> -----Original Message-----
> From: linux-scsi-ow...@vger.kernel.org [mailto:linux-scsi-
> ow...@vger.kernel.org] On Behalf Of Alex Thorlton
> Sent: Tuesday, 02 December, 2014 3:58 PM
...
> We've recently upgraded our big machine up to 6144 cores, and we're
> shaking out a number of bugs related to booting at that large core
> count. Last night I tripped a warning from the lpfc driver that appears
> to be related to a kzalloc that uses the number of cores as part of it's
> size calculation. Here's the backtrace from the warning:
...
> For a little bit more information on exactly what's going wrong, we're
> tripping the warning from lpfc_pci_probe_one_s4 (as you can see from the
> trace). That function calls down to lpfc_sli4_driver_resource_setup,
> which contains the failing kzalloc here:
>
> phba->sli4_hba.cpu_map = kzalloc((sizeof(struct lpfc_vector_map_info) *
> phba->sli4_hba.num_present_cpu),
> GFP_KERNEL);
>
> As mentioned, it looks like we're multiplying the number available cpus
> by that struct size to get an allocation size, which ends up being
> greater than KMALLOC_MAX_SIZE.
>
> Does anyone have any ideas on what could be done to break that
> allocation up into smaller pieces, or to make it in a different way so
> that we avoid this warning?
>
> Any help is greatly appreciated. Thanks!
>
That structure includes an NR_CPU-based maskbits field, which is
probably too big.
include/cpumask.h:
typedef struct cpumask { DECLARE_BITMAP(bits, NR_CPUS); } cpumask_t;
drivers/scsi/lpfc/lpfc_sli4.h:
struct lpfc_vector_map_info {
uint16_t phys_id;
uint16_t core_id;
uint16_t irq;
uint16_t channel_id;
struct cpumask maskbits;
};
maskbits appears to only be used for setting IRQ affinity hints in
drivers/scsi/lpfc_init.c:
for (idx = 0; idx < vectors; idx++) {
cpup = phba->sli4_hba.cpu_map;
cpu = lpfc_find_next_cpu(phba, phys_id);
...
mask = &cpup->maskbits;
cpumask_clear(mask);
cpumask_set_cpu(cpu, mask);
i = irq_set_affinity_hint(phba->sli4_hba.msix_entries[idx].
vector, mask);
In similar code, mpt3sas and lockless hpsa just call get_cpu_mask()
inside the loop:
cpu = cpumask_first(cpu_online_mask);
for (i = 0; i < h->msix_vector; i++) {
rc = irq_set_affinity_hint(h->intr[i], get_cpu_mask(cpu));
cpu = cpumask_next(cpu, cpu_online_mask);
}
get_cpu_mask() uses the global cpu_bit_bitmap array, which is declared
in kernel/cpu.c:
extern const unsigned long
cpu_bit_bitmap[BITS_PER_LONG+1][BITS_TO_LONGS(NR_CPUS)];
That approach should work for lpfc.
---
Rob Elliott HP Server Storage
--
To unsubscribe from this list: send the line "unsubscribe linux-scsi" in
the body of a message to majord...@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
