Get rid of the constant clrsetbits_le32(), instead prepare the whole
content of the register once and write it at the end of the function.
The big plan here is to remove all the CONFIG_HPS_SDR_ macros, hide
them in QTS compatibility layer in board implementation and pass only
a small structure into the driver.
Signed-off-by: Marek Vasut <marex@denx.de>
Remove all parameters of this function, since they are only constants
passed in from another function, so make them local. Also, rename the
function to get_errata_rows() as this is closer to what it does.
Signed-off-by: Marek Vasut <marex@denx.de>
This function is implementing the DDR calibration Stage 3 as
described in Altera EMI_RM 2015.05.04 . The main body of this
function is almost identical to Stage 1.3 (DQ/DQS centering)
for all but two flags -- use_read_test and update_fom. Convert
this function to call rw_mgr_mem_calibrate_dq_dqs_centering()
with the correct flags set to trim down the code duplication.
Moreover, reorder the remnants in the function a little and
convert the function to return either 0 or -EIO in case of
success and failure respectively, to match the common return
value convention.
Signed-off-by: Marek Vasut <marex@denx.de>
Clean up coding style, mostly broken multiline formatting strings
in debug output. Also clean up data types and constify where applicable.
Signed-off-by: Marek Vasut <marex@denx.de>
Fix data types and constify where applicable, fix broken multiline
debug strings and fix comments. No functional change.
Signed-off-by: Marek Vasut <marex@denx.de>
Fix the return value of the function to match common convention
where 0 means success and negative means error. Fix the return
values in case of an error to use errno.h codes.
Signed-off-by: Marek Vasut <marex@denx.de>
The read_group and write_group arguments are the same in all
cases when this function is invoked, just merge them into one
rw_group argument. Also, clean up the function argument data
types and constify them.
Signed-off-by: Marek Vasut <marex@denx.de>
The search_stop_check() always resets the bit_chk to a new value,
so bit_chk will never survive between search_left_edge() and
search_right_edge() invocation. There is thus no need to pass
it into these functions, so make bit_chk local to them.
Signed-off-by: Marek Vasut <marex@denx.de>
Factor out center_dq_windows(), which is common code between
stage 2 and stage 3 of the calibration again and cater for
the minor differences.
Signed-off-by: Marek Vasut <marex@denx.de>
Factor out the get_window_mid_index() for finding the best DQ
window centering. This is again code common to two stages, so
pull it out.
Signed-off-by: Marek Vasut <marex@denx.de>
First of all, don't break strings in the debug_cond() calls across
multiple lines. Next, zap a couple of really odd casts in the code.
Finally, make i into a signed variable, so it is possible to avoid
constructs of the form:
for (i = FOO;; i--) { <code>; if (i == 0) break; }.
Signed-off-by: Marek Vasut <marex@denx.de>
Factor out common code from search_left_edge() and search_right_edge()
which checks whether searching for the window edge should stop. The
code is almost identical, so pull it into separate function and cater
for the minor differences.
Signed-off-by: Marek Vasut <marex@denx.de>
Factor out common code from rw_mgr_mem_calibrate_writes_center()
and rw_mgr_mem_calibrate_vfifo_center() for searching for the
left edge of the window. The code is almost identical, so pull
it into separate function and cater for the minor differences.
Signed-off-by: Marek Vasut <marex@denx.de>
Factor out common code from rw_mgr_mem_calibrate_writes_center()
and rw_mgr_mem_calibrate_vfifo_center() for searching for the
right edge of the window. The code is almost identical, so pull
it into separate function and cater for the minor differences.
Signed-off-by: Marek Vasut <marex@denx.de>
Reorder the content of the function a little and fix the comments
so they at least become full sentences. Constify function args.
Fix the return value to match the common convention of 0 meaning
success.
Signed-off-by: Marek Vasut <marex@denx.de>
The C language has signed types, so make use of them. Fix this
obscene loop, which tries to do away with unsigned type, but
just makes the code more cryptic.
Signed-off-by: Marek Vasut <marex@denx.de>
Clean up the debug output handling at the end of the function
and factor out common function call from the condition.
Signed-off-by: Marek Vasut <marex@denx.de>
Implement common sdr_find_phase_delay() function and convert all
places which use such code to this unified function. No functional
change to the code.
