Re: [PATCH 2/2] powerpc/pseries: Add driver for PAPR SCM regions

On Sat, Oct 13, 2018 at 9:37 AM Dan Williams [off-list ref] wrote:

On Tue, Oct 9, 2018 at 11:21 PM Oliver O'Halloran [off-list ref] wrote:

quoted

Adds a driver that implements support for enabling and accessing PAPR
SCM regions. Unfortunately due to how the PAPR interface works we can't
use the existing of_pmem driver (yet) because:

 a) The guest is required to use the H_SCM_BIND_MEM h-call to add
    add the SCM region to it's physical address space, and
 b) There is currently no mechanism for relating a bare of_pmem region
    to the backing DIMM (or not-a-DIMM for our case).

Both of these are easily handled by rolling the functionality into a
seperate driver so here we are...

Signed-off-by: Oliver O'Halloran <oohall@gmail.com>
---
The alternative implementation here is that we have the pseries code
do the h-calls and craft a pmem-region@<addr> node based on that.
However, that doesn't solve b) and mpe has expressed his dislike of
adding new stuff to the DT at runtime so i'd say that's a non-starter.
---
 arch/powerpc/platforms/pseries/Kconfig    |   7 +
 arch/powerpc/platforms/pseries/Makefile   |   1 +
 arch/powerpc/platforms/pseries/papr_scm.c | 340 ++++++++++++++++++++++++++++++
 3 files changed, 348 insertions(+)
 create mode 100644 arch/powerpc/platforms/pseries/papr_scm.c

diff --git a/arch/powerpc/platforms/pseries/Kconfig b/arch/powerpc/platforms/pseries/Kconfig
index 0c698fd6d491..4b0fcb80efe5 100644
--- a/arch/powerpc/platforms/pseries/Kconfig
+++ b/arch/powerpc/platforms/pseries/Kconfig

@@ -140,3 +140,10 @@ config IBMEBUS
        bool "Support for GX bus based adapters"
        help
          Bus device driver for GX bus based adapters.
+
+config PAPR_SCM
+       depends on PPC_PSERIES && MEMORY_HOTPLUG
+       select LIBNVDIMM
+       tristate "Support for the PAPR Storage Class Memory interface"
+       help
+         Enable access to hypervisor provided storage class memory.

diff --git a/arch/powerpc/platforms/pseries/Makefile b/arch/powerpc/platforms/pseries/Makefile
index 892b27ced973..a43ec843c8e2 100644
--- a/arch/powerpc/platforms/pseries/Makefile
+++ b/arch/powerpc/platforms/pseries/Makefile

@@ -24,6 +24,7 @@ obj-$(CONFIG_IO_EVENT_IRQ)    += io_event_irq.o
 obj-$(CONFIG_LPARCFG)          += lparcfg.o
 obj-$(CONFIG_IBMVIO)           += vio.o
 obj-$(CONFIG_IBMEBUS)          += ibmebus.o
+obj-$(CONFIG_PAPR_SCM)         += papr_scm.o

 ifdef CONFIG_PPC_PSERIES
 obj-$(CONFIG_SUSPEND)          += suspend.o

diff --git a/arch/powerpc/platforms/pseries/papr_scm.c b/arch/powerpc/platforms/pseries/papr_scm.c
new file mode 100644
index 000000000000..87d4dbc18845
--- /dev/null
+++ b/arch/powerpc/platforms/pseries/papr_scm.c

