Re: [PATCH v3 17/27] powerpc/powernv/pmem: Implement the Read Error Log command

From: Frederic Barrat
Date: Thu Mar 05 2020 - 04:33:38 EST


+ if (rc)
+ goto out;
+
+ rc = ocxl_global_mmio_read64(ocxlpmem->ocxl_afu,
+ ocxlpmem-
admin_command.data_offset + 0x28,
+ OCXL_HOST_ENDIAN, &log->wwid[1]);
+ if (rc)
+ goto out;
+
+ rc = ocxl_global_mmio_read64(ocxlpmem->ocxl_afu,
+ ocxlpmem-
admin_command.data_offset + 0x30,
+ OCXL_HOST_ENDIAN, (u64 *)log-
fw_revision);
+ if (rc)
+ goto out;
+ log->fw_revision[8] = '\0';
+
+ buf_length = (user_buf_length < log->buf_size) ?
+ user_buf_length : log->buf_size;
+ for (i = 0; i < buf_length + 0x48; i += 8) {
+ u64 val;
+
+ rc = ocxl_global_mmio_read64(ocxlpmem->ocxl_afu,
+ ocxlpmem-
admin_command.data_offset + i,
+ OCXL_HOST_ENDIAN, &val);
+ if (rc)
+ goto out;
+
+ if (buf_is_user) {
+ if (copy_to_user(&log->buf[i], &val,
sizeof(u64))) {
+ rc = -EFAULT;
+ goto out;
+ }
+ } else
+ log->buf[i] = val;
+ }


I think it could be a bit simplified by keeping the handling of the
user
buffer out of this function. Always call it with a kernel buffer.
And
have only one copy_to_user() call on the ioctl() path. You'd need to
allocate a kernel buf on the ioctl path, but you're already doing it
on
the probe() path, so it should be doable to share code.

Hmm, the problem then is that on the IOCTL side, I'll have to save,
modify, then restore the buf member of struct
ioctl_ocxl_pmem_error_log, which would be uglier.


buf is just an output buffer. All you'd need to do is allocate a kernel buf, like it's already done for the "probe" case in dump_error_log(). And add a global copy_to_user() of the buf at the end of the ioctl path, instead of having multiple smaller copy_to_user() in the loop here.
copy_to_user() is a bit expensive so it's usually better to regroup them. I think it's easy here and make sense since that function is also trying to handle both a kernel and user space bufffers.
But we're not in a critical path, and after this patch, there are others copying out mmio content to user buffers and those don't have a kernel buffer to handle, so the copy_to_user() in a loop makes things easier.
So I guess the conclusion is whatever you think is the easiest...