Signed-off-by: Marek Vasut <marex@denx.de>
If d != 0 after sdr_working_phase() finishes, the else branch
of the condition has no impact on anything at all, since work_end
was already set independently of the value of "d" . Zap this
useless code.
Signed-off-by: Marek Vasut <marex@denx.de>
Fix data types, constify where applicable, fix comments and debug
strings. In case enough failing reads are detected, just return
right away instead of breaking from the loop. This lets us make
the error handling at the end much simpler.
Signed-off-by: Marek Vasut <marex@denx.de>
Pluck out all this VFIFO value counting, which turns out to be
completely unused. Thus, remove it to simplify the code.
Signed-off-by: Marek Vasut <marex@denx.de>
The "d" variable is set to 0 in sdr_backup_phase() and is not used
at all in sdr_nonworking_phase(). Make it local and zap it respectively.
Signed-off-by: Marek Vasut <marex@denx.de>
Wrap dtaps_per_ptap into the sdr_working_phase() function to trim
down the number of params. It's a constant too.
Signed-off-by: Marek Vasut <marex@denx.de>
It is now clear that the max_working_cnt variable is totally unused.
Zap the variable and zap other variables which became unused due to
this change too.
Signed-off-by: Marek Vasut <marex@denx.de>
Pull out the loop for eaching working/non-working DQS enable phase
into a separate function, as this is mostly common code between.
Clean up sdr_working_phase() and sdr_nonworking_phase() while switching
these two functions to the common sdr_find_phase().
Signed-off-by: Marek Vasut <marex@denx.de>
Get rid of found_{begin,end} variables. Instead of breaking out
through all of the loops, just return when the begin/end of the
window is found and be done with it. Also clean up the trailing
conditional expression, which is now much easier.
Signed-off-by: Marek Vasut <marex@denx.de>
Fix the arguments passed to these functions. The bit_chk is
overriden by rw_mgr_mem_calibrate_read_test_all_ranks() which
is invoked by all three sdr_*_phase() functions, so just make
this into local variable.
Signed-off-by: Marek Vasut <marex@denx.de>
Fix the arguments passed to these functions. The grp argument
does not have to be passed via reference, it's never modified
within either of those functions, so make it into a value.
Signed-off-by: Marek Vasut <marex@denx.de>
Reorder the end of the function a little by moving the conditional
debug output around a little. Rename the function from _centre() to
_center(). Document the function in kerneldoc.
Signed-off-by: Marek Vasut <marex@denx.de>
This function is a treasure trove of ad-hoc iterative
implementations of mathematical functions. Replace all
of those with their non-iterative counterpart.
Signed-off-by: Marek Vasut <marex@denx.de>
Clean up the arguments of this function. Most of the pointers
passed into the function are either not needed at all, or can
be passed as value instead of reference. Also fix the broken
multiline debug strings. No functional change.
Signed-off-by: Marek Vasut <marex@denx.de>
This function is only invoked from rw_mgr_mem_calibrate_dqs_enable_calibration()
and at this point, it is just one level of indirection, so wrap the
rw_mgr_mem_calibrate_vfifo_find_dqs_en_phase_sweep_dq_in_delay() into
rw_mgr_mem_calibrate_dqs_enable_calibration() to get rid of the level
of indirection.
Signed-off-by: Marek Vasut <marex@denx.de>
Replace at least one of the loops in this function with call of a
standard function call instead of the ad-hoc implementation. The
other one cannot be replaced, since the delay is incremented for
each group.
Signed-off-by: Marek Vasut <marex@denx.de>
The read_group and write_group params have the same value for all (one)
invocations of this function, just merge them into a single param.
Signed-off-by: Marek Vasut <marex@denx.de>
Start cleaning up this function. In the first part, just fix
the incorrectly broken debug strings and fix return value to
respect the common convention.
Signed-off-by: Marek Vasut <marex@denx.de>
Rework this function such that the code is more readable. Zap
unused parameter "num_tries" while at it. Also wrap parameter
"bit_chk" into this function as it's value is not used outside.
Finally, fix the return value from this function to match the
common expectation, where 0 means success.
