On 2020/12/07 17:16, javier.gonz@xxxxxxxxxxx wrote:
> On 07.12.2020 08:06, Damien Le Moal wrote:
>> On 2020/12/07 16:46, javier.gonz@xxxxxxxxxxx wrote:
>>> On 04.12.2020 23:40, Keith Busch wrote:
>>>> On Fri, Dec 04, 2020 at 11:25:12AM +0000, Damien Le Moal wrote:
>>>>> On 2020/12/04 20:02, SelvaKumar S wrote:
>>>>>> This patchset tries to add support for TP4065a ("Simple Copy Command"),
>>>>>> v2020.05.04 ("Ratified")
>>>>>>
>>>>>> The Specification can be found in following link.
>>>>>> https://nvmexpress.org/wp-content/uploads/NVM-Express-1.4-Ratified-TPs-1.zip
>>>>>>
>>>>>> This is an RFC. Looking forward for any feedbacks or other alternate
>>>>>> designs for plumbing simple copy to IO stack.
>>>>>>
>>>>>> Simple copy command is a copy offloading operation and is used to copy
>>>>>> multiple contiguous ranges (source_ranges) of LBA's to a single destination
>>>>>> LBA within the device reducing traffic between host and device.
>>>>>>
>>>>>> This implementation accepts destination, no of sources and arrays of
>>>>>> source ranges from application and attach it as payload to the bio and
>>>>>> submits to the device.
>>>>>>
>>>>>> Following limits are added to queue limits and are exposed in sysfs
>>>>>> to userspace
>>>>>> - *max_copy_sectors* limits the sum of all source_range length
>>>>>> - *max_copy_nr_ranges* limits the number of source ranges
>>>>>> - *max_copy_range_sectors* limit the maximum number of sectors
>>>>>> that can constitute a single source range.
>>>>>
>>>>> Same comment as before. I think this is a good start, but for this to be really
>>>>> useful to users and kernel components alike, this really needs copy emulation
>>>>> for drives that do not have a native copy feature, similarly to what write zeros
>>>>> handling for instance: if the drive does not have a copy command (simple copy
>>>>> for NVMe or XCOPY for scsi), then the block layer should issue read/write
>>>>> commands to seamlessly execute the copy. Otherwise, this will only serve a small
>>>>> niche for users and will not be optimal for FS and DM drivers that could be
>>>>> simplified with a generic block layer copy functionality.
>>>>>
>>>>> This is my 10 cents though, others may differ about this.
>>>>
>>>> Yes, I agree that copy emulation support should be included with the
>>>> hardware enabled solution.
>>>
>>> Keith, Damien,
>>>
>>> Can we do the block layer emulation with this patchset and then work in
>>> follow-up patchses on (i) the FS interface with F2FS as a first user and
>>> (ii) other HW accelerations such as XCOPY?
>>
>> The initial patchset supporting NVMe simple copy and emulation copy, all under
>> an API that probably will be similar that of dm-kcopyd will cover all block
>> devices. Other hardware native support for copy functions such as scsi extended
>> copy can be added later under the hood without any API changes (or minimal changes).
>
> Sounds good. That we can do. We will add a new patch for this.
>
>>
>> I am not sure what you mean by "FS interface for F2FS": the block layer API for
>> this copy functionality will be what F2FS (and other FSes) will call. That is
>> the interface, no ?
>
> Essentially yes.. I mean adding the F2FS logic and potentially some
> helpers to the block layer to aid GC.
GC is very much special to each FS. SO I do not think adding helpers to the
block layer will have value. We should stick to a pure block copy API for that
layer.
>
>>
>>> For XCOPY, I believe we need to have a separate discussion as much works
>>> is already done that we should align to.
>>
>> I think Martin (added to this thread) and others have looked into it but I do
>> not think that anything made it into the kernel yet.
>
> Exactly. Looking at some of the code posted through time and recalling
> the discussions at LSF/MM, seems like there are a number of things we
> are not addressing here that could be incorporated down the road, such
> as dedicated syscalls / extensions, multi namespace / device support,
> etc.
dm-kcopyd interface supports copy between multiple devices. That of course would
not enable NVMe simple copy use, but that makes the interface generic enough so
that we should not have any problem with other hardware copy functions.
>>
>
--
Damien Le Moal
Western Digital Research
