On Thu, Oct 23, 2025 at 09:54:02PM +0200, Michal Wajdeczko wrote:
>
>
> On 10/22/2025 12:41 AM, Michał Winiarski wrote:
> > Connect the helpers to allow save and restore of VRAM migration data in
> > stop_copy / resume device state.
> >
> > Co-developed-by: Lukasz Laguna <[email protected]>
> > Signed-off-by: Lukasz Laguna <[email protected]>
> > Signed-off-by: Michał Winiarski <[email protected]>
> > ---
> > drivers/gpu/drm/xe/xe_gt_sriov_pf_control.c | 18 ++
> > .../gpu/drm/xe/xe_gt_sriov_pf_control_types.h | 2 +
> > drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c | 222 ++++++++++++++++++
> > drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.h | 6 +
> > .../drm/xe/xe_gt_sriov_pf_migration_types.h | 3 +
> > drivers/gpu/drm/xe/xe_sriov_pf_control.c | 3 +
> > 6 files changed, 254 insertions(+)
> >
> > diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_pf_control.c
> > b/drivers/gpu/drm/xe/xe_gt_sriov_pf_control.c
> > index e7156ad3d1839..680f2de44144b 100644
> > --- a/drivers/gpu/drm/xe/xe_gt_sriov_pf_control.c
> > +++ b/drivers/gpu/drm/xe/xe_gt_sriov_pf_control.c
> > @@ -191,6 +191,7 @@ static const char *control_bit_to_string(enum
> > xe_gt_sriov_control_bits bit)
> > CASE2STR(SAVE_DATA_GUC);
> > CASE2STR(SAVE_DATA_GGTT);
> > CASE2STR(SAVE_DATA_MMIO);
> > + CASE2STR(SAVE_DATA_VRAM);
> > CASE2STR(SAVE_DATA_DONE);
> > CASE2STR(SAVE_FAILED);
> > CASE2STR(SAVED);
> > @@ -832,6 +833,7 @@ static void pf_exit_vf_save_wip(struct xe_gt *gt,
> > unsigned int vfid)
> > pf_escape_vf_state(gt, vfid, XE_GT_SRIOV_STATE_SAVE_DATA_GGTT);
> > pf_escape_vf_state(gt, vfid, XE_GT_SRIOV_STATE_SAVE_DATA_MMIO);
> > pf_escape_vf_state(gt, vfid, XE_GT_SRIOV_STATE_SAVE_DATA_DONE);
> > + pf_escape_vf_state(gt, vfid, XE_GT_SRIOV_STATE_SAVE_DATA_VRAM);
> > }
> > }
> >
> > @@ -885,6 +887,19 @@ static int pf_handle_vf_save_data(struct xe_gt *gt,
> > unsigned int vfid)
> > ret = xe_gt_sriov_pf_migration_mmio_save(gt, vfid);
> > if (ret)
> > return ret;
> > +
> > + pf_enter_vf_state(gt, vfid, XE_GT_SRIOV_STATE_SAVE_DATA_VRAM);
> > + return -EAGAIN;
> > + }
> > +
> > + if (pf_exit_vf_state(gt, vfid, XE_GT_SRIOV_STATE_SAVE_DATA_VRAM)) {
> > + if (xe_gt_sriov_pf_migration_vram_size(gt, vfid) > 0) {
> > + ret = xe_gt_sriov_pf_migration_vram_save(gt, vfid);
> > + if (ret == -EAGAIN)
> > + pf_enter_vf_state(gt, vfid,
> > XE_GT_SRIOV_STATE_SAVE_DATA_VRAM);
> > + if (ret)
> > + return ret;
> > + }
> > }
> >
> > return 0;
> > @@ -1100,6 +1115,9 @@ pf_handle_vf_restore_data(struct xe_gt *gt, unsigned
> > int vfid)
> > case XE_SRIOV_MIGRATION_DATA_TYPE_GUC:
> > ret = xe_gt_sriov_pf_migration_guc_restore(gt, vfid, data);
> > break;
> > + case XE_SRIOV_MIGRATION_DATA_TYPE_VRAM:
> > + ret = xe_gt_sriov_pf_migration_vram_restore(gt, vfid, data);
> > + break;
> > default:
> > xe_gt_sriov_notice(gt, "Skipping VF%u unknown data type: %d\n",
> > vfid, data->type);
> > break;
> > diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_pf_control_types.h
> > b/drivers/gpu/drm/xe/xe_gt_sriov_pf_control_types.h
> > index 9dfcebd5078ac..fba10136f7cc7 100644
> > --- a/drivers/gpu/drm/xe/xe_gt_sriov_pf_control_types.h
> > +++ b/drivers/gpu/drm/xe/xe_gt_sriov_pf_control_types.h
> > @@ -36,6 +36,7 @@
> > * @XE_GT_SRIOV_STATE_SAVE_DATA_GUC: indicates PF needs to save VF GuC
> > migration data.