Signed-off-by: Marek Vasut <marex@denx.de>
This function is called from one single place and it's sole purpose
is to call one single function with slightly modified arguments.
Zap this function to skip this useless intermediate step.
Signed-off-by: Marek Vasut <marex@denx.de>
Just extract this piece of functionality into separate function
to make the code better separated. This matches the division in
Altera documentation, Altera EMI_RM 2015.05.04 , section 1, the
UniPHY Calibration Stages.
Signed-off-by: Marek Vasut <marex@denx.de>
Just extract this piece of functionality into separate function
to make the code better separated. This matches the division in
Altera documentation, Altera EMI_RM 2015.05.04 , section 1, the
UniPHY Calibration Stages.
Signed-off-by: Marek Vasut <marex@denx.de>
Just extract this piece of functionality into separate function
to make the code better separated. This matches the division in
Altera documentation, Altera EMI_RM 2015.05.04 , section 1, the
UniPHY Calibration Stages.
Signed-off-by: Marek Vasut <marex@denx.de>
This is kind of microseries-within-series indent cleanup.
It is clear that the read_group and write_group variables
have the same value, to just make them into one variable
called rw_group. While doing this, constify the variables
as they are constant.
It is likely that this patch has checkpatch warnings, but
for the sake of not breaking the code, these are ignored.
No functional change.
Signed-off-by: Marek Vasut <marex@denx.de>
This is kind of microseries-within-series indent cleanup.
This patch fixes the broken formatting strings in debug_cond()
invocations.
It is likely that this patch has checkpatch warnings, but
for the sake of not breaking the code, these are ignored.
No functional change.
Signed-off-by: Marek Vasut <marex@denx.de>
This is kind of microseries-within-series indent cleanup.
This patch cleans up the handling of grp_calibrated such
that the variable isn't used all over the place, but just
very localy. This allows trimming down the indent issues.
It is likely that this patch has checkpatch warnings, but
for the sake of not breaking the code, these are ignored.
No functional change.
Signed-off-by: Marek Vasut <marex@denx.de>
This is kind of microseries-within-series indent cleanup.
This patch just tweaks the indentation so it is visible
what to do with the grp_calibrated variable.
It is likely that this patch has checkpatch warnings, but
for the sake of not breaking the code, these are ignored.
No functional change.
Signed-off-by: Marek Vasut <marex@denx.de>
This is kind of microseries-within-series indent cleanup.
This patch just tweaks the indentation so it is visible
what is supposed to go where.
It is likely that this patch has checkpatch warnings, but
for the sake of not breaking the code, these are ignored.
No functional change.
Signed-off-by: Marek Vasut <marex@denx.de>
This patch just adds an expanded documentation header to the
aforementioned function. This is needed to make it easier to
match the purpose of this function with the documentation.
No functional change.
Signed-off-by: Marek Vasut <marex@denx.de>
This is kind of microseries-within-series indent cleanup.
Rework the code for the the middle-loop of the mega-loop
this time and deal with the group_failed variable. Instead
of checking if the group failed in the previous calibration
part, just jump to the end of the loop if calibration did
fail and increment the counter. This shaves away one more
level of indent while making the code slightly more readable.
It is likely that this patch has checkpatch warnings, but
for the sake of not breaking the code, these are ignored.
No functional change.
Signed-off-by: Marek Vasut <marex@denx.de>
This is kind of microseries-within-series indent cleanup.
Rework the code for the last loop within the mega-loop
to make it actually readable and not an insane cryptic pile
of indent failure.
It is likely that this patch has checkpatch warnings, but
for the sake of not breaking the code, these are ignored.
No functional change.
Signed-off-by: Marek Vasut <marex@denx.de>
This is kind of microseries-within-series indent cleanup.
Rework the code for the third loop within the middle-loop
of the mega-loop to make it actually readable and not an
insane cryptic pile of indent failure.
It is likely that this patch has checkpatch warnings, but
for the sake of not breaking the code, these are ignored.
No functional change.
Signed-off-by: Marek Vasut <marex@denx.de>
This is kind of microseries-within-series indent cleanup.
Rework the code for the second loop within the middle-loop
of the mega-loop to make it actually readable and not an
insane cryptic pile of indent failure.
