Quoting Imre Deak (2025-11-14 22:22:32-03:00) >On Sat, Nov 15, 2025 at 02:40:24AM +0200, Imre Deak wrote: >> On Fri, Nov 14, 2025 at 04:46:29PM -0300, Gustavo Sousa wrote: >> > Quoting Imre Deak (2025-11-12 14:53:47-03:00) >> > >On Fri, Nov 07, 2025 at 09:05:40PM -0300, Gustavo Sousa wrote: >> > >> Xe3p_LPD has a new feature that allows the driver to allocate at runtime >> > >> the DDI (TC ones) port to drive a legacy connection on the Type-C >> > >> subsystem. This allows better resource utilization, because now there >> > >> is no need to statically reserve ports for legacy connectors on the >> > >> Type-C subsystem. >> > >> >> > >> That said, our driver is not yet ready for the dynamic allocation. >> > >> Thus, as an incremental step, let's add the logic containing the >> > >> required programming sequence for the allocation, but, instead of >> > >> selecting the first available port, we try so use the 1:1 mapping >> > >> expected by the driver today. >> > >> >> > >> Bspec: 68954 >> > >> Co-developed-by: Dnyaneshwar Bhadane <[email protected]> >> > >> Signed-off-by: Dnyaneshwar Bhadane <[email protected]> >> > >> Signed-off-by: Gustavo Sousa <[email protected]> >> > >> --- >> > >> >> > >> NOTE: This patch is still a WIP. There are some opens to resolve here. >> > >> Nevertheless, I'm sending it here for early feedback. >> > >> >> > >> For the HIP-index stuff, I have a local refactor started and need to >> > >> finish it up and send it. >> > >> >> > >> The other open is about concurrent calls to iom_dp_resource_lock(). It >> > >> is likely that we need to have a software lock to prevent concurrent >> > >> access to IOM_DP_HW_RESOURCE_SEMAPHORE from our driver. >> > >> --- >> > >> drivers/gpu/drm/i915/display/intel_display_regs.h | 20 ++- >> > >> drivers/gpu/drm/i915/display/intel_tc.c | 151 >> > >> +++++++++++++++++++++- >> > >> 2 files changed, 169 insertions(+), 2 deletions(-) >> > >> >> > >> diff --git a/drivers/gpu/drm/i915/display/intel_display_regs.h >> > >> b/drivers/gpu/drm/i915/display/intel_display_regs.h >> > >> index 89ea0156ee06..0cf7d43ce210 100644 >> > >> --- a/drivers/gpu/drm/i915/display/intel_display_regs.h >> > >> +++ b/drivers/gpu/drm/i915/display/intel_display_regs.h >> > >> @@ -2908,6 +2908,25 @@ enum skl_power_gate { >> > >> #define DP_PIN_ASSIGNMENT(idx, x) ((x) << ((idx) * 4)) >> > >> /* See enum intel_tc_pin_assignment for the pin assignment field >> > >> values. */ >> > >> >> > >> +/* >> > >> + * FIXME: There is also a definition for this register in >> > >> intel_dkl_phy_regs.h. >> > >> + * We need to consolidate the definitions. >> > >> + */ >> > >> +#define HIP_INDEX_REG0 _MMIO(0x1010a0) >> > >> +#define HIP_168_INDEX_MASK REG_GENMASK(3, 0) >> > >> +#define HIP_168_IOM_RES_MGMT >> > >> REG_FIELD_PREP(HIP_168_INDEX_MASK, 0x1) >> > >> + >> > >> +#define IOM_DP_HW_RESOURCE_SEMAPHORE _MMIO(0x168038) >> > >> +#define IOM_DP_HW_SEMLOCK REG_BIT(31) >> > >> +#define IOM_REQUESTOR_ID_MASK REG_GENMASK(3, >> > >> 0) >> > >> +#define IOM_REQUESTOR_ID_DISPLAY_ENGINE >> > >> REG_FIELD_PREP(IOM_REQUESTOR_ID_MASK, 0x4) >> > >> + >> > >> +#define IOM_DP_RESOURCE_MNG _MMIO(0x16802c) >> > >> +#define IOM_DDI_CONSUMER_SHIFT(tc_port) ((tc_port) * 4) >> > >> +#define IOM_DDI_CONSUMER_MASK(tc_port) (0xf << >> > >> IOM_DDI_CONSUMER_SHIFT(tc_port)) >> > >> +#define IOM_DDI_CONSUMER(tc_port, x) ((x) << >> > >> IOM_DDI_CONSUMER_SHIFT(tc_port)) >> > >> +#define IOM_DDI_CONSUMER_STATIC_TC(tc_port) >> > >> IOM_DDI_CONSUMER(tc_port, 0x8 + (tc_port)) >> > > >> > >It would