+
+ rc = admin_response_handled(ocxlpmem);
+ if (rc)
+ goto out;
+
+out:
+ mutex_unlock(&ocxlpmem->admin_command.lock);
+ return rc;
+
+}
+
+static int ioctl_error_log(struct ocxlpmem *ocxlpmem,
+ struct ioctl_ocxl_pmem_error_log __user *uarg)
+{
+ struct ioctl_ocxl_pmem_error_log args;
+ int rc;
+
+ if (copy_from_user(&args, uarg, sizeof(args)))
+ return -EFAULT;
+
+ rc = read_error_log(ocxlpmem, &args, true);
+ if (rc)
+ return rc;
+
+ if (copy_to_user(uarg, &args, sizeof(args)))
+ return -EFAULT;
+
+ return 0;
+}
+
+static long file_ioctl(struct file *file, unsigned int cmd,
unsigned long args)
+{
+ struct ocxlpmem *ocxlpmem = file->private_data;
+ int rc = -EINVAL;
+
+ switch (cmd) {
+ case IOCTL_OCXL_PMEM_ERROR_LOG:
+ rc = ioctl_error_log(ocxlpmem,
+ (struct ioctl_ocxl_pmem_error_log
__user *)args);
+ break;
+ }
+ return rc;
+}
+
static const struct file_operations fops = {
.owner = THIS_MODULE,
.open = file_open,
.release = file_release,
+ .unlocked_ioctl = file_ioctl,
+ .compat_ioctl = file_ioctl,
};
/**
@@ -527,6 +736,60 @@ static int read_device_metadata(struct
ocxlpmem *ocxlpmem)
return 0;
}
+static const char *decode_error_log_type(u8 error_log_type)
+{
+ switch (error_log_type) {
+ case 0x00:
+ return "general";
+ case 0x01:
+ return "predictive failure";
+ case 0x02:
+ return "thermal warning";
+ case 0x03:
+ return "data loss";
+ case 0x04:
+ return "health & performance";
+ default:
+ return "unknown";
+ }
+}
+
+static void dump_error_log(struct ocxlpmem *ocxlpmem)
+{
+ struct ioctl_ocxl_pmem_error_log log;
+ u32 buf_size;
+ u8 *buf;
+ int rc;
+
+ if (ocxlpmem->admin_command.data_size == 0)
+ return;
+
+ buf_size = ocxlpmem->admin_command.data_size - 0x48;
+ buf = kzalloc(buf_size, GFP_KERNEL);
+ if (!buf)
+ return;
+
+ log.buf = buf;
+ log.buf_size = buf_size;
+
+ rc = read_error_log(ocxlpmem, &log, false);
+ if (rc < 0)
+ goto out;
+
+ dev_warn(&ocxlpmem->dev,
+ "OCXL PMEM Error log: WWID=0x%016llx%016llx LID=0x%x
PRC=%x type=0x%x %s, Uptime=%u seconds timestamp=0x%llx\n",
+ log.wwid[0], log.wwid[1],
+ log.log_identifier, log.program_reference_code,
+ log.error_log_type,
+ decode_error_log_type(log.error_log_type),
+ log.power_on_seconds, log.timestamp);
+ print_hex_dump(KERN_WARNING, "buf", DUMP_PREFIX_OFFSET, 16, 1,
buf,
+ log.buf_size, false);

dev_warn already logs a warning. Isn't KERN_DEBUG more appropriate
for
the hex dump?



The hex dump is associated binary data for the warning, it doesn't
replicate the contents of the message.


My point is not about duplicating, it's about exposing an hexadecimal dump where it makes sense. Those DEBUG and WARNING tags are used for filtering content. For example to know what to display on the console. A warning to mention that a device hits a serious error is perfectly fine. A hexadecimal dump which is going to be meaningless to most everybody is not. The system is not crashing, so it's not like the console is our last hope. I think the dump is debug data and should be tagged as such.

Fred




+
+out:
+ kfree(buf);
+}
+
/**
* probe_function0() - Set up function 0 for an OpenCAPI
persistent memory device
* This is important as it enables templates higher than 0 across
all other functions,
@@ -568,6 +831,7 @@ static int probe(struct pci_dev *pdev, const
struct pci_device_id *ent)
struct ocxlpmem *ocxlpmem;
int rc;
u16 elapsed, timeout;
+ u64 chi;
if (PCI_FUNC(pdev->devfn) == 0)
return probe_function0(pdev);
@@ -667,6 +931,11 @@ static int probe(struct pci_dev *pdev, const
struct pci_device_id *ent)
return 0;
err:
+ if (ocxlpmem &&
+ (ocxlpmem_chi(ocxlpmem, &chi) == 0) &&
+ (chi & GLOBAL_MMIO_CHI_ELA))
+ dump_error_log(ocxlpmem);
+
/*
* Further cleanup is done in the release handler via
free_ocxlpmem()
* This allows us to keep the character device live to handle
IOCTLs to
diff --git a/arch/powerpc/platforms/powernv/pmem/ocxl_internal.h
b/arch/powerpc/platforms/powernv/pmem/ocxl_internal.h
index d2d81fec7bb1..b953ee522ed4 100644
--- a/arch/powerpc/platforms/powernv/pmem/ocxl_internal.h
+++ b/arch/powerpc/platforms/powernv/pmem/ocxl_internal.h
@@ -5,6 +5,7 @@
#include <linux/cdev.h>
#include <misc/ocxl.h>
#include <linux/libnvdimm.h>
+#include <uapi/nvdimm/ocxl-pmem.h>

Can't we limit the extra include to ocxl.c?