It is likely that this patch has checkpatch warnings, but
for the sake of not breaking the code, these are ignored.
No functional change.
Signed-off-by: Marek Vasut <marex@denx.de>
This is kind of microseries-within-series indent cleanup.
Rework the code for the first loop within the middle-loop
of the mega-loop to make it actually readable and not an
insane cryptic pile of indent failure.
It is likely that this patch has checkpatch warnings, but
for the sake of not breaking the code, these are ignored.
No functional change.
Signed-off-by: Marek Vasut <marex@denx.de>
Redo the mega-condition such that if the calibration is to be skipped,
the positive branch of the condition does all the work and returns.
The negative branch, which is in fact the default behavior, is then
converted to a code which is no longer conditional. This trims down
the indent by one level.
Signed-off-by: Marek Vasut <marex@denx.de>
Perform minor coding style cleanup of the mem_skip_calibrate() function,
clean up comments and add kerneldoc. No functional change.
Signed-off-by: Marek Vasut <marex@denx.de>
Turn the big inner if (RW_MGR_MEM_NUMBER_OF_RANKS == ...) conditional
into a switch {} statement instead. No functional change.
Signed-off-by: Marek Vasut <marex@denx.de>
First, invert the logic of the if (odt_mode == ...) conditional to make
the OFF mode harder to miss. It is a short piece of code right at the
end, so move it up.
Also, clean up data types and constify where applicable and clean up
the cs_and_odt_mask assignment. No functional change.
Signed-off-by: Marek Vasut <marex@denx.de>
Clean mem_config() function. First, reorder the math done in the
function such that WLAT and RLAT computation is together. Then,
scrap contradictory comments which do not match the result of the
math at all. Next, extract the mem_precharge_and_activate() call
from the end of the function as it is completely unrelated here.
Finally, rename the function to mem_init_latency().
Signed-off-by: Marek Vasut <marex@denx.de>
Clean the function up slightly by using clrsetbits_le32() to flip
bits in registers instead of cryptic bitmasks. Zap condition checking
for PHY_DEBUG_IN_DEBUG_MODE flag, which is never set. Split the
calibration report into separate debug_mem_calibrate() function.
Signed-off-by: Marek Vasut <marex@denx.de>
Init both structures with zeroes and zap all those zeroing shenanigans
further down in the sdram_calibration_full().
Signed-off-by: Marek Vasut <marex@denx.de>
Clean the function up by getting rid of all the insane XOR-leftshift
combos when assembling register values. While at it, remove all the
ad-hoc variables necessary for this XOR-leftshift voodoo. Finally,
get rid of the iterative division implementation of two constants
and replace it with a DIV_ROUND_UP() macro :-)
Signed-off-by: Marek Vasut <marex@denx.de>
Contemporary CPUs can perform division just fine, use this
functionality and zap another implementation of iterative
division :-)
Signed-off-by: Marek Vasut <marex@denx.de>
Factor out almost common code from rw_mgr_mem_handoff() and
rw_mgr_mem_initialize() into separate rw_mgr_mem_load_user().
No functional change.
Signed-off-by: Marek Vasut <marex@denx.de>
Clean the DQS and OCT parts of the function, clean up the chopped
formatting strings in debug_cond() and slightly improve the code.
Zap group_bgn argument as it is used only in debug messages. Document
the function using kerneldoc. No functional change.
Signed-off-by: Marek Vasut <marex@denx.de>
Apparently, in case of the DQ and DM, the value if the new_delay variable
is calculated, but the value is not used. Zap the entire code which does
calculate the value.
It is not clear to me whether or not the code is doing the right thing
in the first place. Right now, it calls scc_mgr_load_dq() and
scc_mgr_load_dm() respectively, but I suspect it might need to call
scc_mgr_apply_group_dq_out1_delay() and scc_mgr_apply_group_dm_out1_delay()
instead. This is something Altera must investigate.
Signed-off-by: Marek Vasut <marex@denx.de>
First, zap unused argument of the function. Next, clean up
the data types, constify where applicable, clean up comments
and add kerneldoc.
Signed-off-by: Marek Vasut <marex@denx.de>
Move scc_mgr_set_hhp_extras() out of scc_set_bypass_mode() as it
has nothing to do in there. Instead, invoke it from mem_calibrate()
just before invoking scc_set_bypass_mode().
