@@ -66,7 +66,6 @@ static struct rdt_resource *resctrl_event_get_resource(u16 event_num)
static void resctrl_pmu_event_destroy(struct perf_event *event)
{
- struct hw_perf_event *hwc = &event->hw;
u16 event_num = get_event(event);
struct rdt_resource *r;
@@ -74,7 +73,7 @@ static void resctrl_pmu_event_destroy(struct perf_event *event)
if (!r)
return;
- resctrl_arch_mon_ctx_free(r, event_num, hwc->idx);
+ resctrl_arch_mon_ctx_free(r, event_num, event->pmu_private);
}
static int resctrl_pmu_event_init(struct perf_event *event)
@@ -144,9 +143,9 @@ static int resctrl_pmu_event_init(struct perf_event *event)
return -EINVAL;
}
- hwc->idx = resctrl_arch_mon_ctx_alloc_no_wait(r, event_num);
- if (hwc->idx == -ENOSPC)
- return -ENOSPC;
+ event->pmu_private = resctrl_arch_mon_ctx_alloc_no_wait(r, event_num);
+ if (event->pmu_private && IS_ERR(event->pmu_private))
+ return PTR_ERR(event->pmu_private);
event->destroy = resctrl_pmu_event_destroy;
local64_set(&hwc->prev_count, 0);
local64_set(&event->count, 0);
@@ -183,7 +182,8 @@ static void resctrl_pmu_event_update(struct perf_event *event)
prev = local64_read(&hwc->prev_count);
err = resctrl_arch_rmid_read(r, d, closid, rmid,
- event_num, &now, hwc->idx);
+ event_num, &now,
+ event->pmu_private);
if (err)
return;
} while (local64_cmpxchg(&hwc->prev_count, prev, now) != prev);
@@ -318,28 +318,37 @@ struct rdt_resource *resctrl_arch_get_resource(enum resctrl_res_level l)
void *resctrl_arch_mon_ctx_alloc_no_wait(struct rdt_resource *r, int evtid)
{
struct mpam_resctrl_res *res;
- u32 *ret = kmalloc(sizeof(*ret), GFP_KERNEL);
+ u32 *ctx = kmalloc(sizeof(*ctx), GFP_KERNEL);
+ int err;
- if (!ret)
+ if (!ctx)
return ERR_PTR(-ENOMEM);
switch (evtid) {
case QOS_L3_OCCUP_EVENT_ID:
res = container_of(r, struct mpam_resctrl_res, resctrl_res);
- *ret = mpam_alloc_csu_mon(res->class);
- return ret;
+ err = mpam_alloc_csu_mon(res->class);
+ break;
case QOS_L3_MBM_LOCAL_EVENT_ID:
case QOS_L3_MBM_TOTAL_EVENT_ID:
if (mpam_monitors_free_runing)
return mon_is_rmid_idx;
res = container_of(r, struct mpam_resctrl_res, resctrl_res);
- *ret = mpam_alloc_mbwu_mon(res->class);
- return ret;
+ err = mpam_alloc_mbwu_mon(res->class);
+ break;
+ default:
+ err = -EOPNOTSUPP;
+ }
+
+ if (err < 0) {
+ kfree(ctx);
+ return ERR_PTR(err);
}
- return ERR_PTR(-EOPNOTSUPP);
+ *ctx = err;
+ return ctx;
}
void resctrl_arch_mon_ctx_free(struct rdt_resource *r, int evtid,
The return pointer from the resctrl_arch_mon_ctx_alloc_no_wait() function is saved in a 32-bit variable 'hwc->idx' which results in the loss of the upper 32 bits. This, in turn, triggers a kernel panic when attempting to access a corrupted pointer. Use 'event->pmu_private' instead of 'hwc->idx' to resolve the issue. Another reason contributing to the problem due to continuous 'error irq' messages. The resctrl_arch_mon_ctx_alloc_no_wait() function returns a valid PTR when it cannot allocate monitor resources, resulting in an incorrect MON_SEL configuration. To resolve this issue, return an ERR_PTR when monitor allocation fails dmesg: mpam: error irq from msc:0 'Monitor_Range', partid:228, pmg: 0, ris: 0 mpam: error irq from msc:0 'Monitor_Range', partid:228, pmg: 0, ris: 0 mpam: error irq from msc:0 'Monitor_Range', partid:228, pmg: 0, ris: 0 ... mpam: error irq from msc:0 'Monitor_Range', partid:228, pmg: 0, ris: 0 Signed-off-by: Shanker Donthineni <sdonthineni@nvidia.com> --- Changes in v2: -Use 'event->pmu_private' instead of 'hwc->idx' to keep monitor context -Return ERR_PTR if unable to allocate resource in resctrl_arch_mon_ctx_alloc_no_wait() drivers/perf/resctrl_pmu.c | 12 ++++++------ drivers/platform/mpam/mpam_resctrl.c | 23 ++++++++++++++++------- 2 files changed, 22 insertions(+), 13 deletions(-)