Re: [Qemu-block] [PATCH v3 5/6] block/nvme: add support for write zeros

[Maxim Levitsky]

On Fri, 2019-07-05 at 15:33 +0200, Max Reitz wrote:
> On 03.07.19 17:59, Maxim Levitsky wrote:
> > Signed-off-by: Maxim Levitsky <address@hidden>
> > ---
> >  block/nvme.c         | 69 +++++++++++++++++++++++++++++++++++++++++++-
> >  block/trace-events   |  1 +
> >  include/block/nvme.h | 19 +++++++++++-
> >  3 files changed, 87 insertions(+), 2 deletions(-)
> > 
> > diff --git a/block/nvme.c b/block/nvme.c
> > index 152d27b07f..02e0846643 100644
> > --- a/block/nvme.c
> > +++ b/block/nvme.c
> 
> [...]
> 
> > @@ -469,6 +473,11 @@ static void nvme_identify(BlockDriverState *bs, int 
> > namespace, Error **errp)
> >      s->nsze = le64_to_cpu(idns->nsze);
> >      lbaf = &idns->lbaf[NVME_ID_NS_FLBAS_INDEX(idns->flbas)];
> >  
> > +    if (NVME_ID_NS_DLFEAT_WRITE_ZEROS(idns->dlfeat) &&
> > +            NVME_ID_NS_DLFEAT_READ_BEHAVIOR(idns->dlfeat) ==
> > +                    NVME_ID_NS_DLFEAT_READ_BEHAVIOR_ZEROS)
> > +        bs->supported_write_flags |= BDRV_REQ_MAY_UNMAP;
> > +
> 
> This violates the coding style, there should be curly brackets here.
100% agree + I need to see if we can update the checkpatch.pl to catch this.


> 
> >      if (lbaf->ms) {
> >          error_setg(errp, "Namespaces with metadata are not yet supported");
> >          goto out;
> > @@ -763,6 +772,8 @@ static int nvme_file_open(BlockDriverState *bs, QDict 
> > *options, int flags,
> >      int ret;
> >      BDRVNVMeState *s = bs->opaque;
> >  
> > +    bs->supported_write_flags = BDRV_REQ_FUA;
> > +
> >      opts = qemu_opts_create(&runtime_opts, NULL, 0, &error_abort);
> >      qemu_opts_absorb_qdict(opts, options, &error_abort);
> >      device = qemu_opt_get(opts, NVME_BLOCK_OPT_DEVICE);
> > @@ -791,7 +802,6 @@ static int nvme_file_open(BlockDriverState *bs, QDict 
> > *options, int flags,
> >              goto fail;
> >          }
> >      }
> > -    bs->supported_write_flags = BDRV_REQ_FUA;
> 
> Any reason for this movement?

This is because the nvme_identify checks if the underlying namespace
supports 'discarded data reads back as zeros', and in which case it sets the
BDRV_REQ_MAY_UNMAP in bs->supported_write_flags which later allow me to set
'deallocate' bit in the write zeros command which hints the controller
to discard the area.

This was moved to avoid overwriting the value. I could have instead just ored 
the value,
but this way I think is cleaner a bit.



> 
> >      return 0;
> >  fail:
> >      nvme_close(bs);
> > @@ -1085,6 +1095,60 @@ static coroutine_fn int 
> > nvme_co_flush(BlockDriverState *bs)
> >  }
> >  
> >  
> > +static coroutine_fn int nvme_co_pwrite_zeroes(BlockDriverState *bs,
> > +                                              int64_t offset,
> > +                                              int bytes,
> > +                                              BdrvRequestFlags flags)
> > +{
> > +    BDRVNVMeState *s = bs->opaque;
> > +    NVMeQueuePair *ioq = s->queues[1];
> > +    NVMeRequest *req;
> > +
> > +    if (!s->supports_write_zeros) {
> > +        return -ENOTSUP;
> > +    }
> > +
> > +    uint32_t cdw12 = ((bytes >> s->blkshift) - 1) & 0xFFFF;
> 
> Another coding style violation: Variable declarations and other code may
> not be mixed.
Another bug in checkpatch.pl :-)

> 
> > +
> > +    NvmeCmd cmd = {
> > +        .opcode = NVME_CMD_WRITE_ZEROS,
> > +        .nsid = cpu_to_le32(s->nsid),
> > +        .cdw10 = cpu_to_le32((offset >> s->blkshift) & 0xFFFFFFFF),
> > +        .cdw11 = cpu_to_le32(((offset >> s->blkshift) >> 32) & 0xFFFFFFFF),
> > +    };
> > +
> > +    NVMeCoData data = {
> > +        .ctx = bdrv_get_aio_context(bs),
> > +        .ret = -EINPROGRESS,
> > +    };
> 
> [...]
> 
> > diff --git a/include/block/nvme.h b/include/block/nvme.h
> > index 3ec8efcc43..65eb65c740 100644
> > --- a/include/block/nvme.h
> > +++ b/include/block/nvme.h
> > @@ -653,12 +653,29 @@ typedef struct NvmeIdNs {
> >      uint8_t     mc;
> >      uint8_t     dpc;
> >      uint8_t     dps;
> > -    uint8_t     res30[98];
> > +
> > +    uint8_t     nmic;
> > +    uint8_t     rescap;
> > +    uint8_t     fpi;
> > +    uint8_t     dlfeat;
> > +
> > +    uint8_t     res30[94];
> >      NvmeLBAF    lbaf[16];
> >      uint8_t     res192[192];
> >      uint8_t     vs[3712];
> >  } NvmeIdNs;
> >  
> > +
> > +/*Deallocate Logical Block Features*/
> > +#define NVME_ID_NS_DLFEAT_GUARD_CRC(dlfeat)       ((dlfeat) & 0x10)
> > +#define NVME_ID_NS_DLFEAT_WRITE_ZEROS(dlfeat)     ((dlfeat) & 0x04)
> 
> Isn’t it bit 3, i.e. 0x08?
Oops, I haven't noticed that 'read behavier' field is 3 bits and not 2!
Thank you very much. I haven't caught this since my device I tested on doesn't 
support this anyway (dlfeat == 0)

> 
> Max
> 
> > +
> > +#define NVME_ID_NS_DLFEAT_READ_BEHAVIOR(dlfeat)     ((dlfeat) & 0x3)
> > +#define NVME_ID_NS_DLFEAT_READ_BEHAVIOR_UNDEFINED   0
> > +#define NVME_ID_NS_DLFEAT_READ_BEHAVIOR_ZEROS       1
> > +#define NVME_ID_NS_DLFEAT_READ_BEHAVIOR_ONES        2
> > +
> > +
> >  #define NVME_ID_NS_NSFEAT_THIN(nsfeat)      ((nsfeat & 0x1))
> >  #define NVME_ID_NS_FLBAS_EXTENDED(flbas)    ((flbas >> 4) & 0x1)
> >  #define NVME_ID_NS_FLBAS_INDEX(flbas)       ((flbas & 0xf))
> > 
> 


Thank you very very much for the review,
        Best regards,
                Maxim Levitsky