Re: [PATCH 05/15] habanalabs: add command buffer module

From: Oded Gabbay
Date: Mon Jan 28 2019 - 02:53:54 EST


On Sun, Jan 27, 2019 at 8:49 AM Mike Rapoport <rppt@xxxxxxxxxxxxx> wrote:
>
> On Fri, Jan 25, 2019 at 11:47:03PM +0200, Oded Gabbay wrote:
> > On Wed, Jan 23, 2019 at 2:28 PM Mike Rapoport <rppt@xxxxxxxxxxxxx> wrote:
> > >
> > > On Wed, Jan 23, 2019 at 02:00:47AM +0200, Oded Gabbay wrote:
> > > > This patch adds the CB module, which allows the user to create and
> > > > destroy CBs and to map them to the user's process address-space.
> > >
> > > Can you please spell "command buffer" at least first time it's mentioned?
> > fixed
> > >
> > > > A command buffer is a memory blocks that reside in DMA-able address-space
> > > > and is physically contiguous so it can be accessed by the device without
> > > > MMU translation. The command buffer memory is allocated using the
> > > > coherent DMA API.
> > > >
> > > > When creating a new CB, the IOCTL returns a handle of it, and the
> > > > user-space process needs to use that handle to mmap the buffer to get a VA
> > > > in the user's address-space.
> > > >
> > > > Before destroying (freeing) a CB, the user must unmap the CB's VA using the
> > > > CB handle.
> > > >
> > > > Each CB has a reference counter, which tracks its usage in command
> > > > submissions and also its mmaps (only a single mmap is allowed).
> > > >
> > > > The driver maintains a pool of pre-allocated CBs in order to reduce
> > > > latency during command submissions. In case the pool is empty, the driver
> > > > will go to the slow-path of allocating a new CB, i.e. calling
> > > > dma_alloc_coherent.
> > > >
> > > > Signed-off-by: Oded Gabbay <oded.gabbay@xxxxxxxxx>
> > > > ---
> > > > drivers/misc/habanalabs/Makefile | 3 +-
> > > > drivers/misc/habanalabs/command_buffer.c | 414 +++++++++++++++++++++
> > > > drivers/misc/habanalabs/device.c | 43 ++-
> > > > drivers/misc/habanalabs/goya/goya.c | 28 ++
> > > > drivers/misc/habanalabs/habanalabs.h | 95 ++++-
> > > > drivers/misc/habanalabs/habanalabs_drv.c | 2 +
> > > > drivers/misc/habanalabs/habanalabs_ioctl.c | 102 +++++
> > > > include/uapi/misc/habanalabs.h | 62 +++
> > > > 8 files changed, 746 insertions(+), 3 deletions(-)
> > > > create mode 100644 drivers/misc/habanalabs/command_buffer.c
> > > > create mode 100644 drivers/misc/habanalabs/habanalabs_ioctl.c
> > > > create mode 100644 include/uapi/misc/habanalabs.h
>
> [ ... ]
>
> > > > +int hl_cb_create(struct hl_device *hdev, struct hl_cb_mgr *mgr,
> > > > + u32 cb_size, u64 *handle, int ctx_id)
> > > > +{
> > > > + struct hl_cb *cb;
> > > > + bool alloc_new_cb = true;
> > > > + int rc;
> > > > +
> > > > + if (hdev->disabled) {
> > > > + dev_warn_ratelimited(hdev->dev,
> > > > + "Device is disabled !!! Can't create new CBs\n");
> > > > + rc = -EBUSY;
> > > > + goto out_err;
> > > > + }
> > > > +
> > > > + /* Minimum allocation must be PAGE SIZE */
> > > > + if (cb_size < PAGE_SIZE)
> > > > + cb_size = PAGE_SIZE;
> > > > +
> > > > + if (ctx_id == HL_KERNEL_ASID_ID &&
> > > > + cb_size <= hdev->asic_prop.cb_pool_cb_size) {
> > > > +
> > > > + spin_lock(&hdev->cb_pool_lock);
> > > > + if (!list_empty(&hdev->cb_pool)) {
> > > > + cb = list_first_entry(&hdev->cb_pool, typeof(*cb),
> > > > + pool_list);
> > > > + list_del(&cb->pool_list);
> > > > + spin_unlock(&hdev->cb_pool_lock);
> > > > + alloc_new_cb = false;
> > > > + } else {
> > > > + spin_unlock(&hdev->cb_pool_lock);
> > > > + dev_warn_once(hdev->dev, "CB pool is empty\n");
> > >
> > > Isn't it going to be a false alarm when you allocate the cb for the first
> > > time?
> > Why ?
> > The cb_pool list holds a list of available CBs. See hl_cb_pool_init()
> > - it adds newly allocated CBs to this pool list.
> >
> > if (!list_empty(&hdev->cb_pool)) { - this checks whether the
> > pool is not empty so we can take an available CB from it. If the list
> > is empty (hence the pool is empty), we print the warning.
>
> Sorry if it's too much nitpicking, but why the allocation of the first cb
> should be a warning? There's nothing wrong there... Maybe dev_dbg()
> instead?
Yeah, that's a fair point. The issue is I would like to know if we
reach to this state and dev_dbg isn't usually enabled.
Still, I get what you are saying and I'll change this to dev_dbg.

