[PATCH 2/3] drm/amdgpu: add per device user friendly xgmi events for vega20
Kim, Jonathan
Jonathan.Kim at amd.com
Tue Sep 22 03:03:29 UTC 2020
> -----Original Message-----
> From: Kasiviswanathan, Harish <Harish.Kasiviswanathan at amd.com>
> Sent: Monday, September 21, 2020 8:52 PM
> To: Kim, Jonathan <Jonathan.Kim at amd.com>; amd-gfx at lists.freedesktop.org
> Subject: RE: [PATCH 2/3] drm/amdgpu: add per device user friendly xgmi
> events for vega20
>
> [AMD Official Use Only - Internal Distribution Only]
>
> Few comments inline.
>
> -----Original Message-----
> From: Kim, Jonathan <Jonathan.Kim at amd.com>
> Sent: Thursday, September 17, 2020 2:15 PM
> To: amd-gfx at lists.freedesktop.org
> Cc: Kasiviswanathan, Harish <Harish.Kasiviswanathan at amd.com>; Kim,
> Jonathan <Jonathan.Kim at amd.com>; Kim, Jonathan
> <Jonathan.Kim at amd.com>
> Subject: [PATCH 2/3] drm/amdgpu: add per device user friendly xgmi events
> for vega20
>
> Non-outbound data metrics are non useful so mark them as legacy.
> Bucket new perf counters into device and not device ip.
> Bind events to chip instead of IP.
> Report available event counters and not number of hw counter banks.
> Move DF public macros to private since not needed outside of IP version.
>
> v3: attr groups const array is global but attr groups are allocated per device
> which doesn't work and causes problems on memory allocation and de-
> allocation for pmu unregister. Switch to building const attr groups per pmu
> instead to simplify solution.
>
> v2: add comments on sysfs structure and formatting.
>
> Signed-off-by: Jonathan Kim <Jonathan.Kim at amd.com>
> ---
> drivers/gpu/drm/amd/amdgpu/amdgpu.h | 13 -
> drivers/gpu/drm/amd/amdgpu/amdgpu_pmu.c | 341
> ++++++++++++++++++++----
> drivers/gpu/drm/amd/amdgpu/amdgpu_pmu.h | 6 +-
> drivers/gpu/drm/amd/amdgpu/df_v3_6.c | 72 +----
> drivers/gpu/drm/amd/amdgpu/df_v3_6.h | 9 -
> 5 files changed, 304 insertions(+), 137 deletions(-)
>
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu.h
> b/drivers/gpu/drm/amd/amdgpu/amdgpu.h
> index 13f92dea182a..f43dfdd2716a 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu.h
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu.h
> @@ -1279,19 +1279,6 @@ bool amdgpu_device_load_pci_state(struct pci_dev
> *pdev);
>
> #include "amdgpu_object.h"
>
> -/* used by df_v3_6.c and amdgpu_pmu.c */
> -#define AMDGPU_PMU_ATTR(_name, _object)
> \
> -static ssize_t \
> -_name##_show(struct device *dev, \
> - struct device_attribute *attr, \
> - char *page) \
> -{ \
> - BUILD_BUG_ON(sizeof(_object) >= PAGE_SIZE - 1);
> \
> - return sprintf(page, _object "\n"); \
> -} \
> - \
> -static struct device_attribute pmu_attr_##_name = __ATTR_RO(_name)
> -
> static inline bool amdgpu_is_tmz(struct amdgpu_device *adev) {
> return adev->gmc.tmz_enabled;
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_pmu.c
> b/drivers/gpu/drm/amd/amdgpu/amdgpu_pmu.c
> index 1b0ec715c8ba..74fe8fbdc0d1 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_pmu.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_pmu.c
> @@ -27,9 +27,19 @@
> #include <linux/init.h>
