, and then adds swaps where
required. Checked with sparse that all accesses are correct.
MMIO accesses use readl/writel so already are performed LE, but prototypes
now reflect this with __le*.
There were a couple of (debug) instances of DMA pointers being truncated to
32bits which have been fixed too.
Signed-off-by: Matt Evans <matt@ozlabs.org>
Signed-off-by: Sarah Sharp <sarah.a.sharp@linux.intel.com>
This patch disable the optional PM feature inside the Hudson3 platform under
the following conditions:
1. If an isochronous device is connected to xHCI port and is active;
2. Optional PM feature that powers down the internal Bus PLL when the link is
in low power state is enabled.
The PM feature needs to be disabled to eliminate PLL startup delays when the
link comes out of low power state. The performance of DMA data transfer could
be impacted if system delay were encountered and in addition to the PLL start
up delays. Disabling the PM would leave room for unpredictable system delays
in order to guarantee uninterrupted data transfer to isochronous audio or
video stream devices that require time sensitive information. If data in an
audio/video stream was interrupted then erratic audio or video performance
may be encountered.
AMD PLL quirk is already implemented in OHCI/EHCI driver. After moving the
quirk code to pci-quirks.c and export them, xHCI driver can call it directly
without having the quirk implementation in itself.
Signed-off-by: Andiry Xu <andiry.xu@amd.com>
Signed-off-by: Sarah Sharp <sarah.a.sharp@linux.intel.com>
On a resume, when the power is lost during hibernate, the USB core will
call hub_reset_resume for the xHCI USB 2.0 roothub, but not for the USB
3.0 roothub:
[ 164.748310] usb usb1: root hub lost power or was reset
[ 164.748353] usb usb2: root hub lost power or was reset
[ 164.748487] usb usb3: root hub lost power or was reset
[ 164.748488] xhci_hcd 0000:01:00.0: Stop HCD
...
[ 164.870039] hub 4-0:1.0: hub_resume
...
[ 164.870054] hub 3-0:1.0: hub_reset_resume
This causes issues later, because the USB core assumes the USB 3.0 hub
attached to the USB 3.0 roothub is still active. It attempts to queue a
control URB for the external hub, which fails because all the device
slot contexts were released when the USB 3.0 roothub lost power:
[ 164.980044] hub 4-1:1.0: hub_resume
[ 164.980047] xhci_hcd 0000:01:00.0: Get port status returned 0x10101
[ 164.980049] xHCI xhci_urb_enqueue called with unaddressed device
[ 164.980053] hub 3-0:1.0: port 1: status 0101 change 0001
[ 164.980056] hub 4-1:1.0: hub_port_status failed (err = -22)
[ 164.980060] xhci_hcd 0000:01:00.0: `MEM_WRITE_DWORD(3'b000, 32'hffffc90008948440, 32'h202e1, 4'hf);
[ 164.980062] xHCI xhci_urb_enqueue called with unaddressed device
[ 164.980066] xhci_hcd 0000:01:00.0: clear port connect change, actual port 0 status = 0x2e1
[ 164.980069] hub 4-1:1.0: hub_port_status failed (err = -22)
[ 164.980072] xhci_hcd 0000:01:00.0: get port status, actual port 1 status = 0x2a0
[ 164.980074] xHCI xhci_urb_enqueue called with unaddressed device
[ 164.980077] xhci_hcd 0000:01:00.0: Get port status returned 0x100
[ 164.980079] hub 4-1:1.0: hub_port_status failed (err = -22)
[ 164.980082] xHCI xhci_urb_enqueue called with unaddressed device
[ 164.980085] hub 4-1:1.0: hub_port_status failed (err = -22)
[ 164.980088] hub 4-1:1.0: port 4: status 0000 change 0000