@@ -0,0 +1,340 @@
+// SPDX-License-Identifier: GPL-2.0
+
+#define pr_fmt(fmt)    "papr-scm: " fmt
+
+#include <linux/of.h>
+#include <linux/kernel.h>
+#include <linux/module.h>
+#include <linux/ioport.h>
+#include <linux/slab.h>
+#include <linux/ndctl.h>
+#include <linux/libnvdimm.h>
+#include <linux/platform_device.h>
+
+#include <asm/plpar_wrappers.h>
+
+#define BIND_ANY_ADDR (~0ul)
+
+#define PAPR_SCM_DIMM_CMD_MASK \
+       ((1ul << ND_CMD_GET_CONFIG_SIZE) | \
+        (1ul << ND_CMD_GET_CONFIG_DATA) | \
+        (1ul << ND_CMD_SET_CONFIG_DATA))
+
+struct papr_scm_priv {
+       struct platform_device *pdev;
+       struct device_node *dn;
+       uint32_t drc_index;
+       uint64_t blocks;
+       uint64_t block_size;
+       int metadata_size;
+
+       uint64_t bound_addr;
+
+       struct nvdimm_bus_descriptor bus_desc;
+       struct nvdimm_bus *bus;
+       struct nvdimm *nvdimm;
+       struct resource res;
+       struct nd_region *region;
+       struct nd_interleave_set nd_set;
+};
+
+static int drc_pmem_bind(struct papr_scm_priv *p)
+{
+       unsigned long ret[PLPAR_HCALL_BUFSIZE];
+       uint64_t rc, token;
+
+       /*
+        * When the hypervisor cannot map all the requested memory in a single
+        * hcall it returns H_BUSY and we call again with the token until
+        * we get H_SUCCESS. Aborting the retry loop before getting H_SUCCESS
+        * leave the system in an undefined state, so we wait.
+        */
+       token = 0;
+
+       do {
+               rc = plpar_hcall(H_SCM_BIND_MEM, ret, p->drc_index, 0,
+                               p->blocks, BIND_ANY_ADDR, token);
+               token = be64_to_cpu(ret[0]);

Does plpar_hcall sleep? Should there be a cond_resched() here if not?

That's probably a good idea. The implementations I have to test with
never return H_BUSY and I had figured it just indicated that

quoted

+       } while (rc == H_BUSY);
+
+       if (rc) {
+               dev_err(&p->pdev->dev, "bind err: %lld\n", rc);
+               return -ENXIO;
+       }
+
+       p->bound_addr = be64_to_cpu(ret[1]);
+
+       dev_dbg(&p->pdev->dev, "bound drc %x to %pR\n", p->drc_index, &p->res);
+
+       return 0;
+}
+
+static int drc_pmem_unbind(struct papr_scm_priv *p)
+{
+       unsigned long ret[PLPAR_HCALL_BUFSIZE];
+       uint64_t rc, token;
+
+       token = 0;
+
+       /* NB: unbind has the same retry requirements mentioned above */
+       do {
+               rc = plpar_hcall(H_SCM_UNBIND_MEM, ret, p->drc_index,
+                               p->bound_addr, p->blocks, token);
+               token = be64_to_cpu(ret);

...same busy loop comment.