> #include "amdgpu.h"
> #include "amdgpu_pmu.h"
> -#include "df_v3_6.h"
>
> #define PMU_NAME_SIZE 32
> +#define NUM_FORMATS_AMDGPU_PMU 4
> +#define NUM_FORMATS_DF_LEGACY 3
> +#define NUM_EVENTS_DF_LEGACY 8
> +#define NUM_EVENTS_VEGA20_XGMI 2
> +#define NUM_EVENTS_VEGA20_MAX NUM_EVENTS_VEGA20_XGMI
> +
> +struct amdgpu_pmu_event_attribute {
> + struct device_attribute attr;
> + const char *event_str;
> + unsigned int type;
> +};
>
> /* record to keep track of pmu entry per pmu type per device */ struct
> amdgpu_pmu_entry { @@ -37,10 +47,74 @@ struct amdgpu_pmu_entry {
> struct amdgpu_device *adev;
> struct pmu pmu;
> unsigned int pmu_perf_type;
> + struct attribute_group fmt_attr_group;
> + struct amdgpu_pmu_event_attribute *fmt_attr;
> + struct attribute_group evt_attr_group;
> + struct amdgpu_pmu_event_attribute *evt_attr;
> };
>
> +static ssize_t amdgpu_pmu_event_show(struct device *dev,
> + struct device_attribute *attr, char *buf) {
> + struct amdgpu_pmu_event_attribute *amdgpu_pmu_attr;
> +
> + amdgpu_pmu_attr = container_of(attr, struct
> amdgpu_pmu_event_attribute,
> + attr);
> +
> + if (!amdgpu_pmu_attr->type)
> + return sprintf(buf, "%s\n", amdgpu_pmu_attr->event_str);
> +
> + return sprintf(buf, "%s,type=0x%x\n",
> + amdgpu_pmu_attr->event_str, amdgpu_pmu_attr-
> >type); }
> +
> static LIST_HEAD(amdgpu_pmu_list);
>
> +/*
> + * Event formatting is global to all amdgpu events under sysfs folder
> + * /sys/bus/event_source/devices/amdgpu_<dev_num> where dev_num is the
> + * primary device index. Registered events can be found in subfolder "events"
> + * and formatting under subfolder "format".
> + *
> + * Formats "event", "instance", and "umask" are currently used by xGMI
> +but can
> + * be for generalized for other IP usage. If format naming is
> +insufficient
> + * for newly registered IP events, append to the list below and handle
> +the
> + * perf events hardware configuration (see hwc->config) as required by the IP.
> + *
> + * Format "type" indicates IP type generated on pmu registration (see
> + * init_pmu_by_type) so non-legacy events omit this in the per-chip
> +event
> + * list (e.g. vega20_events).
> + */
> +static const char *amdgpu_pmu_formats[NUM_FORMATS_AMDGPU_PMU][2]
> = {
> + { "event", "config:0-7" },
> + { "instance", "config:8-15" },
> + { "umask", "config:16-23"},
> + { "type", "config:56-63"}
> +};
> +
> +/* Vega20 events */
> +static const char *vega20_events[NUM_EVENTS_VEGA20_MAX][2] = {
> + { "xgmi_link0_data_outbound", "event=0x7,instance=0x46,umask=0x2"
> },
> + { "xgmi_link1_data_outbound", "event=0x7,instance=0x47,umask=0x2"
> } };
> +
> +/* All df_vega20_* items are DEPRECATED. Use vega20_ items above
> +instead. */ static const char
> *df_vega20_formats[NUM_FORMATS_DF_LEGACY][2] = {
> + { "event", "config:0-7" },
> + { "instance", "config:8-15" },
> + { "umask", "config:16-23"}
> +};
> +
> +static const char *df_vega20_events[NUM_EVENTS_DF_LEGACY][2] = {
> + { "cake0_pcsout_txdata", "event=0x7,instance=0x46,umask=0x2" },
> + { "cake1_pcsout_txdata", "event=0x7,instance=0x47,umask=0x2" },
> + { "cake0_pcsout_txmeta", "event=0x7,instance=0x46,umask=0x4" },