[ 164.980091] xHCI xhci_urb_enqueue called with unaddressed device
[ 164.980094] hub 4-1:1.0: activate --> -22
[ 164.980113] xHCI xhci_urb_enqueue called with unaddressed device
[ 164.980117] hub 4-1:1.0: hub_port_status failed (err = -22)
[ 164.980119] xHCI xhci_urb_enqueue called with unaddressed device
[ 164.980123] hub 4-1:1.0: can't resume port 4, status -22
[ 164.980126] hub 4-1:1.0: port 4 status ffff.ffff after resume, -22
[ 164.980129] usb 4-1.4: can't resume, status -22
[ 164.980131] hub 4-1:1.0: logical disconnect on port 4
This causes issues when a USB 3.0 hard drive is attached to the external
USB 3.0 hub when the system is hibernated:
[ 6249.849653] sd 8:0:0:0: [sdb] Unhandled error code
[ 6249.849659] sd 8:0:0:0: [sdb] Result: hostbyte=DID_ERROR driverbyte=DRIVER_OK
[ 6249.849663] sd 8:0:0:0: [sdb] CDB: Read(10): 28 00 00 00 2a 08 00 00 02 00
[ 6249.849671] end_request: I/O error, dev sdb, sector 10760
Make sure to inform the USB core that *both* xHCI roothubs lost power.
Signed-off-by: Sarah Sharp <sarah.a.sharp@linux.intel.com>
Currently, when resetting a device, xHCI driver disables all but one
endpoints and frees their rings, but leaves alone any streams that
might have been allocated. Later, when users try to free allocated
streams, we oops in xhci_setup_no_streams_ep_input_ctx() because
ep->ring is NULL.
Let's free not only rings but also stream data as well, so that
calling free_streams() on a device that was reset will be safe.
This should be queued for stable trees back to 2.6.35.
Reviewed-by: Micah Elizabeth Scott <micah@vmware.com>
Signed-off-by: Dmitry Torokhov <dtor@vmware.com>
Signed-off-by: Sarah Sharp <sarah.a.sharp@linux.intel.com>
Cc: stable@kernel.org
* 'usb-next' of git://git.kernel.org/pub/scm/linux/kernel/git/gregkh/usb-2.6: (172 commits)
USB: Add support for SuperSpeed isoc endpoints
xhci: Clean up cycle bit math used during stalls.
xhci: Fix cycle bit calculation during stall handling.
xhci: Update internal dequeue pointers after stalls.
USB: Disable auto-suspend for USB 3.0 hubs.
USB: Remove bogus USB_PORT_STAT_SUPER_SPEED symbol.
xhci: Return canceled URBs immediately when host is halted.
xhci: Fixes for suspend/resume of shared HCDs.
xhci: Fix re-init on power loss after resume.
xhci: Make roothub functions deal with device removal.
xhci: Limit roothub ports to 15 USB3 & 31 USB2 ports.
xhci: Return a USB 3.0 hub descriptor for USB3 roothub.
xhci: Register second xHCI roothub.
xhci: Change xhci_find_slot_id_by_port() API.
xhci: Refactor bus suspend state into a struct.
xhci: Index with a port array instead of PORTSC addresses.
USB: Set usb_hcd->state and flags for shared roothubs.
usb: Make core allocate resources per PCI-device.
usb: Store bus type in usb_hcd, not in driver flags.
usb: Change usb_hcd->bandwidth_mutex to a pointer.
...
When the xHCI host controller is halted, it won't respond to commands
placed on the command ring. So if an URB is cancelled after the first
roothub is deallocated, it will try to place a stop endpoint command on
the command ring, which will fail. The command watchdog timer will fire
after five seconds, and the host controller will be marked as dying, and
all URBs will be completed.
Add a flag to the xHCI's internal state variable for when the host
controller is halted. Immediately return the canceled URB if the host
controller is halted.
