On Mon, Mar 4, 2019 at 2:04 PM Igor Konopko [off-list ref] wrote:

On 04.03.2019 13:22, Hans Holmberg wrote:

quoted

On Mon, Mar 4, 2019 at 12:44 PM Javier González [off-list ref] wrote:

quoted

quoted

On 4 Mar 2019, at 12.30, Hans Holmberg [off-list ref] wrote:

On Mon, Mar 4, 2019 at 10:05 AM Javier González [off-list ref] wrote:

quoted

quoted

On 27 Feb 2019, at 18.14, Igor Konopko [off-list ref] wrote:

Current lightnvm and pblk implementation does not care
about NVMe max data transfer size, which can be smaller
than 64*K=256K. This patch fixes issues related to that.

Could you describe *what* issues you are fixing?

I'm fixing the issue related to controllers which NVMe max data transfer
size is lower that 256K (for example 128K, which happens for existing
NVMe controllers and NVMe spec compliant). Such a controllers are not
able to handle command which contains 64 PPAs, since the the size of
DMAed buffer will be above the capabilities of such a controller.

Thanks for the explanation, that would be great to have in the commit message.

quoted

quoted

quoted

quoted

quoted

Signed-off-by: Igor Konopko <redacted>
---
drivers/lightnvm/core.c      | 9 +++++++--
drivers/nvme/host/lightnvm.c | 1 +
include/linux/lightnvm.h     | 1 +
3 files changed, 9 insertions(+), 2 deletions(-)

diff --git a/drivers/lightnvm/core.c b/drivers/lightnvm/core.c
index 5f82036fe322..c01f83b8fbaf 100644
--- a/drivers/lightnvm/core.c
+++ b/drivers/lightnvm/core.c

@@ -325,6 +325,7 @@ static int nvm_create_tgt(struct nvm_dev *dev, struct nvm_ioctl_create *create)
      struct nvm_target *t;
      struct nvm_tgt_dev *tgt_dev;
      void *targetdata;
+     unsigned int mdts;
      int ret;

      switch (create->conf.type) {

@@ -412,8 +413,12 @@ static int nvm_create_tgt(struct nvm_dev *dev, struct nvm_ioctl_create *create)
      tdisk->private_data = targetdata;
      tqueue->queuedata = targetdata;

-     blk_queue_max_hw_sectors(tqueue,
-                     (dev->geo.csecs >> 9) * NVM_MAX_VLBA);
+     mdts = (dev->geo.csecs >> 9) * NVM_MAX_VLBA;
+     if (dev->geo.mdts) {
+             mdts = min_t(u32, dev->geo.mdts,
+                             (dev->geo.csecs >> 9) * NVM_MAX_VLBA);
+     }
+     blk_queue_max_hw_sectors(tqueue, mdts);

      set_capacity(tdisk, tt->capacity(targetdata));
      add_disk(tdisk);

diff --git a/drivers/nvme/host/lightnvm.c b/drivers/nvme/host/lightnvm.c
index b759c25c89c8..b88a39a3cbd1 100644
--- a/drivers/nvme/host/lightnvm.c
+++ b/drivers/nvme/host/lightnvm.c

@@ -991,6 +991,7 @@ int nvme_nvm_register(struct nvme_ns *ns, char *disk_name, int node)
      geo->csecs = 1 << ns->lba_shift;
      geo->sos = ns->ms;
      geo->ext = ns->ext;
+     geo->mdts = ns->ctrl->max_hw_sectors;

      dev->q = q;
      memcpy(dev->name, disk_name, DISK_NAME_LEN);

diff --git a/include/linux/lightnvm.h b/include/linux/lightnvm.h
index 5d865a5d5cdc..d3b02708e5f0 100644
--- a/include/linux/lightnvm.h
+++ b/include/linux/lightnvm.h

@@ -358,6 +358,7 @@ struct nvm_geo {
      u16     csecs;          /* sector size */
      u16     sos;            /* out-of-band area size */
      bool    ext;            /* metadata in extended data buffer */
+     u32     mdts;           /* Max data transfer size*/

      /* device write constrains */
      u32     ws_min;         /* minimum write size */
--

2.17.1

I see where you are going with this and I partially agree, but none of
the OCSSD specs define a way to define this parameter. Thus, adding this
behavior taken from NVMe in Linux can break current implementations. Is
this a real life problem for you? Or this is just for NVMe “correctness”?

Javier

Hmm.Looking into the 2.0 spec what it says about vector reads:

(figure 28):"The number of Logical Blocks (NLB): This field indicates
the number of logical blocks to be read. This is a 0’s based value.
Maximum of 64 LBAs is supported."

You got the max limit covered, and the spec  does not say anything
about the minimum number of LBAs to support.

Matias: any thoughts on this?

Javier: How would this patch break current implementations?

Say an OCSSD controller that sets mdts to a value under 64 or does not
set it at all (maybe garbage). Think you can get to one pretty quickly...

So we cant make use of a perfectly good, standardized, parameter
because some hypothetical non-compliant device out there might not
provide a sane value?

quoted

quoted

Igor: how does this patch fix the mdts restriction? There are no
checks on i.e. the gc read path that ensures that a lower limit than
NVM_MAX_VLBA is enforced.

This is the other part where the implementation breaks.

No, it just does not fix it.

over-and-out,
Hans

IO requests issued from both garbage collector and writer thread are
upper limiter by pblk->max_write_pgs variable. This variable is
calculated based on queue_max_hw_sectors() already in pblk.

User reads on the other hand are splitted by block layer based on
queue_max_hw_sectors() too.

Mea culpa, right you are!

Is there any other path which I'm missing, which will still tries to
issue IOs with higher IO size? Or am I missing sth else in my logic?

I missed the adjustments of max_write_pgs in pblk_core_init. We should be good.

Thanks,
Hans