> + { "cake1_pcsout_txmeta", "event=0x7,instance=0x47,umask=0x4" },
> + { "cake0_ftiinstat_reqalloc", "event=0xb,instance=0x46,umask=0x4" },
> + { "cake1_ftiinstat_reqalloc", "event=0xb,instance=0x47,umask=0x4" },
> + { "cake0_ftiinstat_rspalloc", "event=0xb,instance=0x46,umask=0x8" },
> + { "cake1_ftiinstat_rspalloc", "event=0xb,instance=0x47,umask=0x8" },
> +};
>
> /* initialize perf counter */
> static int amdgpu_perf_event_init(struct perf_event *event) @@ -73,7 +147,8
> @@ static void amdgpu_perf_start(struct perf_event *event, int flags)
> hwc->state = 0;
>
> switch (pe->pmu_perf_type) {
> - case PERF_TYPE_AMDGPU_DF:
> + case PERF_TYPE_AMDGPU_DF_LEGACY:
> + case PERF_TYPE_AMDGPU_XGMI:
> if (!(flags & PERF_EF_RELOAD)) {
> target_cntr = pe->adev->df.funcs->pmc_start(pe-
> >adev,
> hwc->config, 0 /* unused */,
> @@ -108,7 +183,8 @@ static void amdgpu_perf_read(struct perf_event
> *event)
> prev = local64_read(&hwc->prev_count);
>
> switch (pe->pmu_perf_type) {
> - case PERF_TYPE_AMDGPU_DF:
> + case PERF_TYPE_AMDGPU_DF_LEGACY:
> + case PERF_TYPE_AMDGPU_XGMI:
> pe->adev->df.funcs->pmc_get_count(pe->adev,
> hwc->config, hwc->idx,
> &count);
> break;
> @@ -133,7 +209,8 @@ static void amdgpu_perf_stop(struct perf_event *event,
> int flags)
> return;
>
> switch (pe->pmu_perf_type) {
> - case PERF_TYPE_AMDGPU_DF:
> + case PERF_TYPE_AMDGPU_DF_LEGACY:
> + case PERF_TYPE_AMDGPU_XGMI:
> pe->adev->df.funcs->pmc_stop(pe->adev, hwc->config, hwc-
> >idx,
> 0);
> break;
> @@ -160,10 +237,15 @@ static int amdgpu_perf_add(struct perf_event
> *event, int flags)
> struct amdgpu_pmu_entry,
> pmu);
>
> + if (pe->pmu_perf_type == PERF_TYPE_AMDGPU_MAX)
> + pe->pmu_perf_type = (hwc->config >>
> AMDGPU_PERF_TYPE_SHIFT) &
> +
> AMDGPU_PERF_TYPE_MASK;
> +
> event->hw.state = PERF_HES_UPTODATE | PERF_HES_STOPPED;
>
> switch (pe->pmu_perf_type) {
> - case PERF_TYPE_AMDGPU_DF:
> + case PERF_TYPE_AMDGPU_DF_LEGACY:
> + case PERF_TYPE_AMDGPU_XGMI:
> target_cntr = pe->adev->df.funcs->pmc_start(pe->adev,
> hwc->config, 0 /* unused */,
> 1 /* add counter */);
> @@ -197,7 +279,8 @@ static void amdgpu_perf_del(struct perf_event *event,
> int flags)
> amdgpu_perf_stop(event, PERF_EF_UPDATE);
>
> switch (pe->pmu_perf_type) {
> - case PERF_TYPE_AMDGPU_DF:
> + case PERF_TYPE_AMDGPU_DF_LEGACY:
> + case PERF_TYPE_AMDGPU_XGMI:
> pe->adev->df.funcs->pmc_stop(pe->adev, hwc->config, hwc-
> >idx,
> 1);
> break;
> @@ -208,25 +291,83 @@ static void amdgpu_perf_del(struct perf_event
> *event, int flags)
> perf_event_update_userpage(event);
> }
>
> -/* vega20 pmus */
> -
> -/* init pmu tracking per pmu type */
> -static int init_pmu_by_type(struct amdgpu_device *adev,
> - const struct attribute_group *attr_groups[],
> - char *pmu_type_name, char *pmu_file_prefix,
> - unsigned int pmu_perf_type,
> - unsigned int num_counters)
> +static void amdgpu_pmu_create_attributes(struct attribute_group
> *attr_group,
> + struct amdgpu_pmu_event_attribute
> *pmu_attr,
> + const char *events[][2],
> + int s_offset,
> + int e_offset,
> + unsigned int type)
>
> Is s_offset really required. I see 0 is passed in all cases.