Signed-off-by: Sarah Sharp <sarah.a.sharp@linux.intel.com>
Make sure the HCD_FLAG_HW_ACCESSIBLE flag is mirrored by both roothubs,
since it refers to whether the shared hardware is accessible. Make sure
each bus is marked as suspended by setting usb_hcd->state to
HC_STATE_SUSPENDED when the PCI host controller is resumed.
Signed-off-by: Sarah Sharp <sarah.a.sharp@linux.intel.com>
When a host controller has lost power during a suspend, we must
reinitialize it. Now that the xHCI host has two roothubs, xhci_run() and
xhci_stop() expect to be called with both usb_hcd structures. Be sure
that the re-initialization code in xhci_resume() mirrors the process the
USB PCI probe function uses.
Signed-off-by: Sarah Sharp <sarah.a.sharp@linux.intel.com>
This patch changes the xHCI driver to allocate two roothubs. This touches
the driver initialization and shutdown paths, roothub emulation code, and
port status change event handlers. This is a rather large patch, but it
can't be broken up, or it would break git-bisect.
Make the xHCI driver register its own PCI probe function. This will call
the USB core to create the USB 2.0 roothub, and then create the USB 3.0
roothub. This gets the code for registering a shared roothub out of the
USB core, and allows other HCDs later to decide if and how many shared
roothubs they want to allocate.
Make sure the xHCI's reset method marks the xHCI host controller's primary
roothub as the USB 2.0 roothub. This ensures that the high speed bus will
be processed first when the PCI device is resumed, and any USB 3.0 devices
that have migrated over to high speed will migrate back after being reset.
This ensures that USB persist works with these odd devices.
The reset method will also mark the xHCI USB2 roothub as having an
integrated TT. Like EHCI host controllers with a "rate matching hub" the
xHCI USB 2.0 roothub doesn't have an OHCI or UHCI companion controller.
It doesn't really have a TT, but we'll lie and say it has an integrated
TT. We need to do this because the USB core will reject LS/FS devices
under a HS hub without a TT.
Other details:
-------------
The roothub emulation code is changed to return the correct number of
ports for the two roothubs. For the USB 3.0 roothub, it only reports the
USB 3.0 ports. For the USB 2.0 roothub, it reports all the LS/FS/HS
ports. The code to disable a port now checks the speed of the roothub,
and refuses to disable SuperSpeed ports under the USB 3.0 roothub.
The code for initializing a new device context must be changed to set the
proper roothub port number. Since we've split the xHCI host into two
roothubs, we can't just use the port number in the ancestor hub. Instead,
we loop through the array of hardware port status register speeds and find
the Nth port with a similar speed.
The port status change event handler is updated to figure out whether the
port that reported the change is a USB 3.0 port, or a non-SuperSpeed port.
Once it figures out the port speed, it kicks the proper roothub.
The function to find a slot ID based on the port index is updated to take
into account that the two roothubs will have over-lapping port indexes.
It checks that the virtual device with a matching port index is the same
speed as the passed in roothub.
There's also changes to the driver initialization and shutdown paths:
1. Make sure that the xhci_hcd pointer is shared across the two
usb_hcd structures. The xhci_hcd pointer is allocated and the
registers are mapped in when xhci_pci_setup() is called with the
primary HCD. When xhci_pci_setup() is called with the non-primary
HCD, the xhci_hcd pointer is stored.
2. Make sure to set the sg_tablesize for both usb_hcd structures. Set
the PCI DMA mask for the non-primary HCD to allow for 64-bit or 32-bit
DMA. (The PCI DMA mask is set from the primary HCD further down in
the xhci_pci_setup() function.)
3. Ensure that the host controller doesn't start kicking khubd in
response to port status changes before both usb_hcd structures are
registered. xhci_run() only starts the xHC running once it has been
called with the non-primary roothub. Similarly, the xhci_stop()
function only halts the host controller when it is called with the
non-primary HCD. Then on the second call, it resets and cleans up the
MSI-X irqs.