Yes, there are no consumers referred to in ocxl_interal.[hc]

Completely unrelated, but ocxl.c contains most of the code for this
driver. We should consider renaming it to ocxlpmem.c or something
along
those lines, since it does a lot more than just interfacing with the
opencapi interface. And would avoid confusion with an other already
existing ocxl.c file.


Ok, my thinking was that it's already in a pmem directory, but I can
see arguments both ways.


#include <linux/mm.h>
#define LABEL_AREA_SIZE (1UL << PA_SECTION_SHIFT)
diff --git a/include/uapi/nvdimm/ocxl-pmem.h
b/include/uapi/nvdimm/ocxl-pmem.h
new file mode 100644
index 000000000000..b10f8ac0c20f
--- /dev/null
+++ b/include/uapi/nvdimm/ocxl-pmem.h
@@ -0,0 +1,46 @@
+/* SPDX-License-Identifier: GPL-2.0+ WITH Linux-syscall-note */
+/* Copyright 2017 IBM Corp. */
+#ifndef _UAPI_OCXL_SCM_H
+#define _UAPI_OCXL_SCM_H
+
+#include <linux/types.h>
+#include <linux/ioctl.h>
+
+#define OCXL_PMEM_ERROR_LOG_ACTION_RESET (1 << (32-32))
+#define OCXL_PMEM_ERROR_LOG_ACTION_CHKFW (1 << (53-32))
+#define OCXL_PMEM_ERROR_LOG_ACTION_REPLACE (1 << (54-32))
+#define OCXL_PMEM_ERROR_LOG_ACTION_DUMP (1 << (55-32))
+
+#define OCXL_PMEM_ERROR_LOG_TYPE_GENERAL (0x00)
+#define OCXL_PMEM_ERROR_LOG_TYPE_PREDICTIVE_FAILURE (0x01)
+#define OCXL_PMEM_ERROR_LOG_TYPE_THERMAL_WARNING (0x02)
+#define OCXL_PMEM_ERROR_LOG_TYPE_DATA_LOSS (0x03)
+#define OCXL_PMEM_ERROR_LOG_TYPE_HEALTH_PERFORMANCE (0x04)
+
+struct ioctl_ocxl_pmem_error_log {
+ __u32 log_identifier; /* out */
+ __u32 program_reference_code; /* out */
+ __u32 action_flags; /* out, recommended course of action */
+ __u32 power_on_seconds; /* out, Number of seconds the
controller has been on when the error occurred */
+ __u64 timestamp; /* out, relative time since the current IPL */
+ __u64 wwid[2]; /* out, the NAA formatted WWID associated with
the controller */
+ char fw_revision[8+1]; /* out, firmware revision as null
terminated text */

The 8+1 size will make the compiler add some padding here. Are we
confident that all the compilers, at least on powerpc, will do the
same
thing and we can guarantee a kernel ABI? I would play it safe and
have a
discussion with folks who understand compilers better.


I'll add some explicit padding.



+ __u16 buf_size; /* in/out, buffer size provided/required.
+ * If required is greater than provided, the
buffer
+ * will be truncated to the amount provided. If
its
+ * less, then only the required bytes will be
populated.
+ * If it is 0, then there are no more error log
entries.
+ */
+ __u8 error_log_type;
+ __u8 reserved1;
+ __u32 reserved2;
+ __u64 reserved3[2];
+ __u8 *buf; /* pointer to output buffer */
+};
+
+/* ioctl numbers */
+#define OCXL_PMEM_MAGIC 0x5C

Randomly picked?
See (and add entry in) Documentation/userspace-api/ioctl/ioctl-
number.rst

Ok


Fred



+/* SCM devices */
+#define IOCTL_OCXL_PMEM_ERROR_LOG _IOWR(OCXL_PMEM
_MAGIC, 0x01, struct ioctl_ocxl_pmem_error_log)
+
+#endif /* _UAPI_OCXL_SCM_H */