Right now it's 0 because we only have xGMI events. To add other event types to the events array, s_offset is needed because we have to call this function per event type in /* other events can be added here */ (I should probably update this comment to reflect that). So pmu_attr needs to know where it left off from the last event type creation per pmu registration. For formats, offsets aren't necessary because they cover all event types so it's called only once per ASIC. Maybe to avoid misuse on that point, I'll split amdgpu_pmu_create_attribute into amdgpu_pmu_create_event_attrs_by_type and have a new amdgpu_pmu_create_format_attrs be a wrapper on amdgpu_pmu_create_event_attrs_by_type with s_offset and type set to 0.
>
>
> - char pmu_name[PMU_NAME_SIZE];
> - struct amdgpu_pmu_entry *pmu_entry;
> - int ret = 0;
> + int i;
> +
> + pmu_attr += s_offset;
> +
> + for (i = s_offset; i < e_offset; i++) {
> + attr_group->attrs[i] = &pmu_attr->attr.attr;
> + sysfs_attr_init(&pmu_attr->attr.attr);
> + pmu_attr->attr.attr.name = events[i][0];
> + pmu_attr->attr.attr.mode = 0444;
> + pmu_attr->attr.show = amdgpu_pmu_event_show;
> + pmu_attr->event_str = events[i][1];
> + pmu_attr->type = type;
> + pmu_attr++;
> + }
> +}
>
> - pmu_entry = kzalloc(sizeof(struct amdgpu_pmu_entry), GFP_KERNEL);
> +static int amdgpu_pmu_alloc_pmu_attrs(
> + struct attribute_group *fmt_attr_group,
> + struct amdgpu_pmu_event_attribute
> **fmt_attr,
> + int fmt_num_attrs,
> + struct attribute_group *evt_attr_group,
> + struct amdgpu_pmu_event_attribute
> **evt_attr,
> + int evt_num_attrs)
> +{
> + *fmt_attr = kcalloc(fmt_num_attrs, sizeof(**fmt_attr), GFP_KERNEL);
>
> - if (!pmu_entry)
> + if (!(*fmt_attr))
> return -ENOMEM;
>
> - pmu_entry->adev = adev;
> + fmt_attr_group->attrs = kcalloc(fmt_num_attrs + 1,
> + sizeof(*fmt_attr_group->attrs), GFP_KERNEL);
> +
> + if (!fmt_attr_group->attrs)
> + goto err_fmt_attr_grp;
> +
> + *evt_attr = kcalloc(evt_num_attrs, sizeof(**evt_attr), GFP_KERNEL);
> +
> + if (!(*evt_attr))
> + goto err_evt_attr;
> +
> + evt_attr_group->attrs = kcalloc(evt_num_attrs + 1,
> + sizeof(*evt_attr_group->attrs), GFP_KERNEL);
> +
> + if (!evt_attr_group->attrs)
> + goto err_evt_attr_grp;
> +
> + return 0;
> +err_evt_attr_grp:
> + kfree(*evt_attr);
> +err_evt_attr:
> + kfree(fmt_attr_group->attrs);
> +err_fmt_attr_grp:
> + kfree(*fmt_attr);
> + return -ENOMEM;
> +}
> +
> +/* init pmu tracking per pmu type */
> +static int init_pmu_by_type(struct amdgpu_pmu_entry *pmu_entry,
> + char *pmu_type_name, char *pmu_file_prefix,
> + unsigned int pmu_perf_type)
>
> I feel this function could be a simpler ASIC independent helper function. If you
> check all asic_types instead of just VEGA20 in amdgpu_pmu_init(), this function
> could be something like this init_pmu_by_type(struct amdgpu_pmu_entry
> *pmu_entry,
> char *pmu_type_name, char *pmu_file_prefix,
> unsigned int pmu_perf_type,
> const char *events[][2], int num_of_events,
> const char *fmts[][2], int mum_of_events); This way
> you could probably avoid all "is_legacy ?" cases
My first comment should answer why it's necessary to arbitrate by asic_type in init_pmu_by_type. We don't know how to create attributes per event type per registration unless we know the ASIC. Your suggestion would limit us to a single event type per registration, similar to how vega20 DF legacy is constructed - amdgpu_df_<num>, which is what we're trying to get away from. We need to dynamically set the "type" field per event because we can't re-register the pmu attribute groups later with new events (throws EEXIST). This is why Vega20 registers twice (single event type per pmu as legacy and the new xgmi events as per event type per pmu) while other ASICS will only register once.