> > * @XE_GT_SRIOV_STATE_SAVE_DATA_GGTT: indicates PF needs to save VF GGTT
> > migration data.
> > * @XE_GT_SRIOV_STATE_SAVE_DATA_MMIO: indicates PF needs to save VF MMIO
> > migration data.
> > + * @XE_GT_SRIOV_STATE_SAVE_DATA_VRAM: indicates PF needs to save VF VRAM
> > migration data.
> > * @XE_GT_SRIOV_STATE_SAVE_DATA_DONE: indicates that all migration data
> > was produced by Xe.
> > * @XE_GT_SRIOV_STATE_SAVE_FAILED: indicates that VF save operation has
> > failed.
> > * @XE_GT_SRIOV_STATE_SAVED: indicates that VF data is saved.
> > @@ -82,6 +83,7 @@ enum xe_gt_sriov_control_bits {
> > XE_GT_SRIOV_STATE_SAVE_DATA_GUC,
> > XE_GT_SRIOV_STATE_SAVE_DATA_GGTT,
> > XE_GT_SRIOV_STATE_SAVE_DATA_MMIO,
> > + XE_GT_SRIOV_STATE_SAVE_DATA_VRAM,
> > XE_GT_SRIOV_STATE_SAVE_DATA_DONE,
> > XE_GT_SRIOV_STATE_SAVE_FAILED,
> > XE_GT_SRIOV_STATE_SAVED,
> > diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c
> > b/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c
> > index 41335b15ffdbe..2c6a86d98ee31 100644
> > --- a/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c
> > +++ b/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c
> > @@ -17,6 +17,7 @@
> > #include "xe_gt_sriov_printk.h"
> > #include "xe_guc_buf.h"
> > #include "xe_guc_ct.h"
> > +#include "xe_migrate.h"
> > #include "xe_sriov.h"
> > #include "xe_sriov_migration_data.h"
> > #include "xe_sriov_pf_migration.h"
> > @@ -485,6 +486,220 @@ int xe_gt_sriov_pf_migration_mmio_restore(struct
> > xe_gt *gt, unsigned int vfid,
> > return pf_restore_vf_mmio_mig_data(gt, vfid, data);
> > }
> >
> > +/**
> > + * xe_gt_sriov_pf_migration_vram_size() - Get the size of VF VRAM
> > migration data.
> > + * @gt: the &xe_gt
> > + * @vfid: the VF identifier
> > + *
> > + * This function is for PF only.
> > + *
> > + * Return: size in bytes or a negative error code on failure.
> > + */
> > +ssize_t xe_gt_sriov_pf_migration_vram_size(struct xe_gt *gt, unsigned int
> > vfid)
> > +{
> > + if (gt != xe_root_mmio_gt(gt_to_xe(gt)))
>
> probably you mean
>
> if (!xe_gt_is_main_type(gt))
Yes - I'll use it.