be simpler to define the above without the shift, i.e. as 8 + >> > >tc_port. >> > >> > You mean something like this? >> > >> > #define IOM_DDI_CONSUMER_STATIC_TC(tc_port) (0x8 + (tc_port)) >> > >> > ? >> > >> > Yeah... Looking at the code, we wouldn't need to shift back when >> > defining expected_consumer. >> >> Yes. >> >> > >> > > >> > >> + >> > >> #define _TCSS_DDI_STATUS_1 0x161500 >> > >> #define _TCSS_DDI_STATUS_2 0x161504 >> > >> #define TCSS_DDI_STATUS(tc) >> > >> _MMIO(_PICK_EVEN(tc, \ >> > >> @@ -2946,5 +2965,4 @@ enum skl_power_gate { >> > >> #define MTL_TRDPRE_MASK REG_GENMASK(7, 0) >> > >> >> > >> >> > >> - >> > >> #endif /* __INTEL_DISPLAY_REGS_H__ */ >> > >> diff --git a/drivers/gpu/drm/i915/display/intel_tc.c >> > >> b/drivers/gpu/drm/i915/display/intel_tc.c >> > >> index 7e17ca018748..3c333999bbe4 100644 >> > >> --- a/drivers/gpu/drm/i915/display/intel_tc.c >> > >> +++ b/drivers/gpu/drm/i915/display/intel_tc.c >> > >> @@ -9,6 +9,7 @@ >> > >> >> > >> #include "i915_reg.h" >> > >> #include "intel_atomic.h" >> > >> +#include "intel_bios.h" >> > >> #include "intel_cx0_phy_regs.h" >> > >> #include "intel_ddi.h" >> > >> #include "intel_de.h" >> > >> @@ -25,6 +26,9 @@ >> > >> #include "intel_modeset_lock.h" >> > >> #include "intel_tc.h" >> > >> >> > >> +#define IOM_DP_RES_SEMAPHORE_LOCK_TIMEOUT_US 10 >> > >> +#define IOM_DP_RES_SEMAPHORE_RETRY_TIMEOUT_US 10000 >> > > >> > >The above param names should make it clear how poll_timeout_us() uses >> > >them (i.e. stg like sleep vs. timeout instead of lock_timeout vs. >> > >retry_timeout), but probably even clearer to just drop the defines and >> > >inline the values in the call. >> > >> > Ack. I going with the latter. >> > >> > > >> > >> + >> > >> enum tc_port_mode { >> > >> TC_PORT_DISCONNECTED, >> > >> TC_PORT_TBT_ALT, >> > >> @@ -1200,6 +1204,143 @@ static void xelpdp_tc_phy_get_hw_state(struct >> > >> intel_tc_port *tc) >> > >> __tc_cold_unblock(tc, domain, tc_cold_wref); >> > >> } >> > >> >> > >> +static void iom_res_mgmt_prepare_reg_access(struct intel_display >> > >> *display) >> > >> +{ >> > >> + /* >> > >> + * IOM resource management registers live in the 2nd 4KB page >> > >> of IOM >> > >> + * address space. So we need to configure HIP_INDEX_REG0 with >> > >> the >> > >> + * correct index. >> > >> + * >> > >> + * FIXME: We need to have this and dekel PHY implementation >> > >> using a >> > >> + * common abstraction to access registers on the HIP-indexed >> > >> ranges, and >> > >> + * this function would then be dropped. >> > >> + */ >> > >> + intel_de_rmw(display, HIP_INDEX_REG0, >> > >> + HIP_168_INDEX_MASK, HIP_168_IOM_RES_MGMT); >> > > >> > >This matches what intel_dkl_phy.c does, that one also taking the >> > >required lock around the access. At one point the intel_dkl_phy >> > >file/interface could be renamed to intel_tc_reg or similar accordingly. >> > >> > I have already started a local series that introduces the "HIP-index >> > based registers" abstraction. I need to go back to finish it up. >> > >> > The basic idea is that both intel_dkl_phy.c and this IOM stuff would use >> > functions specific to accessing registers behind the HIP-based ranges. >> >> Using intel_hip_reg instead of intel_tc_reg is probably better, but I >> still think the current interface should be just renamed, instead of >> adding a new interface and making the current >> intel_dkl_phy_read/write/rmw() use that new interface. >> >> I went ahead and put the above together now: >> https://github.com/ideak/linux/commits/hip-reg >> >> > >> > To give an idea, here is a copy/paste of the kerneldoc I currently have >> > in that WIP series: >> > >> > | diff --git a/drivers/gpu/drm/i915/display/intel_hip_idx.c >> > b/drivers/gpu/drm/i915/display/intel_hip_idx.c >> > | new file mode 100644 >> > | index 000000000000..ff2492b8275d >> > | --- /dev/null >> > | +++ b/drivers/gpu/drm/i915/display/intel_hip_idx.c >> > | @@ -0,0 +1,110 @@ >> > | +// SPDX-License-Identifier: MIT >> > | +/* >> > | + * Copyright (C) 2025 Intel Corporation >> > | + */ >> > | + >> > | +/** >> > | + * DOC: Display HIP-indexed register access >> > | + * >> > | + * Display MMIO mapping for offsets in [0x168000,0x16ffff] are >> > governed by >> > | + * configurations in the HIP_INDEX registers provided by the hardware. >> > | + * >> > | + * Usually each of the valid 4KB range in that space will be mapped >> > to some IP >> > | + * block, providing access to registers of that IP. The HIP_INDEX >> > registers >> > | + * expose an 8-bit index value for each 4KB range. Those indices >> > provide a way >> > | + * to access data that is beyond the initial 4KB block provided by >> > the target >> > | + * IP. >> > | + * >> > | + * As an example, say that the range [0x16a000,0x16afff] is mapped to >> > some >> > | + * sub-IP that contains a 8KB register file. To access the initial >> > 4KB of that >> > | + * register file, we would need to set the index respective to >> > | + * [0x16a000,0x16afff] in HIP_INDEX to 0; to access data in the >> > second 4KB >> > | + * window, we would need to set the index to 1. >> > | + * >> > | + * For simple read, write or rmw operations on a single register, the >> > | + * functions intel_hip_idx_reg_read(), intel_hip_idx_reg_write() and >> > | + * intel_hip_idx_reg_rmw() can be used, which will call >> > intel_hip_idx_lock() >> > | + * and intel_hip_idx_unlock() internally. >> > | + * >> > | + * For other scenarios, then it is necessary to wrap the register >> > accesses >> > | + * with explicit calls to intel_hip_idx_lock() and >> > intel_hip_idx_unlock(), and >> > | + * use the MMIO functions provided by intel_de.h. For the latter, the >> > function >> > | + * intel_hip_idx_reg_to_i915_reg() needs to be used in order to >> > provide the >> > | + * correct reg argument to those functions. >> > | + */ >> > | (...) >> > >> > > >> > >> +} >> > >> + >> > >> +/* >> > >> + * FIXME: This function also needs to avoid concurrent accesses from >> > >> the driver >> > >> + * itself, possibly via a software lock. >> > >> + */ >> > >> +static int iom_dp_resource_lock(struct intel_tc_port *tc) >> > >> +{ >> > >> + struct intel_display *display = to_intel_display(tc->dig_port); >> > >> + u32 val = IOM_DP_HW_SEMLOCK | IOM_REQUESTOR_ID_DISPLAY_ENGINE; >> > >> + int ret; >> > >> + >> > >> + iom_res_mgmt_prepare_reg_access(display); >> > >> + ret = poll_timeout_us(intel_de_write(display, >> > >> IOM_DP_HW_RESOURCE_SEMAPHORE, val), >> > >> + (intel_de_read(display, >> > >> IOM_DP_HW_RESOURCE_SEMAPHORE) & val) == val, >> > > >> > >This happens to work, but it's more future proof/bspec conformant to >> > >properly mask the requestor ID field when reading it back. >> > >> > Agreed. >> > >> > > >> > >> + IOM_DP_RES_SEMAPHORE_LOCK_TIMEOUT_US, >> > >> + IOM_DP_RES_SEMAPHORE_RETRY_TIMEOUT_US, >> > >> false); >> > >> + >> > >> + if (ret) >> > >> + drm_err(display->drm, "Port %s: timeout trying to lock >> > >> IOM semaphore\n", >> > >> + tc->port_name); >> > >> + >> > >> + return ret; >> > >> +} >> > >> + >> > >> +static void