[PATCH v2 1/2] lib: sbi: Wait in sbi_hsm_hart_wait() till the hart is prepared to start
Evgenii Shatokhin
e.shatokhin at yadro.com
Fri Mar 3 07:01:23 PST 2023
On 02.03.2023 07:07, Anup Patel wrote:
> On Tue, Feb 28, 2023 at 9:26 PM Evgenii Shatokhin <e.shatokhin at yadro.com> wrote:
>>
>> On 28.02.2023 08:06, Anup Patel wrote:
>>> On Wed, Feb 22, 2023 at 2:39 AM Evgenii Shatokhin <e.shatokhin at yadro.com> wrote:
>>>>
>>>> "wfi" instruction is not guaranteed to wait for an interrupt to be received
>>>> by the hart, it is just a hint for the CPU. According to RISC-V Privileged
>>>> Architectures spec. v20211203, even an implementation of "wfi" as "nop" is
>>>> legal.
>>>>
>>>> As a result, a secondary (non-boot) hart waiting in the while loop in
>>>> sbi_hsm_hart_wait() might leave the loop and continue even if it got no
>>>> IPI or it got an IPI unrelated to sbi_hsm_hart_start(). This could lead to
>>>> the following race condition when booting Linux, for example:
>>>>
>>>> Boot hart (#0) Secondary hart (#1)
>>>> runs Linux startup code waits in sbi_hsm_hart_wait()
>>>>
>>>> sbi_ecall(SBI_EXT_HSM,
>>>> SBI_EXT_HSM_HART_START,
>>>> ...)
>>>> enters sbi_hsm_hart_start()
>>>> sets state of hart #1 to START_PENDING
>>>> leaves sbi_hsm_hart_wait()
>>>> runs to the end of init_warmboot()
>>>> returns to scratch->next_addr
>>>> (next_addr can be garbage here)
>>>>
>>>> sets next_addr, etc. for hart #1
>>>> (no good: hart #1 has already left)
>>>>
>>>> sends IPI to hart #1
>>>> (no good either)
>>>>
>>>> If this happens, the secondary hart jumps to a wrong next_addr at the end
>>>> of init_warmboot(), which leads to a system hang or crash.
>>>>
>>>> To reproduce the issue more reliably, one could add a delay in
>>>> sbi_hsm_hart_start() after setting the hart's state but before sending
>>>> IPI to that hart:
>>>>
>>>> hstate = atomic_cmpxchg(&hdata->state, SBI_HSM_STATE_STOPPED,
>>>> SBI_HSM_STATE_START_PENDING);
>>>> ...
>>>> + sbi_timer_mdelay(10);
>>>> init_count = sbi_init_count(hartid);
>>>> rscratch->next_arg1 = arg1;
>>>> rscratch->next_addr = saddr;
>>>>
>>>> The issue can be reproduced, for example, in a QEMU VM with '-machine virt'
>>>> and 2 or more CPUs, with Linux as the guest OS.
>>>
>>> Good catch. This is a valid issue which needs to be fixed.
>>>
>>>>
>>>> This patch adds an auxiliary per-hart state variable, aux_state, as well as
>>>> its accessor functions, to synchronize execution of the boot hart and the
>>>> starting secondary hart.
>>>>
>>>> Initially, aux_state is 0: sbi_scratch_alloc_offset() zeroes the allocated
>>>> memory. The boot hart sets it to 1 after it has set next_addr, etc., for the
>>>> secondary hart, then sends the IPI to the latter.
>>>>
>>>> The secondary hart waits in sbi_hsm_hart_wait() both for aux_state to become
>>>> non-zero and for an IPI to happen. This should be safe even if the secondary
>>>> hart somehow gets an IPI unrelated to sbi_hsm_hart_start().
>>>>
>>>> Memory barriers are used to make sure the writes to next_addr, etc., in the
>>>> boot hart are visible to the secondary hart if it reads non-zero aux_state.
>>>>
>>>> Only one bit of aux_state is actually used here, so, this per-hart variable
>>>> can be used to hold other data in the future, if needed.
>>>
>>> Adding aux_state is an overly complicated solution in my opinion.
>>>
>>> Instead, you can simply do the following:
>>>
>>> rscratch->next_arg1 = arg1;
>>> rscratch->next_addr = saddr;
>>> rscratch->next_mode = smode;
>>>
>>> /*
>>> * Barrier to ensure that the above writes have happened before
>>> * updating the HART state.
>>> */
>>> smp_wmb();
>>>
>>> Before "hdata = sbi_scratch_offset_ptr(rscratch, hart_data_offset);" in
>>> sbi_hsm_hart_start().
>>
>> True, the proposed fix it is more complicated than I would like it to
>> be, but there is a reason.