Signed-off-by: Sarah Sharp <sarah.a.sharp@linux.intel.com>
There are several variables in the xhci_hcd structure that are related to
bus suspend and resume state. There are a couple different port status
arrays that are accessed by port index. Move those variables into a
separate structure, xhci_bus_state. Stash that structure in xhci_hcd.
When we have two roothhubs that can be suspended and resumed separately,
we can have two xhci_bus_states, and index into the port arrays in each
structure with the fake roothub port index (not the real hardware port
index).
Signed-off-by: Sarah Sharp <sarah.a.sharp@linux.intel.com>
Make sure to call into the USB core's link, unlink, and giveback URB
functions with the usb_hcd pointer found by using urb->dev->bus. This
will avoid confusion later, when the xHCI driver will deal with URBs from
two separate buses (the USB 3.0 roothub and the faked USB 2.0 roothub).
Assume xhci_urb_dequeue() will be called with the proper usb_hcd.
Signed-off-by: Sarah Sharp <sarah.a.sharp@linux.intel.com>
The USB core will set hcd->state to HC_STATE_RUNNING before calling
xhci_run, so there's no point in setting it twice. The USB core also
doesn't pay attention to HC_STATE_RUNNING on the resume path anymore; it
uses HCD_RH_RUNNING(), which looks at hcd->flags & (1U <<
HCD_FLAG_RH_RUNNING. Therefore, it's safe to remove the state set in
xhci_bus_resume().
Signed-off-by: Sarah Sharp <sarah.a.sharp@linux.intel.com>
The xHCI driver doesn't ever test hcd->state for HC_STATE_HALT. The USB
core recently stopped using it internally, so there's no point in setting
it in the driver. We still need to set HC_STATE_RUNNING in order to make
it past the USB core's hcd->state check in register_roothub().
Signed-off-by: Sarah Sharp <sarah.a.sharp@linux.intel.com>
xHCI 1.0 spec specifies the xHC shall halt within 16ms after software clears
Run/Stop bit. In xHCI 0.96 spec the time limit is 16 microframes (2ms), it's
too short and often cause dmesg shows "Host controller not halted, aborting
reset." message when rmmod xhci-hcd.
Modify the time limit to comply with xHCI 1.0 specification and prevents the
warning message showing when remove xhci-hcd.
Signed-off-by: Andiry Xu <andiry.xu@amd.com>
Signed-off-by: Sarah Sharp <sarah.a.sharp@linux.intel.com>
Set hcd->state = HC_STATE_SUSPENDED if there is a power loss during system
resume or the system is hibernated, otherwise leave it be. The variable
old_state is redundant and made an unreachable code path, so remove it.
Signed-off-by: Andiry Xu <andiry.xu@amd.com>
Signed-off-by: Sarah Sharp <sarah.a.sharp@linux.intel.com>
The test of placing a number of command no-ops on the command ring and
counting the number of no-op events that were generated was only used
during the initial xHCI driver bring up. This test is no longer used, so
delete it.
Signed-off-by: Sarah Sharp <sarah.a.sharp@linux.intel.com>
Functions that are not used outsde of the module they are defined
should be marked as static.
Signed-off-by: Dmitry Torokhov <dtor@vmware.com>
Signed-off-by: Sarah Sharp <sarah.a.sharp@linux.intel.com>
xhci->ir_set points to __iomem region, but xhci_print_ir_set accepts
plain struct xhci_intr_reg * causing multiple sparse warning at call
sites and inside the fucntion when we try to read that memory.
Instead of adding __iomem qualifier to the argument let's rework the
function so it itself gets needed register set from xhci and prints
it.
Signed-off-by: Dmitry Torokhov <dtor@vmware.com>
Signed-off-by: Sarah Sharp <sarah.a.sharp@linux.intel.com>
When xhci_discover_or_reset_device() is called after a host controller
power loss, the virtual device may need to be reallocated. Make sure
xhci_alloc_dev() uses GFP_NOIO. This avoid causing a deadlock by allowing
the kernel to flush pending I/O while reallocating memory for a virtual
device for a USB mass storage device that's holding the backing store for
dirty memory buffers.
