Re: [PATCH V6 4/5] perf/x86/intel/uncore: Fix PMON enumeration with NUMA disabled

From: Chen, Zide

Date: Thu Apr 02 2026 - 17:32:04 EST

On 4/1/2026 7:48 PM, Mi, Dapeng wrote:
>
> On 4/2/2026 4:25 AM, Chen, Zide wrote:
>>
>> On 3/30/2026 6:26 PM, Mi, Dapeng wrote:
>>> On 3/31/2026 5:24 AM, Zide Chen wrote:
>>>> When NUMA is disabled on a NUMA-capable platform, UPI and M3UPI PMON
>>>> units are not enumerated.
>>>>
>>>> In this case, pcibus_to_node() always returns NUMA_NO_NODE, causing
>>>> uncore_device_to_die() to return -1 for all PCI devices. As a result,
>>>> the corresponding PMON units are not added to the RB tree.
>>>>
>>>> These PMON units are per-die resources, and their utility when NUMA is
>>>> disabled is limited. The driver does not prohibit their use, and the
>>>> enumeration should still work correctly.
>>>>
>>>> Fix this by using uncore_pcibus_to_dieid(), which works regardless of
>>>> whether NUMA is enabled. This requires calling
>>>> snbep_pci2phy_map_init() in spr_uncore_pci_init().
>>>>
>>>> Since pci_init() is called before mmio_init(), remove the redundant
>>>> snbep_pci2phy_map_init() call from spr_uncore_mmio_init(). If
>>>> snbep_pci2phy_map_init() fails, uncore driver should be bailed out,
>>>> so the fallback path in spr_uncore_mmio_init() can be removed.
>>>>
>>>> Signed-off-by: Zide Chen <zide.chen@xxxxxxxxx>
>>>> ---
>>>> V6:
>>>> - Split from patch v5 3/4.
>>>> - Remove the redundant call in spr_uncore_mmio_init().
>>>> - Update commit messages.
>>>> ---
>>>> arch/x86/events/intel/uncore.c | 1 +
>>>> arch/x86/events/intel/uncore_snbep.c | 26 +++++++++++---------------
>>>> 2 files changed, 12 insertions(+), 15 deletions(-)
>>>>
>>>> diff --git a/arch/x86/events/intel/uncore.c b/arch/x86/events/intel/uncore.c
>>>> index 786bd51a0d89..e9cc1ba921c5 100644
>>>> --- a/arch/x86/events/intel/uncore.c
>>>> +++ b/arch/x86/events/intel/uncore.c
>>>> @@ -67,6 +67,7 @@ int uncore_die_to_segment(int die)
>>>> return bus ? pci_domain_nr(bus) : -EINVAL;
>>>> }
>>>>
>>>> +/* Note: This API can only be used when NUMA information is available. */
>>>> int uncore_device_to_die(struct pci_dev *dev)
>>>> {
>>>> int node = pcibus_to_node(dev->bus);
>>>> diff --git a/arch/x86/events/intel/uncore_snbep.c b/arch/x86/events/intel/uncore_snbep.c
>>>> index 8ee06d4659bb..73da1e88e286 100644
>>>> --- a/arch/x86/events/intel/uncore_snbep.c
>>>> +++ b/arch/x86/events/intel/uncore_snbep.c
>>>> @@ -6415,7 +6415,7 @@ static void spr_update_device_location(int type_id)
>>>>
>>>> while ((dev = pci_get_device(PCI_VENDOR_ID_INTEL, device, dev)) != NULL) {
>>>>
>>>> - die = uncore_device_to_die(dev);
>>>> + die = uncore_pcibus_to_dieid(dev->bus);
>>>> if (die < 0)
>>>> continue;
>>>>
>>>> @@ -6439,6 +6439,10 @@ static void spr_update_device_location(int type_id)
>>>>
>>>> int spr_uncore_pci_init(void)
>>>> {
>>>> + int ret = snbep_pci2phy_map_init(0x3250, SKX_CPUNODEID, SKX_GIDNIDMAP, true);
>>>> + if (ret)
>>>> + return ret;
>>>> +
>>>> /*
>>>> * The discovery table of UPI on some SPR variant is broken,
>>>> * which impacts the detection of both UPI and M3UPI uncore PMON.
>>>> @@ -6460,21 +6464,13 @@ int spr_uncore_pci_init(void)
>>>>
>>>> void spr_uncore_mmio_init(void)
>>>> {
>>>> - int ret = snbep_pci2phy_map_init(0x3250, SKX_CPUNODEID, SKX_GIDNIDMAP, true);
>>>> + uncore_mmio_uncores = uncore_get_uncores(UNCORE_ACCESS_MMIO,
>>>> + UNCORE_SPR_MMIO_EXTRA_UNCORES,
>>>> + spr_mmio_uncores,
>>>> + UNCORE_SPR_NUM_UNCORE_TYPES,
>>>> + spr_uncores);
>>>>
>>>> - if (ret) {
>>>> - uncore_mmio_uncores = uncore_get_uncores(UNCORE_ACCESS_MMIO, 0, NULL,
>>>> - UNCORE_SPR_NUM_UNCORE_TYPES,
>>>> - spr_uncores);
>>>> - } else {
>>>> - uncore_mmio_uncores = uncore_get_uncores(UNCORE_ACCESS_MMIO,
>>>> - UNCORE_SPR_MMIO_EXTRA_UNCORES,
>>>> - spr_mmio_uncores,
