This overwrote the first 2 instructions of crt0 in the bootloader!
I'm really not sure how this *didn't* cause a fatal exception.
This address isn't special as far as I know, so just move it to the
TCSM by making it a static variable.
Change-Id: I58e1486804aeb2b68325e8de2aa1874c97abef19
The abiflags data is only used to communicate ABI information to a
program loader -- you can see what info is stored with readelf -A.
Dropping it shaves 24 bytes off of every binary (including plugins).
Change-Id: Iae78eeffe5c840ff67717707fb94821d24aac8ec
It never worked, and hasn't compiled in something like a decade, Given
the HW capabilities (limited onboard flash, no expandability) there's
really no point in trying to fix/complete it.
Change-Id: I7d175089840396f8891645bd10010d730dd5bfdc
They were never finished, never saw any release ever, and haven't
compiled for the better part of a decade. Given their HW capabilities [1],
they are not worth trying to fix.
[1] 1-2MB RAM, ~256MB onboard flash, no expandability
Change-Id: I7b2a5806d687114c22156bb0458d4a10a9734190
There's absolutely no way for gpio_config() to get called from two
different threads due to the co-operative threading model, and it
is unsafe to call from IRQ context no matter what we do.
Change-Id: I58f7d1f68c7a414610bb020e26b774cb1015a3b0
Using a macro to put each function in its own .icode-based section
allows us to put the functions in IRAM _and_ have linker GC. This
removes a troublesome #ifdef BOOTLOADER_SPL on the X1000 target.
Change-Id: Ia7b59778f5c36b7970dee4280547e434a1f4fc5a
This only required a minor patch to the usb-designware driver due
to DMA requiring physical addresses -- on the X1000, these differ
from virtual addresses so we have to do the usual conversion.
Both the mass storage and HID drivers work, but there are a few
issues so this can't be considered 100% stable yet.
- Mass storage might not be detected properly on insertion,
and USB has to be replugged before it shows up
- HID driver may occasionally panic or hang the machine
Change-Id: Ia3ce7591d5928ec7cbca7953abfef01bdbd873ef
- Added register names to reduce usage of magic numbers
- Added function to control max charging current, needed for USB
- Corrected comment about axp173, since FiiO M3K has an axp192
Change-Id: I6604ce8d44e5a2ee84061cf042d17ccc4734ac57
After conducting some simplistic tests, I found that the power usage
did not appear to be affected by the CPU frequency.
I tested by playing back a 44.1 KHz FLAC file on single track repeat,
and measured current with the AXP173's battery discharge current ADC.
The button and LCD backlights were set to always on. Headphones were
unplugged and the volume was muted to eliminate any influence from
the headphone amp.
On average the current usage was between 78-81 mA at 1008 MHz, 252 MHz,
and 112 MHz. If anything, 1008 MHz drew _less_ current than the lower
frequencies, by about 1-3 mA.
A possible explanation for this, assuming it's not just a bias of the
test, is that the CPU idle state saves so much power that it's better
to maximize the real time that the CPU spends idling. More systematic
testing is needed to confirm this.
Change-Id: I527473e8c4c12bc1e94f8d4e849fecc108022abe
There's no point including this in normal builds: the stats are not
used for anything, they are not really of interest to anyone except
developers, and add a small overhead to the kernel tick.
Change-Id: I1b4f67cc62d11d634a8cec279dca513dd10eea96
Initializing the clocks in the SPL brings Rockbox in line with
how the FiiO M3K's original SPL works. It's likely other X1000
devices do this too.
There was a logic error in the previous setup: the code falsely
assumed that DDR memory would always be running from MPLL, but
it would be switched to APLL by the bootloader. Rockbox would
then try to re-init APLL, albeit with the same parameters. Maybe
this was the cause of the boot hang on some units.
Change-Id: I64064585e491bbdf1e95fe9428c91a9314f2a917
What we really want is to avoid any interrupts being generated
before the drivers which handle them are properly initialized.
Intead of trashing all GPIOs, search for the problem pins and
fix them, leaving the others alone.
This fixes the M3K's button light flickering on boot and should
stop the M3K from entering a potentially confusing "dead" state
where all the lights are off but the CPU is still on.
Change-Id: I13a6da0f0950190396bff5d6e8c343c668e8fea1
SPL is now designed so core X1000 code is in control of the boot,
under the reasonable assumption that the device boots from flash.
It should not be too hard to adapt to other X1000 ports.
The biggest functional change is that the SPL can now read/write
the flash, under the control of a host computer. The SPL relies
on the boot ROM for USB communication, so the host has to execute
the SPL multiple times following a protocol.