I think the problem is the pmu_perf_type arg is getting mixed up with per-event type with the abuse of the enum. I'll change this to make that clear.
>
>
> +{
> + const struct attribute_group *attr_groups[] = {
> + &pmu_entry->fmt_attr_group,
> + &pmu_entry->evt_attr_group,
> + NULL
> + };
> + char pmu_name[PMU_NAME_SIZE];
> + bool is_legacy = pmu_perf_type == PERF_TYPE_AMDGPU_DF_LEGACY;
> + int ret = 0, num_events = 0;
> +
> pmu_entry->pmu = (struct pmu){
> .event_init = amdgpu_perf_event_init,
> .add = amdgpu_perf_add,
> @@ -237,59 +378,157 @@ static int init_pmu_by_type(struct amdgpu_device
> *adev,
> .task_ctx_nr = perf_invalid_context,
> };
>
> - pmu_entry->pmu.attr_groups = attr_groups;
> + switch (pmu_entry->adev->asic_type) {
> + case CHIP_VEGA20:
> + ret = amdgpu_pmu_alloc_pmu_attrs(&pmu_entry-
> >fmt_attr_group,
> + &pmu_entry->fmt_attr,
> + is_legacy ?
> NUM_FORMATS_DF_LEGACY :
> +
> NUM_FORMATS_AMDGPU_PMU,
> + &pmu_entry->evt_attr_group,
> + &pmu_entry->evt_attr,
> + is_legacy ? NUM_EVENTS_DF_LEGACY
> :
> +
> NUM_EVENTS_VEGA20_MAX);
> +
> + if (ret)
> + goto err_out;
> +
> + amdgpu_pmu_create_attributes(&pmu_entry-
> >fmt_attr_group,
> + pmu_entry->fmt_attr,
> + is_legacy ? df_vega20_formats :
> +
> amdgpu_pmu_formats, 0,
> + is_legacy ?
> NUM_FORMATS_DF_LEGACY :
> +
> NUM_FORMATS_AMDGPU_PMU,
> + 0);
> +
> + amdgpu_pmu_create_attributes(&pmu_entry->evt_attr_group,
> + pmu_entry->evt_attr,
> + is_legacy ? df_vega20_events :
> + vega20_events, 0,
> + is_legacy ? NUM_EVENTS_DF_LEGACY
> :
> +
> NUM_EVENTS_VEGA20_XGMI,
> + is_legacy ?
> PERF_TYPE_AMDGPU_DF_LEGACY :
> +
> PERF_TYPE_AMDGPU_XGMI);
> + num_events += is_legacy ? NUM_EVENTS_DF_LEGACY :
> +
> NUM_EVENTS_VEGA20_XGMI;
> +
>
> Are you using += here in preparation for future events. Because this variable
> num_events is used only once in this function
Yes. I'll rename this to total_num_events to make that clear.
