[Intel-xe] [RFC PATCH] drm/xe: Add uapi for dumpable bos

[Maarten Lankhorst]

Hey,

On 2023-10-06 20:29, Souza, Jose wrote:
> On Fri, 2023-10-06 at 19:18 +0200, maarten.lankhorst at linux.intel.com wrote:
>> From: Maarten Lankhorst <dev at lankhorst.se>
>>
>> Add XE_GEM_CREATE_FLAG_DUMPABLE for use with VM_BIND.
>> For BO's with vram regions, XE_GEM_CREATE_FLAG_NEEDS_VISIBLE_VRAM
>> also needs to be specified.
>>
>> After this, we can use the flag XE_VM_BIND_FLAG_DUMPABLE to notify
>> devcoredump that this mapping should be dumped.
>>
>> This is not hooked up, but the uapi should be ready before merging.
>> If needed, we can relax the restrictions later.
>>
>> It's also not yet certain whether the BO contents will be preserved,
>> or a reference to the BO. The latter allows us to skip a whole lot of
>> tricky locking. This requires userspace to destroy all bo's that were
>> marked dumpable though, but I feel this is a cleaner solution from
>> a kernel perspective.
>>
>> Signed-off-by: Maarten Lankhorst <maarten.lankhorst at linux.intel.com>
>> ---
>>   drivers/gpu/drm/xe/xe_bo.c |  9 +++++++++
>>   drivers/gpu/drm/xe/xe_bo.h |  1 +
>>   drivers/gpu/drm/xe/xe_vm.c | 17 ++++++++++++++---
>>   include/uapi/drm/xe_drm.h  | 25 +++++++++++++++++++++++++
>>   4 files changed, 49 insertions(+), 3 deletions(-)
>>
>> diff --git a/drivers/gpu/drm/xe/xe_bo.c b/drivers/gpu/drm/xe/xe_bo.c
>> index 61789c0e88fb..0de0bd5eb825 100644
>> --- a/drivers/gpu/drm/xe/xe_bo.c
>> +++ b/drivers/gpu/drm/xe/xe_bo.c
>> @@ -1776,6 +1776,7 @@ int xe_gem_create_ioctl(struct drm_device *dev, void *data,
>>   			 ~(XE_GEM_CREATE_FLAG_DEFER_BACKING |
>>   			   XE_GEM_CREATE_FLAG_SCANOUT |
>>   			   XE_GEM_CREATE_FLAG_NEEDS_VISIBLE_VRAM |
>> +			   XE_GEM_CREATE_FLAG_DUMPABLE |
>>   			   xe->info.mem_region_mask)))
>>   		return -EINVAL;
>>   
>> @@ -1810,6 +1811,14 @@ int xe_gem_create_ioctl(struct drm_device *dev, void *data,
>>   		bo_flags |= XE_BO_NEEDS_CPU_ACCESS;
>>   	}
>>   
>> +	if (args->flags & XE_GEM_CREATE_FLAG_DUMPABLE) {
>> +		if (XE_IOCTL_DBG(xe, (bo_flags & XE_BO_CREATE_VRAM_MASK &&
>> +				 !(bo_flags & XE_GEM_CREATE_FLAG_NEEDS_VISIBLE_VRAM))))
>> +			return -EINVAL;
>> +
>> +		bo_flags |= XE_BO_NEEDS_CPU_ACCESS | XE_BO_DUMPABLE;
> 
> It is not viable to just ignore VMAs that can't be mapped in CPU during dump?
> Because it will not be viable to dump real applications in small PCIe bar systems, it would run out of CPU mappable memory.
> 
> And if UMD really wants a dump of a small test that fits into the CPU mappable memory it can set XE_GEM_CREATE_FLAG_NEEDS_VISIBLE_VRAM and get the
> whole dump.
> 
> So we would be able to drop the need of XE_GEM_CREATE_FLAG_DUMPABLE.
It would be doable, we could evict the BO's to sysmem if we only take a 
reference and dump each bo.

Taking a reference during crash would require the userspace app to 
recreate all dumpable bo's after a hang, or it will risk overwriting
the state at the time of the dump.

~Maarten