[Intel-gfx] [PATCH v2 5/8] drm/i915/dp: rewrite DP 2.0 128b/132b link training based on errata
Ville Syrjälä
ville.syrjala at linux.intel.com
Tue Feb 8 09:39:11 UTC 2022
On Tue, Feb 08, 2022 at 11:17:22AM +0200, Jani Nikula wrote:
> On Fri, 04 Feb 2022, Ville Syrjälä <ville.syrjala at linux.intel.com> wrote:
> > On Thu, Feb 03, 2022 at 11:03:54AM +0200, Jani Nikula wrote:
> >> The DP 2.0 errata completely overhauls the 128b/132b link training, with
> >> no provisions for backward compatibility with the original DP 2.0
> >> specification.
> >>
> >> The changes are too intrusive to consider reusing the same code for both
> >> 8b/10b and 128b/132b, mainly because the LTTPR channel equalisation is
> >> done concurrently instead of serialized.
> >>
> >> NOTES:
> >>
> >> * It's a bit unclear when to wait for DP_INTERLANE_ALIGN_DONE and
> >> per-lane DP_LANE_SYMBOL_LOCKED. Figure xx4 in the SCR implies the
> >> LANEx_CHANNEL_EQ_DONE sequence may end with either 0x77,0x77,0x85 *or*
> >> 0x33,0x33,0x84 (for four lane configuration in DPCD 0x202..0x204)
> >> i.e. without the above bits set. Text elsewhere seems contradictory or
> >> incomplete.
> >>
> >> * We read entire link status (6 bytes) everywhere instead of individual
> >> DPCD addresses.
> >>
> >> * There are some subtle ambiguities or contradictions in the order of
> >> some DPCD access and TPS signal enables/disables. It's also not clear
> >> whether these are significant.
> >>
> >> v2:
> >> - Always try one last time after timeouts to avoid races (Ville)
> >> - Extend timeout to cover the entire LANEx_EQ_DONE sequence (Ville)
> >> - Also check for eq interlane align done in LANEx_CDS_DONE Sequence (Ville)
> >> - Check for Intra-hop status before link training
> >>
> >> Cc: Uma Shankar <uma.shankar at intel.com>
> >> Cc: Ville Syrjälä <ville.syrjala at linux.intel.com>
> >> Signed-off-by: Jani Nikula <jani.nikula at intel.com>
> >> ---
> >> .../drm/i915/display/intel_dp_link_training.c | 279 +++++++++++++++++-
> >> 1 file changed, 278 insertions(+), 1 deletion(-)
> >>
> >> diff --git a/drivers/gpu/drm/i915/display/intel_dp_link_training.c b/drivers/gpu/drm/i915/display/intel_dp_link_training.c
> >> index 4e507aa75a03..cc2b82d9114c 100644
> >> --- a/drivers/gpu/drm/i915/display/intel_dp_link_training.c
> >> +++ b/drivers/gpu/drm/i915/display/intel_dp_link_training.c
> >> @@ -1102,6 +1102,277 @@ intel_dp_link_train_all_phys(struct intel_dp *intel_dp,
> >> return ret;
> >> }
> >>
> >> +
> >> +/*
> >> + * 128b/132b DP LANEx_EQ_DONE Sequence (DP 2.0 E11 3.5.2.16.1)
> >> + */
> >> +static bool
> >> +intel_dp_128b132b_lane_eq(struct intel_dp *intel_dp,
> >> + const struct intel_crtc_state *crtc_state)
> >> +{
> >> + struct intel_encoder *encoder = &dp_to_dig_port(intel_dp)->base;
> >> + struct drm_i915_private *i915 = to_i915(encoder->base.dev);
> >> + u8 link_status[DP_LINK_STATUS_SIZE];
> >> + int delay_us;
> >> + int try, max_tries = 20;
> >> + unsigned long deadline;
> >> + bool timeout = false;
> >> +
> >> + /*
> >> + * Reset signal levels. Start transmitting 128b/132b TPS1.
> >> + *
> >> + * Put DPRX and LTTPRs (if any) into intra-hop AUX mode by writing TPS1
> >> + * in DP_TRAINING_PATTERN_SET.