>
> + /* other events can be added here */
> +
> + break;
> + default:
> + ret = -ENODEV;
> + goto err_out;
> + };
> +
> + pmu_entry->pmu.attr_groups = kmemdup(attr_groups,
> sizeof(attr_groups),
> + GFP_KERNEL);
> +
> + if (!pmu_entry->pmu.attr_groups)
> + goto err_attr_group;
> +
> pmu_entry->pmu_perf_type = pmu_perf_type;
> - snprintf(pmu_name, PMU_NAME_SIZE, "%s_%d",
> - pmu_file_prefix, adev_to_drm(adev)-
> >primary->index);
> + snprintf(pmu_name, PMU_NAME_SIZE, "%s_%d", pmu_file_prefix,
> + adev_to_drm(pmu_entry->adev)->primary-
> >index);
>
> ret = perf_pmu_register(&pmu_entry->pmu, pmu_name, -1);
>
> - if (ret) {
> - kfree(pmu_entry);
> - pr_warn("Error initializing AMDGPU %s PMUs.\n",
> pmu_type_name);
> - return ret;
> - }
> + if (ret)
> + goto err_register;
>
> pr_info("Detected AMDGPU %s Counters. # of Counters = %d.\n",
> - pmu_type_name, num_counters);
> + pmu_type_name, num_events);
>
> Should the "# of Counters" be changes to "# of Events"
Then maybe we should do "# of Perf Events" instead to be explicit that we're using perf_events for non-legacy cases. I think there are a bunch of other events tracked that don't use perf_events in our driver.
Thanks,
Jon
>
> list_add_tail(&pmu_entry->entry, &amdgpu_pmu_list);
>
> return 0;
> +err_register:
> + kfree(pmu_entry->pmu.attr_groups);
> +err_attr_group:
> + kfree(pmu_entry->fmt_attr_group.attrs);
> + kfree(pmu_entry->fmt_attr);
> + kfree(pmu_entry->evt_attr_group.attrs);
> + kfree(pmu_entry->evt_attr);
> +err_out:
> + pr_warn("Error initializing AMDGPU %s PMUs.\n", pmu_type_name);
> + return ret;
> +}
> +
> +/* destroy all pmu data associated with target device */ void
> +amdgpu_pmu_fini(struct amdgpu_device *adev) {
> + struct amdgpu_pmu_entry *pe, *temp;
> +
> + list_for_each_entry_safe(pe, temp, &amdgpu_pmu_list, entry) {
> + if (pe->adev != adev)
> + continue;
> + list_del(&pe->entry);
> + perf_pmu_unregister(&pe->pmu);
> + kfree(pe->pmu.attr_groups);
> + kfree(pe->fmt_attr_group.attrs);
> + kfree(pe->fmt_attr);
> + kfree(pe->evt_attr_group.attrs);
> + kfree(pe->evt_attr);
> + kfree(pe);
> + }
> +}
> +
> +static bool amdgpu_pmu_is_supported(struct amdgpu_device *adev) {
> + return adev->asic_type == CHIP_VEGA20;
> }
>
> /* init amdgpu_pmu */
> int amdgpu_pmu_init(struct amdgpu_device *adev) {
> int ret = 0;
> + struct amdgpu_pmu_entry *pmu_entry, *pmu_entry_legacy;
>
> - switch (adev->asic_type) {
> - case CHIP_VEGA20:
> - /* init df */
> - ret = init_pmu_by_type(adev, df_v3_6_attr_groups,
> - "DF", "amdgpu_df",
> PERF_TYPE_AMDGPU_DF,
> - DF_V3_6_MAX_COUNTERS);
> -
> - /* other pmu types go here*/
> - break;
> - default:
> + if (!amdgpu_pmu_is_supported(adev))
> return 0;
> - }
>
> - return 0;
> -}
> + if (adev->asic_type == CHIP_VEGA20) {
> + pmu_entry_legacy = kzalloc(sizeof(struct amdgpu_pmu_entry),
> + GFP_KERNEL);
>
> + if (!pmu_entry_legacy)
> + return -ENOMEM;
>
