GitLab

The current PSCI implementation can apply certain optimizations upon the
assumption that all PSCI participants are cache-coherent.

  - Skip performing cache maintenance during power-up.

  - Skip performing cache maintenance during power-down:

    At present, on the power-down path, CPU driver disables caches and
    MMU, and performs cache maintenance in preparation for powering down
    the CPU. This means that PSCI must perform additional cache
    maintenance on the extant stack for correct functioning.

    If all participating CPUs are cache-coherent, CPU driver would
    neither disable MMU nor perform cache maintenance. The CPU being
    powered down, therefore, remain cache-coherent throughout all PSCI
    call paths. This in turn means that PSCI cache maintenance
    operations are not required during power down.

  - Choose spin locks instead of bakery locks:

    The current PSCI implementation must synchronize both cache-coherent
    and non-cache-coherent participants. Mutual exclusion primitives are
    not guaranteed to function on non-coherent memory. For this reason,
    the current PSCI implementation had to resort to bakery locks.

    If all participants are cache-coherent, the implementation can
    enable MMU and data caches early, and substitute bakery locks for
    spin locks. Spin locks make use of architectural mutual exclusion
    primitives, and are lighter and faster.

The optimizations are applied when HW_ASSISTED_COHERENCY build option is
enabled, as it's expected that all PSCI participants are cache-coherent
in those systems.

Change-Id: Iac51c3ed318ea7e2120f6b6a46fd2db2eae46ede
Signed-off-by: Jeenu Viswambharan <jeenu.viswambharan@arm.com>

The PSCI implementation performs cache maintenance operations on its
data structures to ensure their visibility to both cache-coherent and
non-cache-coherent participants. These cache maintenance operations
can be skipped if all PSCI participants are cache-coherent. When
HW_ASSISTED_COHERENCY build option is enabled, we assume PSCI
participants are cache-coherent.

For usage abstraction, this patch introduces wrappers for PSCI cache
maintenance and barrier operations used for state coordination: they are
effectively NOPs when HW_ASSISTED_COHERENCY is enabled, but are
applied otherwise.

Also refactor local state usage and associated cache operations to make
it clearer.

Change-Id: I77f17a90cba41085b7188c1345fe5731c99fad87
Signed-off-by: Jeenu Viswambharan <jeenu.viswambharan@arm.com>

ARM Trusted Firmware keeps certain data structures in a memory region
with non-cacheable attributes (termed as "coherent memory") to keep data
coherent with observers that are cache-coherent, and those not. These
data structures pertain to power management and mutual exclusion. Using
coherent memory also costs at least an additional page to map memory
with special memory attributes.

On systems with hardware-assisted coherency, all CPUs that participate
in power management and mutual exclusion are cache-coherent, obviating
the need for special memory attributes for such data structures.
Instead, they can be placed in normal memory, along with rest of data.

On systems with hardware-assisted coherency, where build option
HW_ASSISTED_COHERENCY will be set, also having USE_COHERENT_MEMORY
enabled only wastes a page of memory without any
benefit. Therefore, with HW_ASSISTED_COHERENCY set to 1, require that
USE_COHERENT_MEMORY is explicitly set to 0.

Change-Id: I5101657ae6b1a46278069f23e2d88ee5cbd98efa
Signed-off-by: Jeenu Viswambharan <jeenu.viswambharan@arm.com>

At present, warm-booted CPUs keep their caches disabled when enabling
MMU, and remains so until they enter coherency later.

On systems with hardware-assisted coherency, for which
HW_ASSISTED_COHERENCY build flag would be enabled, warm-booted CPUs can
have both caches and MMU enabled at once.

Change-Id: Icb0adb026e01aecf34beadf49c88faa9dd368327
Signed-off-by: Jeenu Viswambharan <jeenu.viswambharan@arm.com>

The boolean build option HW_ASSISTED_COHERENCY is introduced to enable
various optimizations in ARM Trusted Software, when built for such
systems. It's set to 0 by default.

