[ Upstream commit 7ce98fb6c53d2311b3e9faae90b1a9c1a96534db ]
SBE fifo operations should be allowed while the SBE is in any of the
"IPL" states. Operations should succeed in this state.
Fixes: 9f4a8a2d7f fsi/sbefifo: Add driver for the SBE FIFO
Reviewed-by: Joel Stanley <joel@jms.id.au>
Tested-by: Alistair Popple <alistair@popple.id.au>
Signed-off-by: Eddie James <eajames@linux.ibm.com>
Link: https://lore.kernel.org/r/1561575415-3282-1-git-send-email-eajames@linux.ibm.com
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Signed-off-by: Sasha Levin <sashal@kernel.org>
[ Upstream commit 371975b0b07520c85098652d561639837a60a905 ]
Change d1dcd67825 re-worked the struct fsi_slave initialisation in
fsi_slave_init, but introduced a few inconsitencies: the slave->dev is
now registered through cdev_device_add, but we may kfree() the device
out from underneath the cdev registration. We may also leave an IDA
allocated.
This change fixes the error paths, so that we kfree() only before the
device is registered with the core code. We also move the smode write to
before we start creating proper devices, as it's the most likely to
fail. We also remove the IDA-allocated minor on error, and properly
clean up the of_node.
Fixes: d1dcd67825 ("fsi: Add cfam char devices")
Reported-by: Lei YU <mine260309@gmail.com>
Tested-by: John Wang <wangzqbj@inspur.com>
Signed-off-by: Jeremy Kerr <jk@ozlabs.org>
Signed-off-by: Joel Stanley <joel@jms.id.au>
Signed-off-by: Sasha Levin <sashal@kernel.org>
[ Upstream commit 9f4c2b516b4f031e3cd0e45957f4150b3c1a083d ]
Subtracting the offset delta from four-byte alignment lead to wrapping
of the requested length where `count` is less than `off`. Generalise the
length handling to enable and optimise aligned access sizes for all
offset and size combinations. The new formula produces the following
results for given offset and count values:
offset count | length
--------------+-------
0 1 | 1
0 2 | 2
0 3 | 2
0 4 | 4
0 5 | 4
1 1 | 1
1 2 | 1
1 3 | 1
1 4 | 1
1 5 | 1
2 1 | 1
2 2 | 2
2 3 | 2
2 4 | 2
2 5 | 2
3 1 | 1
3 2 | 1
3 3 | 1
3 4 | 1
3 5 | 1
We might need something like this for the cfam chardevs as well, for
example we don't currently implement any alignment restrictions /
handling in the hardware master driver.
Signed-off-by: Andrew Jeffery <andrew@aj.id.au>
Signed-off-by: Joel Stanley <joel@jms.id.au>
Link: https://lore.kernel.org/r/20191108051945.7109-6-joel@jms.id.au
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Signed-off-by: Sasha Levin <sashal@kernel.org>
commit 8919dfcb31161fae7d607bbef5247e5e82fd6457 upstream.
The scom driver currently fails out of operations if certain system
errors are flagged in the status register; system checkstop, special
attention, or recoverable error. These errors won't impact the ability
of the scom engine to perform operations, so the driver should continue
under these conditions.
Also, don't do a PIB reset for these conditions, since it won't help.
Fixes: 6b293258cd ("fsi: scom: Major overhaul")
Signed-off-by: Eddie James <eajames@linux.ibm.com>
Cc: stable <stable@vger.kernel.org>
Acked-by: Jeremy Kerr <jk@ozlabs.org>
Acked-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Signed-off-by: Joel Stanley <joel@jms.id.au>
Link: https://lore.kernel.org/r/20190827041249.13381-1-jk@ozlabs.org
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Otherwise cronus putmem fails istep and BML fails to upload skiboot
To do that, we still use our one-page command buffer for small commands
for speed, and for anything bigger, with a limit of 1MB plus a page,
we vmalloc a temporary buffer.
The limit was chosen because Cronus will break up any data transfer
into 1M chunks (the extra page is for the command header).
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Reviewed-by: Andrew Jeffery <andrew@aj.id.au>
The chardev conversion forgot to copy the fsi_dev,
silly mistake, compounded by a testing mistake on
my side, this specific driver wasn't being tested
properly.