>
> > + return 0;
> > +
> > + return xe_gt_sriov_pf_config_get_lmem(gt, vfid);
> > +}
> > +
> > +static struct dma_fence *__pf_save_restore_vram(struct xe_gt *gt, unsigned
> > int vfid,
> > + struct xe_bo *vram, u64
> > vram_offset,
> > + struct xe_bo *sysmem, u64
> > sysmem_offset,
> > + size_t size, bool save)
> > +{
> > + struct dma_fence *ret = NULL;
> > + struct drm_exec exec;
> > + int err;
> > +
> > + drm_exec_init(&exec, DRM_EXEC_INTERRUPTIBLE_WAIT, 0);
> > + drm_exec_until_all_locked(&exec) {
> > + err = drm_exec_lock_obj(&exec, &vram->ttm.base);
> > + drm_exec_retry_on_contention(&exec);
> > + if (err) {
> > + ret = ERR_PTR(err);
> > + goto err;
> > + }
> > +
> > + err = drm_exec_lock_obj(&exec, &sysmem->ttm.base);
> > + drm_exec_retry_on_contention(&exec);
> > + if (err) {
> > + ret = ERR_PTR(err);
> > + goto err;
> > + }
> > + }
> > +
> > + ret = xe_migrate_vram_copy_chunk(vram, vram_offset, sysmem,
> > sysmem_offset, size,
> > + save ? XE_MIGRATE_COPY_TO_SRAM :
> > XE_MIGRATE_COPY_TO_VRAM);
> > +
> > +err:
> > + drm_exec_fini(&exec);
> > +
> > + return ret;
> > +}
> > +
> > +static int pf_save_vram_chunk(struct xe_gt *gt, unsigned int vfid,
> > + struct xe_bo *src_vram, u64 src_vram_offset,
> > + size_t size)
> > +{
> > + struct xe_sriov_migration_data *data;
> > + struct dma_fence *fence;
> > + int ret;
> > +
> > + data = xe_sriov_migration_data_alloc(gt_to_xe(gt));
> > + if (!data)
> > + return -ENOMEM;
> > +
> > + ret = xe_sriov_migration_data_init(data, gt->tile->id, gt->info.id,
> > + XE_SRIOV_MIGRATION_DATA_TYPE_VRAM,
> > + src_vram_offset, size);
> > + if (ret)
> > + goto fail;
> > +
> > + fence = __pf_save_restore_vram(gt, vfid,
> > + src_vram, src_vram_offset,
> > + data->bo, 0, size, true);
> > +
> > + ret = dma_fence_wait_timeout(fence, false, 5 * HZ);
> > + dma_fence_put(fence);
> > + if (!ret) {
> > + ret = -ETIME;
> > + goto fail;
> > + }
> > +
> > + pf_dump_mig_data(gt, vfid, data);
> > +
> > + ret = xe_gt_sriov_pf_migration_save_produce(gt, vfid, data);
> > + if (ret)
> > + goto fail;
> > +
> > + return 0;
> > +
> > +fail:
> > + xe_sriov_migration_data_free(data);
> > + return ret;
> > +}
> > +
> > +#define VF_VRAM_STATE_CHUNK_MAX_SIZE SZ_512M
> > +static int pf_save_vf_vram_mig_data(struct xe_gt *gt, unsigned int vfid)
> > +{
> > + struct xe_gt_sriov_migration_data *migration = pf_pick_gt_migration(gt,
> > vfid);
> > + loff_t *offset = &migration->vram_save_offset;
> > + struct xe_bo *vram;
> > + size_t vram_size, chunk_size;
> > + int ret;
> > +
> > + vram = xe_gt_sriov_pf_config_get_lmem_obj(gt, vfid);
> > + if (!vram)
> > + return -ENXIO;
>
> no error message ?
Well, this should be converted to assert after we block re-provisioning,
so I don't think it's needed (or know what to print out :) ).
>
> > +
> > + vram_size = xe_bo_size(vram);
> > + chunk_size = min(vram_size - *offset, VF_VRAM_STATE_CHUNK_MAX_SIZE);
>
> what if *offset > vram_size ?
We control the offset - I'll add an assert.