> >> + */
> >> + if (!intel_dp_reset_link_train(intel_dp, crtc_state, DP_PHY_DPRX,
> >> + DP_TRAINING_PATTERN_1)) {
> >> + drm_err(&i915->drm,
> >> + "[ENCODER:%d:%s] Failed to start 128b/132b TPS1\n",
> >> + encoder->base.base.id, encoder->base.name);
> >> + return false;
> >> + }
> >> +
> >> + delay_us = drm_dp_128b132b_read_aux_rd_interval(&intel_dp->aux);
> >> +
> >> + /* Read the initial TX FFE settings. */
> >> + if (drm_dp_dpcd_read_link_status(&intel_dp->aux, link_status) < 0) {
> >> + drm_err(&i915->drm,
> >> + "[ENCODER:%d:%s] Failed to read TX FFE presets\n",
> >> + encoder->base.base.id, encoder->base.name);
> >> + return false;
> >> + }
> >> +
> >> + /* Update signal levels and training set as requested. */
> >> + intel_dp_get_adjust_train(intel_dp, crtc_state, DP_PHY_DPRX, link_status);
> >> + if (!intel_dp_update_link_train(intel_dp, crtc_state, DP_PHY_DPRX)) {
> >> + drm_err(&i915->drm,
> >> + "[ENCODER:%d:%s] Failed to set initial TX FFE settings\n",
> >> + encoder->base.base.id, encoder->base.name);
> >> + return false;
> >> + }
> >> +
> >> + /* Start transmitting 128b/132b TPS2. */
> >> + if (!intel_dp_set_link_train(intel_dp, crtc_state, DP_PHY_DPRX,
> >> + DP_TRAINING_PATTERN_2)) {
> >> + drm_err(&i915->drm,
> >> + "[ENCODER:%d:%s] Failed to start 128b/132b TPS2\n",
> >> + encoder->base.base.id, encoder->base.name);
> >> + return false;
> >> + }
> >> +
> >> + /* Time budget for the LANEx_EQ_DONE Sequence */
> >> + deadline = jiffies + msecs_to_jiffies(400);
> >
> > Didn't we have a msecs_to_jiffies_timeout() that adds an extra
> > jiffy to make sure we don't bail too early?
>
> Hrmh, still local in i915_utils.h after all these years.
>
> >
> >> +
> >> + for (try = 0; try < max_tries; try++) {
> >> + usleep_range(delay_us, 2 * delay_us);
> >> +
> >> + /*
> >> + * The delay may get updated. The transmitter shall read the
> >> + * delay before link status during link training.
> >> + */
> >> + delay_us = drm_dp_128b132b_read_aux_rd_interval(&intel_dp->aux);
> >> +
> >> + if (drm_dp_dpcd_read_link_status(&intel_dp->aux, link_status) < 0) {
> >> + drm_err(&i915->drm,
> >> + "[ENCODER:%d:%s] Failed to read link status\n",
> >> + encoder->base.base.id, encoder->base.name);
> >> + return false;
> >> + }
> >> +
> >> + if (drm_dp_128b132b_link_training_failed(link_status)) {
> >> + intel_dp_dump_link_status(intel_dp, DP_PHY_DPRX, link_status);
> >> + drm_err(&i915->drm,
> >> + "[ENCODER:%d:%s] Downstream link training failure\n",
> >> + encoder->base.base.id, encoder->base.name);
> >> + return false;
> >> + }
> >> +
> >> + if (drm_dp_128b132b_lane_channel_eq_done(link_status, crtc_state->lane_count)) {
> >> + drm_dbg_kms(&i915->drm,
> >> + "[ENCODER:%d:%s] Lane channel eq done\n",
> >> + encoder->base.base.id, encoder->base.name);
> >> + break;
> >> + }
> >
> > The state diagrame has thow two steps in reverse order. I suppose it
> > doens't matter but probably better if we don't deviate too much without
> > a good reason.
>
> I thought the fail bit should probably have priority. What if you have
> both eq done *and* lt failed set, for whatever reason?
Yeah, I guess from that angle it might make sense to check
the fail bit first. Unless there's some crazy device out there
that succeeds in link training but accidentally sets the fail bit
anyway.
>
> >
> >> +
> >> + if (timeout) {
> >> + intel_dp_dump_link_status(intel_dp, DP_PHY_DPRX, link_status);
> >> + drm_err(&i915->drm,
> >> + "[ENCODER:%d:%s] Lane channel eq timeout\n",
> >> + encoder->base.base.id, encoder->base.name);
> >> + return false;
> >> + }
> >> +
> >> + if (time_after(jiffies, deadline))
> >> + timeout = true; /* try one last time after deadline */
> >
> > Is there a reason we can't do this just before drm_dp_dpcd_read_link_status()
> > so we don't have to pass the timeout status from one loop iteration to
> > the next?