>>>> - UNCORE_SPR_NUM_UNCORE_TYPES,
>>>> - spr_uncores);
>>>> -
>>>> - spr_uncore_imc_free_running.num_boxes = uncore_type_max_boxes(uncore_mmio_uncores, UNCORE_SPR_IMC) / 2;
>>>> - }
>>>> + spr_uncore_imc_free_running.num_boxes = uncore_type_max_boxes(uncore_mmio_uncores, UNCORE_SPR_IMC) / 2;
>>> I'm not sure if we can directly remove the snbep_pci2phy_map_init() call
>>> here. In theory, the snbep_pci2phy_map_init() call in spr_uncore_pci_init()
>>> could fail and then spr_uncore_mmio_init() doesn't know it and directly
>>> initializes MMIO PMU, then it could lead to the MMIO initialization fails.
>>
>> Yes, this is true. But I would argue that the fix in this patch is
>> correct, and the issue you pointed out is not new: the uncore driver
>> registers a PMU device without guaranteeing it's functioning.
>>
>> This is because the Intel uncore driver employs a lazy init approach.
>> And when init_box() fails, it doesn't unregister the inaccessible PMU
>> devices. For example, intel_generic_uncore_mmio_init_box() could fail
>> for a number of reasons, making all associated PMU devices non-functional.
>>
>> Originally the uncore driver tried to enumerate PCI/MSR/MMIO uncore
>> independently, but evolving hardware complexity makes this more
>> challenging. This patch is just one example, IMC Freerunning is
>> MMIO-accessed but relies on PCI devices to read the die-specific MMIO
>> base address. Explicitly gating sysfs node creation with PCI init code
>> in mmio_init() is neither clean nor reliable.
>>
>> To fix it, it seems reasonable to have init_box() return int and
>> unregister the PMU device if deemed inaccessible — similar to what
>> perf_event_ibs_init() does.
>>
>> --- a/arch/x86/events/intel/uncore.h
>> +++ b/arch/x86/events/intel/uncore.h
>> @@ -129,7 +129,7 @@ struct intel_uncore_type {
>> #define events_group attr_groups[2]
>>
>> struct intel_uncore_ops {
>> - void (*init_box)(struct intel_uncore_box *);
>> + int (*init_box)(struct intel_uncore_box *);
>>
>> --- a/arch/x86/events/intel/uncore.c
>> +++ b/arch/x86/events/intel/uncore.c
>> @@ -1155,7 +1155,8 @@ static int uncore_pci_pmu_register(struct pci_dev
>> *pdev,
>> box->dieid = die;
>> box->pci_dev = pdev;
>> box->pmu = pmu;
>> - uncore_box_init(box);
>> + ret = uncore_box_init(box);
>> + if (ret)
>> + return ret;
>>
>> @@ -1598,8 +1599,10 @@ static int uncore_box_ref(struct
>> intel_uncore_type **types,
>> pmu = type->pmus;
>> for (i = 0; i < type->num_boxes; i++, pmu++) {
>> box = pmu->boxes[id];
>> - if (box && box->cpu >= 0 &&
>> atomic_inc_return(&box->refcnt) == 1)
>> - uncore_box_init(box);
>> + if (box && box->cpu >= 0 &&
>> atomic_inc_return(&box->refcnt) == 1)
>> + if (uncore_box_init(box))
>> + uncore_pmu_unregister(pmu);
>
> Yes, I like this idea. The return value of init_box() should always be
> checked. I'm not quite sure if there are other resources need to be cleaned
> besides unregistering the corresponding uncore pmu, please double check.
> Thanks.

I'm thinking of removing this patch from this series and putting it
together with the init_box() changes, where it will be a complete fix
and I need more time to double check all the init_box() callbacks.

>>
>>
>>> Currently the PCI, CPU and MMIO initialization are totally independent,
>>> only when the 3 types initialization all fail, then uncore PMU can abort.
>>>
>>> ```
>>>
>>> if (uncore_init->pci_init) {
>>> pret = uncore_init->pci_init();
>>> if (!pret)
>>> pret = uncore_pci_init();
>>> }
>>>
>>> if (uncore_init->cpu_init) {
>>> uncore_init->cpu_init();
>>> cret = uncore_cpu_init();
>>> }
>>>
>>> if (uncore_init->mmio_init) {
>>> uncore_init->mmio_init();
>>> mret = uncore_mmio_init();
>>> }
>>>
>>> if (cret && pret && mret) {
>>> ret = -ENODEV;
>>> goto free_discovery;
>>> }
>>> ```
>>>
>>>
>>>> }
>>>>
>>>> /* end of SPR uncore support */