Signed-off-by: Marek Vasut <marex@denx.de>
Minor coding style cleanup for this function. Furthermore, move
ad-hoc debug_cond() calls from the only location from where this
function is invoked into this actual function.
Signed-off-by: Marek Vasut <marex@denx.de>
Make this function more readable, no functional change. Also, zap the
forward declaration, which is no longer needed.
Signed-off-by: Marek Vasut <marex@denx.de>
Implement universal scc_mgr_set_all_ranks() function and convert
various ad-hoc implementations of similar functionality to use
this single function. Document the function in kerneldoc.
Signed-off-by: Marek Vasut <marex@denx.de>
Implement unified scc_mgr_set() function and convert all those
9 scc_mgr_set_*() ad-hoc functions to call this one function.
Signed-off-by: Marek Vasut <marex@denx.de>
This patch just puts functions which look similar next to each
other, so they can be sorted out. No functional change.
Signed-off-by: Marek Vasut <marex@denx.de>
Turn the insides of these functions into trivial clrsetbits_le32()
and fix the data type of their argument to reflect it's actual size.
Signed-off-by: Marek Vasut <marex@denx.de>
Remove the remaining invocations of sdr_get_addr() and the function
itself. This makes the code a bit less cryptic.
Signed-off-by: Marek Vasut <marex@denx.de>
The debug messages missed proper newlines and/or spaces in them.
Fix the formatting.
Signed-off-by: Marek Vasut <marex@denx.de>
Cc: Chin Liang See <clsee@altera.com>
Cc: Dinh Nguyen <dinguyen@altera.com>
Cc: Tom Rini <trini@konsulko.com>
It is the configuration data that should go into the register,
not the register mask, just like the surrounding code does it.
Fix this typo.
Signed-off-by: Marek Vasut <marex@denx.de>
Cc: Chin Liang See <clsee@altera.com>
Cc: Dinh Nguyen <dinguyen@altera.com>
Cc: Tom Rini <trini@konsulko.com>
Move the structure prototype from sdram.h header file into sdram.c
source file, since it is used only there and for local purpose only.
There is no point in having it global.
While at this move, fix the data types in the structure from uintNN_t
to uNN and fix the coding style a bit.
Signed-off-by: Marek Vasut <marex@denx.de>
This patch enables the SDRAM controller that is used on Altera's SoCFPGA
family. This patch configures the SDRAM controller based on a configuration
file that is generated from the Quartus tool, sdram_config.h.
Signed-off-by: Dinh Nguyen <dinguyen@opensource.altera.com>
By default the bstopre value has been set to 0x100, used to be 1/4
value of refint. Modern DDR has increased the refresh time. Adjust
to 1/4 of refresh interval dynamically. Individual board can still
override this value in board ddr file, or to use auto-precharge.
Signed-off-by: York Sun <yorksun@freescale.com>
This patch adds the DDR3 setup and training code taken from the Marvell
U-Boot repository. This code used to be included as a binary (bin_hdr)
into the Armada A38x boot image. Not linked with the main U-Boot. With this
code addition and the serdes/PHY setup code, the Armada A38x support
in mainline U-Boot is finally self-contained. So the complete image
for booting can be built from mainline U-Boot. Without any additional
external inclusion.
Note:
This code has undergone many hours (days!) of coding-style cleanup and
refactoring. It still is not checkpatch clean though, I'm afraid. As the
factoring of the code has so many levels of indentation that many lines
are longer than 80 chars.
Signed-off-by: Stefan Roese <sr@denx.de>
With the upcoming addition of the Armada 38x DDR support, which is not
compatible to the Armada XP DDR init code, we need to introduce a new
directory infrastructure. To support multiple Marvell DDR controller.
This will be the new structure:
drivers/ddr/marvell/axp
Supporting Armada XP (AXP) devices (and perhaps Armada 370)
drivers/ddr/marvell/a38x
Supporting Armada 38x devices (and perhaps Armada 39x)
Signed-off-by: Stefan Roese <sr@denx.de>
In case SPD address changes between board revisions, updating SPD
address can be called from board file.
