Re: [Qemu-devel] [RFC] introduce a dynamic library to expose qemu block API

[Wenchao Xia <xiawenc@linux.vnet.ibm.com>]

于 2012-7-24 2:15, Blue Swirl 写道:
> On Wed, Jul 18, 2012 at 8:51 AM, Wenchao Xia <xiawenc@linux.vnet.ibm.com> wrote:
>>    Hi, following is API draft, prototypes were taken from qemu/block.h,
>> and the API prefix is changed frpm bdrv to qbdrvs, to declare related
>> object is BlockDriverState, not BlockDriver. One issue here is it may
>> require include block_int.h, which is not LGPL2 licensed yet.
>>    API format is kept mostly the same with qemu generic block layer, to
>> make it easier for implement, and easier to make qemu migrate on it if
>> possible.
>>
>>
>> /* structure init and uninit */
>> BlockDriverState *qbdrvs_new(const char *device_name);
>> void qbdrvs_delete(BlockDriverState *bs);
>>
>>
>> /* file open and close */
>> int qbdrvs_open(BlockDriverState *bs, const char *filename, int flags,
>>                BlockDriver *drv);
>
> How are the errors passed?
>
    Returning negative value should indicate errors. I plan to insert
a qblock-err.h to define all number there, and then a function of
const char *qberr2str(int err) to translate error.

> Alternative version with file descriptor or struct FILE instead of
> filename might become useful but can be added later.
>
   Agree.
>> void qbdrvs_close(BlockDriverState *bs);
>> int qbdrvs_img_create(const char *filename, const char *fmt,
>>                      const char *base_filename, const char *base_fmt,
>>                      char *options, uint64_t img_size, int flags);
>
> 'const char *options'
>
>>
>>
>> /* sync access */
>> int qbdrvs_read(BlockDriverState *bs, int64_t sector_num,
>>                uint8_t *buf, int nb_sectors);
>> int qbdrvs_write(BlockDriverState *bs, int64_t sector_num,
>>                 const uint8_t *buf, int nb_sectors);
>
> Do we want to use sectors here? How about just raw byte offsets and
> number of bytes? I'd leave any hardware details out and just provide
> file semantics (open/read/write/close). Future QEMU refactorings could
> make supporting HW info inconvenient. If HW details (geometry etc., VM
> state) are needed, there should be a separate API.
>
   not sure why original API used sector instead of  raw address,
if no special requirement for sector I think raw address is better.

> Vectored I/O might be useful too.
>
>>
>>
>> /* info retrieve */
>> //sector, size and geometry info
>> int qbdrvs_get_info(BlockDriverState *bs, BlockDriverInfo *bdi);
>
> Currently BlockDriverInfo does not look very useful for outside users,
> with the exception of dirty state. How about accessors instead:
> bool qbdrvs_is_dirty(BlockDriverState *bs);
>
> Also the format (QCOW) is needed so that the user can use backing file
> functions:
> const char *qbdrvs_get_format(BlockDriverState *bs);
>
   I hope to package the info into a structure, such as
#define QBLOCK_INFO_SIZE (128)
struct QblockInfo {
     int dirty;
     char *format;
     int reserved[RESERVE_SPACE];
};
assert sizeof(QblockInfo) == QBLOCK_INFO_SIZE in every version. This
will make caller's code more clean.

>> int64_t qbdrvs_getlength(BlockDriverState *bs);
>> int64_t qbdrvs_get_allocated_file_size(BlockDriverState *bs);
>> void qbdrvs_get_geometry(BlockDriverState *bs, uint64_t *nb_sectors_ptr);
>
> Also this function shouldn't be needed if we don't give HW details
> with this API.
>
     OK.

>> //image type
>> const char *qbdrvs_get_format_name(BlockDriverState *bs);
>> //backing file info
>> void qbdrvs_get_backing_filename(BlockDriverState *bs,
>>                                 char *filename, int filename_size);
>> void qbdrvs_get_full_backing_filename(BlockDriverState *bs,
>>                                      char *dest, size_t sz);
>
> These are specific to QCOW etc., so
> bool qbdrvs_has_backing_files(BlockDriverState *bs)?
>
   How about using return value to indicate it?

int qbdrvs_get_backing_filename(BlockDriverState *bs, char *filename, 
int filename_size);

>>
>>
>> /* advanced image content access */
>> int qbdrvs_is_allocated(BlockDriverState *bs, int64_t sector_num, int
>> nb_sectors,
>>                        int *pnum);
>> int qbdrvs_discard(BlockDriverState *bs, int64_t sector_num, int
>> nb_sectors);
>
> Again, some files do not have a concept of allocation.
>
   Maybe return value could indicate it.

>> int qbdrvs_has_zero_init(BlockDriverState *bs);
>>
>>
>>> Il 16/07/2012 10:16, Wenchao Xia ha scritto:
>>>>>
>>>>>
>>>>     Really thanks for the investigation, I paid quite sometime to dig out
>>>> which license is compatible to LGPL, this have sorted it out.
>>>>     The coroutine and structure inside is quite a challenge.
>>>
>>>
>>> Coroutines are really just a small complication in the program flow if
>>> all you support is synchronous access to files (i.e. no HTTP etc.).
>>> Their usage should be completely transparent.
>>>
>>>> What about
>>>> provide the library first in nbd + sync access, and waiting for the
>>>> library employer response? If it is good to use, then replace implement
>>>> code to native qemu block layer code, change code's license, while keep
>>>> API unchanged.
>>>
>>>
>>> You can start by proposing the API.
>>>
>>> Paolo
>>>
>>
>>
>> --
>> Best Regards
>>
>> Wenchao Xia
>>
>>
>>
>


-- 
Best Regards

Wenchao Xia