This patch should be queued for the 2.6.37 stable tree.
Signed-off-by: Sarah Sharp <sarah.a.sharp@linux.intel.com>
Cc: stable@kernel.org
Synchronize the interrupts instead of free them in xhci_suspend(). This will
prevent a double free when the host is suspended and then the card removed.
Set the flag hcd->msix_enabled when using MSI-X, and check the flag in
suspend_common(). MSI-X synchronization will be handled by xhci_suspend(),
and MSI/INTx will be synchronized in suspend_common().
This patch should be queued for the 2.6.37 stable tree.
Reported-by: Matthew Garrett <mjg@redhat.com>
Signed-off-by: Andiry Xu <andiry.xu@amd.com>
Signed-off-by: Sarah Sharp <sarah.a.sharp@linux.intel.com>
Cc: stable@kernel.org
We have been having problems with the USB-IF Gold Tree tests when plugging
and unplugging devices from the tree. I have seen that the reset-device
and configure-endpoint commands, which are invoked from
xhci_discover_or_reset_device() and xhci_configure_endpoint(), will sometimes
time out.
After much debugging, I determined that the commands themselves do not actually
time out, but rather their completion events do not get delivered to the right
place.
This happens when the command ring has just wrapped around, and it's enqueue
pointer is left pointing to the link TRB. xhci_discover_or_reset_device() and
xhci_configure_endpoint() use the enqueue pointer directly as their command
TRB pointer, without checking whether it's pointing to the link TRB.
When the completion event arrives, if the command TRB is pointing to the link
TRB, the check against the command ring dequeue pointer in
handle_cmd_in_cmd_wait_list() fails, so the completion inside the command does
not get signaled.
The patch below fixes the timeout problem for me.
This should be queued for the 2.6.35 and 2.6.36 stable trees.
Signed-off-by: Paul Zimmerman <paulz@synopsys.com>
Signed-off-by: Sarah Sharp <sarah.a.sharp@linux.intel.com>
Cc: stable@kernel.org
Andiry's xHCI bus suspend patch introduced the possibly of a host
controller replaying old commands on the command ring, if the host
successfully restores the registers after a resume.
After a resume from suspend, the xHCI driver must restore the registers,
including the command ring pointer. I had suggested that Andiry set the
command ring pointer to the current command ring dequeue pointer, so that
the driver wouldn't have to zero the command ring.
Unfortunately, setting the command ring pointer to the current dequeue
pointer won't work because the register assumes the pointer is 64-byte
aligned, and TRBs on the command ring are 16-byte aligned. The lower
seven bits will always be masked off, leading to the written pointer being
up to 3 TRBs behind the intended pointer.
Here's a log excerpt. On init, the xHCI driver places a vendor-specific
command on the command ring:
[ 215.750958] xhci_hcd 0000:01:00.0: Vendor specific event TRB type = 48
[ 215.750960] xhci_hcd 0000:01:00.0: NEC firmware version 30.25
[ 215.750962] xhci_hcd 0000:01:00.0: Command ring deq = 0x3781e010 (DMA)
When we resume, the command ring dequeue pointer to be written should have
been 0x3781e010. Instead, it's 0x3781e000:
[ 235.557846] xhci_hcd 0000:01:00.0: // Setting command ring address to 0x3781e001
[ 235.557848] xhci_hcd 0000:01:00.0: `MEM_WRITE_DWORD(3'b000, 64'hffffc900100bc038, 64'h3781e001, 4'hf);
[ 235.557850] xhci_hcd 0000:01:00.0: `MEM_WRITE_DWORD(3'b000, 32'hffffc900100bc020, 32'h204, 4'hf);
[ 235.557866] usb usb9: root hub lost power or was reset
(I can't see the results of this bug because the xHCI restore always fails
on this box, and the xHCI driver re-allocates everything.)