Signed-off-by: York Sun <yorksun at freescale.com>
Signed-off-by: Prabhakar Kushwaha <prabhakar at freescale.com>
The macro to select the I2C address for ECC bus-width detection
was defined incorrectly for the Marvell DB-MV784MP-GP board. This
patch changes the macro to the correct value to fix this issue.
Signed-off-by: Stefan Roese <sr@denx.de>
This erratum only applies to general purpose DDR controllers in LS2.
It shouldn't be applied to DP-DDR controller. Check DDRC versoin number
before applying workaround.
Signed-off-by: York Sun <yorksun@freescale.com>
Add built-in memory test to catch errors after DDR is initialized, before
any other transactions. To enable this test, define CONFIG_FSL_DDR_BIST.
An environmental variable "ddr_bist" is checked before starting test.
It takes a while (several seconds) depending on system memory size.
Signed-off-by: York Sun <yorksun@freescale.com>
CS0 was not allowed to be empty by u-boot driver in the past to simplify
the driver. This may be inconvenient for some debugging. This patch lifts
the restrictions. Controller interleaving still requires CS0 populated.
Signed-off-by: York Sun <yorksun@freescale.com>
Add/update registers for DDR4, including DQ mappings. Allow raw timing
method used for all controllers. Update mode_9 register to 0x500 for
improved stability. Check DDR controller version number individually
in case a SoC has multiple DDR controllers of different versions.
Increase read-write turnaround for DDR4 high speeds.
Signed-off-by: York Sun <yorksun@freescale.com>
According to the MPC8555/MPC8541 reference manual the SS_EN (source
synchronous enable) bit in the DDR_SDRAM_CLK_CNLT register must be set
during initialization.
>From section 9.4.1.8 of that manual:
Source synchronous enable. This bit field must be set during
initialization. See Section 9.6.1, "DDR SDRAM Initialization
Sequence," details.
0 - Reserved
1 - The address and command are sent to the DDR SDRAMs source
synchronously.
In addition, Freescale application note AN2805 is also very clear that
this bit must be set.
This patch reverts a change introduced by commit
457caecdbc.
Testing Done:
Compiled targets CONFIG_TARGET_MPC8555CDS and CONFIG_TARGET_MPC8541CDS
and inspected the generated assembly code to verify the SS_EN bit was being
set. There is one extra instruction emitted:
fff9b774: 65 29 80 00 oris r9,r9,32768
Compiled the CONFIG_TARGET_MPC8548CDS target and verified that no
additional instructions were emitted related to this patch.
Booted an image on a MPC8541 based board successfully.
Signed-off-by: Curt Brune <curt@cumulusnetworks.com>
Reviewed-by: York Sun <yorksun@freescale.com>
Add sync of refresh for multiple DDR controllers. DDRC initialization
needs to complete first. Code is re-ordered to keep refresh close.
Signed-off-by: York Sun <yorksun@freescale.com>
Controller number is passed for function calls to support individual
DDR clock, depending on SoC implementation. It is backward compatible
with exising platforms. Multiple clocks have been verifyed on LS2085A
emulator.
Signed-off-by: York Sun <yorksun@freescale.com>
On ZeBu emulator, CAS to preamble overrides need to be set to
satisfy the timing. This only impact platforms with CONFIG_EMU.
These should be set before MEM_EN is set.
Signed-off-by: York Sun <yorksun@freescale.com>
This patch adds the DDR3 setup and training code taken from the Marvell
U-Boot repository. This code used to be included as a binary (bin_hdr)
into the AXP boot image. Not linked with the main U-Boot. With this code
addition and the following serdes/PHY setup code, the Armada-XP support
in mainline U-Boot is finally self-contained. So the complete image
for booting can be built from mainline U-Boot. Without any additional
external inclusion. Hopefully other MVEBU SoC's will follow here.
Support for some SoC's has been removed in this version. This is:
MV_MSYS:
The code referred to by the MV_MSYS define is currently unused. And its
not really planned to support this in mainline. So lets remove it to
make the code clearer and increase the readability.
MV88F68XX (A38x):
The code referred to by the MV88F68XX define (A38x) is currently unused.
And its partial and not sufficient for this device in this stage.