>
> The point is to check one last time after timeout has passed, like you
> suggested in previous review, and I agreed.
Sure but why can't it be something more like?
timeout = time_after();
read_status();
if (bad)
bail;
if (timeout)
bail;
I think we have it more like that in wait_for()/etc.
>
> >
> >> +
> >> + /* Update signal levels and training set as requested. */
> >> + intel_dp_get_adjust_train(intel_dp, crtc_state, DP_PHY_DPRX, link_status);
> >> + if (!intel_dp_update_link_train(intel_dp, crtc_state, DP_PHY_DPRX)) {
> >> + drm_err(&i915->drm,
> >> + "[ENCODER:%d:%s] Failed to update TX FFE settings\n",
> >> + encoder->base.base.id, encoder->base.name);
> >> + return false;
> >> + }
> >> + }
> >> +
> >> + if (try == max_tries) {
> >> + intel_dp_dump_link_status(intel_dp, DP_PHY_DPRX, link_status);
> >> + drm_err(&i915->drm,
> >> + "[ENCODER:%d:%s] Max loop count reached\n",
> >> + encoder->base.base.id, encoder->base.name);
> >> + return false;
> >> + }
> >> +
> >> + for (;;) {
> >> + if (drm_dp_128b132b_eq_interlane_align_done(link_status)) {
> >> + drm_dbg_kms(&i915->drm,
> >> + "[ENCODER:%d:%s] Interlane align done\n",
> >> + encoder->base.base.id, encoder->base.name);
> >> + break;
> >> + }
> >
> > Not a big fan of pasing the link_status between loops. Can't
> > we just read the status here at the start of the loop always?
>
> I guess. It can lead to an extra read if interlane align gets done
> already in the previous loop.
>
> Combined with the timeout flag from the previous loop, the
> implementation proposed here passes CDS if the last attempt after
> timeout has passed has interlane align done, but does not read anything
> more here.
I don't think optimizing away one dpcd read should be very important
here. Less stuff to think about if we don't have to pass data
from one loop to the next IMO. If nothing gets passed then nothing
can be stale/etc.
>
> >
> >> +
> >> + if (timeout) {
> >> + intel_dp_dump_link_status(intel_dp, DP_PHY_DPRX, link_status);
> >> + drm_err(&i915->drm,
> >> + "[ENCODER:%d:%s] Interlane align timeout\n",
> >> + encoder->base.base.id, encoder->base.name);
> >> + return false;
> >> + }
> >> +
> >> + if (time_after(jiffies, deadline))
> >> + timeout = true; /* try one last time after deadline */
> >> +
> >> + usleep_range(2000, 3000);
> >> +
> >> + if (drm_dp_dpcd_read_link_status(&intel_dp->aux, link_status) < 0) {
> >> + drm_err(&i915->drm,
> >> + "[ENCODER:%d:%s] Failed to read link status\n",
> >> + encoder->base.base.id, encoder->base.name);
> >> + return false;
> >> + }
> >> +
> >> + if (drm_dp_128b132b_link_training_failed(link_status)) {
> >> + intel_dp_dump_link_status(intel_dp, DP_PHY_DPRX, link_status);
> >> + drm_err(&i915->drm,
> >> + "[ENCODER:%d:%s] Downstream link training failure\n",
> >> + encoder->base.base.id, encoder->base.name);
> >> + return false;
> >> + }
> >> + }
> >> +
> >> + return true;
> >> +}
> >> +
> >> +/*
> >> + * 128b/132b DP LANEx_CDS_DONE Sequence (DP 2.0 E11 3.5.2.16.2)
> >> + */
> >> +static bool
> >> +intel_dp_128b132b_lane_cds(struct intel_dp *intel_dp,
> >> + const struct intel_crtc_state *crtc_state,
> >> + int lttpr_count)
> >> +{
> >> + struct intel_encoder *encoder = &dp_to_dig_port(intel_dp)->base;
> >> + struct drm_i915_private *i915 = to_i915(encoder->base.dev);
> >> + u8 link_status[DP_LINK_STATUS_SIZE];
> >> + unsigned long deadline;
> >> + bool timeout = false;
> >> +
> >> + if (drm_dp_dpcd_writeb(&intel_dp->aux, DP_TRAINING_PATTERN_SET,
> >> + DP_TRAINING_PATTERN_2_CDS) != 1) {
> >> + drm_err(&i915->drm,
> >> + "[ENCODER:%d:%s] Failed to start 128b/132b TPS2 CDS\n",