>>
>> Should we worry about the scenario when two or more harts enter
>> sbi_hsm_hart_start() with the same target hartid? It must not happen
>> during normal operation of the OS, but, in theory, errors in the
>> higher-level software could lead to this.
>>
>> Setting hstate to SBI_HSM_STATE_START_PENDING atomically and checking
>> its previous values before changing *rscratch in sbi_hsm_hart_start()
>> guarantees that the current hart has exclusive access to *rscratch after
>> that. If we move rscratch->next_arg1 = arg1, etc., before hdata =
>> sbi_scratch_offset_ptr(...), it would be possible for two harts to race
>> and write different values to *rscratch for the same target hart. It is
>> very unlikely to happen, but still.
>>
>> If it is somehow guaranteed that the current hart has exclusive access
>> to *rscratch even at the beginning of sbi_hsm_hart_start(), then yes, we
>> can drop aux_state, etc., and use a simpler fix.
>>
>> smp_wmb() will then be paired with rmb() in atomic_read(&hdata->state)
>> from sbi_hsm_hart_wait() and it would make sure the writes are seen in
>> correct order.
>>
>> So, should we worry about misbehaving OSes which could allow 2 or more
>> harts to enter sbi_hsm_hart_start() for the same target hart in parallel?
>>
>> What do you think?
>
> I see your point but it is better to tackle the situation of 2 or more
> harts entering sbi_hsm_hart_start() rather than aux_state which
> less readable.
>
> How about this ?
>
> 1) Re-base on-top of the below patch
> 2) Update this patch to:
> a) Move "rscratch->next_*" updates before
> "hdata = sbi_scratch_offset_ptr(rscratch, hart_data_offset);"
> in sbi_hsm_hart_start()
> b) Add "start_ticket" which needs to be acquired (0->1) before
> "rscratch->next_*" updates in sbi_hsm_hart_start() and
> if a HART fails to acquire it then sbi_hsm_hart_start() fails
> c) Read the "scratch->next_" data and release "start_ticket" (1->0)
> in sbi_hsm_hart_start_finish() before doing sbi_hart_switch_mode().
>
If 'start_ticket' is a global, rather than a per-hart variable, no hart
would be allowed to run sbi_hsm_hart_start() till the secondary hart
finishes init_warm_startup() and releases the ticket. This could prevent
the boot hart to call sbi_hsm_hart_start() for other harts, e.g. if the
boot hart executes sbi_hsm_hart_start() quickly, while the secondary
hart is slow.
I think, it would be sufficient to prevent entering sbi_hsm_hart_start()
only for the *same* target hart rather than for *all* target harts.
If it is OK to add fields to 'struct sbi_hsm_data', we could add
'start_ticket' there, instead of 'aux_data', and use it as you suggested.
Would that be OK?
Regards,
Evgenii
> Regards,
> Anup
>
> diff --git a/include/sbi/sbi_hsm.h b/include/sbi/sbi_hsm.h
> index c0b4830..4b5601b 100644
> --- a/include/sbi/sbi_hsm.h
> +++ b/include/sbi/sbi_hsm.h
> @@ -66,7 +66,8 @@ int sbi_hsm_hart_start(struct sbi_scratch *scratch,
> u32 hartid, ulong saddr, ulong smode, ulong arg1);
> int sbi_hsm_hart_stop(struct sbi_scratch *scratch, bool exitnow);
> void sbi_hsm_hart_resume_start(struct sbi_scratch *scratch);
> -void sbi_hsm_hart_resume_finish(struct sbi_scratch *scratch);
> +void __noreturn sbi_hsm_hart_resume_finish(struct sbi_scratch *scratch,
> + u32 hartid);
> int sbi_hsm_hart_suspend(struct sbi_scratch *scratch, u32 suspend_type,
> ulong raddr, ulong rmode, ulong arg1);
> bool sbi_hsm_hart_change_state(struct sbi_scratch *scratch, long oldstate,
> @@ -76,6 +77,7 @@ int sbi_hsm_hart_get_state(const struct sbi_domain
> *dom, u32 hartid);
> int sbi_hsm_hart_interruptible_mask(const struct sbi_domain *dom,
> ulong hbase, ulong *out_hmask);
> void __sbi_hsm_suspend_non_ret_save(struct sbi_scratch *scratch);