So lets remove it to make the code clearer and increase the readability.
MV88F66XX (ALP):
The code referred to by the MV88F66XX define is currently unused. And its
not really planned to support this in mainline. So lets remove it to
make the code clearer and increase the readability.
MV88F78X60_Z1:
The code referred to by the MV88F78X60_Z1 define is currently unused. As the
Z1 revision of the AXP is not supported in mainline anymore.
So lets remove it to make the code clearer and increase the readability.
Remove support for Z1 & A0 AXP revisions (steppings). The current stepping
is B0 and this is the only one that is actively supported in this code
version.
Tested on AXP using a SPD DIMM setup on the Marvell DB-MV784MP-GP board and
on a custom fixed DDR configuration board (maxbcm).
Note:
This code has undergone many hours of coding-style cleanup and refactoring.
It still is not checkpatch clean though, I'm afraid. As the factoring of the
code has so many levels of indentation that many lines are longer than 80
chars. This might be some task to tackly later on.
Signed-off-by: Stefan Roese <sr@denx.de>
Reviewed-by: Luka Perkov <luka.perkov@sartura.hr>
Internal memory controller counters can reach a bad state after
training in DDR4 mode if accumulated ECC or DBI mode is eanbled.
Signed-off-by: York Sun <yorksun@freescale.com>
For DDR controller version 4.7 or newer, MRC_CYC (mode register set
cycle time) is max(tMRD, tMOD). tMRD is 4nCK, or 8nCK (RDIMM). tMOD
is max(12nCK, 15ns) according to JEDEC spec.
DDR4 is not affected by this change.
Signed-off-by: York Sun <yorksun@freescale.com>
With the introducing of generic board and ARM-based cores, current
deep sleep framework doesn't work anymore.
This patch will convert the current framework to adapt this change.
Basically it does:
1. Converts all the Freescale's DDR driver to support deep sleep.
2. Added basic framework support for ARM-based and PPC-based
cores separately.
Signed-off-by: Tang Yuantian <Yuantian.Tang@freescale.com>
Reviewed-by: York Sun <yorksun@freescale.com>
Some UDIMMs have faulty SPD with wrong mapping for DQ[36-39].
Using raw card spec in case this error is detected.
Signed-off-by: York Sun <yorksun@freescale.com>
Increase write-to-write and read-to-read turnaround time for two-slot DDR
configurations. Previously only quad-rank and two dual-rank configurations
have this additional turnaround time. A recent test on two single-rank
DIMMs shows the shorter additional turnaround time is also needed.
Signed-off-by: York Sun <yorksun@freescale.com>
U-Boot has never cared about the type when we get max/min of two
values, but Linux Kernel does. This commit gets min, max, min3, max3
macros synced with the kernel introducing type checks.
Many of references of those macros must be fixed to suppress warnings.
We have two options:
- Use min, max, min3, max3 only when the arguments have the same type
(or add casts to the arguments)
- Use min_t/max_t instead with the appropriate type for the first
argument
Signed-off-by: Masahiro Yamada <yamada.m@jp.panasonic.com>
Acked-by: Pavel Machek <pavel@denx.de>
Acked-by: Lukasz Majewski <l.majewski@samsung.com>
Tested-by: Lukasz Majewski <l.majewski@samsung.com>
[trini: Fixup arch/blackfin/lib/string.c]
Signed-off-by: Tom Rini <trini@ti.com>
When accumulated ECC is enabled, the DQ_MAP for ECC[4:7] needs to be set
to 0, i.e. 0->0, 1->1, etc., required by controller logic, even these pins
are not actually connected.
Also fix a bug when reading from DDR register to use proper accessor for
correct endianess.
Signed-off-by: York Sun <yorksun@freescale.com>
The driver was written using old DDR3 spec which only covers low speeds.
The value would be suboptimal for higher speeds. Fix both timing according
to latest DDR3 spec, remove tCKE as an config option.
Signed-off-by: York Sun <yorksun@freescale.com>
U-boot has been initializing DDR for the main memory. The presumption
is the memory stays as a big continuous block, either linear or
interleaved. This change is to support putting some DDR controllers
to separated space without counting into main memory. The standalone
memory controller could use different number of DIMM slots.
