Re: [Xen-devel] [PATCH 1/4] x86: move syscall trampolines off the stack

["Jan Beulich" <JBeulich@xxxxxxxx>]

>>> On 19.05.15 at 18:59, <andrew.cooper3@xxxxxxxxxx> wrote:
> On 18/05/15 13:46, Jan Beulich wrote:
>> @@ -616,6 +653,24 @@ static void cpu_smpboot_free(unsigned in
>>      free_cpumask_var(per_cpu(cpu_sibling_mask, cpu));
>>      free_cpumask_var(per_cpu(cpu_core_mask, cpu));
>>  
>> +    if ( per_cpu(stubs.addr, cpu) )
>> +    {
>> +        unsigned long mfn = per_cpu(stubs.mfn, cpu);
>> +        unsigned char *stub_page = map_domain_page(mfn);
>> +        unsigned int i;
>> +
>> +        memset(stub_page + (cpu & (STUBS_PER_PAGE - 1)) * STUB_BUF_SIZE,
>> +               0xcc, STUB_BUF_SIZE);
>> +        for ( i = 0; i < STUBS_PER_PAGE; ++i )
>> +            if ( stub_page[i * STUB_BUF_SIZE] != 0xcc)
>> +                break;
> 
> There is a race condition between allocate and free, as
> write_stub_trampoline() is written by the cpu itself.  Just finding 0xcc
> here doesn't mean there isn't a different cpu in the process of coming
> up and intending to use the stub page.

No, there is no race afaict: percpu_traps_init() gets called before
a CPU is marked online and hence before __cpu_up() returns, and
thus before the CPU hotplug lock gets dropped, so there can't be
any CPUs going down concurrently.

Jan


_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxx
http://lists.xen.org/xen-devel