>
> > +
> > + ret = pf_save_vram_chunk(gt, vfid, vram, *offset, chunk_size);
> > + if (ret)
> > + goto fail;
> > +
> > + *offset += chunk_size;
> > +
> > + xe_bo_put(vram);
> > +
> > + if (*offset < vram_size)
> > + return -EAGAIN;
> > +
> > + return 0;
> > +
> > +fail:
> > + xe_bo_put(vram);
> > + xe_gt_sriov_err(gt, "Failed to save VF%u VRAM data (%pe)\n", vfid,
> > ERR_PTR(ret));
> > + return ret;
> > +}
> > +
> > +static int pf_restore_vf_vram_mig_data(struct xe_gt *gt, unsigned int vfid,
> > + struct xe_sriov_migration_data *data)
> > +{
> > + u64 end = data->hdr.offset + data->hdr.size;
> > + struct dma_fence *fence;
> > + struct xe_bo *vram;
> > + size_t size;
> > + int ret = 0;
> > +
> > + vram = xe_gt_sriov_pf_config_get_lmem_obj(gt, vfid);
> > + if (!vram)
> > + return -ENXIO;
>
> no error message ? other errors are reported
Same as above.
>
> > +
> > + size = xe_bo_size(vram);
> > +
> > + if (end > size || end < data->hdr.size) {
> > + ret = -EINVAL;
> > + goto err;
> > + }
> > +
> > + pf_dump_mig_data(gt, vfid, data);
> > +
> > + fence = __pf_save_restore_vram(gt, vfid, vram, data->hdr.offset,
> > + data->bo, 0, data->hdr.size, false);
> > + ret = dma_fence_wait_timeout(fence, false, 5 * HZ);
>
> define this timeout at least as macro (if not as helper function, as this
> might be platform/settings specific)
I'll use a macro for now.
>
> > + dma_fence_put(fence);
> > + if (!ret) {
> > + ret = -ETIME;
> > + goto err;
> > + }
> > +
> > + return 0;
> > +err:
> > + xe_bo_put(vram);
> > + xe_gt_sriov_err(gt, "Failed to restore VF%u VRAM data (%pe)\n", vfid,
> > ERR_PTR(ret));
> > + return ret;
> > +}
> > +
> > +/**
> > + * xe_gt_sriov_pf_migration_vram_save() - Save VF VRAM migration data.
> > + * @gt: the &xe_gt
> > + * @vfid: the VF identifier (can't be 0)
> > + *
> > + * This function is for PF only.
> > + *
> > + * Return: 0 on success or a negative error code on failure.
> > + */
> > +int xe_gt_sriov_pf_migration_vram_save(struct xe_gt *gt, unsigned int vfid)
> > +{
> > + xe_gt_assert(gt, IS_SRIOV_PF(gt_to_xe(gt)));
> > + xe_gt_assert(gt, vfid != PFID);
> > + xe_gt_assert(gt, vfid <= xe_sriov_pf_get_totalvfs(gt_to_xe(gt)));
> > +
> > + return pf_save_vf_vram_mig_data(gt, vfid);
> > +}
> > +
> > +/**
> > + * xe_gt_sriov_pf_migration_vram_restore() - Restore VF VRAM migration
> > data.
> > + * @gt: the &xe_gt
> > + * @vfid: the VF identifier (can't be 0)
> > + *
> > + * This function is for PF only.
> > + *
> > + * Return: 0 on success or a negative error code on failure.
> > + */
> > +int xe_gt_sriov_pf_migration_vram_restore(struct xe_gt *gt, unsigned int
> > vfid,
> > + struct xe_sriov_migration_data *data)
> > +{
> > + xe_gt_assert(gt, IS_SRIOV_PF(gt_to_xe(gt)));
> > + xe_gt_assert(gt, vfid != PFID);
> > + xe_gt_assert(gt, vfid <= xe_sriov_pf_get_totalvfs(gt_to_xe(gt)));
> > +
> > + return pf_restore_vf_vram_mig_data(gt, vfid, data);
> > +}
> > +
> > +/**
> > + * xe_gt_sriov_pf_migration_save_init() - Initialize per-GT migration
> > related data.