> >> + encoder->base.base.id, encoder->base.name);
> >> + return false;
> >> + }
> >> +
> >> + /* Time budget for the LANEx_CDS_DONE Sequence */
> >> + deadline = jiffies + msecs_to_jiffies((lttpr_count + 1) * 20);
> >> +
> >> + for (;;) {
> >> + usleep_range(2000, 3000);
> >> +
> >> + if (drm_dp_dpcd_read_link_status(&intel_dp->aux, link_status) < 0) {
> >> + drm_err(&i915->drm,
> >> + "[ENCODER:%d:%s] Failed to read link status\n",
> >> + encoder->base.base.id, encoder->base.name);
> >> + return false;
> >> + }
> >> +
> >> + if (drm_dp_128b132b_eq_interlane_align_done(link_status) &&
> >> + drm_dp_128b132b_cds_interlane_align_done(link_status) &&
> >> + drm_dp_128b132b_lane_symbol_locked(link_status, crtc_state->lane_count)) {
> >> + drm_dbg_kms(&i915->drm,
> >> + "[ENCODER:%d:%s] CDS interlane align done\n",
> >> + encoder->base.base.id, encoder->base.name);
> >> + break;
> >> + }
> >> +
> >> + if (drm_dp_128b132b_link_training_failed(link_status)) {
> >> + intel_dp_dump_link_status(intel_dp, DP_PHY_DPRX, link_status);
> >> + drm_err(&i915->drm,
> >> + "[ENCODER:%d:%s] Downstream link training failure\n",
> >> + encoder->base.base.id, encoder->base.name);
> >> + return false;
> >> + }
> >> +
> >> + if (timeout) {
> >> + intel_dp_dump_link_status(intel_dp, DP_PHY_DPRX, link_status);
> >> + drm_err(&i915->drm,
> >> + "[ENCODER:%d:%s] CDS timeout\n",
> >> + encoder->base.base.id, encoder->base.name);
> >> + return false;
> >> + }
> >> +
> >> + if (time_after(jiffies, deadline))
> >> + timeout = true; /* try one last time after deadline */
> >> + }
> >> +
> >> + /* FIXME: Should DP_TRAINING_PATTERN_DISABLE be written first? */
> >> + if (intel_dp->set_idle_link_train)
> >> + intel_dp->set_idle_link_train(intel_dp, crtc_state);
> >> +
> >> + return true;
> >> +}
> >> +
> >> +/*
> >> + * 128b/132b link training sequence. (DP 2.0 E11 SCR on link training.)
> >> + */
> >> +static bool
> >> +intel_dp_128b132b_link_train(struct intel_dp *intel_dp,
> >> + const struct intel_crtc_state *crtc_state,
> >> + int lttpr_count)
> >> +{
> >> + struct drm_i915_private *i915 = dp_to_i915(intel_dp);
> >> + struct intel_connector *connector = intel_dp->attached_connector;
> >> + struct intel_encoder *encoder = &dp_to_dig_port(intel_dp)->base;
> >> + bool passed = false;
> >> + u8 sink_status;
> >> +
> >> + /* FIXME: We should possibly do this earlier. */
> >> + if (drm_dp_dpcd_readb(&intel_dp->aux, DP_SINK_STATUS, &sink_status) != 1) {
> >> + drm_dbg_kms(&i915->drm, "Failed to read sink status\n");
> >> + return false;
> >> + }
> >> +
> >> + if (sink_status & DP_INTRA_HOP_AUX_REPLY_INDICATION) {
> >> + drm_dbg_kms(&i915->drm, "Previous link training in progress\n");
> >> + return false;
> >> + }
> >> +
> >> + if (intel_dp_128b132b_lane_eq(intel_dp, crtc_state) &&
> >> + intel_dp_128b132b_lane_cds(intel_dp, crtc_state, lttpr_count))
> >> + passed = true;
> >
> > I think we are supposed to wait for the intra-hop AUX bit to go low
> > here.
>
> Where's that said in the spec? I thought we're supposed to wait *before*
> starting link training.
I think it was mentioned there somewhere. Also we have to wait
before we can allow normal AUX activity anyway (assuming we had
that part hookd up).
>
> Also, I think the bit might only go low after we've stopped link
> training, i.e. once the higher level calls intel_dp_stop_link_train()
> which sets DP_TRAINING_PATTERN_DISABLE. So the wait, if any, would need
> to be in that function.
Yeah that might be right spot.
--
Ville Syrjälä
Intel
More information about the Intel-gfx
mailing list