[lvc-project] [PATCH v2] The value may overflow
Mathieu Desnoyers
mathieu.desnoyers at efficios.com
Tue Sep 5 22:36:40 MSK 2023
On 9/5/23 15:27, Paul E. McKenney wrote:
> On Tue, Sep 05, 2023 at 09:40:51AM -0700, Paul E. McKenney wrote:
>> On Tue, Sep 05, 2023 at 10:34:43AM -0400, Mathieu Desnoyers wrote:
>>> On 9/5/23 10:15, David Laight wrote:
>>>> ...
>>>>> That would instead be more than 512-16=496 CPUs, correct? 496 CPUs would
>>>>> only require a 31-bit shift, which should be OK, but 497 would require
>>>>> a 32-bit shift, which would result in sign extension. If it turns out
>>>>> that sign extension is OK, then we should get in trouble at 513 CPUs,
>>>>> which would result in a 33-bit shift (and is that even defined in C?).
>>>>
>>>> Not quite right :-)
>>>>
>>>> (1 << 31) is int and negative, that gets sign extended before
>>>> being converted to 'unsigned long' - so has the top 33 bits set.
>>
>> Good point, thank you for the correction.
>>
>>>> (1 << 32) is undefined, the current x86 cpu ignore the high
>>>> shift bits so it is (1 << 0).
>>
>> Which would cause it to attempt to invoke SRCU callbacks on the
>> lowest-numbered CPU associated with that srcu_node structure.
>>
>>> Yes, I was about to reply the same thing. A shift of 31 is buggy,
>>> because shifting 1 << 31 raises the sign bit, which sets the top 33
>>> bits when cast to unsigned long. A shift of 1 << 32 is undefined,
>>> with for instance x86 choosing to ignore the top bit.
>>>
>>> But in order to have a 1 << 31 shift from this expression:
>>>
>>> sdp->grpmask = 1 << (cpu - sdp->mynode->grplo);
>>>
>>> I *think* we need the group to have 32 cpus or more (indexed within
>>> the group from grplo to grplo + 31 (both inclusive)).
>>>
>>> So as soon as we have one group with 32 cpus, the problem should show
>>> up. With FANOUT_LEAF=16, we can have 15 groups of 31 cpus and 1
>>> group of 32 cpus, for:
>>>
>>> 15*31 + 32 = 497 cpus.
>>>
>>> AFAIU, this would be the minimum value for smp_processor_id()+1 which
>>> triggers this issue.
>>
>> By default, there are 16 CPUs per leaf srcu_node structure. Each leaf
>> srcu_node structure takes up one bit in the parent srcu_node structures'
>> srcu_data_have_cbs[] array. Up to 30 bits is OK, 31 bits is questionable,
>> and 32 bits and larger erroneous.
>>
>> This is the situation as I see it (assuming dense CPU numbering):
>>
>> # Leaf srcu_node Largest
>> structures #CPUs CPU # Status
>>
>> 0-30 0-480 479 Good
>> 31 481-496 495 Questionable
>> 32- 497- 496+ Bad.
>>
>> Tree SRCU differs from Tree RCU in its operation, so this bug should
>> not hold up SRCU grace periods. It might instead cause SRCU callbacks
>> to be ignored (which would admittedly look quite similar to the user).
>>
>> My attempts to cheat the numbering system ran up against the limited
>> height of the srcu_node tree.
>>
>> But there is still the question of why this has not been seen in the
>> wild, given that there really are systems with more than 479 CPUs
>> out there. One possibility is the call to srcu_schedule_cbs_sdp()
>> from srcu_funnel_gp_start(). But it does not seem likely that this
>> would happen all that often, as it requires back-to-back grace periods
>> and then some.
>>
>> Maybe people with large systems boot with srcutree.convert_to_big=0?
>>
>> Adding Laurent for his thoughts.
>>
>> Aha!
>>
>> Here is what makes it work, given David's description of 1<<32:
>>
>> static void srcu_schedule_cbs_snp(struct srcu_struct *ssp, struct srcu_node *snp,
>> unsigned long mask, unsigned long delay)
>> {
>> int cpu;
>>
>> for (cpu = snp->grplo; cpu <= snp->grphi; cpu++) {
>> if (!(mask & (1 << (cpu - snp->grplo))))
>> continue;
>> srcu_schedule_cbs_sdp(per_cpu_ptr(ssp->sda, cpu), delay);
>> }
>> }
>>
>> As long as at least one bit is set in the result of 1<<N, and as long
>> as the compiler always does the same thing for a given N, then this loop
>> will make the right thing happen.
