[Intel-gfx] [PATCH v2] drm/i915/tgl: Add sysfs interface to control class-of-service
Kumar Valsan, Prathap
prathap.kumar.valsan at intel.com
Tue Aug 27 14:17:31 UTC 2019
On Mon, Aug 26, 2019 at 09:39:48AM +0100, Chris Wilson wrote:
> Quoting Prathap Kumar Valsan (2019-08-26 00:35:27)
> > To provide shared last-level-cache isolation to cpu workloads running
> > concurrently with gpu workloads, the gpu allocation of cache lines needs
> > to be restricted to certain ways. Currently GPU hardware supports four
> > class-of-service(CLOS) levels and there is an associated way-mask for
> > each CLOS.
> >
> > Hardware supports reading supported way-mask configuration for GPU using
> > a bios pcode interface. The supported way-masks and the one currently
> > active is communicated to userspace via a sysfs file--closctrl. Admin user
> > can then select a new mask by writing the mask value to the file.
> >
> > Note of Caution: Restricting cache ways using this mechanism presents a
> > larger attack surface for side-channel attacks.
> >
> > Example usage:
> > The active way-mask is highlighted within square brackets.
> > > cat /sys/class/drm/card0/closctrl
> > [0xffff] 0xff00 0xc000 0x8000
> >
> > CLOS0 is currently active.
> >
> > > echo 0x8000 > /sys/class/drm/card0/closctrl
> > > cat /sys/class/drm/card0/closctrl
> > 0xffff 0xff00 0xc000 [0x8000]
> >
> > CLOS3 is currently active
> >
> > Signed-off-by: Prathap Kumar Valsan <prathap.kumar.valsan at intel.com>
> > ---
> > Changes in v2:
> > Declare closctrl_show and closctrl_store as static functions.
> > drivers/gpu/drm/i915/gt/intel_mocs.c | 57 ++++++++++++++++++++---
> > drivers/gpu/drm/i915/gt/intel_mocs.h | 1 +
> > drivers/gpu/drm/i915/i915_drv.h | 8 ++++
> > drivers/gpu/drm/i915/i915_reg.h | 1 +
> > drivers/gpu/drm/i915/i915_sysfs.c | 67 ++++++++++++++++++++++++++++
> > 5 files changed, 129 insertions(+), 5 deletions(-)
> >
> > diff --git a/drivers/gpu/drm/i915/gt/intel_mocs.c b/drivers/gpu/drm/i915/gt/intel_mocs.c
> > index 728704bbbe18..dd13e61944fd 100644
> > --- a/drivers/gpu/drm/i915/gt/intel_mocs.c
> > +++ b/drivers/gpu/drm/i915/gt/intel_mocs.c
> > @@ -26,6 +26,7 @@
> > #include "intel_gt.h"
> > #include "intel_mocs.h"
> > #include "intel_lrc.h"
> > +#include "intel_sideband.h"
> >
> > /* structures required */
> > struct drm_i915_mocs_entry {
> > @@ -51,6 +52,7 @@ struct drm_i915_mocs_table {
> > #define LE_SCF(value) ((value) << 14)
> > #define LE_COS(value) ((value) << 15)
> > #define LE_SSE(value) ((value) << 17)
> > +#define LE_COS_MASK GENMASK(16, 15)
> >
> > /* Defines for the tables (LNCFMOCS0 - LNCFMOCS31) - two entries per word */
> > #define L3_ESC(value) ((value) << 0)
> > @@ -408,10 +410,13 @@ void intel_mocs_init_engine(struct intel_engine_cs *engine)
> > unused_value);
> > }
> >
> > -static void intel_mocs_init_global(struct intel_gt *gt)
> > +void intel_mocs_init_global(struct intel_gt *gt)
> > {
> > + struct drm_i915_private *i915 = gt->i915;
> > struct intel_uncore *uncore = gt->uncore;
> > struct drm_i915_mocs_table table;
> > + unsigned int active_clos;
> > + u32 value, unused_value;
> > unsigned int index;
> >
> > GEM_BUG_ON(!HAS_GLOBAL_MOCS_REGISTERS(gt->i915));
> > @@ -422,20 +427,31 @@ static void intel_mocs_init_global(struct intel_gt *gt)
> > if (GEM_DEBUG_WARN_ON(table.size > table.n_entries))
> > return;
> >
> > - for (index = 0; index < table.size; index++)
> > + active_clos = atomic_read(&i915->clos.active_clos);
> > +
> > + for (index = 0; index < table.size; index++) {
> > + value = table.table[index].control_value;
> > + value &= ~LE_COS_MASK;
> > + value |= FIELD_PREP(LE_COS_MASK, active_clos);
> > +
> > intel_uncore_write(uncore,
> > GEN12_GLOBAL_MOCS(index),
> > - table.table[index].control_value);
> > + value);
> > + }
> >
> > /*
> > * Ok, now set the unused entries to the invalid entry (index 0). These
> > * entries are officially undefined and no contract for the contents and
> > * settings is given for these entries.