>
> > > > + }
> > > > + }
> > > > +
> > > > + if (alloc_new_cb) {
> > > > + cb = hl_cb_alloc(hdev, cb_size, ctx_id);
> > > > + if (!cb) {
> > > > + rc = -ENOMEM;
> > > > + goto out_err;
> > > > + }
> > > > + }
> > > > +
> > > > + cb->hdev = hdev;
> > > > + cb->ctx_id = ctx_id;
> > > > +
> > > > + spin_lock(&mgr->cb_lock);
> > > > + rc = idr_alloc(&mgr->cb_handles, cb, 1, 0, GFP_ATOMIC);
> > >
> > > It seems the ID will remain dangling if the cb is reused.
> >
> > I'm not sure what you mean by this comment. Reused by whom ? in how
> > fashion it is reused ?
>
> Sorry if I didn't explain it more clearly.
> If the case the cb is reused, you anyway call idr_alloc() and overwrite the
> previous value of cb->id and it never gets idr_remove()'ed
I don't think that is the case.
Please look at hl_cb_destroy(). There, we do the idr_remove and then
we kref_put the CB. In it's release code path, we check if this is a
CB from pool, and if so, we return it to the pool. When it will be
alloc'ed again, it will get a new id.
The problem in this patch is that hl_cb_destroy is not used yet for
CB's from the pool because the command submission code which use that
comes at a later patch, so indeed it might be confusing. But if you
will take a look at the entire code and check when hl_cb_destroy is
called I think you will agree with me.
But if you still think otherwise, please tell me. I might be missing
something here.

Thanks,
Oded

>
> > >
> > > > + spin_unlock(&mgr->cb_lock);
> > > > +
> > > > + if (rc < 0) {
> > > > + dev_err(hdev->dev, "Failed to allocate IDR for a new CB\n");
> > > > + goto release_cb;
> > > > + }
> > > > +
> > > > + cb->id = rc;
> > > > +
> > > > + kref_init(&cb->refcount);
> > > > + spin_lock_init(&cb->lock);
> > > > +
> > > > + /*
> > > > + * idr is 32-bit so we can safely OR it with a mask that is above
> > > > + * 32 bit
> > > > + */
> > > > + *handle = cb->id | HL_MMAP_CB_MASK;
> > > > + *handle <<= PAGE_SHIFT;
> > > > +
> > > > + return 0;
> > > > +
> > > > +release_cb:
> > > > + cb_do_release(hdev, cb);
> > > > +out_err:
> > > > + *handle = 0;
> > > > +
> > > > + return rc;
> > > > +}
> > > > +
>
> --
> Sincerely yours,
> Mike.
>