iom_dp_resource_unlock(struct intel_tc_port *tc) >> > >> +{ >> > >> + struct intel_display *display = to_intel_display(tc->dig_port); >> > >> + >> > >> + iom_res_mgmt_prepare_reg_access(display); >> > >> + intel_de_write(display, IOM_DP_HW_RESOURCE_SEMAPHORE, >> > >> IOM_REQUESTOR_ID_DISPLAY_ENGINE); >> > >> +} >> > >> + >> > >> +static bool xe3p_tc_iom_allocate_ddi(struct intel_tc_port *tc, bool >> > >> allocate) >> > >> +{ >> > >> + struct intel_display *display = to_intel_display(tc->dig_port); >> > >> + struct intel_digital_port *dig_port = tc->dig_port; >> > >> + enum tc_port tc_port = intel_encoder_to_tc(&dig_port->base); >> > >> + u32 val; >> > >> + u32 consumer; >> > >> + u32 expected_consumer; >> > >> + bool ret; >> > >> + >> > >> + if (DISPLAY_VER(display) < 35) >> > >> + return true; >> > >> + >> > >> + if (tc->mode != TC_PORT_LEGACY) >> > >> + return true; >> > >> + >> > >> + if >> > >> (!intel_bios_encoder_supports_dyn_port_over_tc(dig_port->base.devdata)) >> > > >> > >dedicated_external is stored separately in dig_port, why the "related" >> > >dyn_port_over_tc flag isn't? >> > >> > The only reason dedicated_external is stored is because VBT data is >> > already freed by the time intel_encoder_is_tc() is called in the driver >> > unbind path. In the future we should fix the VBT lifetime issue in >> > order to be able to drop the dedicated_external member of dig_port. >> > >> > > >> > >> + return true; >> > >> + >> > >> + if (iom_dp_resource_lock(tc)) >> > >> + return false; >> > >> + >> > >> + val = intel_de_read(display, IOM_DP_RESOURCE_MNG); >> > >> + >> > >> + consumer = val & IOM_DDI_CONSUMER_MASK(tc_port); >> > >> + consumer >>= IOM_DDI_CONSUMER_SHIFT(tc_port); >> > >> + >> > >> + /* >> > >> + * Bspec instructs to select first available DDI, but our >> > >> driver is not >> > >> + * ready for such dynamic allocation yet. For now, we force a >> > >> "static" >> > >> + * allocation: map the physical port (where HPD happens) to the >> > >> + * encoder's DDI (logical TC port, represented by tc_port). >> > >> + */ >> > >> + expected_consumer = IOM_DDI_CONSUMER_STATIC_TC(tc_port); >> > >> + expected_consumer >>= IOM_DDI_CONSUMER_SHIFT(tc_port); >> >> One more thing occured to me: why can't this allocate any free DDI? IOW >> does the address of DDI_BUF_CTL (aka DDI_CTL_DE) used for tc_port depend >> on which DDI gets allocated (or is there any other dependency on which >> DDI got allocated)? AFAICS there is no such dependency and the above >> address would be DDI_BUF_CTL(encoder->port) regardless of the allocated >> DDI. In that case any free DDI could be allocated here. > >Ok, checking this again, DDI_BUF_CTL etc. DDI register addresses will >depend on the allocated DDI. So nvm the above, the mapping needs to >stay 1:1 for now until all the DDI reg accesses are converted to index >the registers with the allocated DDI index.
As far as I understand this, especially after talking with Windows folks, the allocated DDI will define the port index for the whole programming, including the registers used to program the PHY - and the hardware would take care of routing to the correct PHY. Thus, it appears we would need to do the allocation at hotplug time, like saying "this PHY will be driven by DDI x". One of the reasons I think we can't allocate a free DDI at the moment is that the driver is expecting a 1:1 static mapping for HPD interations. We will neeed to make the driver aware of the mapping in order to use the correct encoder when handling HPD events. -- Gustavo Sousa