Change-Id: I638390da6e1718fe024dcf5b402e07084f1eb014
Signed-off-by: Jeenu Viswambharan <jeenu.viswambharan@arm.com>

Remove redundant assert

Clarify errata ERRATA_A53_836870 documentation

Fix TSPD implementation of STD SMC ABORT

build: Use separate CFLAGS for tools

RK3399 ARM TF clean up 20170210

Tegra changes from downstream v2

Sorry to miss the security configuration for SRAM, if we don't support
it, somebody may modify the comment of SRAM in the non-secure space.
Let's fix this issue.
Signed-off-by: Xing Zheng <zhengxing@rock-chips.com>

This fixes code that set a tFC value in a register using the tRFC
value instead.
Signed-off-by: Derek Basehore <dbasehore@chromium.org>

The F1 CAS latency setting was not bit shifted, which resulted in
setting the DRAM additive latency value instead.
Signed-off-by: Derek Basehore <dbasehore@chromium.org>

On resume, we use the DFS hardware to switch frequency index,
followed by a full training sequence on that index. Leaving
the DFS training modules enabled causes issues with the full
training done at resume. We also only needs these enabled
during a call to ddr_set_rate during runtime, so there's no
issue disabling them at the end of ddr_set_rate.
Signed-off-by: Derek Basehore <dbasehore@chromium.org>
Signed-off-by: Xing Zheng <zhengxing@rock-chips.com>

This moves the setting of the DQS drive strength to the M0 to minimize
the impact on DDR transactions. We need to have the DQS drive strength
changed for data training, which is triggered by the M0, but it also
needs to be changed back when data training is finished.
Signed-off-by: Derek Basehore <dbasehore@chromium.org>

This removes an optimization to not recalculate parameters if the
frequency index being switched to hold the next frequency. This is
because some registers do not have a copy per frequency index, so this
optimization might be causing problems.
Signed-off-by: Derek Basehore <dbasehore@chromium.org>

We were getting far off values on resume for the RX_CAL_DQS values.
This saves and restores the values for suspend/resume until the root
of the problem is figured out
Signed-off-by: Derek Basehore <dbasehore@chromium.org>

These macros were accidentally deleted in a previous cleanup. This
slipped through because the code using them is currently unused, but
that may change in the future.
Signed-off-by: Julius Werner <jwerner@chromium.org>

This patch shuffles the M0 Makefile flags around a bit trying to make
their purpose clearer and remove duplication. Since all three build
steps (compiling, assembling, linking) actually call GCC, remove the
misleading aliases $(AS) and $(LD) to avoid confusion that those tools
might be called directly. Split flags into a common group that has
meaning for all three steps and separate variables specific to each
step. Remove -nostartfiles which is a strict subset of -nostdlib.

Also add explicit parameters for -mfloat-abi=soft, -fomit-frame-pointer
and -fno-common. If omitted these settings depend on the toolchain's
built-in default and cause various problems if they resolve to
unexpected values.
Signed-off-by: Julius Werner <jwerner@chromium.org>

The goal is that make clear the secure and SoC codes. Now cleaning them
will help secure code extensions for RK3399 in the future.
Signed-off-by: Xing Zheng <zhengxing@rock-chips.com>

Maybe the coreboot will reference the BL31 parameters (e.g the TZRAM_BASE
and TZRAM_SIZE for DDR secure regions), we can split them and don't have
to hardcode the range in two places.
Signed-off-by: Xing Zheng <zhengxing@rock-chips.com>

Move the BL31 loaded base address 0x10000 to 0x1000, and configure
the the memory range 0~1MB is secure, the goal is that make sure
the BL31 image will be not modified.
Signed-off-by: Xing Zheng <zhengxing@rock-chips.com>

So far, there are more and more features are supported on the RK3399,
meanwhile, these features are increasingly being defined and intertwined.
It's time to clean up and make them clearer.
Signed-off-by: Xing Zheng <zhengxing@rock-chips.com>

This removes waiting for vblank on the M0 during ddrfreq transitions.
That will now be done in the kernel to allow scheduling to be done on
the CPU core that changes the ddr frequency. Waiting for vblank in
the M0 would have the CPU core that waits on the M0 spin looping for
up to 16ms (1 frame for the display).
Signed-off-by: Derek Basehore <dbasehore@chromium.org>

we will set PMU_CRU_GATEDIS_CON0 when idle port, it will enable
all clock, for save power consumption, we need to restore old value
when finish it.
Signed-off-by: Lin Huang <hl@rock-chips.com>

