r14 is not supposed to be clobbered by functions. Switch
to r12 and call GET_GOT when needed. This will allow u-boot
to loose the -ffixed-r14 gcc option.
Signed-off-by: Joakim Tjernlund <Joakim.Tjernlund@transmode.se>
Using the GOT in IRQ handlers requires r14 to be -ffixed-r14.
Avoid this by relocatate transfer_to_handler too.
This will allow to free up r14 later on.
Signed-off-by: Joakim Tjernlund <Joakim.Tjernlund@transmode.se>
NULL is an absolute value and should not be relocated.
After this correction code like:
void weak_fun(void) __attribute__((weak));
printf("weak_fun:%p\n", weak_fun);
will still print null after relocation.
Signed-off-by: Joakim Tjernlund <Joakim.Tjernlund@transmode.se>
This ensures that subsequent accesses properly hit the new window.
The dcbi during the NAND loop was accidentally working around this;
it's no longer necessary, as the cache is not enabled.
Reported-by: Suchit Lepcha <Suchit.Lepcha@freescale.com>
Signed-off-by: Scott Wood <scottwood@freescale.com>
Signed-off-by: Kim Phillips <kim.phillips@freescale.com>
Currently there are in excess of 100 bytes located at the beginning of the image
built by start.S that are not being utilized. This patch moves a few functions
into this part of the image. This will create a greater number of *available*
bytes that can be used by board specific code in NAND builds and will decrease
the size of the assembled code in other builds.
Signed-off-by: Ron Madrid <ron_madrid@sbcglobal.net>
Signed-off-by: Kim Phillips <kim.phillips@freescale.com>
At least some revisions of the 8313, and possibly other chips, do not
wait for all pages of the initial 4K NAND region to be loaded before
beginning execution; thus, we wait for it before branching out of the
first NAND page.
This fixes warm reset problems when booting from NAND on 8313erdb.
Signed-off-by: Scott Wood <scottwood@freescale.com>
Acked-by: Kim Phillips <kim.phillips@freescale.com>
Cleans up some latent issues with the data cache control so that
dcache_enable() and dcache_disable() will work reliably (after
unlock_ram_in_cache() has been called)
Signed-off-by: Nick Spence <nick.spence@freescale.com>
Signed-off-by: Kim Phillips <kim.phillips@freescale.com>
Record the Arbiter Event Register values and optionally display them.
The Arbiter Event Register can record the type and effective address of
an arbiter error, even through an HRESET. This patch stores the values in
the global data structure.
Display of the Arbiter Event registers immediately after the RSR value
can be enabled with defines. The Arbiter values will only be displayed
if an arbiter event has occured since the last Power On Reset, and either
of the following defines exist:
#define CONFIG_DISPLAY_AER_BRIEF - display only the arbiter address and
and type register values
#define CONFIG_DISPLAY_AER_FULL - display and interpret the arbiter
event register values
Address Only transactions are one of the trapped events that can register
as an arbiter event. They occur with some cache manipulation instructions
if the HID0_ABE (Address Broadcast Enable) is set and the memory region
has the MEMORY_COHERENCE WIMG bit set. Setting:
#define CONFIG_MASK_AER_AO - prevents the arbiter from recording address
only events, so that it can still capture
other real problems.
Signed-off-by: Nick Spence <nick.spence@freescale.com>
Signed-off-by: Kim Phillips <kim.phillips@freescale.com>
This is needed in unlock_ram_in_cache() because it is called from C and
will corrupt the small data area anchor that is kept in R2.
lock_ram_in_cache() is modified similarly as good coding practice, but
is not called from C.
Signed-off-by: Nick Spence <nick.spence@freescale.com>
Signed-off-by: Kim Phillips <kim.phillips@freescale.com>
Note that with older board revisions, NAND boot may only work after a
power-on reset, and not after a warm reset. I don't have a newer board
to test on; if you have a board with a 33MHz crystal, please let me know
if it works after a warm reset.
Signed-off-by: Scott Wood <scottwood@freescale.com>
Change all code that conditionally operates on high bat
registers (that is, BATs 4-7) to look at CONFIG_HIGH_BATS
instead of the myriad ways this is done now. Define the option
for every config for which high bats are supported (and
enabled by early boot, on parts where they're not always
enabled)
Signed-off-by: Becky Bruce <becky.bruce@freescale.com>
This reverts commit 70431e8a73 which has
proven problematic getting right from the start at least on 83xx and
4xx.
Signed-off-by: Kim Phillips <kim.phillips@freescale.com>
Remove a few absolute references to CFG_MONITOR_BASE for ppc/mpc83xx
and use GOT relative reference.
Signed-off-by: Joakim Tjernlund <Joakim.Tjernlund@transmode.se>
Signed-off-by: Kim Phillips <kim.phillips@freescale.com>
Put the version (and magic) after the HRCW. This puts it in a fixed
location in flash, not at the start of flash but as close as we can get.
Signed-off-by: Jerry Van Baren <vanbaren@cideas.com>
Replace all instances of CFG_IMMRBAR with CFG_IMMR, so that the 83xx
tree matches the other 8xxx trees.
Signed-off-by: Timur Tabi <timur@freescale.com>
The 834x rev1.x silicon has one CPU5 errata.
The issue is when the data cache locked with
HID0[DLOCK], the dcbz instruction looks like no-op inst.
The right behavior of the data cache is when the data cache
Locked with HID0[DLOCK], the dcbz instruction allocates
new tags in cache.
The 834x rev3.0 and later and 8360 have not this bug inside.
So, when 834x rev3.0/8360 are working with ECC, the dcbz
instruction will corrupt the stack in cache, the processor will
checkstop reset.
However, the 834x rev1.x can work with ECC with these code,
because the sillicon has this cache bug. The dcbz will not
corrupt the stack in cache.
Really, it is the fault code running on fault sillicon.
This patch fix the incorrect dcbz operation. Instead of
CPU FP writing to initialise the ECC.
CHANGELOG:
* Fix the incorrect dcbz operation instead of CPU FP
writing to initialise the ECC memory. Otherwise, it
will corrupt the stack in cache, The processor will checkstop
reset.
Signed-off-by: Dave Liu <daveliu@freescale.com>
CHANGELOG:
* On 83xx systems, use the CFG_FLASH_SIZE macro to program the LBC local access
window registers, instead of using a hard-coded value of 8MB.
Signed-off-by: Timur Tabi <timur@freescale.com>