> > */
> > + unused_value = table.table[0].control_value;
> > + unused_value &= ~LE_COS_MASK;
> > + unused_value |= FIELD_PREP(LE_COS_MASK, active_clos);
> > +
> > for (; index < table.n_entries; index++)
> > intel_uncore_write(uncore,
> > GEN12_GLOBAL_MOCS(index),
> > - table.table[0].control_value);
> > + unused_value);
> > }
> >
> > static int emit_mocs_control_table(struct i915_request *rq,
> > @@ -625,10 +641,41 @@ int intel_mocs_emit(struct i915_request *rq)
> > return 0;
> > }
> >
> > +static void intel_read_clos_way_mask(struct intel_gt *gt)
> > +{
> > + struct drm_i915_private *i915 = gt->i915;
> > + struct drm_i915_mocs_table table;
> > + int ret, i;
> > + u32 val;
> > +
> > + if (!get_mocs_settings(gt, &table))
> > + return;
> > +
> > + /* COS is same for all entries */
> > + atomic_set(&i915->clos.active_clos,
> > + FIELD_GET(LE_COS_MASK, get_entry_control(&table, 0)));
> > + for (i = 0; i < NUM_OF_CLOS; i++) {
> > + val = i;
> > + ret = sandybridge_pcode_read(i915,
> > + ICL_PCODE_LLC_COS_WAY_MASK_INFO,
> > + &val, NULL);
> > + if (ret) {
> > + DRM_ERROR("Mailbox read error = %d\n", ret);
> > + return;
> > + }
> > +
> > + i915->clos.way_mask[i] = val;
> > + }
> > +
> > + i915->clos.support_way_mask_read = true;
> > +}
> > +
> > void intel_mocs_init(struct intel_gt *gt)
> > {
> > intel_mocs_init_l3cc_table(gt);
> >
> > - if (HAS_GLOBAL_MOCS_REGISTERS(gt->i915))
> > + if (HAS_GLOBAL_MOCS_REGISTERS(gt->i915)) {
> > + intel_read_clos_way_mask(gt);
> > intel_mocs_init_global(gt);
> > + }
> > }
> > diff --git a/drivers/gpu/drm/i915/gt/intel_mocs.h b/drivers/gpu/drm/i915/gt/intel_mocs.h
> > index 2ae816b7ca19..e64e1b104753 100644
> > --- a/drivers/gpu/drm/i915/gt/intel_mocs.h
> > +++ b/drivers/gpu/drm/i915/gt/intel_mocs.h
> > @@ -53,6 +53,7 @@ struct i915_request;
> > struct intel_engine_cs;
> > struct intel_gt;
> >
> > +void intel_mocs_init_global(struct intel_gt *gt);
> > void intel_mocs_init(struct intel_gt *gt);
> > void intel_mocs_init_engine(struct intel_engine_cs *engine);
> >
> > diff --git a/drivers/gpu/drm/i915/i915_drv.h b/drivers/gpu/drm/i915/i915_drv.h
> > index b42651a387d9..0e250416c5a9 100644
> > --- a/drivers/gpu/drm/i915/i915_drv.h
> > +++ b/drivers/gpu/drm/i915/i915_drv.h
> > @@ -1621,6 +1621,14 @@ struct drm_i915_private {
> > bool distrust_bios_wm;
> > } wm;
> >
> > + /* Last Level Cache Class of Service */
> > + struct {
> > + bool support_way_mask_read;
> > + atomic_t active_clos;
> > +#define NUM_OF_CLOS 4
> > + u16 way_mask[NUM_OF_CLOS];
> > + } clos;
> > +
> > struct dram_info {
> > bool valid;
> > bool is_16gb_dimm;
> > diff --git a/drivers/gpu/drm/i915/i915_reg.h b/drivers/gpu/drm/i915/i915_reg.h
> > index 02e1ef10c47e..399acb7a36d8 100644
> > --- a/drivers/gpu/drm/i915/i915_reg.h
