Re: [Qemu-devel] [PATCH 1/2] block/nvme: add support for write zeros

[Maxim Levitsky]

On Tue, 2019-08-27 at 18:22 -0400, John Snow wrote:
> Without a commit message, I have no real hope of reviewing this. I was
> CC'd, though, so I'll give it a blind shot.
> 
> We want to add write_zeroes support for block/nvme, but I can't really
> verify any of that is correct or working without a unit test, a spec, or
> some instructions to help me verify any of this does what it looks like
> it does.

Its a bit problematic to have unit tests for this specific driver,
because it is the first and the only driver which is a driver in the
real sense of the word, that is it works directly with host hardware,
replacing the in-kernel driver.

I guess some unit tests could be done using a nested VM or so.


Yea, I was initially confused as well as what write zeros does, but it
does _exactly_ what it says, just writes zeros to the block you ask
for. So basically just like a write command, but without data.

Plus optionally if the drive  declares that after discard,
the blocks read as zeros, you can enable 'DEALOC' bit in write zeros

command to specify that in addition to filling the block with zeros
it would be nice to discard its flash data as well (this is still a hint
though).


> 
> On 8/25/19 3:15 AM, Maxim Levitsky wrote:
> > Signed-off-by: Maxim Levitsky <address@hidden>
> > ---
> >  block/nvme.c         | 72 +++++++++++++++++++++++++++++++++++++++++++-
> >  block/trace-events   |  1 +
> >  include/block/nvme.h | 19 +++++++++++-
> >  3 files changed, 90 insertions(+), 2 deletions(-)
> > 
> > diff --git a/block/nvme.c b/block/nvme.c
> > index 5be3a39b63..f8bd11e19a 100644
> > --- a/block/nvme.c
> > +++ b/block/nvme.c
> > @@ -111,6 +111,8 @@ typedef struct {
> >      uint64_t max_transfer;
> >      bool plugged;
> >  
> > +    bool supports_write_zeros;
> > +
> 
> I suppose the spelling of "zeroes" is not as established as I thought it
> was. Actually, what's worse is that the NVME writers apparently couldn't
> decide what to name it themselves either:
> 
> "Write Zeroes" has 23 hits.
> "Write Zeros" has just two, in Figure 114 for the Identify NS Data.
> 
> Oh, in QEMU we're not much better:
> 
> jhuston@probe (review) ~/s/qemu> git grep -i zeros | wc -l
> 265
> jhuston@probe (review) ~/s/qemu> git grep -i zeroes | wc -l
> 747
> 
> I'm going to suggest that we use 'zeroes' as the spelling here, to match
> the existing 'pwrite_zeroes', and then otherwise to match the NVME
> spec's usual spelling.
No problem I didn't notice the two spelling to be honest.


> 
> >      CoMutex dma_map_lock;
> >      CoQueue dma_flush_queue;
> >  
> > @@ -421,6 +423,7 @@ static void nvme_identify(BlockDriverState *bs, int 
> > namespace, Error **errp)
> >      NvmeIdNs *idns;
> >      NvmeLBAF *lbaf;
> >      uint8_t *resp;
> > +    uint16_t oncs;
> >      int r;
> >      uint64_t iova;
> >      NvmeCmd cmd = {
> > @@ -458,6 +461,9 @@ static void nvme_identify(BlockDriverState *bs, int 
> > namespace, Error **errp)
> >      s->max_transfer = MIN_NON_ZERO(s->max_transfer,
> >                            s->page_size / sizeof(uint64_t) * s->page_size);
> >  
> > +    oncs = le16_to_cpu(idctrl->oncs);
> > +    s->supports_write_zeros = (oncs & NVME_ONCS_WRITE_ZEROS) != 0;
> > +
> 
> For other reviewers: oncs is "Optional NVM Command Support".
> 
> I think it's better to say `!!(oncs & NVME_ONCS_WRITE_ZEROES)` to remove
> doubt over the width of the bitmask.
No problem.