>>
>> But even that is not relied upon, because the calling code looks like
>> this:
>>
>> spin_lock_irq_rcu_node(snp);
>> cbs = false;
>> last_lvl = snp >= sup->level[rcu_num_lvls - 1];
>> if (last_lvl)
>> cbs = ss_state < SRCU_SIZE_BIG || snp->srcu_have_cbs[idx] == gpseq;
>> snp->srcu_have_cbs[idx] = gpseq;
>> rcu_seq_set_state(&snp->srcu_have_cbs[idx], 1);
>> sgsne = snp->srcu_gp_seq_needed_exp;
>> if (srcu_invl_snp_seq(sgsne) || ULONG_CMP_LT(sgsne, gpseq))
>> WRITE_ONCE(snp->srcu_gp_seq_needed_exp, gpseq);
>> if (ss_state < SRCU_SIZE_BIG)
>> mask = ~0;
>> else
>> mask = snp->srcu_data_have_cbs[idx];
>> snp->srcu_data_have_cbs[idx] = 0;
>> spin_unlock_irq_rcu_node(snp);
>> if (cbs)
>> srcu_schedule_cbs_snp(ssp, snp, mask, cbdelay);
>>
>> So that last_lvl check means that the srcu_schedule_cbs_snp() function
>> is invoked only for leaf srcu_node structures. Which by default limit
>> the shift to 16.
>>
>> So this bug appears to have no effect for default RCU setups, even on very
>> large 64-bit systems, which is consistent with field experience. Even if
>> this is the case, it still should be fixed, to avoid confusion if nothing
>> else. Or just in case someone decides to set CONFIG_RCU_FANOUT_LEAF=32.
>> Which actually happened the other day due to someone trusting ChatGPT's
>> opinion about RCU Kconfig options...
>
> And I therefore queued Denis's v3 patch with an edited commit log.
> Of course, if anyone sees some other way that the bug could manifest
> other than in a 64-bit kernel built with CONFIG_RCU_FANOUT_LEAF greater
> than 30 on a system with at least 31 CPUs, please let me know so that
> I can adjust.
>
> Thanx, Paul
>
> ------------------------------------------------------------------------
>
> commit ed083b0e22f1396dee3599896249a3f218845298
> Author: Denis Arefev <arefev at swemel.ru>
> Date: Mon Sep 4 15:21:14 2023 +0300
>
> Fix srcu_struct node grpmask overflow on 64-bit systems
>
> The value of an arithmetic expression 1 << (cpu - sdp->mynode->grplo)
AFAIU, the overflow resides in the "bitwise expression" and not
the arithmetic expression.
Other than this, please add my
Reviewed-by: Mathieu Desnoyers <mathieu.desnoyers at efficios.com>
> is subject to overflow due to a failure to cast operands to a larger
> data type before performing arithmetic.
>
> The maximum result of this subtraction is defined by the RCU_FANOUT_LEAF
> Kconfig option, which on 64-bit systems defaults to 16 (resulting in a
> maximum shift of 15), but which can be set up as high as 64 (resulting
> in a maximum shift of 63). A value of 31 can result in sign extension,
> resulting in 0xffffffff80000000 instead of the desired 0x80000000.
> A value of 31 or greater triggers undefined behavior per the C standard.
>
> This bug has not been known to cause issues because almost all kernels
> take the default CONFIG_RCU_FANOUT_LEAF=16. Furthermore, as long as
> a given compiler gives a deterministic result for 1<<N for N>=32, the
> code correctly invokes all SRCU callbacks, albeit wasting CPU time along
> the way.
>
> This commit therefore substitutes the correct 1UL for the buggy 1.
>
> Found by Linux Verification Center (linuxtesting.org) with SVACE.
>
> Signed-off-by: Denis Arefev <arefev at swemel.ru>
> Cc: Mathieu Desnoyers <mathieu.desnoyers at efficios.com>
> Cc: David Laight <David.Laight at aculab.com>
> Signed-off-by: Paul E. McKenney <paulmck at kernel.org>
>
> diff --git a/kernel/rcu/srcutree.c b/kernel/rcu/srcutree.c
> index 833a8f848a90..5602042856b1 100644
> --- a/kernel/rcu/srcutree.c
> +++ b/kernel/rcu/srcutree.c
> @@ -223,7 +223,7 @@ static bool init_srcu_struct_nodes(struct srcu_struct *ssp, gfp_t gfp_flags)
> snp->grplo = cpu;
> snp->grphi = cpu;
> }
> - sdp->grpmask = 1 << (cpu - sdp->mynode->grplo);
> + sdp->grpmask = 1UL << (cpu - sdp->mynode->grplo);
> }
> smp_store_release(&ssp->srcu_sup->srcu_size_state, SRCU_SIZE_WAIT_BARRIER);
> return true;
> @@ -835,7 +835,7 @@ static void srcu_schedule_cbs_snp(struct srcu_struct *ssp, struct srcu_node *snp
> int cpu;
>
> for (cpu = snp->grplo; cpu <= snp->grphi; cpu++) {
> - if (!(mask & (1 << (cpu - snp->grplo))))
> + if (!(mask & (1UL << (cpu - snp->grplo))))
> continue;
> srcu_schedule_cbs_sdp(per_cpu_ptr(ssp->sda, cpu), delay);
> }
--
Mathieu Desnoyers
EfficiOS Inc.
https://www.efficios.com
More information about the lvc-project
mailing list