[PATCH 2/3] drm/xe: Update xe_migrate_vram to support compression

[Matthew Auld]

On 15/07/2025 09:37, Matthew Auld wrote:
> On 14/07/2025 18:33, Matthew Brost wrote:
>> While SVM does not currently support compression, other users of
>> xe_migrate_vram (e.g., devcoredump) expect the data to be read back
>> uncompressed. Update xe_migrate_vram to support compressed data.
>>
>> Cc: stable at vger.kernel.org
>> Fixes: 9c44fd5f6e8a ("drm/xe: Add migrate layer functions for SVM 
>> support")
>> Signed-off-by: Matthew Brost <matthew.brost at intel.com>
>> ---
>>   drivers/gpu/drm/xe/xe_migrate.c | 31 ++++++++++++++++++++++++-------
>>   1 file changed, 24 insertions(+), 7 deletions(-)
>>
>> diff --git a/drivers/gpu/drm/xe/xe_migrate.c b/drivers/gpu/drm/xe/ 
>> xe_migrate.c
>> index ba1cff2e4cda..936daa2b363d 100644
>> --- a/drivers/gpu/drm/xe/xe_migrate.c
>> +++ b/drivers/gpu/drm/xe/xe_migrate.c
>> @@ -1613,7 +1613,8 @@ static struct dma_fence *xe_migrate_vram(struct 
>> xe_migrate *m,
>>                        unsigned long len,
>>                        unsigned long sram_offset,
>>                        dma_addr_t *sram_addr, u64 vram_addr,
>> -                     const enum xe_migrate_copy_dir dir)
>> +                     const enum xe_migrate_copy_dir dir,
>> +                     bool needs_ccs_emit)
>>   {
>>       struct xe_gt *gt = m->tile->primary_gt;
>>       struct xe_device *xe = gt_to_xe(gt);
>> @@ -1623,10 +1624,12 @@ static struct dma_fence 
>> *xe_migrate_vram(struct xe_migrate *m,
>>       u64 src_L0_ofs, dst_L0_ofs;
>>       struct xe_sched_job *job;
>>       struct xe_bb *bb;
>> -    u32 update_idx, pt_slot = 0;
>> +    u32 update_idx, pt_slot = 0, flush_flags = 0;
>>       unsigned long npages = DIV_ROUND_UP(len + sram_offset, PAGE_SIZE);
>>       unsigned int pitch = len >= PAGE_SIZE && !(len & ~PAGE_MASK) ?
>>           PAGE_SIZE : 4;
>> +    bool use_comp_pat = xe_device_has_flat_ccs(xe) &&
>> +        GRAPHICS_VER(xe) >= 20 && dir == XE_MIGRATE_COPY_TO_SRAM;
>>       int err;
>>       if (drm_WARN_ON(&xe->drm, (len & XE_CACHELINE_MASK) ||
>> @@ -1637,6 +1640,8 @@ static struct dma_fence *xe_migrate_vram(struct 
>> xe_migrate *m,
>>       batch_size += pte_update_cmd_size(len);
>>       batch_size += EMIT_COPY_DW;
>> +    if (needs_ccs_emit)
>> +        batch_size += EMIT_COPY_CCS_DW;
>>       bb = xe_bb_new(gt, batch_size, use_usm_batch);
>>       if (IS_ERR(bb)) {
>> @@ -1652,7 +1657,7 @@ static struct dma_fence *xe_migrate_vram(struct 
>> xe_migrate *m,
>>           dst_L0_ofs = xe_migrate_vram_ofs(xe, vram_addr, false);
>>       } else {
>> -        src_L0_ofs = xe_migrate_vram_ofs(xe, vram_addr, false);
>> +        src_L0_ofs = xe_migrate_vram_ofs(xe, vram_addr, use_comp_pat);
>>           dst_L0_ofs = xe_migrate_vm_addr(pt_slot, 0) + sram_offset;
>>       }
>> @@ -1661,6 +1666,17 @@ static struct dma_fence *xe_migrate_vram(struct 
>> xe_migrate *m,
>>       emit_copy(gt, bb, src_L0_ofs, dst_L0_ofs, len, pitch);
>> +    if (needs_ccs_emit) {
>> +        if (dir == XE_MIGRATE_COPY_TO_VRAM)
>> +            flush_flags = xe_migrate_ccs_copy(m, bb, src_L0_ofs,
>> +                              false, dst_L0_ofs,
>> +                              true, len, 0, true);
>> +        else
>> +            flush_flags = xe_migrate_ccs_copy(m, bb, src_L0_ofs,
>> +                              true, dst_L0_ofs,
>> +                              false, len, 0, true);
>> +    }
> 
> I think we can drop this and anything related to needs_ccs_emit.  In 
> theory we should only need the use_comp_pat change. IIUC this path is 
> VRAM only, and only xe2+ can be decompressed in the KMD (dg2 is no-go), 
> and for that we don't manage the raw CCS state.

So xe_migrate_ccs_copy() is not actually decompressing anything, but 
rather copying the raw CCS state around, which I think we use for manual 
save/restore during swap and things like that to preserve the 
compression state.

I think we do also use xe_migrate_ccs_copy() to intitially clear the CCS 
state, but I think that has already happened somewhere else? Or do we 
need it here, if we say go from tt -> vram on platforms like dg2?

> 
>> +
>>       job = xe_bb_create_migration_job(m->q, bb,
>>                        xe_migrate_batch_base(m, use_usm_batch),
>>                        update_idx);
>> @@ -1669,7 +1685,7 @@ static struct dma_fence *xe_migrate_vram(struct 
>> xe_migrate *m,
>>           goto err;
>>       }
>> -    xe_sched_job_add_migrate_flush(job, 0);
>> +    xe_sched_job_add_migrate_flush(job, flush_flags);
>>       mutex_lock(&m->job_mutex);
>>       xe_sched_job_arm(job);
>> @@ -1708,7 +1724,7 @@ struct dma_fence *xe_migrate_to_vram(struct 
>> xe_migrate *m,
>>                        u64 dst_addr)
>>   {
>>       return xe_migrate_vram(m, npages * PAGE_SIZE, 0, src_addr, 
>> dst_addr,
>> -                   XE_MIGRATE_COPY_TO_VRAM);
>> +                   XE_MIGRATE_COPY_TO_VRAM, false);
>>   }
>>   /**
>> @@ -1729,7 +1745,7 @@ struct dma_fence *xe_migrate_from_vram(struct 
>> xe_migrate *m,
>>                          dma_addr_t *dst_addr)
>>   {
>>       return xe_migrate_vram(m, npages * PAGE_SIZE, 0, dst_addr, 
>> src_addr,
>> -                   XE_MIGRATE_COPY_TO_SRAM);
>> +                   XE_MIGRATE_COPY_TO_SRAM, false);
>>   }
>>   static void xe_migrate_dma_unmap(struct xe_device *xe, dma_addr_t 
>> *dma_addr,
>> @@ -1890,7 +1906,8 @@ int xe_migrate_access_memory(struct xe_migrate 
>> *m, struct xe_bo *bo,
>>                         dma_addr + current_page,
>>                         vram_addr, write ?
>>                         XE_MIGRATE_COPY_TO_VRAM :
>> -                      XE_MIGRATE_COPY_TO_SRAM);
>> +                      XE_MIGRATE_COPY_TO_SRAM,
>> +                      xe_bo_needs_ccs_pages(bo));
>>           if (IS_ERR(__fence)) {
>>               if (fence)
>>                   dma_fence_wait(fence, false);
>