> 
> >      memset(resp, 0, 4096);
> >  
> >      cmd.cdw10 = 0;
> > @@ -470,6 +476,12 @@ 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;
> > +    }
> > +
> >      if (lbaf->ms) {
> >          error_setg(errp, "Namespaces with metadata are not yet supported");
> >          goto out;
> > @@ -764,6 +776,8 @@ static int nvme_file_open(BlockDriverState *bs, QDict 
> > *options, int flags,
> >      int ret;
> >      BDRVNVMeState *s = bs->opaque;
> >  
> > +    bs->supported_write_flags = BDRV_REQ_FUA;
> > +
> 
> Is this a related change?
Yes. this tells qemu that you can 'force unit access' on write zeros
command, which is a flag (non optional thankfully) that tells
the drive to actually write to the flash, and not just cache the result
in some of its caches.

> 
> >      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);
> > @@ -792,7 +806,6 @@ static int nvme_file_open(BlockDriverState *bs, QDict 
> > *options, int flags,
> >              goto fail;
> >          }
> >      }
> > -    bs->supported_write_flags = BDRV_REQ_FUA;
> >      return 0;
> >  fail:
> >      nvme_close(bs);
> > @@ -1086,6 +1099,60 @@ 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];
> 
> I think it'd be slick to name the queues, but that's not related to this
> patch.
You can say that again. There is plenty todo in that driver.
This driver kind of has support to allocate more that one hardware IO queue,
but it only allocate 1 queue. It is actually quite optimal from my work on 
nvme-mdev.
Note that [1] is because the [0] is the admin queue, so technically it allocates
2 queues, one admin and one IO.

> 
> > +    NVMeRequest *req;
> > +
> > +    uint32_t cdw12 = ((bytes >> s->blkshift) - 1) & 0xFFFF;
> > +
> > +    if (!s->supports_write_zeros) {
> > +        return -ENOTSUP;
> > +    }
> > +
> > +    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,
> > +    };
> > +
> > +    if (flags & BDRV_REQ_MAY_UNMAP) {
> > +        cdw12 |= (1 << 25);
> > +    }
> > +
> > +    if (flags & BDRV_REQ_FUA) {
> > +        cdw12 |= (1 << 30);
> > +    }
> > +
> > +    cmd.cdw12 = cpu_to_le32(cdw12);
> > +
> > +    trace_nvme_write_zeros(s, offset, bytes, flags);
> > +    assert(s->nr_queues > 1);
> > +    req = nvme_get_free_req(ioq);
> > +    assert(req);
> > +
> > +    nvme_submit_command(s, ioq, req, &cmd, nvme_rw_cb, &data);
> > +
> > +    data.co = qemu_coroutine_self();
> > +    while (data.ret == -EINPROGRESS) {
> > +        qemu_coroutine_yield();
> > +    }
> > +
> > +    trace_nvme_rw_done(s, true, offset, bytes, data.ret);
> > +    return data.ret;
> > +}
> > +
> > +
> >  static int nvme_reopen_prepare(BDRVReopenState *reopen_state,
> >                                 BlockReopenQueue *queue, Error **errp)
> >  {
> > @@ -1190,6 +1257,9 @@ static BlockDriver bdrv_nvme = {
> >  
> >      .bdrv_co_preadv           = nvme_co_preadv,
> >      .bdrv_co_pwritev          = nvme_co_pwritev,
> > +
> > +    .bdrv_co_pwrite_zeroes    = nvme_co_pwrite_zeroes,
> > +
> >      .bdrv_co_flush_to_disk    = nvme_co_flush,
> >      .bdrv_reopen_prepare      = nvme_reopen_prepare,
> >  
> > diff --git a/block/trace-events b/block/trace-events
> > index 04209f058d..8209fbd0c7 100644
> > --- a/block/trace-events
> > +++ b/block/trace-events
> > @@ -149,6 +149,7 @@ nvme_submit_command_raw(int c0, int c1, int c2, int c3, 
> > int c4, int c5, int c6,
> >  nvme_handle_event(void *s) "s %p"
> >  nvme_poll_cb(void *s) "s %p"
> >  nvme_prw_aligned(void *s, int is_write, uint64_t offset, uint64_t bytes, 
> > int flags, int niov) "s %p is_write %d offset %"PRId64" bytes %"PRId64" 
> > flags %d niov %d"
> > +nvme_write_zeros(void *s, uint64_t offset, uint64_t bytes, int flags) "s 
> > %p offset %"PRId64" bytes %"PRId64" flags %d"
> 
> I was told once that we wanted trace events to match the name of the
> function they occurred within whenever possible.
> 
> Maybe we haven't really been very good about enforcing that.
I didn't knew that.


