The device bit that indicates that the device is capable of I/O
with private pages lets avoid excessive copying in the Hyper-V
SCSI driver.

Set that bit equal to the confidential external memory one to
not bounce buffer

Signed-off-by: Roman Kisel <romank@linux.microsoft.com>
---
 drivers/scsi/storvsc_drv.c | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/drivers/scsi/storvsc_drv.c b/drivers/scsi/storvsc_drv.c
index a8614e54544e..f647f8fc2f8f 100644
--- a/drivers/scsi/storvsc_drv.c
+++ b/drivers/scsi/storvsc_drv.c
@@ -686,6 +686,8 @@ static void handle_sc_creation(struct vmbus_channel *new_sc)
 	struct vmstorage_channel_properties props;
 	int ret;
 
+	dev->use_priv_pages_for_io = new_sc->confidential_external_memory;
+
 	stor_device = get_out_stor_device(device);
 	if (!stor_device)
 		return;
-- 
2.43.0

On Tue, Apr 08, 2025 at 05:08:35PM -0700, Roman Kisel wrote:
> The device bit that indicates that the device is capable of I/O
> with private pages lets avoid excessive copying in the Hyper-V
> SCSI driver.
> 
> Set that bit equal to the confidential external memory one to
> not bounce buffer

Drivers have absolutely no business telling this.  The need for bounce
buffering or not is a platform/IOMMU decision and not one specific to
a certain device or driver.

On 4/9/2025 3:53 AM, Christoph Hellwig wrote:
> On Tue, Apr 08, 2025 at 05:08:35PM -0700, Roman Kisel wrote:
>> The device bit that indicates that the device is capable of I/O
>> with private pages lets avoid excessive copying in the Hyper-V
>> SCSI driver.
>>
>> Set that bit equal to the confidential external memory one to
>> not bounce buffer
> 
> Drivers have absolutely no business telling this.  The need for bounce
> buffering or not is a platform/IOMMU decision and not one specific to
> a certain device or driver.

Seemed to work although I cannot claim nothing is going to be broken
ever. It did appear from the code that one could have this per-device
bit.

As I understand, you're saying this is architecturally broken. Do you
think a broader set of changes would improve the implementation?

> 

-- 
Thank you,
Roman