> -/* destroy all pmu data associated with target device */ -void
> amdgpu_pmu_fini(struct amdgpu_device *adev) -{
> - struct amdgpu_pmu_entry *pe, *temp;
> + pmu_entry_legacy->adev = adev;
> + pmu_entry_legacy->fmt_attr_group.name = "format";
> + pmu_entry_legacy->fmt_attr_group.attrs = NULL;
> + pmu_entry_legacy->evt_attr_group.name = "events";
> + pmu_entry_legacy->evt_attr_group.attrs = NULL;
>
> - list_for_each_entry_safe(pe, temp, &amdgpu_pmu_list, entry) {
> - if (pe->adev == adev) {
> - list_del(&pe->entry);
> - perf_pmu_unregister(&pe->pmu);
> - kfree(pe);
> + ret = init_pmu_by_type(pmu_entry_legacy, "DF", "amdgpu_df",
> +
> PERF_TYPE_AMDGPU_DF_LEGACY);
> +
> + if (ret) {
> + kfree(pmu_entry_legacy);
> + return ret;
> }
> }
> +
> + pmu_entry = kzalloc(sizeof(struct amdgpu_pmu_entry), GFP_KERNEL);
> +
> + if (!pmu_entry) {
> + if (adev->asic_type == CHIP_VEGA20)
> + amdgpu_pmu_fini(adev);
> + return -ENOMEM;
> + }
> +
> + pmu_entry->adev = adev;
> + pmu_entry->fmt_attr_group.name = "format";
> + pmu_entry->fmt_attr_group.attrs = NULL;
> + pmu_entry->evt_attr_group.name = "events";
> + pmu_entry->evt_attr_group.attrs = NULL;
> +
> + ret = init_pmu_by_type(pmu_entry, "Event", "amdgpu",
> +
> PERF_TYPE_AMDGPU_MAX);
> + if (ret) {
> + if (adev->asic_type == CHIP_VEGA20)
> + amdgpu_pmu_fini(adev);
> + kfree(pmu_entry);
> +
> + }
> +
> + return ret;
> }
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_pmu.h
> b/drivers/gpu/drm/amd/amdgpu/amdgpu_pmu.h
> index 7dddb7160a11..0d214abe720e 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_pmu.h
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_pmu.h
> @@ -27,10 +27,14 @@
> #define _AMDGPU_PMU_H_
>
> enum amdgpu_pmu_perf_type {
> - PERF_TYPE_AMDGPU_DF = 0,
> + PERF_TYPE_AMDGPU_DF_LEGACY = 0,
> + PERF_TYPE_AMDGPU_XGMI,
> PERF_TYPE_AMDGPU_MAX
> };
>
> +#define AMDGPU_PERF_TYPE_SHIFT 56
> +#define AMDGPU_PERF_TYPE_MASK 0xff
> +
> int amdgpu_pmu_init(struct amdgpu_device *adev); void
> amdgpu_pmu_fini(struct amdgpu_device *adev);
>
> diff --git a/drivers/gpu/drm/amd/amdgpu/df_v3_6.c
> b/drivers/gpu/drm/amd/amdgpu/df_v3_6.c
> index 0ca6e176acb0..6e57ae95f997 100644
> --- a/drivers/gpu/drm/amd/amdgpu/df_v3_6.c
> +++ b/drivers/gpu/drm/amd/amdgpu/df_v3_6.c
> @@ -30,71 +30,17 @@
> #define DF_3_6_SMN_REG_INST_DIST 0x8
> #define DF_3_6_INST_CNT 8
>
> -static u32 df_v3_6_channel_number[] = {1, 2, 0, 4, 0, 8, 0,
> - 16, 32, 0, 0, 0, 2, 4, 8};
> -
> -/* init df format attrs */
> -AMDGPU_PMU_ATTR(event, "config:0-7");
> -AMDGPU_PMU_ATTR(instance, "config:8-15");
> -AMDGPU_PMU_ATTR(umask, "config:16-23");
> -
> -/* df format attributes */
> -static struct attribute *df_v3_6_format_attrs[] = {
> - &pmu_attr_event.attr,
> - &pmu_attr_instance.attr,
> - &pmu_attr_umask.attr,
> - NULL
> -};
> -
> -/* df format attribute group */
> -static struct attribute_group df_v3_6_format_attr_group = {
> - .name = "format",