quoted

+       } while (rc == H_BUSY);
+
+       if (rc)
+               dev_err(&p->pdev->dev, "unbind error: %lld\n", rc);
+
+       return !!rc;
+}
+
+static int papr_scm_meta_get(struct papr_scm_priv *p,
+                       struct nd_cmd_get_config_data_hdr *hdr)
+{
+       unsigned long data[PLPAR_HCALL_BUFSIZE];
+       int64_t ret;
+
+       if (hdr->in_offset >= p->metadata_size || hdr->in_length != 1)
+               return -EINVAL;
+
+       ret = plpar_hcall(H_SCM_READ_METADATA, data, p->drc_index,
+                       hdr->in_offset, 1);
+
+       if (ret == H_PARAMETER) /* bad DRC index */
+               return -ENODEV;
+       if (ret)
+               return -EINVAL; /* other invalid parameter */
+
+       hdr->out_buf[0] = data[0] & 0xff;
+
+       return 0;
+}
+
+static int papr_scm_meta_set(struct papr_scm_priv *p,
+                       struct nd_cmd_set_config_hdr *hdr)
+{
+       int64_t ret;
+
+       if (hdr->in_offset >= p->metadata_size || hdr->in_length != 1)
+               return -EINVAL;
+
+       ret = plpar_hcall_norets(H_SCM_WRITE_METADATA,
+                       p->drc_index, hdr->in_offset, hdr->in_buf[0], 1);
+
+       if (ret == H_PARAMETER) /* bad DRC index */
+               return -ENODEV;
+       if (ret)
+               return -EINVAL; /* other invalid parameter */
+
+       return 0;
+}
+
+int papr_scm_ndctl(struct nvdimm_bus_descriptor *nd_desc, struct nvdimm *nvdimm,
+               unsigned int cmd, void *buf, unsigned int buf_len, int *cmd_rc)
+{
+       struct nd_cmd_get_config_size *get_size_hdr;
+       struct papr_scm_priv *p;
+
+       /* Only dimm-specific calls are supported atm */
+       if (!nvdimm)
+               return -EINVAL;
+
+       p = nvdimm_provider_data(nvdimm);
+
+       switch (cmd) {
+       case ND_CMD_GET_CONFIG_SIZE:
+               get_size_hdr = buf;
+
+               get_size_hdr->status = 0;
+               get_size_hdr->max_xfer = 1;
+               get_size_hdr->config_size = p->metadata_size;
+               *cmd_rc = 0;
+               break;
+
+       case ND_CMD_GET_CONFIG_DATA:
+               *cmd_rc = papr_scm_meta_get(p, buf);
+               break;
+
+       case ND_CMD_SET_CONFIG_DATA:
+               *cmd_rc = papr_scm_meta_set(p, buf);
+               break;
+
+       default:
+               return -EINVAL;
+       }
+
+       dev_dbg(&p->pdev->dev, "returned with cmd_rc = %d\n", *cmd_rc);
+
+       return 0;
+}
+
+static const struct attribute_group *region_attr_groups[] = {
+       &nd_region_attribute_group,
+       &nd_device_attribute_group,
+       &nd_mapping_attribute_group,
+       &nd_numa_attribute_group,
+       NULL,
+};
+
+static const struct attribute_group *bus_attr_groups[] = {
+       &nvdimm_bus_attribute_group,
+       NULL,
+};
+
+static const struct attribute_group *papr_scm_dimm_groups[] = {
+       &nvdimm_attribute_group,
+       &nd_device_attribute_group,
+       NULL,
+};
+
+static int papr_scm_nvdimm_init(struct papr_scm_priv *p)
+{
+       struct device *dev = &p->pdev->dev;
+       struct nd_mapping_desc mapping;
+       struct nd_region_desc ndr_desc;
+       unsigned long dimm_flags;
+
+       p->bus_desc.ndctl = papr_scm_ndctl;
+       p->bus_desc.module = THIS_MODULE;
+       p->bus_desc.of_node = p->pdev->dev.of_node;
+       p->bus_desc.attr_groups = bus_attr_groups;
+       p->bus_desc.provider_name = kstrdup(p->pdev->name, GFP_KERNEL);
+
+       if (!p->bus_desc.provider_name)
+               return -ENOMEM;
+
+       p->bus = nvdimm_bus_register(NULL, &p->bus_desc);
+       if (!p->bus) {
+               dev_err(dev, "Error creating nvdimm bus %pOF\n", p->dn);
+               return -ENXIO;
+       }
+
+       dimm_flags = 0;
+       set_bit(NDD_ALIASING, &dimm_flags);
+
+       p->nvdimm = nvdimm_create(p->bus, p, papr_scm_dimm_groups,
+                               dimm_flags, PAPR_SCM_DIMM_CMD_MASK, 0, NULL);

Looks good, although I'm just about to push out commits that change
this function signature to take a 'security_ops' pointer. If you need
a stable branch to base this on, let me know.

quoted

+       if (!p->nvdimm) {
+               dev_err(dev, "Error creating DIMM object for %pOF\n", p->dn);
+               goto err;
+       }
+
+       /* now add the region */
+
+       memset(&mapping, 0, sizeof(mapping));
+       mapping.nvdimm = p->nvdimm;
+       mapping.start = p->res.start;
+       mapping.size = p->blocks * p->block_size; // XXX: potential overflow?

Ideally, should be nvdimm relative addresses. Since you have just a
single DIMM per-region I would expect this to be 0 to
nvdimm->region_size. That said I don't think anything breaks if you
make the nvdimm address equal to the system-physical address, just
wanted to point out the intent. This only matters when you have
multiple DIMMs per region and need to record the per-DIMM contribution
in the labels on each DIMM.

bleh

That's what I assumed originally. I wasn't sure so I looked at what
the NFIT driver did and confused myself by assuming the address field
of acpi_nfit_memory_map was an SPA.

Other than that looks ok to me:

Acked-by: Dan Williams <redacted>

...just the matter of what to do about function signature change.