After commit b7d2b38, the system suffers from -Warray-bounds warning with -Wall:
chip/mpfs_opensbi.c: In function 'mpfs_hart_to_scratch':
chip/mpfs_opensbi.c:251:26: warning: array subscript hartid is outside array bounds of 'sbi_scratch_holder_t[0]' {aka 'struct sbi_scratch_holder_s[]'} [-Warray-bounds]
251 | return (unsigned long)(&g_scratches[hartid].scratch);
Fix it by reverting back to what is was earlier. g_scratches shouldn't be in the bss
region that would be zeroed out.
Signed-off-by: Eero Nurkkala <eero.nurkkala@offcode.fi>
After the commit b8b541f, the system would not start up properly
with certain coldboot conditions. For example, if the OpenSBI picked
hart4 as the coldboot hart in preference to hart3, the system would
get stuck due to stack corruption. OpenSBI uses a lottery mechanism
to pick the coldboot hart.
Also fix g_scratches area in such a manner than it will not get
initialized to zero. If several harts initialize the area to zero, there's
danger the stack pointer gets wiped out.
Now any coldboot hart works.
Signed-off-by: Eero Nurkkala <eero.nurkkala@offcode.fi>
Currently the IHC (Inter Hart Communication) depends on OpenAMP and
rptun. However, the bootloader portion of the IHC doesn't need
either of them. Now they are wasting a lot of bootloader space.
Reorganize the bootloader portion into a separate file 'mpfs_ihc_sbi.c'.
This file contains the OpenSBI vendor extensions, or the only required
functionalities for the bootloader. On the other hand, 'mpfs_ihc.c'
contains the non-bootloader code.
This patch also makes it possible to utilize 2 RPMSG channels. This
has been tested so that 2 separate NuttXs on harts 1 and 2 communicate
with Linux kernel that runs on harts 3 and 4.
New configuration files are added as well:
- rpmsg-ch1: sample config for RPMSG
- rpmsg-ch2: sample config for another RPMSG channel
- rpmsg-sbi: sample bootloader config for RPMSG/OpenSBI
Signed-off-by: Eero Nurkkala <eero.nurkkala@offcode.fi>
This reverts commit ea9144bda8.
The commit made Icicle MPFS DDR useless. Revert the change for now.
Signed-off-by: Eero Nurkkala <eero.nurkkala@offcode.fi>
Make a separate config flag for enabling L2 cache. This is on by
default when compiling a standalone/bootloader configuration, but
can also be disabled for special cases, such as memory testing
Signed-off-by: Jukka Laitinen <jukkax@ssrc.tii.ae>
Run-time check for L3 page table size, to ensure it is large enough
to map all of the kernel memory.
NOTE: The check has to be run-time, as KFLASH_SIZE/KSRAM_SIZE are really
linker relocation symbols, and thus cannot be utilized compile-time.
L3 table maps 2MB of memory, this brings an implicit requirement for
any L3 region to be aligned to 2MB. This commit adds build time sanity
checks to ensure this requirement is met.
For other SvXX architectures the boundary requirement (might be) is
different.
This allows properly using 10/100Mbps also with 1G phy. Some gigabit PHYs
come out of reset with 1G advertisement enabled, causing other devices to
set up link with 1G. If, after this, the link is set to 10/100 on the mpfs,
the link won't work.
Signed-off-by: Jukka Laitinen <jukkax@ssrc.tii.ae>
Simplify the clock mode from the board.h -files. Also make the
SD clock definable as well.
Co-authored-by: Petro Karashchenko <petro.karashchenko@gmail.com>
Signed-off-by: Eero Nurkkala <eero.nurkkala@offcode.fi>
Some related products, such as Aries m100pfs, don't support eMMC
speeds up to 200MHz. Thus, provide option to select slower clock.
This has only to do with the clocking, no CMD6 is sent to select
high speed modes.
Signed-off-by: Eero Nurkkala <eero.nurkkala@offcode.fi>
The L3 mapping function was just way too simplistic. Depending on memory
configuration it either works or not.
Noticed that with icicle:pnsh the software crashes due to instruction
page fault, reason is the map_region() implementation that does not
work for regions that are not aligned to 2MB (the L2 page size).
Implemented an extremely simplistic page table allocator for the L3
references, that should once and for all get rid of the L3 mapping issue.
NOTE: gran_alloc() cannot be used at this point, it is too early for it.