> - .attrs = df_v3_6_format_attrs,
> -};
> +/* Defined in global_features.h as FTI_PERFMON_VISIBLE */
> +#define DF_V3_6_MAX_COUNTERS 4
>
> -/* df event attrs */
> -AMDGPU_PMU_ATTR(cake0_pcsout_txdata,
> - "event=0x7,instance=0x46,umask=0x2");
> -AMDGPU_PMU_ATTR(cake1_pcsout_txdata,
> - "event=0x7,instance=0x47,umask=0x2");
> -AMDGPU_PMU_ATTR(cake0_pcsout_txmeta,
> - "event=0x7,instance=0x46,umask=0x4");
> -AMDGPU_PMU_ATTR(cake1_pcsout_txmeta,
> - "event=0x7,instance=0x47,umask=0x4");
> -AMDGPU_PMU_ATTR(cake0_ftiinstat_reqalloc,
> - "event=0xb,instance=0x46,umask=0x4");
> -AMDGPU_PMU_ATTR(cake1_ftiinstat_reqalloc,
> - "event=0xb,instance=0x47,umask=0x4");
> -AMDGPU_PMU_ATTR(cake0_ftiinstat_rspalloc,
> - "event=0xb,instance=0x46,umask=0x8");
> -AMDGPU_PMU_ATTR(cake1_ftiinstat_rspalloc,
> - "event=0xb,instance=0x47,umask=0x8");
> -
> -/* df event attributes */
> -static struct attribute *df_v3_6_event_attrs[] = {
> - &pmu_attr_cake0_pcsout_txdata.attr,
> - &pmu_attr_cake1_pcsout_txdata.attr,
> - &pmu_attr_cake0_pcsout_txmeta.attr,
> - &pmu_attr_cake1_pcsout_txmeta.attr,
> - &pmu_attr_cake0_ftiinstat_reqalloc.attr,
> - &pmu_attr_cake1_ftiinstat_reqalloc.attr,
> - &pmu_attr_cake0_ftiinstat_rspalloc.attr,
> - &pmu_attr_cake1_ftiinstat_rspalloc.attr,
> - NULL
> -};
> -
> -/* df event attribute group */
> -static struct attribute_group df_v3_6_event_attr_group = {
> - .name = "events",
> - .attrs = df_v3_6_event_attrs
> -};
> +/* get flags from df perfmon config */
> +#define DF_V3_6_GET_EVENT(x) (x & 0xFFUL)
> +#define DF_V3_6_GET_INSTANCE(x) ((x >> 8) & 0xFFUL)
> +#define DF_V3_6_GET_UNITMASK(x) ((x >> 16) & 0xFFUL)
> +#define DF_V3_6_PERFMON_OVERFLOW 0xFFFFFFFFFFFFULL
>
> -/* df event attr groups */
> -const struct attribute_group *df_v3_6_attr_groups[] = {
> - &df_v3_6_format_attr_group,
> - &df_v3_6_event_attr_group,
> - NULL
> -};
> +static u32 df_v3_6_channel_number[] = {1, 2, 0, 4, 0, 8, 0,
> + 16, 32, 0, 0, 0, 2, 4, 8};
>
> static uint64_t df_v3_6_get_fica(struct amdgpu_device *adev,
> uint32_t ficaa_val)
> diff --git a/drivers/gpu/drm/amd/amdgpu/df_v3_6.h
> b/drivers/gpu/drm/amd/amdgpu/df_v3_6.h
> index 76998541bc30..2505c7ef258a 100644
> --- a/drivers/gpu/drm/amd/amdgpu/df_v3_6.h
> +++ b/drivers/gpu/drm/amd/amdgpu/df_v3_6.h
> @@ -35,15 +35,6 @@ enum DF_V3_6_MGCG {
> DF_V3_6_MGCG_ENABLE_63_CYCLE_DELAY = 15 };
>
> -/* Defined in global_features.h as FTI_PERFMON_VISIBLE */
> -#define DF_V3_6_MAX_COUNTERS 4
> -
> -/* get flags from df perfmon config */
> -#define DF_V3_6_GET_EVENT(x) (x & 0xFFUL)
> -#define DF_V3_6_GET_INSTANCE(x) ((x >> 8) & 0xFFUL)
> -#define DF_V3_6_GET_UNITMASK(x) ((x >> 16) & 0xFFUL)
> -#define DF_V3_6_PERFMON_OVERFLOW 0xFFFFFFFFFFFFULL
> -
> extern const struct attribute_group *df_v3_6_attr_groups[]; extern const
> struct amdgpu_df_funcs df_v3_6_funcs;
>
> --
> 2.17.1
More information about the amd-gfx
mailing list