On 06/27/2017 11:01 AM, Mark Rutland wrote:
> On Mon, Jun 26, 2017 at 03:32:44PM -0700, Florian Fainelli wrote:
>> From: Brian Norris <computersforpeace@xxxxxxxxx>
>>
>> This commit adds support for the Broadcom STB S2/S3/S5 suspend states on
>> ARM based SoCs.
>>
>> This requires quite a lot of code in order to deal with the different HW
>> blocks that need to be quiesced during suspend:
>>
>> - DDR PHY SHIM
>> - DDR memory controller and sequencer
>> - control processor
>>
>> The final steps of the suspend execute in an on-chip SRAM and there is a
>> little bit of assembly code in order to shut down the DDR PHY PLL and
>> then go into a wfi loop until a wake-up even occurs. Conversely the
>> resume part involves waiting for the DDR PHY PLL to come back up and
>> resume executions where we left.
>>
>> For S3, because of our memory hashing (actual hashing code not included
>> for simplicity, and is bypassed) we need to relocate the writable
>> variables (stack) into SRAM shortly before suspending in order to leave
>> the DRAM untouched and create a reliable hash of its contents.
>>
>> This code has been contributed by Brian Norris initially and has been
>> incrementally fixed and updated to support new chips by a lot of people.
>
> [...]
>
>> +static int (*brcmstb_pm_do_s2_sram)(void __iomem *aon_ctrl_base,
>> + void __iomem *ddr_phy_pll_status);
>> +
>> +static int brcmstb_init_sram(struct device_node *dn)
>> +{
>> + void __iomem *sram;
>> + struct resource res;
>> + int ret;
>> +
>> + ret = of_address_to_resource(dn, 0, &res);
>> + if (ret)
>> + return ret;
>> +
>> + /* Cached, executable remapping of SRAM */
>> +#ifdef CONFIG_ARM
>> + sram = __arm_ioremap_exec(res.start, resource_size(&res), true);
>> +#else
>> + sram = __ioremap(res.start, resource_size(&res), PAGE_KERNEL_EXEC);
>> +#endif
>
> ... so we're mapping the SRAM as executable via the page tables (which
> themselves live in memory, and are accessed asynchronously by the MMU) ...
>
>> +static void *brcmstb_pm_copy_to_sram(void *fn, size_t len)
>> +{
>> + unsigned int size = ALIGN(len, FNCPY_ALIGN);
>> +
>> + if (ctrl.boot_sram_len < size) {
>> + pr_err("standby code will not fit in SRAM\n");
>> + return NULL;
>> + }
>> +
>> + return fncpy(ctrl.boot_sram, fn, size);
>> +}
>> +
>> +/*
>> + * S2 suspend/resume picks up where we left off, so we must execute carefully
>> + * from SRAM, in order to allow DDR to come back up safely before we continue.
>> + */
>> +static int brcmstb_pm_s2(void)
>> +{
>> + /* A previous S3 can set a value hazardous to S2, so make sure. */
>> + if (ctrl.s3entry_method == 1) {
>> + shimphy_set((PWRDWN_SEQ_NO_SEQUENCING <<
>> + SHIMPHY_PAD_S3_PWRDWN_SEQ_SHIFT),
>> + ~SHIMPHY_PAD_S3_PWRDWN_SEQ_MASK);
>> + ddr_ctrl_set(false);
>> + }
>> +
>> + brcmstb_pm_do_s2_sram = brcmstb_pm_copy_to_sram(&brcmstb_pm_do_s2,
>> + brcmstb_pm_do_s2_sz);
>> + if (!brcmstb_pm_do_s2_sram)
>> + return -EINVAL;
>> +
>> + return brcmstb_pm_do_s2_sram(ctrl.aon_ctrl_base,
>> + ctrl.memcs[0].ddr_phy_base +
>> + ctrl.pll_status_offset);
>> +}
>
> ... here we copy the function into that executable SRAM, and then try to
> execute it, with the MMU on ...