Fixes: d8f4587655 "fsi: scom: Convert to use the new chardev"
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
The bus scanning process isn't terribly good at parallel attempts
at rescanning the same bus. Let's have a per-master mutex protecting
the scanning process.
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
This aims to deprecate the "raw" sysfs file used for directly
accessing the CFAM and instead use a char device like the
other sub drivers.
Since it reworks the slave creation code and adds a cfam device
type, we also use the opportunity to convert the attributes
to attribute groups and add a couple more.
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
This converts FSI scom to use the new fsi-core controlled
chardev allocator and use a real cdev instead of a miscdev.
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
This converts FSI sbefifo to use the new fsi-core controlled
chardev allocator and use a real cdev instead of a miscdev.
One side effect is to fix the object lifetime by removing
the use of devm_kzalloc() for something that contains kobjects,
and using proper reference counting.
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
The various FSI devices (sbefifo, occ, scom, more to come)
currently use misc devices.
This is problematic as the minor device space for misc is
limited and there can be a lot of them. Also it limits our
ability to move them to a dedicated /dev/fsi directory or
to be smart about device naming and numbering.
It also means we have IDAs on every single of these drivers
This creates a common fsi "device_type" for the optional
/dev/fsi grouping and a dev_t allocator for all FSI devices.
"Legacy" devices get to use a backward compatible numbering
scheme (as long as chip id <16 and there's only one copy
of a given unit type per chip).
A single major number and a single IDA are shared for all
FSI devices.
This doesn't convert the FSI device drivers to use the new
scheme yet, they will be converted individually.
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
s390 defines a global dump_trace() symbol. Rename ours to
dump_ucode_trace() to avoid a collision in build tests.
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
In case memory resources for *fw* were allocated, release them
before return.
Addresses-Coverity-ID: 1472044 ("Resource leak")
Fixes: 6a794a27da ("fsi: master-ast-cf: Add new FSI master using Aspeed ColdFire")
Signed-off-by: Gustavo A. R. Silva <gustavo@embeddedor.com>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Then reading the RTAG/RCRC "registers" from the coprocessor after
a command is complete, mask out the top bits, only keep the relevant
bits. Microcode v5 will leave garbage in those top bits as a
result of a performance optimization.
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
---
Some of the exit path missed the unlock. Move the mutex to
an outer function to avoid the problem completely
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
The Aspeed AST2x00 can contain a ColdFire v1 coprocessor which
is currently unused on OpenPower systems.
This adds an alternative to the fsi-master-gpio driver that
uses that coprocessor instead of bit banging from the ARM
core itself. The end result is about 4 times faster.
The firmware for the coprocessor and its source code can be
found at https://github.com/ozbenh/cf-fsi and is system specific.
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
This moves the definitions for various protocol details
(message & response codes, delays etc...) out of
fsi-master-gpio.c to fsi-master.h in order to share them
with other master implementations.
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Reviewed-by: Joel Stanley <joel@jms.id.au>
The embedded struct device needs a release function to be
able to successfully remove the driver.
We remove the devm_gpiod_put() as they are unnecessary
(the resources will be released automatically) and because
fsi_master_unregister() will cause the master structure to
be freed.
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Reviewed-by: Joel Stanley <joel@jms.id.au>
In the error path of fsi_master_register(), we currently
use device_unregister(). This will cause the last reference
to the structure to be dropped, thus freeing the enclosing
structure, which isn't what the callers want.
Use device_del() instead so that we return to the caller
with a refcount of 1. The caller can then assume that it
must use put_device() after a call to fsi_master_register()
regardless of whether the latter suceeded or failed.
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Some definitions are generic to the FSI protocol or any
give master implementation. Rename them to remove the
"GPIO" prefix in preparation for moving them to a common
header.
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Reviewed-by: Joel Stanley <joel@jms.id.au>
# Conflicts:
# drivers/fsi/fsi-master-gpio.c
This adds a few more tracepoints that have proven useful when
debugging issues with the FSI bus.
This also makes echo_delay() use clock_zeros() instead of
open-code it in order to share the tracepoint.
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Reviewed-by: Joel Stanley <joel@jms.id.au>
What the driver called "FSI_GPIO_PRIME_SLAVE_CLOCKS" is what
the FSI spec calls tSendDelay and should be 16 clocks by
default.