Signed-off-by: York Sun <yorksun@freescale.com>
JEDEC spec allows DRAM vendors to use prime DQ for write leveling. This
is not an issue unless some DQ pins are not connected. If a platform uses
regular DIMMs but with reduced DDR ECC pins, the prime DQ may end up on
those floating pins for the second rank. The workaround is to use a known
good chip select for this purpose.
Signed-off-by: York Sun <yorksun@freescale.com>
Reading DDR register should use ddr_in32() for proper endianess.
This patch fixes incorrect waiting time for ARM platforms.
Signed-off-by: York Sun <yorksun@freescale.com>
If less than 8 ECC pins are used for DDR data bus width smaller than 64
bits, the 8-bit ECC code will be transmitted/received across several beats,
and it will be used to check 64-bits of data once 8-bits of ECC are
accumulated.
Signed-off-by: York Sun <yorksun@freescale.com>
The field wrtord_bg should add 2 clocks if on the fly chop is enabled,
according to DDR controller manual for DDR4.
Signed-off-by: York Sun <yorksun@freescale.com>
Previously the driver was only tested on Power SoCs. Different barrier
instructions are needed for ARM SoCs.
Signed-off-by: York Sun <yorksun@freescale.com>
When the DDR controller is initialized below a junction temperature of
0°C and then operated above a junction temperature of 65°C, the DDR
controller may cause receive data errors, resulting ECC errors and/or
corrupted data. This erratum applies to the following SoCs and their
variants: MPC8536, MPC8569, MPC8572, P1010, P1020, P1021, P1022, P1023,
P2020.
Signed-off-by: York Sun <yorksun@freescale.com>
When T104x soc wakes up from deep sleep, control is passed to the
primary core that starts executing uboot. After re-initialized some
IP blocks, like DDRC, kernel will take responsibility to continue
to restore environment it leaves before.
Signed-off-by: Tang Yuantian <Yuantian.Tang@freescale.com>
Reviewed-by: York Sun <yorksun@freescale.com>
This is a theoretical possible out of bounds error in DDR driver. Adding
check before using array index. Also change some runtime conditions to
pre-compiling conditions.
Signed-off-by: York Sun <yorksun@freescale.com>
Reviewed-by: York Sun <yorksun@freescale.com>
Freescale LayerScape SoCs support controller interleaving on 256 byte size.
This interleaving is mandoratory.
Signed-off-by: York Sun <yorksun@freescale.com>
DDR base address has been the same from the view of core and DDR
controllers. This has changed for Freescale ARM-based SoCs. Controllers
setup DDR memory in a contiguous space and cores view it at separated
locations.
Signed-off-by: York Sun <yorksun@freescale.com>
Initially it was believed the DDR controller on Freescale ARM would have
big endian. But some platform will have little endian.
Signed-off-by: York Sun <yorksun@freescale.com>
Existing workaround only handles one RDIMM on reference design. In case
of two RDIMMs being used, the workaround requires two separate writes to
DDR_SDRAM_MD_CNTL register.
This patch also restores two debug registers changed by the workaround.
Signed-off-by: York Sun <yorksun@freescale.com>
CC: Ben Collins <ben.c@servergy.com>
CC: James Yang <James.Yang@freescale.com>
The DRAM base has been zero for Power SoCs. It could be non-zero
for ARM SoCs. Use a macro instead of hard-coding to zero.
Signed-off-by: York Sun <yorksun@freescale.com>
Some DDR registers' fields have expanded to accommodate larger values.
These changes are backward compatible. Some fields are removed for newer
DDR controllers. Writing to those fields are safely ignored.
TIMING_CFG_2 register is fixed. Additive latency is added to RD_TO_PRE
automatically. It was a misunderstanding in commit c360ceac.
Signed-off-by: York Sun <yorksun@freescale.com>
Fix ccsr_ddr structure to avoid using typedef. Combine DDR2 and DDR3
structure for 83xx, 85xx and 86xx.
Signed-off-by: York Sun <yorksun@freescale.com>
Freescale DDR driver has been used for mpc83xx, mpc85xx, mpc86xx SoCs.
The similar DDR controllers will be used for ARM-based SoCs.
Signed-off-by: York Sun <yorksun@freescale.com>