If the system starts up using the SD-card, and later wants
to switch to use the eMMC, it's not going to happen. SD-card
will be kept selected within the FPGA making the transition
fail.
It's been possible to switch from eMMC to SD, but now the
transition is possible for both directions.
Signed-off-by: Eero Nurkkala <eero.nurkkala@offcode.fi>
Issuing the following commands doesn't succeed:
- conn
- disconn
- conn
This USB driver doesn't even disconnect without this patch.
If the USB driver has been started from another hart, closing it
will not disable the PLIC interrupt. This means it's possible many
harts get the USB interrupt and thus make the USB look very unstable.
Fix these problems by disconnecting the USB via the USB_POWER
register at shutdown and disabling the interrupt at shutdown. Also
clear the software internals via the mpfs_sw_setup() for the conn /
disconect loop to succeed.
Signed-off-by: Eero Nurkkala <eero.nurkkala@offcode.fi>
The underlying hardware supports 9 endpoints:
- EP0
- 4x IN EPs
- 4x OUT EPs
Currently the driver assumes every EP number is unique. This limits
the amount of EPs to 1 + 4 = 5. Utilize the EPs in such a manner
that all may be used.
Also fix a few error handling related bugs. Update the composite
driver to match the current situation as well.
Signed-off-by: Eero Nurkkala <eero.nurkkala@offcode.fi>
mpfs_write_tx_fifo() gets stuck in the following case:
- CDCACM is used
- ttyACM0 is opened and then closed from the remote end,
such as Linux or Windows
- data is written into ttyACM0 from NuttX
- tx fifo will never get empty and the system is stuck
Fix this by issuing an error code if the transmit fifo doesn't
proceed as expected. The error code is then propagated into
higher level keeping the system functional.
Signed-off-by: Eero Nurkkala <eero.nurkkala@offcode.fi>
nxsig_usleep() will wait for the next timer tick which is way
too much here. It's not sleeping 100 us, but rather, near 1/60 s.
This causes severe performance problems. Fix this by polling the
register for a while if the remote end is busy.
Signed-off-by: Eero Nurkkala <eero.nurkkala@offcode.fi>
- Clarify the macros MSSIO_EC_DEFAULT and MSSIO_EC_USB_DEFAULT
- Remove PULLDOWN bit from MSSIO_EC_DEFAULT, it was on by accident
- Fix some EC configuration macros; DRVSTR was wrong, clean up the others
- Define GPIO_PULLUP and GPIO_PULLDOWN like on many other platforms
Signed-off-by: Jukka Laitinen <jukkax@ssrc.tii.ae>
Since the commit cf22dd8 (related to OpenAMP update), the notifyid
is no longer NOTIFY_ALL, but the vq id.
Utilize the vq id now properly as it's being provided. However,
vq id 0 generates action.
Signed-off-by: Eero Nurkkala <eero.nurkkala@offcode.fi>
OpenSBI vendor extension calls must not cause scheduling, as they're
part of M-mode trap handling. Thus, comment out nxsig_usleep() as
well, which is occasionally taken and crashes the system in that
case. Fix this by commenting out lines that have the potential to
cause scheduling.
Signed-off-by: Eero Nurkkala <eero.nurkkala@offcode.fi>
It is not really needed; g_hart_stacks is only used during SBI init as
a temporary stack area. We can use the scratch area buffers for this, as
the scratch areas define almost 4K of extra space, which is used for
exception stacks anyway.
The L3 table address was calculated incorrectly. For every 2MiB of
mapped memory, an offset of 4KiB is needed from the base of the L3
table. The old calculation failed if paddr was not aligned to a 2MiB
boundary.
RX_FIFO_ADDRs and TX_FIFO_ADDR were misconfigured. These addresses
overlapped causing data corruption during high USB loads. For
example, data corruption was present during the following conditions:
1. Composite USB driver was used (CDC/ACM + Mass storage)
2. /dev/ttyACM0 was accessed instantly from Linux side when
starting up.
3. Training data was sent to /dev/ttyACM0 from NuttX from the
very beginning periodically.
It was observed that while Mass storage was negotiating, sometimes
data sent from NuttX to Linux via CDC/ACM was corrupt, although it
was sent properly on the TX fifo.
Also, don't access TXCSRL_REG_EPN_TX_FIFO_NE_MASK for EP0 as it's
not applicable.
Signed-off-by: Eero Nurkkala <eero.nurkkala@offcode.fi>