> 
> Oh well.
> 
> >  nvme_qiov_unaligned(const void *qiov, int n, void *base, size_t size, int 
> > align) "qiov %p n %d base %p size 0x%zx align 0x%x"
> >  nvme_prw_buffered(void *s, uint64_t offset, uint64_t bytes, int niov, int 
> > is_write) "s %p offset %"PRId64" bytes %"PRId64" niov %d is_write %d"
> >  nvme_rw_done(void *s, int is_write, uint64_t offset, uint64_t bytes, int 
> > ret) "s %p is_write %d offset %"PRId64" bytes %"PRId64" ret %d"
> > diff --git a/include/block/nvme.h b/include/block/nvme.h
> > index 3ec8efcc43..1f5b406344 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];
> 
> I wouldn't call this "res30" anymore, because now it starts at the 34th
> byte instead of the 30th.

True. That  is one of the nice things about keeping a separate userspace nvme 
driver (
the nvme-mdev was made to prevent this), that you need to duplicate everything,
and that includes the spec structures that already exist and up to date in the 
kernel.

> 
> >      NvmeLBAF    lbaf[16];
> >      uint8_t     res192[192];
> >      uint8_t     vs[3712];
> >  } NvmeIdNs;
> >  
> 
> Pre-existing, but what are any of these names supposed to mean?
> 
> (I imagine they match the spec, but where?...)
lbaf = LBA formats, these are up to 16 LBA formats, where 
each of them specifies the LBA block size, metadata size,
and hint about how fast that format is.
Basically this allows the user to 'low level format' the drive
to several sector sizes (usually 512 and 4K), plus enable/disable metadata
(extra data that drive can store per sector for checksumming purposes)

vs = vendor specific.

> 
> Leaving me some breadcrumbs would greatly reduce the time it takes
> someone who doesn't already know NVME to review this, and I suspect
> you've looked them up recently, so leaving little notes in the cover
> letter at least for relevant sections is very nice for hardware spec
> patches like this.
> 
> > +
> > +/*Deallocate Logical Block Features*/
> 
> ah. "dlfeat" --> "Deallocate Logical (Block) Features".
> 
> From here:
> 
> NVME Express 1.3, Figure 114: Identify - Identify Namespace Data
> Structure, NVM Command Set Specific
> 
> > +#define NVME_ID_NS_DLFEAT_GUARD_CRC(dlfeat)       ((dlfeat) & 0x10)
> 
> Not used in this patch?
Yep, this thing is metadata related which we don't support yet.

> 
> > +#define NVME_ID_NS_DLFEAT_WRITE_ZEROS(dlfeat)     ((dlfeat) & 0x08)
> > +
> > +#define NVME_ID_NS_DLFEAT_READ_BEHAVIOR(dlfeat)     ((dlfeat) & 0x7)
> > +#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))
> > 
> 
> Seems good otherwise, but I didn't trace the actual execution of the new
> command too far -- I'll assume it works. :)

I tested it a bit on the drives I have, seems to work.

Best regards,
        Maxim Levitsky