> -void sbi_hsm_prepare_next_jump(struct sbi_scratch *scratch, u32 hartid);
> +void __noreturn sbi_hsm_hart_start_finish(struct sbi_scratch *scratch,
> + u32 hartid);
>
> #endif
> diff --git a/lib/sbi/sbi_hsm.c b/lib/sbi/sbi_hsm.c
> index 3eeeb58..100b8c0 100644
> --- a/lib/sbi/sbi_hsm.c
> +++ b/lib/sbi/sbi_hsm.c
> @@ -110,7 +110,8 @@ int sbi_hsm_hart_interruptible_mask(const struct
> sbi_domain *dom,
> return 0;
> }
>
> -void sbi_hsm_prepare_next_jump(struct sbi_scratch *scratch, u32 hartid)
> +void __noreturn sbi_hsm_hart_start_finish(struct sbi_scratch *scratch,
> + u32 hartid)
> {
> struct sbi_hsm_data *hdata = sbi_scratch_offset_ptr(scratch,
> hart_data_offset);
> @@ -118,6 +119,9 @@ void sbi_hsm_prepare_next_jump(struct sbi_scratch
> *scratch, u32 hartid)
> if (!__sbi_hsm_hart_change_state(hdata, SBI_HSM_STATE_START_PENDING,
> SBI_HSM_STATE_STARTED))
> sbi_hart_hang();
> +
> + sbi_hart_switch_mode(hartid, scratch->next_arg1, scratch->next_addr,
> + scratch->next_mode, false);
> }
>
> static void sbi_hsm_hart_wait(struct sbi_scratch *scratch, u32 hartid)
> @@ -381,7 +385,8 @@ void sbi_hsm_hart_resume_start(struct sbi_scratch *scratch)
> hsm_device_hart_resume();
> }
>
> -void sbi_hsm_hart_resume_finish(struct sbi_scratch *scratch)
> +void __noreturn sbi_hsm_hart_resume_finish(struct sbi_scratch *scratch,
> + u32 hartid)
> {
> struct sbi_hsm_data *hdata = sbi_scratch_offset_ptr(scratch,
> hart_data_offset);
> @@ -396,6 +401,10 @@ void sbi_hsm_hart_resume_finish(struct
> sbi_scratch *scratch)
> * the warm-boot sequence.
> */
> __sbi_hsm_suspend_non_ret_restore(scratch);
> +
> + sbi_hart_switch_mode(hartid, scratch->next_arg1,
> + scratch->next_addr,
> + scratch->next_mode, false);
> }
>
> int sbi_hsm_hart_suspend(struct sbi_scratch *scratch, u32 suspend_type,
> diff --git a/lib/sbi/sbi_init.c b/lib/sbi/sbi_init.c
> index bc60a42..dcca2c8 100644
> --- a/lib/sbi/sbi_init.c
> +++ b/lib/sbi/sbi_init.c
> @@ -355,12 +355,11 @@ static void __noreturn init_coldboot(struct
> sbi_scratch *scratch, u32 hartid)
> init_count = sbi_scratch_offset_ptr(scratch, init_count_offset);
> (*init_count)++;
>
> - sbi_hsm_prepare_next_jump(scratch, hartid);
> - sbi_hart_switch_mode(hartid, scratch->next_arg1, scratch->next_addr,
> - scratch->next_mode, false);
> + sbi_hsm_hart_start_finish(scratch, hartid);
> }
>
> -static void init_warm_startup(struct sbi_scratch *scratch, u32 hartid)
> +static void __noreturn init_warm_startup(struct sbi_scratch *scratch,
> + u32 hartid)
> {
> int rc;
> unsigned long *init_count;
> @@ -412,10 +411,11 @@ static void init_warm_startup(struct sbi_scratch
> *scratch, u32 hartid)
> init_count = sbi_scratch_offset_ptr(scratch, init_count_offset);
> (*init_count)++;
>
> - sbi_hsm_prepare_next_jump(scratch, hartid);
> + sbi_hsm_hart_start_finish(scratch, hartid);
> }
>
> -static void init_warm_resume(struct sbi_scratch *scratch)
> +static void __noreturn init_warm_resume(struct sbi_scratch *scratch,
> + u32 hartid)
> {
> int rc;
>
> @@ -429,7 +429,7 @@ static void init_warm_resume(struct sbi_scratch *scratch)
> if (rc)
> sbi_hart_hang();
>
> - sbi_hsm_hart_resume_finish(scratch);
> + sbi_hsm_hart_resume_finish(scratch, hartid);
> }
>
> static void __noreturn init_warmboot(struct sbi_scratch *scratch, u32 hartid)
> @@ -443,13 +443,9 @@ static void __noreturn init_warmboot(struct
> sbi_scratch *scratch, u32 hartid)
> sbi_hart_hang();
>
> if (hstate == SBI_HSM_STATE_SUSPENDED)
> - init_warm_resume(scratch);
> + init_warm_resume(scratch, hartid);
> else
> init_warm_startup(scratch, hartid);
> -
> - sbi_hart_switch_mode(hartid, scratch->next_arg1,
> - scratch->next_addr,
> - scratch->next_mode, false);
> }
>
> static atomic_t coldboot_lottery = ATOMIC_INITIALIZER(0);
>
More information about the opensbi
mailing list