The fix is to zero the command ring and put the software and hardware
enqueue and dequeue pointer back to the beginning of the ring. We do this
before the system suspends, to be paranoid and prevent the BIOS from
starting the host without clearing the command ring pointer, which might
cause the host to muck with stale memory. (The pointer isn't required to
be in the suspend power well, but it could be.) The command ring pointer
is set again after the host resumes.
Signed-off-by: Sarah Sharp <sarah.a.sharp@linux.intel.com>
Tested-by: Andiry Xu <andiry.xu@amd.com>
Jiri Slaby reports spinlock is held while calling kmalloc(GFP_KERNEL)
and request_irq() in xhci_resume().
Release the spinlock when setup interrupt.
Reported-by: Jiri Slaby <jirislaby@gmail.com>
Signed-off-by: Andiry Xu <andiry.xu@amd.com>
Signed-off-by: Sarah Sharp <sarah.a.sharp@linux.intel.com>
This reverts commit ef821ae70f.
The correct thing to do is to drop the spinlock, not change
the GFP flag here.
Thanks to Sarah for pointing out I shouldn't have taken this patch in
the first place.
Cc: Sarah Sharp <sarah.a.sharp@linux.intel.com>
Cc: David Sterba <dsterba@suse.cz>
Signed-off-by: Greg Kroah-Hartman <gregkh@suse.de>
coccinelle check scripts/coccinelle/locks/call_kern.cocci found that
in drivers/usb/host/xhci.c an allocation with GFP_KERNEL is done
with locks held:
xhci_resume
spin_lock_irq(xhci->lock)
xhci_setup_msix
kmalloc(GFP_KERNEL)
Change it to GFP_ATOMIC.
Signed-off-by: David Sterba <dsterba@suse.cz>
CC: Sarah Sharp <sarah.a.sharp@linux.intel.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@suse.de>
Fix this error when CONFIG_PM is not enabled:
drivers/usb/host/xhci.c:675: error: implicit declaration of function 'usb_root_hub_lost_power'
Wrap xhci_suspend() and xhci_resume() into an ifdef CONFIG_PM, along with
the functions that only they call -- xhci_save_registers() and
xhci_restore_registers().
Reported-by: Randy Dunlap <randy.dunlap@oracle.com>
Signed-off-by: Sarah Sharp <sarah.a.sharp@linux.intel.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@suse.de>
This patch implements the PCI suspend/resume.
Please refer to xHCI spec for doing the suspend/resume operation.
For S3, CSS/SRS in USBCMD is used to save/restore the internal state.
However, an error maybe occurs while restoring the internal state.
In this case, it means that HC internal state is wrong and HC will be
re-initialized.
Signed-off-by: Libin Yang <libin.yang@amd.com>
Signed-off-by: Dong Nguyen <dong.nguyen@amd.com>
Signed-off-by: Andiry Xu <andiry.xu@amd.com>
Signed-off-by: Sarah Sharp <sarah.a.sharp@linux.intel.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@suse.de>
Add software trigger USB device suspend resume function hook.
Do port suspend & resume in terms of xHCI spec.
Port Suspend:
Stop all endpoints via Stop Endpoint Command with Suspend (SP) flag set.
Place individual ports into suspend mode by writing '3' for Port Link State
(PLS) field into PORTSC register. This can only be done when the port is in
Enabled state. When writing, the Port Link State Write Strobe (LWS) bit shall
be set to '1'.
Allocate an xhci_command and stash it in xhci_virt_device to wait completion for
the last Stop Endpoint Command. Use the Suspend bit in TRB to indicate the Stop
Endpoint Command is for port suspend. Based on Sarah's suggestion.
Port Resume:
Write '0' in PLS field, device will transition to running state.
Ring an endpoints' doorbell to restart it.