>
>> +#define SWAP_STACK(new_sp, saved_sp) \
>> + __asm__ __volatile__ ( \
>> + "mov %[save], sp\n" \
>> + "mov sp, %[new]\n" \
>> + : [save] "=&r" (saved_sp) \
>> + : [new] "r" (new_sp) \
>> + )
>> +
>> +/*
>> + * S3 mode resume to the bootloader before jumping back to Linux, so we can be
>> + * a little less careful about running from DRAM.
>> + */
>> +static int brcmstb_pm_do_s3(unsigned long sp)
>> +{
>> + unsigned long save_sp;
>> + int ret;
>> +
>> + /* Move to a new stack */
>> + SWAP_STACK(sp, save_sp);
>> +
>> + /* should not return */
>> + ret = brcmstb_pm_s3_finish();
>> +
>> + SWAP_STACK(save_sp, sp);
>> +
>> + pr_err("Could not enter S3\n");
>> +
>> + return ret;
>> +}
>
> The compiler can, and almost certainly will spill stack variables. You
> cannot swap the stack safely with inline asm like this. If you need to
> do this, you need to perform the whole swap-call-swap sequence in a
> single inline asm block.
Makes sense, I don't think we saw problems with that, but you are right,
better safe than sorry.
>
> [...]
>
>> +ENTRY(brcmstb_pm_do_s2)
>> + stmfd sp!, {r4-r11, lr}
>> + mov AON_CTRL_REG, r0
>> + mov DDR_PHY_STATUS_REG, r1
>> +
>> + /* Flush memory transactions */
>> + dsb
>> +
>> + /* Cache DDR_PHY_STATUS_REG translation */
>> + ldr r0, [DDR_PHY_STATUS_REG]
>> +
>> + /* power down request */
>> + ldr r0, =PM_S2_COMMAND
>> + ldr r1, =0
>> + str r1, [AON_CTRL_REG, #AON_CTRL_PM_CTRL]
>> + ldr r1, [AON_CTRL_REG, #AON_CTRL_PM_CTRL]
>> + str r0, [AON_CTRL_REG, #AON_CTRL_PM_CTRL]
>> + ldr r0, [AON_CTRL_REG, #AON_CTRL_PM_CTRL]
>> +
>> + /* Wait for interrupt */
>> + wfi
>> + nop
>> +
>> + /* Bring MEMC back up */
>> +1: ldr r0, [DDR_PHY_STATUS_REG]
>> + ands r0, #1
>> + beq 1b
>> +
>> + /* Power-up handshake */
>> + ldr r0, =1
>> + str r0, [AON_CTRL_REG, #AON_CTRL_HOST_MISC_CMDS]
>> + ldr r0, [AON_CTRL_REG, #AON_CTRL_HOST_MISC_CMDS]
>> +
>> + ldr r0, =0
>> + str r0, [AON_CTRL_REG, #AON_CTRL_PM_CTRL]
>> + ldr r0, [AON_CTRL_REG, #AON_CTRL_PM_CTRL]
>> +
>> + /* Return to caller */
>> + ldr r0, =0
>> + ldmfd sp!, {r4-r11, pc}
>> +
>> + ENDPROC(brcmstb_pm_do_s2)
>
> What happens to any page table walks or speculative fetches that occur
> in the window where the DDR is off?
We actually saw a page table walk miss, which is why we cache the
DDR_PHY_STATUS_REG translation by accessing it prior to suspending
(confirmed by seeing a TLB miss using the PMU). In practice this
resulted in an abnormally longer resume delay (200ms vs. several
microseconds) but no crashes. Other than that, I recall now that Russell
suggested making the SRAM __arm_ioremap_exec() non-cacheable for safety,
I will do that in v3.
>
> If the former can be stalled, the CPU can deadlock. If you can any sort
> of external abort as a result of either, the core will go into nowhere
> land trying to handle it.
>
> I do not think this is safe.
Well, this code has been used in production for millions of
suspend/resume cycles and has proven to be robust enough, I don't object
the scenario you are describing could exist, but I am not sure we could
come up with an alternative which is not incredibly complex.
A possibly different approach I suppose could be to relocate this code
into an identify mapping and do a late MMU off, suspend, followed by DDR
bring-up and MMU on, that sounds much more complex, and we would have to
run this through our regression test suite for several thousands of
cycles to confirm the validity of this approach.
Thanks for the feedback!
--
Florian