Change-Id: I3ffaa00e4bf191e043c9df0e2e64d15193ff42c9
Detection at startup is proving to be unreliable. Even if card is not
present at startup, upon insertion it will sort itself out properly.
Change-Id: I9ee90b724c90c530a39264f698c200a48aa72b1d
Including direct use of the external SD card mount
Known issue: If SD card is inserted at startup, it must be
ejected and reinserted to be registered.
Change-Id: I5f420160bda32135cbb088c1e8b04b6e3a73018e
The X3's line out is a bit hot, at ~4.3Vpp, so allow it to be backed off.
(On my X3, backing it off to -6dB brings Vpp down to ~3.4V)
Change-Id: Iea38ef1c6a1b183d0f8fb4eaf2bf9ed6b350a532
- Proper error codes are now returned from all functions. These codes will
be used by a host-side flash tool for error reporting.
- nand_erase_block() was replaced by nand_erase_bytes(). The caller can't
know how big an eraseblock is with the current API, so next best thing
is to verify the correct alignment inside the call and reject the erase
if it isn't properly aligned.
- Fixed typo in nandcmd_block_erase() which would cause an SFC error to be
interpreted as success. Yikes.
Change-Id: Id4ac9b44fa7fc2fcb81ff19ba730df78457c0383
headphone ADC thread stack was slightly too small. Bump it up a bit.
(it was _perfectly_ sized for the prior older toolchain+optimization flags...)
Change-Id: I2ca67c2b85c54f879892a31e281d7696f893389c
Use of IF_COP_CORE was mistakenly introduced as part of 89acde6af2,
effectively short-circuiting multiple tests resulting in the code
paths always being executed, on both cores.
Use the correct macro, so per-CPU paths are handled properly.
Change-Id: Id346cf759fc1b06b7d56694d7af1f469caf785a4
This appears to finally fix the issue
turns out the status register we were writing was only for the CPU
COP cache flush wiped out the CPU cache
--
Added some defines to cut down on the magic numbers
Added some comments explaining such
Set the address to full 20 bit address
0x1FFFFF which is then left shifted 11 internally -- somewhere around 4GB?
Link explains the cache status bits
https://daniel.haxx.se/sansa/memory_controller.txt
Change-Id: I57b7187c2f71a5b54ce145bf3a21ed492a8993cb
Enable its use in the jz47xx MIPS targets.
(accidently committed g#3249 before making these changes)
Change-Id: I1791946f632901f0c7a94b04b009671aa0d71717
* PREV/NEXT now swapped so they do what is expected in most contexts
* List and setting context retains prior behavior
* Enable the ADC that reads the headset remote and map the keys.
* As ADC-based remote "events" arrive as press/release pairs,
delay the button release.
Change-Id: I22d4eac3bfe1573b50eca795cf377bdafdeb5336
affects all hiby targets, fiiom3k, and ibasso dx50/dx90
As well as deduplicating a small pile of code, this also implements
hysteresis so we're not doing a sysfs read/lookup multiple times
back-to-back every time the power management tick fires.
Change-Id: I2f7672acbb36341becf67e07960c24c681270d09
Previously these were placed in DRAM, which is overwritten by RoLo
when it loads a new image, but RoLo must call commit_discard_idcache()
after loading the image.
Change-Id: I5dcc4ca711b774166f83c668695edbcabfab2604
The filesystem API often passes in unaligned receive buffers, and some
code (eg BMP reader) processes data in-place, leading to data loss when
we dropped the cache.
(And document exactly what we're doing, so we don't go through this again
at $future_date)
Change-Id: If47a7f2148a5a1a43777f0bd3be1bdfe8239e91e
In fixing the original bug I tried to optimize discard_dcache_range()
to minimize writeback and inadvertently introduced a second bug, which
typically ends in a TLB refill panic.
It occurs only if the range fits within one cache line, and when both
the start and end of the range are not aligned to a cache line. This
causes ptr to be incremented and end to be decremented, so ptr > end,
and the loop can't terminate.
Change-Id: Ibaac072f1369268d3327d534ad08ef9dcee3db65
- The range-based cache operations on MIPS were broken and only worked
properly when BOTH the address and size were multiples of the cache
line size. If this was not the case, the last cache line of the range
would not be touched!
Fix is to align start/end pointers to cache lines before iterating.
- To my knowledge all MIPS processors have a cache, so I enabled
HAVE_CPU_CACHE_ALIGN by default. This also allows mmu-mips.c to use
the CACHEALIGN_UP/DOWN macros.
- Make jz4760/system-target.h define its cache line size properly.
Change-Id: I1fcd04a59791daa233b9699f04d5ac1cc6bacee7