> > +++ b/drivers/gpu/drm/i915/i915_reg.h
> > @@ -8845,6 +8845,7 @@ enum {
> > #define ICL_PCODE_MEM_SUBSYSYSTEM_INFO 0xd
> > #define ICL_PCODE_MEM_SS_READ_GLOBAL_INFO (0x0 << 8)
> > #define ICL_PCODE_MEM_SS_READ_QGV_POINT_INFO(point) (((point) << 16) | (0x1 << 8))
> > +#define ICL_PCODE_LLC_COS_WAY_MASK_INFO 0x1d
> > #define GEN6_PCODE_READ_D_COMP 0x10
> > #define GEN6_PCODE_WRITE_D_COMP 0x11
> > #define HSW_PCODE_DE_WRITE_FREQ_REQ 0x17
> > diff --git a/drivers/gpu/drm/i915/i915_sysfs.c b/drivers/gpu/drm/i915/i915_sysfs.c
> > index d8a3b180c084..b53143416396 100644
> > --- a/drivers/gpu/drm/i915/i915_sysfs.c
> > +++ b/drivers/gpu/drm/i915/i915_sysfs.c
> > @@ -34,6 +34,7 @@
> > #include "i915_sysfs.h"
> > #include "intel_pm.h"
> > #include "intel_sideband.h"
> > +#include "gt/intel_mocs.h"
> >
> > static inline struct drm_i915_private *kdev_minor_to_i915(struct device *kdev)
> > {
> > @@ -257,6 +258,62 @@ static const struct bin_attribute dpf_attrs_1 = {
> > .private = (void *)1
> > };
> >
> > +static ssize_t closctrl_show(struct device *kdev,
> > + struct device_attribute *attr, char *buf)
> > +{
> > + struct drm_i915_private *dev_priv = kdev_minor_to_i915(kdev);
> > + ssize_t len = 0;
> > + int i;
> > +
> > + for (i = 0; i < NUM_OF_CLOS; i++) {
> > + if (i == atomic_read(&dev_priv->clos.active_clos))
>
> Reading an atomic more than once is liable to give you different
> answers.
>
> > + len += snprintf(buf + len, PAGE_SIZE, "%s0x%x%s ",
> > + "[", dev_priv->clos.way_mask[i], "]");
> > + else
> > + len += snprintf(buf + len, PAGE_SIZE, "0x%x ",
> > + dev_priv->clos.way_mask[i]);
> > + }
> > + len += snprintf(buf + len, PAGE_SIZE, "\n");
> > +
> > + return len;
> > +}
> > +
> > +static ssize_t closctrl_store(struct device *kdev,
> > + struct device_attribute *attr,
> > + const char *buf, size_t count)
> > +{
> > + struct drm_i915_private *dev_priv = kdev_minor_to_i915(kdev);
> > + u8 active_clos, clos_index;
> > + bool valid_mask = false;
> > + ssize_t ret;
> > + u16 way_mask;
> > +
> > + ret = kstrtou16(buf, 0, &way_mask);
> > + if (ret)
> > + return ret;
> > +
> > + active_clos = atomic_read(&dev_priv->clos.active_clos);
> > +
> > + if (dev_priv->clos.way_mask[active_clos] == way_mask)
> > + return count;
> > +
> > + for (clos_index = 0; clos_index < NUM_OF_CLOS; clos_index++) {
> > + if (dev_priv->clos.way_mask[clos_index] == way_mask) {
> > + atomic_set(&dev_priv->clos.active_clos, clos_index);
> > + valid_mask = true;
> > + break;
> > + }
> > + }
>
> How is this serialised against multiple users changing the setting?
>
Should have been using mutex to serialize. Will fix.
> This is not an atomic operation, atomic_t should have been a warning.
> -Chris
More information about the Intel-gfx
mailing list