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Reviewed-by: Joel Stanley <joel@jms.id.au>
Those values control the amount of "dummy" clocks between commands and
between a command and its response.
This adds a way to configure them from sysfs (to be later extended to
defaults in the device-tree). The default remains 16 (the HW default).
This is only supported if the backend supports the new link_config()
callback to configure the generation of those delays.
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Reviewed-by: Joel Stanley <joel@jms.id.au>
---
Move fsi_slave_set_smode() and its helpers to before it's
first user and remove the corresponding forward declaration.
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Reviewed-by: Joel Stanley <joel@jms.id.au>
"dev" is dereferences before it's checked.
Reported-by: Dan Carpenter <dan.carpenter@oracle.com>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
The driver calls of_platform_device_create() which is only available
if OF_ADDRESS is enabled. When building sparc64 images, this results
in
ERROR: "of_platform_device_create" [drivers/fsi/fsi-sbefifo.ko] undefined!
Fixes: 9f4a8a2d7f ("fsi/sbefifo: Add driver for the SBE FIFO")
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Signed-off-by: Guenter Roeck <linux@roeck-us.net>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
There was no unlock of the FFDC mutex.
Fixes: 9f4a8a2d7f ("fsi/sbefifo: Add driver for the SBE FIFO")
Signed-off-by: Eddie James <eajames@linux.vnet.ibm.com>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
This was too hard to split ... this adds a number of features
to the SCOM user interface:
- Support for indirect SCOMs
- read()/write() interface now handle errors and retries
- New ioctl() "raw" interface for use by debuggers
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Reviewed-by: Eddie James <eajames@linux.vnet.ibm.com>
Reviewed-by: Alistair Popple <alistair@popple.id.au>
Add a few more register and bit definitions, also define and use
SCOM_READ_CMD (which is 0 but it makes the code clearer)
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Reviewed-by: Eddie James <eajames@linux.vnet.ibm.com>
Use the proper annotated type __be32 and fixup the
accessor used for get_scom()
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Reviewed-by: Eddie James <eajames@linux.vnet.ibm.com>
Otherwise, multiple clients can open the driver and attempt
to access the PIB at the same time, thus clobbering each other
in the process.
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Reviewed-by: Eddie James <eajames@linux.vnet.ibm.com>
fsi-master-hub.c:128:13: warning: incorrect type in assignment (different base types)
fsi-master-hub.c:128:13: expected unsigned int [unsigned] [usertype] cmd
fsi-master-hub.c:128:13: got restricted __be32 [usertype] <noident>
fsi-master-hub.c:208:13: warning: incorrect type in assignment (different base types)
fsi-master-hub.c:208:13: expected restricted __be32 [addressable] [assigned] [usertype] reg
fsi-master-hub.c:208:13: got int
Signed-off-by: Joel Stanley <joel@jms.id.au>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
fsi-sbefifo.c:547:58: warning: incorrect type in argument 2 (different base types)
fsi-sbefifo.c:547:58: expected restricted __be32 [usertype] *word
fsi-sbefifo.c:547:58: got unsigned int *<noident>
fsi-sbefifo.c:635:16: warning: incorrect type in assignment (different base types)
fsi-sbefifo.c:635:16: expected unsigned int [unsigned] <noident>
fsi-sbefifo.c:635:16: got restricted __be32 [usertype] <noident>
fsi-sbefifo.c:636:16: warning: incorrect type in assignment (different base types)
fsi-sbefifo.c:636:16: expected unsigned int [unsigned] <noident>
fsi-sbefifo.c:636:16: got restricted __be32 [usertype] <noident>
Signed-off-by: Joel Stanley <joel@jms.id.au>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
This driver provides an in-kernel and a user API for accessing
the command FIFO of the SBE (Self Boot Engine) of the POWER9
processor, via the FSI bus.
It provides an in-kernel interface to submit command and receive
responses, along with a helper to locate and analyse the response
status block. It's a simple synchronous submit() type API.
The user interface uses the write/read interface that an earlier
version of this driver already provided, however it has some
specific limitations in order to keep the driver simple and
avoid using up a lot of kernel memory:
- The user should perform a single write() with the command and
a single read() to get the response (with a buffer big enough
to hold the entire response).
