On Wed, 10 Feb 2021 18:17:25 +0000
Jonathan Cameron <jonathan.came...@huawei.com> wrote:
> On Tue, 9 Feb 2021 16:02:54 -0800
> Ben Widawsky <ben.widaw...@intel.com> wrote:
>
> > From: Dan Williams <dan.j.willi...@intel.com>
> >
> > Create the /sys/bus/cxl hierarchy to enumerate:
> >
> > * Memory Devices (per-endpoint control devices)
> >
> > * Memory Address Space Devices (platform address ranges with
> > interleaving, performance, and persistence attributes)
> >
> > * Memory Regions (active provisioned memory from an address space device
> > that is in use as System RAM or delegated to libnvdimm as Persistent
> > Memory regions).
> >
> > For now, only the per-endpoint control devices are registered on the
> > 'cxl' bus. However, going forward it will provide a mechanism to
> > coordinate cross-device interleave.
> >
> > Signed-off-by: Dan Williams <dan.j.willi...@intel.com>
> > Signed-off-by: Ben Widawsky <ben.widaw...@intel.com>
>
> One stray header, and a request for a tiny bit of reordering to
> make it easier to chase through creation and destruction.
>
> Either way with the header move to earlier patch I'm fine with this one.
>
> Reviewed-by: Jonathan Cameron <jonathan.came...@huawei.com>
Actually thinking more on this, what is the justification for the
complexity + overhead of a percpu_refcount vs a refcount
I don't think this is a high enough performance path for it to matter.
Perhaps I'm missing a usecase where it does?
Jonathan
>
> > ---
> > Documentation/ABI/testing/sysfs-bus-cxl | 26 ++
> > .../driver-api/cxl/memory-devices.rst | 17 +
> > drivers/cxl/Makefile | 3 +
> > drivers/cxl/bus.c | 29 ++
> > drivers/cxl/cxl.h | 4 +
> > drivers/cxl/mem.c | 301 +++++++++++++++++-
> > 6 files changed, 378 insertions(+), 2 deletions(-)
> > create mode 100644 Documentation/ABI/testing/sysfs-bus-cxl
> > create mode 100644 drivers/cxl/bus.c
> >
>
>
> > diff --git a/drivers/cxl/cxl.h b/drivers/cxl/cxl.h
> > index 745f5e0bfce3..b3c56fa6e126 100644
> > --- a/drivers/cxl/cxl.h
> > +++ b/drivers/cxl/cxl.h
> > @@ -3,6 +3,7 @@
> >
> > #ifndef __CXL_H__
> > #define __CXL_H__
> > +#include <linux/range.h>
>
> Why is this coming in now? Feels like it should have been in earlier
> patch that started using struct range
>
> >
> > #include <linux/bitfield.h>
> > #include <linux/bitops.h>
> > @@ -55,6 +56,7 @@
> > (FIELD_GET(CXLMDEV_RESET_NEEDED_MASK, status) != \
> > CXLMDEV_RESET_NEEDED_NOT)
> >
> > +struct cxl_memdev;
> > /**
> > * struct cxl_mem - A CXL memory device
> > * @pdev: The PCI device associated with this CXL device.
> > @@ -72,6 +74,7 @@
> > struct cxl_mem {
> > struct pci_dev *pdev;
> > void __iomem *regs;
> > + struct cxl_memdev *cxlmd;
> >
> > void __iomem *status_regs;
> > void __iomem *mbox_regs;
> > @@ -90,4 +93,5 @@ struct cxl_mem {
> > } ram;
> > };
> >
> > +extern struct bus_type cxl_bus_type;
> > #endif /* __CXL_H__ */
> > diff --git a/drivers/cxl/mem.c b/drivers/cxl/mem.c
> > index 0a868a15badc..8bbd2495e237 100644
> > --- a/drivers/cxl/mem.c
> > +++ b/drivers/cxl/mem.c
> > @@ -1,11 +1,36 @@
> >
>
> > +
> > +static void cxl_memdev_release(struct device *dev)
> > +{
> > + struct cxl_memdev *cxlmd = to_cxl_memdev(dev);
> > +
> > + percpu_ref_exit(&cxlmd->ops_active);
> > + ida_free(&cxl_memdev_ida, cxlmd->id);
> > + kfree(cxlmd);
> > +}
> > +
> ...
