On 4/17/2023 10:56 AM, Teres Alexis, Alan Previn wrote:
On Mon, 2023-04-10 at 09:10 -0700, Ceraolo Spurio, Daniele wrote:
alan:snip
+int
+intel_gsc_uc_heci_cmd_submit_nonpriv(struct intel_gsc_uc *gsc,
+ struct intel_context *ce,
+ struct intel_gsc_heci_non_priv_pkt *pkt,
+ u32 *cmd, int timeout_ms)
+{
+ struct intel_engine_cs *eng;
We always use "engine" for engine_cs variables. IMO it's better to stick
to that here as well for consistency across the code.
alan: will do
+ struct i915_gem_ww_ctx ww;
+ struct i915_request *rq;
+ int err, trials = 0;
+
Is the assumption that the caller is holding a wakeref already?
Otherwise we're going to need and engine_pm_get() here (assuming it
doesn't interfere with any locking, otherwise it has to be in the caller)
alan: right now the only places this can get called from is via intel_pxp_gsccs_create_session or
intel_pxp_gsccs_end_arb_fw_session. These functions are either being called by intel_pxp_start
or intel_pxp_end. intel_pxp_start calls intel_runtime_pm_get_if_in_use indirectly from the
session-worker and while intel_pxp_end takes an explicit intel_runtime_pm_get (since it is
for suspend/shutdown cleanup and doesn't use the worker). I'm assuming runtime_pm works right?
we have a similar logic across the paths from ADL version where we dont take explicit
engine_pm_get for the termination via VDBOX because its part of the same code paths.
rpm_get works for the power management side, but not for "activeness"
tracking, for which we need engine_pm_get. However, I've just realized
that the context_pin contains an engine_pm_get, so we're covered.
Therefore, with the other minor comments addressed, this is:
Reviewed-by: Daniele Ceraolo Spurio <daniele.ceraolospurio@xxxxxxxxx>
Daniele
alan:snip
+ err = i915_vma_move_to_active(pkt->bb_vma, rq, EXEC_OBJECT_WRITE);
nit: I don't think we need EXEC_OBJECT_WRITE for the bb as we're not
going to write it.
alan: yes - will remove. (had accidentally kept above flag from offline
debugging version of the code that had additional store dwords into bb).
+ if (err)
+ goto out_rq;
+ err = i915_vma_move_to_active(pkt->heci_pkt_vma, rq, EXEC_OBJECT_WRITE);
+ if (err)
+ goto out_rq;
+
+ eng = rq->context->engine;
+ if (eng->emit_init_breadcrumb) {
+ err = eng->emit_init_breadcrumb(rq);
+ if (err)
+ goto out_rq;
+ }
+
+ err = eng->emit_bb_start(rq, i915_vma_offset(pkt->bb_vma), PAGE_SIZE, 0);
+ if (err)
+ goto out_rq;
+
+ err = ce->engine->emit_flush(rq, 0);
+ if (err)
+ drm_err(&gsc_uc_to_gt(gsc)->i915->drm,
+ "Failed emit-flush for gsc-heci-non-priv-pkterr=%d\n", err);
+
+out_rq:
+ i915_request_get(rq);
+
+ if (unlikely(err))
+ i915_request_set_error_once(rq, err);
+
+ i915_request_add(rq);
+
+ if (!err) {
+ if (i915_request_wait(rq, I915_WAIT_INTERRUPTIBLE,
+ msecs_to_jiffies(timeout_ms)) < 0)
+ err = -ETIME;
+ }
+
+ i915_request_put(rq);
+
+out_unpin_ce:
+ intel_context_unpin(ce);
+out_ww:
+ if (err == -EDEADLK) {
+ err = i915_gem_ww_ctx_backoff(&ww);
+ if (!err) {
+ if (++trials < 10)
+ goto retry;
+ else
+ err = EAGAIN;
+ }
+ }
+ i915_gem_ww_ctx_fini(&ww);
+
+ return err;
+}
diff --git a/drivers/gpu/drm/i915/gt/uc/intel_gsc_uc_heci_cmd_submit.h b/drivers/gpu/drm/i915/gt/uc/intel_gsc_uc_heci_cmd_submit.h
index 3d56ae501991..3addce861854 100644
--- a/drivers/gpu/drm/i915/gt/uc/intel_gsc_uc_heci_cmd_submit.h
+++ b/drivers/gpu/drm/i915/gt/uc/intel_gsc_uc_heci_cmd_submit.h
@@ -8,7 +8,10 @@
#include <linux/types.h>
+struct i915_vma;
+struct intel_context;
struct intel_gsc_uc;
+
struct intel_gsc_mtl_header {
u32 validity_marker;
#define GSC_HECI_VALIDITY_MARKER 0xA578875A
@@ -47,7 +50,7 @@ struct intel_gsc_mtl_header {
* we distinguish the flags using OUTFLAG or INFLAG
*/
u32 flags;
-#define GSC_OUTFLAG_MSG_PENDING 1
+#define GSC_OUTFLAG_MSG_PENDING 1
Nit: this change on the define is not really needed
sure - will fix.
Daniele