Ref: USB device remote wake need another patch to implement. For details of
how USB subsystem do power management, please see:
Documentation/usb/power-management.txt
Signed-off-by: Crane Cai <crane.cai@amd.com>
Signed-off-by: Libin Yang <libin.yang@amd.com>
Signed-off-by: Andiry Xu <andiry.xu@amd.com>
Signed-off-by: Sarah Sharp <sarah.a.sharp@linux.intel.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@suse.de>
xHCI driver uses hardware assigned device address. This may cause device
address conflict in certain cases.
Use kernel assigned address for devices under xHCI. Store the xHC assigned
address locally in xHCI driver.
Signed-off-by: Andiry Xu <andiry.xu@amd.com>
Signed-off-by: Sarah Sharp <sarah.a.sharp@linux.intel.com>
Rename xhci_reset_device() to xhci_discover_or_reset_device().
If xhci_discover_or_reset_device() is called to reset a device which does
not exist or does not match the udev, it calls xhci_alloc_dev() to
re-allocate the device.
This would prevent the reset device failure, possibly due to the xHC restore
error during S3/S4 resume.
Signed-off-by: Andiry Xu <andiry.xu@amd.com>
Signed-off-by: Sarah Sharp <sarah.a.sharp@linux.intel.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@suse.de>
Add a pointer to udev in struct xhci_virt_device. When allocate a new
virt_device, make the pointer point to the corresponding udev.
Modify xhci_check_args(), check if virt_dev->udev matches the target udev,
to make sure command is issued to the right device.
Signed-off-by: Andiry Xu <andiry.xu@amd.com>
Signed-off-by: Sarah Sharp <sarah.a.sharp@linux.intel.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@suse.de>
Most of the work for interrupt handling is done in xhci-ring.c, so it makes
sense to move the functions that are first called when an interrupt happens
(xhci_irq() or xhci_msi_irq()) into xhci-ring.c, so that the compiler can better
optimize them.
Shorten some lines to make it pass checkpatch.
Signed-off-by: Sarah Sharp <sarah.a.sharp@linux.intel.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@suse.de>
I've been using perf to measure the top symbols while transferring 1GB of data
on a USB 3.0 drive with dd. This is using the raw disk with /dev/sdb, with a
block size of 1K.
During performance testing, the top symbol was xhci_triad_to_transfer_ring(), a
function that should return immediately if streams are not enabled for an
endpoint. It turned out that the functions to find the endpoint ring was
defined in xhci-mem.c and used in xhci-ring.c and xhci-hcd.c. I moved a copy of
xhci_triad_to_transfer_ring() and xhci_urb_to_transfer_ring() into xhci-ring.c
and declared them static. I also made a static version of
xhci_urb_to_transfer_ring() in xhci.c.
This improved throughput on a 1GB read of the raw disk with dd from
186MB/s to 195MB/s, and perf reported sampling the xhci_triad_to_transfer_ring()
0.06% of the time, rather than 9.26% of the time.
Signed-off-by: Sarah Sharp <sarah.a.sharp@linux.intel.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@suse.de>
Add urb_priv data structure to xHCI driver. This structure allows multiple
xhci TDs to be linked to one urb, which is essential for isochronous
transfer. For non-isochronous urb, only one TD is needed for one urb;
for isochronous urb, the TD number for the urb is equal to
urb->number_of_packets.
The length field of urb_priv indicates the number of TDs in the urb.
The td_cnt field indicates the number of TDs already processed by xHC.
When td_cnt matches length, the urb can be given back to usbcore.
When an urb is dequeued or cancelled, add all the unprocessed TDs to the
endpoint's cancelled_td_list. When process a cancelled TD, increase
td_cnt field. When td_cnt equals urb_priv->length, giveback the
cancelled urb.
Signed-off-by: Andiry Xu <andiry.xu@amd.com>
Signed-off-by: Sarah Sharp <sarah.a.sharp@linux.intel.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@suse.de>
