Re: [PATCH V3 13/13] HV/Storvsc: Add Isolation VM support for storvsc driver

[Tianyu Lan <ltykernel@gmail.com>]

On 8/20/2021 12:32 PM, hch@lst.de wrote:
> On Thu, Aug 19, 2021 at 06:17:40PM +0000, Michael Kelley wrote:
>>> +#define storvsc_dma_map(dev, page, offset, size, dir) \
>>> +	dma_map_page(dev, page, offset, size, dir)
>>> +
>>> +#define storvsc_dma_unmap(dev, dma_range, dir)		\
>>> +		dma_unmap_page(dev, dma_range.dma,	\
>>> +			       dma_range.mapping_size,	\
>>> +			       dir ? DMA_FROM_DEVICE : DMA_TO_DEVICE)
>>> +
>>
>> Each of these macros is used only once.  IMHO, they don't
>> add a lot of value.  Just coding dma_map/unmap_page()
>> inline would be fine and eliminate these lines of code.
> 
> Yes, I had the same thought when looking over the code.  Especially
> as macros tend to further obsfucate the code (compared to actual helper
> functions).
> 
>>> +				for (i = 0; i < request->hvpg_count; i++)
>>> +					storvsc_dma_unmap(&device->device,
>>> +						request->dma_range[i],
>>> +						request->vstor_packet.vm_srb.data_in == READ_TYPE);
>>
>> I think you can directly get the DMA direction as request->cmd->sc_data_direction.
> 
> Yes.
> 
>>>
>>> @@ -1824,6 +1848,13 @@ static int storvsc_queuecommand(struct Scsi_Host *host, struct scsi_cmnd *scmnd)
>>>   		payload->range.len = length;
>>>   		payload->range.offset = offset_in_hvpg;
>>>
>>> +		cmd_request->dma_range = kcalloc(hvpg_count,
>>> +				 sizeof(*cmd_request->dma_range),
>>> +				 GFP_ATOMIC);
>>
>> With this patch, it appears that storvsc_queuecommand() is always
>> doing bounce buffering, even when running in a non-isolated VM.
>> The dma_range is always allocated, and the inner loop below does
>> the dma mapping for every I/O page.  The corresponding code in
>> storvsc_on_channel_callback() that does the dma unmap allows for
>> the dma_range to be NULL, but that never happens.
> 
> Maybe I'm missing something in the hyperv code, but I don't think
> dma_map_page would bounce buffer for the non-isolated case.  It
> will just return the physical address.

Yes, the swiotlb_force mode isn't enabled in non-isolated VM and so
dma_page_page() returns the physical address directly.

> 
>>> +		if (!cmd_request->dma_range) {
>>> +			ret = -ENOMEM;
>>
>> The other memory allocation failure in this function returns
>> SCSI_MLQUEUE_DEVICE_BUSY.   It may be debatable as to whether
>> that's the best approach, but that's a topic for a different patch.  I
>> would suggest being consistent and using the same return code
>> here.
> 
> Independent of if SCSI_MLQUEUE_DEVICE_BUSY is good (it it a common
> pattern in SCSI drivers), ->queuecommand can't return normal
> negative errnos.  It must return the SCSI_MLQUEUE_* codes or 0.
> We should probably change the return type of the method definition
> to a suitable enum to make this more clear.

Yes, will update. Thanks.

> 
>>> +				if (offset_in_hvpg) {
>>> +					payload->range.offset = dma & ~HV_HYP_PAGE_MASK;
>>> +					offset_in_hvpg = 0;
>>> +				}
>>
>> I'm not clear on why payload->range.offset needs to be set again.
>> Even after the dma mapping is done, doesn't the offset in the first
>> page have to be the same?  If it wasn't the same, Hyper-V wouldn't
>> be able to process the PFN list correctly.  In fact, couldn't the above
>> code just always set offset_in_hvpg = 0?
> 
> Careful.  DMA mapping is supposed to keep the offset in the page, but
> for that the DMA mapping code needs to know what the device considers a
> "page".  For that the driver needs to set the min_align_mask field in
> struct device_dma_parameters.

The default allocate unit of swiotlb bounce is IO_TLB_SIZE(2k).
Otherwise, I find some scsi request cmd's length is less than 100byte.
Keep a small unit can avoid wasting bounce buffer and just need to
update the offset.


> 
>>
>> The whole approach here is to do dma remapping on each individual page
>> of the I/O buffer.  But wouldn't it be possible to use dma_map_sg() to map
>> each scatterlist entry as a unit?  Each scatterlist entry describes a range of
>> physically contiguous memory.  After dma_map_sg(), the resulting dma
>> address must also refer to a physically contiguous range in the swiotlb
>> bounce buffer memory.   So at the top of the "for" loop over the scatterlist
>> entries, do dma_map_sg() if we're in an isolated VM.  Then compute the
>> hvpfn value based on the dma address instead of sg_page().  But everything
>> else is the same, and the inner loop for populating the pfn_arry is unmodified.
>> Furthermore, the dma_range array that you've added is not needed, since
>> scatterlist entries already have a dma_address field for saving the mapped
>> address, and dma_unmap_sg() uses that field.
> 
> Yes, I think dma_map_sg is the right thing to use here, probably even
> for the non-isolated case so that we can get the hv drivers out of their
> little corner and into being more like a normal kernel driver.  That
> is, use the scsi_dma_map/scsi_dma_unmap helpers, and then iterate over
> the dma addresses one page at a time using for_each_sg_dma_page.
> 

I wonder whether we may introduce a new API scsi_dma_map_with_callback. 
Caller provides a callback and run callback in sg loop of 
dma_direct_map_sg(). Caller need to update some data structure in the sg 
loop. Here is such case that driver needs to populate 
payload->range.pfn_array[]. This is why I don't use dma_map_sg() here.


>>
>> One thing:  There's a maximum swiotlb mapping size, which I think works
>> out to be 256 Kbytes.  See swiotlb_max_mapping_size().  We need to make
>> sure that we don't get a scatterlist entry bigger than this size.  But I think
>> this already happens because you set the device->dma_mask field in
>> Patch 11 of this series.  __scsi_init_queue checks for this setting and
>> sets max_sectors to limits transfers to the max mapping size.
> 
> Indeed.
>