As rk3399 TRM1.1 document show, when set PMU_CRU_GATEDIS_CON0/1
register, it need set the write_mask bit (bit16 ~ bit31), but as
we test, it not need it. So need to correct the setting way, otherwise
it will set wrong value to this register.
Signed-off-by: Lin Huang <hl@rock-chips.com>

We found that the DUT will be hanged if we don't set the bit_1 of the
PMUCRU_GATEDIS_CON0. But, from the TRM, there is weird that the bit_1
is set the clk_center1_gating_dis, not clk_pmum0_gating_dis. Is the
TRM incorrect? We need to check it with the IC team and re-clean the
commit message and explain it tomorrow.
Signed-off-by: Xing Zheng <zhengxing@rock-chips.com>

This patch do following things:
1. Request hresetn_cm0s_pmu_req first then request
   poresetn_cm0s_pmu_req during M0 enable.
2. Do not diable M0 clock for ddr dvfs.
3. Correct the clk_pmum0_gating_dis bit, it is BIT0 not BIT1
4. do not set/clear hclk_noc_pmu_en in M0 code, it does not relate
   to the M0 clock.
Signed-off-by: Lin Huang <hl@rock-chips.com>
Signed-off-by: Derek Basehore <dbasehore@chromium.org>

When vop is disabled and we read the vop register the system will
hang, so check vop status when we wait for the DMA finish flag to
avoid this sitiuation. This is done by checking for standby, DMA stop
mode, and disabled window states. Any one of these will prevent the
DMA finish flag from triggering.
Signed-off-by: Lin Huang <hl@rock-chips.com>
Signed-off-by: Derek Basehore <dbasehore@chromium.org>

There is system timer in m0, we can use it to implement a set of
stopwatch functions for measuring timeouts.
Signed-off-by: Lin Huang <hl@rock-chips.com>
Signed-off-by: Derek Basehore <dbasehore@chromium.org>

The phy pll needs to get 2X frequency to the DDR, so set the
pll_postdiv to 0.
Signed-off-by: Lin Huang <hl@rock-chips.com>
Signed-off-by: Derek Basehore <dbasehore@chromium.org>

For ddr dfs stable, We need to enable ddr CA training
when do ddr dfs.
Signed-off-by: Lin Huang <hl@rock-chips.com>

This fixes a hang with setting the DRAM rate based on a race condition
with the M0 which sets the DRAM rate. The AP can also starve the M0,
so this also delays the AP reads to the DONE parameter for the M0.
Signed-off-by: Derek Basehore <dbasehore@chromium.org>

This enables per CS training at 666MHz and above for ddrfreq per
vendor recommendation. Since the threshold was used for latency was
the same value, this also adds a new value for that.
Signed-off-by: Derek Basehore <dbasehore@chromium.org>

This patch sets the frequency configuration of the next DRAM DFS index
to the configuration of the current index. This does not perform a
frequency transition. It just configures registers so the training on
resume for both indices will be correct.
Signed-off-by: Derek Basehore <dbasehore@chromium.org>
Signed-off-by: Xing Zheng <zhengxing@rock-chips.com>

We used dcf do ddr frequency scaling, but we just include a dcf
binary, it hard to maintain later, we have M0 compile flow in ATF,
and M0 can also work for ddr frequency scaling, so let's use it.
Signed-off-by: Lin Huang <hl@rock-chips.com>
Signed-off-by: Xing Zheng <zhengxing@rock-chips.com>

This makes the file consistently use tabs instead of mixing in spaces.
Signed-off-by: Derek Basehore <dbasehore@chromium.org>

Since the ATF project, we usually use the mmio_read_32 and
mmio_write_32. And the mmio_write_32, the firse parameter
is ADDR, the second is VALUE. In order to style consistency:

1/ rename readl/writel to mmio_read_32/mmio_write_32
2/ for keeping the same with mmio_write_32 in the ATF project,
   swap the order of the parameters for M0 mmio_write_32
Signed-off-by: Xing Zheng <zhengxing@rock-chips.com>
Signed-off-by: Lin Huang <hl@rock-chips.com>

when dram frequency below 260MHz, phy master dll may unlock, so
let phy master dll working at dll bypass mode when frequency is
below 260MHz.
Signed-off-by: Lin Huang <hl@rock-chips.com>