>
> > +static int cxl_mem_add_memdev(struct cxl_mem *cxlm)
> > +{
> > + struct pci_dev *pdev = cxlm->pdev;
> > + struct cxl_memdev *cxlmd;
> > + struct device *dev;
> > + struct cdev *cdev;
> > + int rc;
> > +
> > + cxlmd = kzalloc(sizeof(*cxlmd), GFP_KERNEL);
> > + if (!cxlmd)
> > + return -ENOMEM;
> > + init_completion(&cxlmd->ops_dead);
> > +
> > + /*
> > + * @cxlm is deallocated when the driver unbinds so operations
> > + * that are using it need to hold a live reference.
> > + */
> > + cxlmd->cxlm = cxlm;
> > + rc = percpu_ref_init(&cxlmd->ops_active, cxlmdev_ops_active_release, 0,
> > + GFP_KERNEL);
> > + if (rc)
> > + goto err_ref;
> > +
> > + rc = ida_alloc_range(&cxl_memdev_ida, 0, CXL_MEM_MAX_DEVS, GFP_KERNEL);
> > + if (rc < 0)
> > + goto err_id;
> > + cxlmd->id = rc;
> > +
> > + dev = &cxlmd->dev;
> > + device_initialize(dev);
> > + dev->parent = &pdev->dev;
> > + dev->bus = &cxl_bus_type;
> > + dev->devt = MKDEV(cxl_mem_major, cxlmd->id);
> > + dev->type = &cxl_memdev_type;
> > + dev_set_name(dev, "mem%d", cxlmd->id);
> > +
> > + cdev = &cxlmd->cdev;
> > + cdev_init(cdev, &cxl_memdev_fops);
> > +
> > + rc = cdev_device_add(cdev, dev);
> > + if (rc)
> > + goto err_add;
> > +
> > + return devm_add_action_or_reset(dev->parent, cxlmdev_unregister,
> > cxlmd);
>
> This had me scratching my head. The cxlmdev_unregister() if called normally
> or in the _or_reset() results in
>
> percpu_ref_kill(&cxlmd->ops_active);
> cdev_device_del(&cxlmd->cdev, dev);
> wait_for_completion(&cxlmd->ops_dead);
> cxlmd->cxlm = NULL;
> put_device(dev);
> /* If last ref this will result in */
> percpu_ref_exit(&cxlmd->ops_active);
> ida_free(&cxl_memdev_ida, cxlmd->id);
> kfree(cxlmd);
>
> So it's doing all the correct things but not necessarily
> in the obvious order.
>
> For simplicity of review perhaps it's worth reordering probe a bit
> to get the ida immediately after the cxlmd alloc and
> for the cxlmdev_unregister() perhaps reorder the cdev_device_del()
> before the percpu_ref_kill().
>
> Trivial obvious as the ordering has no affect but makes it
> easy for reviewers to tick off setup vs tear down parts.
>
> > +
> > +err_add:
> > + ida_free(&cxl_memdev_ida, cxlmd->id);
> > +err_id:
> > + /*
> > + * Theoretically userspace could have already entered the fops,
> > + * so flush ops_active.
> > + */
> > + percpu_ref_kill(&cxlmd->ops_active);
> > + wait_for_completion(&cxlmd->ops_dead);
> > + percpu_ref_exit(&cxlmd->ops_active);
> > +err_ref:
> > + kfree(cxlmd);
> > +
> > + return rc;
> > +}
> > +
>
>
>
>
_______________________________________________
Linux-nvdimm mailing list -- linux-nvdimm@lists.01.org
To unsubscribe send an email to linux-nvdimm-le...@lists.01.org