> > + * @gt: the &xe_gt
> > + * @vfid: the VF identifier (can't be 0)
> > + */
> > +void xe_gt_sriov_pf_migration_save_init(struct xe_gt *gt, unsigned int
> > vfid)
> > +{
> > + pf_pick_gt_migration(gt, vfid)->vram_save_offset = 0;
> > +}
> > +
> > /**
> > * xe_gt_sriov_pf_migration_size() - Total size of migration data from all
> > components within a GT.
> > * @gt: the &xe_gt
> > @@ -522,6 +737,13 @@ ssize_t xe_gt_sriov_pf_migration_size(struct xe_gt
> > *gt, unsigned int vfid)
> > size += sizeof(struct xe_sriov_pf_migration_hdr);
> > total += size;
> >
> > + size = xe_gt_sriov_pf_migration_vram_size(gt, vfid);
> > + if (size < 0)
> > + return size;
> > + else if (size > 0)
>
> "else" not needed
Ok.
>
> > + size += sizeof(struct xe_sriov_pf_migration_hdr);
> > + total += size;
> > +
> > return total;
> > }
> >
> > diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.h
> > b/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.h
> > index 24a233c4cd0bb..ca518eda5429f 100644
> > --- a/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.h
> > +++ b/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.h
> > @@ -27,6 +27,12 @@ ssize_t xe_gt_sriov_pf_migration_mmio_size(struct xe_gt
> > *gt, unsigned int vfid);
> > int xe_gt_sriov_pf_migration_mmio_save(struct xe_gt *gt, unsigned int
> > vfid);
> > int xe_gt_sriov_pf_migration_mmio_restore(struct xe_gt *gt, unsigned int
> > vfid,
> > struct xe_sriov_migration_data *data);
> > +ssize_t xe_gt_sriov_pf_migration_vram_size(struct xe_gt *gt, unsigned int
> > vfid);
> > +int xe_gt_sriov_pf_migration_vram_save(struct xe_gt *gt, unsigned int
> > vfid);
> > +int xe_gt_sriov_pf_migration_vram_restore(struct xe_gt *gt, unsigned int
> > vfid,
> > + struct xe_sriov_migration_data *data);
> > +
> > +void xe_gt_sriov_pf_migration_save_init(struct xe_gt *gt, unsigned int
> > vfid);
> >
> > ssize_t xe_gt_sriov_pf_migration_size(struct xe_gt *gt, unsigned int vfid);
> >
> > diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration_types.h
> > b/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration_types.h
> > index 75d8b94cbbefb..39a940c9b0a4b 100644
> > --- a/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration_types.h
> > +++ b/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration_types.h
> > @@ -16,6 +16,9 @@
> > struct xe_gt_sriov_migration_data {
> > /** @ring: queue containing VF save / restore migration data */
> > struct ptr_ring ring;
> > +
> > + /** @vram_save_offset: offset within VRAM, used for chunked VRAM save */
>
> "last saved offset" ?
Ok.
Thanks,
-Michał
> > + loff_t vram_save_offset;
> > };
> >
> > #endif
> > diff --git a/drivers/gpu/drm/xe/xe_sriov_pf_control.c
> > b/drivers/gpu/drm/xe/xe_sriov_pf_control.c
> > index c2768848daba1..aac8ecb861545 100644
> > --- a/drivers/gpu/drm/xe/xe_sriov_pf_control.c
> > +++ b/drivers/gpu/drm/xe/xe_sriov_pf_control.c
> > @@ -5,6 +5,7 @@
> >
> > #include "xe_device.h"
> > #include "xe_gt_sriov_pf_control.h"
> > +#include "xe_gt_sriov_pf_migration.h"
> > #include "xe_sriov_migration_data.h"
> > #include "xe_sriov_pf_control.h"
> > #include "xe_sriov_printk.h"
> > @@ -171,6 +172,8 @@ int xe_sriov_pf_control_trigger_save_vf(struct
> > xe_device *xe, unsigned int vfid)
> > return ret;
> >
> > for_each_gt(gt, xe, id) {
> > + xe_gt_sriov_pf_migration_save_init(gt, vfid);
> > +
> > ret = xe_gt_sriov_pf_control_trigger_save_vf(gt, vfid);
> > if (ret)
> > return ret;
>
