Change Details

Some counter arrays are initialized before APs are started. For instance, vmcounter_startup() runs at SI_SUB_KMEM, before SI_SUB_CPU. Prior to r336020, counter_u64_alloc() initialized the counters using an smp_rendezvous(). For early initializations, this ran before smp_started was set, so we'd only initialize the counter for CPU 0, and apparently we relied on the rest of the array already being zeroed(!). With r336020 we now initialize the counters using a CPU_FOREACH() loop, and this revealed a problem. On arm64, we do not set the bit for cpuid 0 in all_cpus until SI_SUB_CPU, so the CPU_FOREACH() loop skips CPU 0 (because CPU_ABSENT(0) is true) when initializing some counters. Thus the trash value 0xdeadc0de is left as the counter value. Fix the problem by using M_ZERO with UMA, and modify UMA to explicitly iterate over all cpuids. I do not believe that we can defer initialization of vm_cnt to after SI_SUB_CPU.

Some counter arrays are initialized before APs are started. For instance, vmcounter_startup() runs at SI_SUB_KMEM, before SI_SUB_CPU. Prior to r336020, counter_u64_alloc() initialized the counters using an smp_rendezvous(). For early initializations, this ran before smp_started was set, so we'd only initialize the counter for CPU 0, and apparently we relied on the rest of the counters already being zeroed(!). With r336020 we now initialize the counters using a CPU_FOREACH() loop, and this revealed a problem. On arm64, we do not set the bit for cpuid 0 in all_cpus until SI_SUB_CPU, so the CPU_FOREACH() loop skips CPU 0 (because CPU_ABSENT(0) is true) when initializing some counters. Thus the trash value 0xdeadc0de is left as the counter value. Fix the problem by using M_ZERO with UMA, and modify UMA to explicitly iterate over all cpuids. I do not believe that we can defer initialization of vm_cnt to after SI_SUB_CPU.

Some counter arrays are initialized before APs are started. For instance, vmcounter_startup() runs at SI_SUB_KMEM, before SI_SUB_CPU. Prior to r336020, counter_u64_alloc() initialized the counters using an smp_rendezvous(). For early initializations, this ran before smp_started was set, so we'd only initialize the counter for CPU 0, and apparently we relied on the rest of the arraycounters already being zeroed(!). With r336020 we now initialize the counters using a CPU_FOREACH() loop, and this revealed a problem. On arm64, we do not set the bit for cpuid 0 in all_cpus until SI_SUB_CPU, so the CPU_FOREACH() loop skips CPU 0 (because CPU_ABSENT(0) is true) when initializing some counters. Thus the trash value 0xdeadc0de is left as the counter value. Fix the problem by using M_ZERO with UMA, and modify UMA to explicitly iterate over all cpuids. I do not believe that we can defer initialization of vm_cnt to after SI_SUB_CPU.