- On a write() the command is simply "stored" into a kernel buffer,
it is submitted as one operation on the subsequent read(). This
allows to have the code write directly from the FIFO into the user
buffer and avoid hogging the SBE between the write() and read()
syscall as it's critical that the SBE be freed asap to respond
to the host. An extra write() will simply replace the previously
written command.
- A write of a single 4 bytes containing the value 0x52534554
in big endian will trigger a reset request. No read is necessary,
the write() call will return when the reset has been acknowledged
or times out.
- The command is limited to 4K bytes.
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Tested-by: Joel Stanley <joel@jms.id.au>
---
The PIB reset causes problems for the running P9 chip. The reset
shouldn't be performed by this driver.
Signed-off-by: Eddie James <eajames@linux.vnet.ibm.com>
Reviewed-by: Christopher Bostic <cbostic@linux.vnet.ibm.com>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Tested-by: Joel Stanley <joel@jms.id.au>
We currently use a spinlock (bit_lock) around operations that clock bits
out of the FSI bus, and a mutex to protect against simultaneous access
to the master.
This means that bit_lock isn't needed for mutual exlusion, only to
prevent timing issues when clocking bits out.
To reflect this, this change converts bit_lock to just the
local_irq_save/restore operation.
Signed-off-by: Jeremy Kerr <jk@ozlabs.org>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Tested-by: Joel Stanley <joel@jms.id.au>
Remove calls to the empty and useless fsi_master_gpio_error()
function, and report CRC errors as "FSI_ERR_NO_SLAVE" when
reading an all 1's response.
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Tested-by: Joel Stanley <joel@jms.id.au>
The FSI protocol defines two modes of recovery from CRC errors,
this implements both:
- If the device returns an ECRC (it detected a CRC error in the
command), then we simply issue the command again.
- If the master detects a CRC error in the response, we send
an E_POLL command which requests a resend of the response
without actually re-executing the command (which could otherwise
have unwanted side effects such as dequeuing a FIFO twice).
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Reviewed-by: Christopher Bostic <cbostic@linux.vnet.ibm.com>
Tested-by: Joel Stanley <joel@jms.id.au>
---
Note: This was actually tested by removing some of my fixes, thus
causing us to hit occasional CRC errors during high LPC activity.
FSI CFAMs support shorter commands that use a relative (or same) address
as the last. This change introduces a last_addr to the master state, and
uses it for subsequent reads/writes, and performs relative addressing
when a subsequent read/write is in range.
Signed-off-by: Jeremy Kerr <jk@ozlabs.org>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Reviewed-by: Christopher Bostic <cbostic@linux.vnet.ibm.com>
Tested-by: Joel Stanley <joel@jms.id.au>
For implementing relative addressing mode, we'll need to build a command
that is coherent with CFAM state. To do that, include the
build_command_* functions in the locked section of read/write/term.
Signed-off-by: Jeremy Kerr <jk@ozlabs.org>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Tested-by: Joel Stanley <joel@jms.id.au>
Most SoC GPIO implementations, including the Aspeed one, have
synchronizers on the GPIO inputs. This means that the value
read from a GPIO is a couple of clocks old, from whatever clock
source feeds those synchronizers.
In practice, this means that in no-delay mode, we are using a
value that can potentially be a bit too old and too close to
the clock edge establishing the data on the other side of the link.
The voltage converters we use on some systems make this worse
and sensitive to things like voltage fluctuations etc... This is,
we believe, the cause of occasional CRC errors encountered during
heavy activity on the LPC bus.
This is fixed by introducing a dummy GPIO read before the actual
data read. It slows down SBEFIFO by about 15% (less than any delay
primitive) and the end result is so far solid.
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Reviewed-by: Christopher Bostic <cbostic@linux.vnet.ibm.com>
Tested-by: Joel Stanley <joel@jms.id.au>
FSI_GPIO_DPOLL_CLOCKS is the number of clocks before sending
a DPOLL command after receiving a BUSY status. It should be
at least tSendDelay (16 clocks).
According to comments in the code, it needs to also be at least
21 clocks due to HW issues.
It's currently 100 clocks which impacts performances negatively
in some cases. Reduces it in half to 50 clocks which seems to
still be solid.
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Reviewed-by: Christopher Bostic <cbostic@linux.vnet.ibm.com>
Tested-by: Joel Stanley <joel@jms.id.au>
