Re: [4/5] pseries: Implement memory hotplug add in the kernel

On 09/17/2014 02:07 AM, Michael Ellerman wrote:

On Mon, 2014-09-15 at 15:32 -0500, Nathan Fontenot wrote:

quoted

This patch adds the ability to do memory hotplug adding in the kernel.

Currently the hotplug add/remove of memory is handled by the drmgr
command. The drmgr command performs the add/remove by performing
some work in user-space and making requests to the kernel to handle 
other pieces. By moving all of the work to the kernel we can do the
add and remove faster, and provide a common place to do memory hotplug
for both the PowerVM and PowerKVM environments.

Signed-off-by: Nathan Fontenot <redacted>
---
 arch/powerpc/platforms/pseries/hotplug-memory.c |  170 +++++++++++++++++++++++
 1 file changed, 170 insertions(+)

diff --git a/arch/powerpc/platforms/pseries/hotplug-memory.c b/arch/powerpc/platforms/pseries/hotplug-memory.c
index 0e60e15..b254773 100644
--- a/arch/powerpc/platforms/pseries/hotplug-memory.c
+++ b/arch/powerpc/platforms/pseries/hotplug-memory.c

@@ -17,6 +17,7 @@
 #include <linux/vmalloc.h>
 #include <linux/memory.h>
 #include <linux/memory_hotplug.h>
+#include <linux/slab.h>
 
 #include <asm/firmware.h>
 #include <asm/machdep.h>

@@ -24,6 +25,8 @@
 #include <asm/sparsemem.h>
 #include <asm/rtas.h>
 
+#include "pseries.h"
+
 DEFINE_MUTEX(dlpar_mem_mutex);
 
 unsigned long pseries_memory_block_size(void)

@@ -69,6 +72,53 @@ unsigned long pseries_memory_block_size(void)
 	return memblock_size;
 }
 
+static void dlpar_free_drconf_property(struct property *prop)
+{
+	kfree(prop->name);
+	kfree(prop->value);
+	kfree(prop);
+}
+
+static struct property *dlpar_clone_drconf_property(struct device_node *dn)
+{
+	struct property *prop, *new_prop;
+
+	prop = of_find_property(dn, "ibm,dynamic-memory", NULL);
+	if (!prop)
+		return NULL;
+
+	new_prop = kzalloc(sizeof(*new_prop), GFP_KERNEL);
+	if (!new_prop)
+		return NULL;
+
+	new_prop->name = kstrdup(prop->name, GFP_KERNEL);
+	new_prop->value = kmalloc(prop->length + 1, GFP_KERNEL);
+	if (!new_prop->name || !new_prop->value) {
+		dlpar_free_drconf_property(new_prop);
+		return NULL;
+	}
+
+	memcpy(new_prop->value, prop->value, prop->length);
+	new_prop->length = prop->length;
+	*(((char *)new_prop->value) + new_prop->length) = 0;

It's not a string, is it?

No, property->value is a void*. I'll drop that line of code.

quoted

+	return new_prop;
+}
+
+static struct memory_block *lmb_to_memblock(struct of_drconf_cell *lmb)
+{
+	unsigned long section_nr;
+	struct mem_section *mem_sect;
+	struct memory_block *mem_block;
+	u64 phys_addr = be64_to_cpu(lmb->base_addr);
+
+	section_nr = pfn_to_section_nr(PFN_DOWN(phys_addr));
+	mem_sect = __nr_to_section(section_nr);
+
+	mem_block = find_memory_block(mem_sect);
+	return mem_block;
+}
+
 #ifdef CONFIG_MEMORY_HOTREMOVE
 static int pseries_remove_memory(u64 start, u64 size)
 {

@@ -155,13 +205,133 @@ static inline int pseries_remove_mem_node(struct device_node *np)
 }
 #endif /* CONFIG_MEMORY_HOTREMOVE */
 
+static int dlpar_add_one_lmb(struct of_drconf_cell *lmb)
+{
+	struct memory_block *mem_block;
+	u64 phys_addr;
+	unsigned long pages_per_block;
+	unsigned long block_sz;
+	int nid, sections_per_block;
+	int rc;
+
+	phys_addr = be64_to_cpu(lmb->base_addr);

of_drconf_cell needs endian annotations.

Yes it does. I can include a patch to update the struct.

quoted

+	block_sz = memory_block_size_bytes();
+	sections_per_block = block_sz / MIN_MEMORY_BLOCK_SIZE;
+	pages_per_block = PAGES_PER_SECTION * sections_per_block;
+
+	if (phys_addr & ((pages_per_block << PAGE_SHIFT) - 1))
+		return -EINVAL;
+
+	nid = memory_add_physaddr_to_nid(phys_addr);
+	rc = add_memory(nid, phys_addr, block_sz);
+	if (rc)
+		return rc;
+
+	rc = memblock_add(phys_addr, block_sz);
+	if (rc) {
+		remove_memory(nid, phys_addr, block_sz);
+		return rc;
+	}
+
+	mem_block = lmb_to_memblock(lmb);
+	if (!mem_block) {
+		remove_memory(nid, phys_addr, block_sz);
+		return -EINVAL;
+	}

That could all use a lot of comments. ie. why do we have to add it twice?

We don't actually add it twice, though I can see how one could think
that based on the names of the routines called. I'll add comments to 
clarify this in v2 of the patch.

memory_add_physaddr_to_nid(), this doesn't add anything despite its naming.
The routine finds the node id for the specified physical address.

add_memory(), this actually adds the memory.

memblock_add(), this informs the memory block information tracking about the
newly added memory. Why this is not done as part of add_memory I don't know.

quoted

+	rc = device_online(&mem_block->dev);
+	put_device(&mem_block->dev);
+	if (rc)
+		remove_memory(nid, phys_addr, block_sz);
+
+	return rc;
+}
+
+static int dlpar_memory_add(struct pseries_hp_errorlog *hp_elog)
+{
+	struct of_drconf_cell *lmb;
+	struct device_node *dn;
+	struct property *prop;
+	uint32_t entries, *p;

*p should be __be32.

Yes.

quoted

+	int i, lmbs_to_add;
+	int lmbs_added = 0;
+	int rc = -EINVAL;

Don't pre-initialise your rc variables.

I did this here for a reason. When asking to add memory by drc_index it
is possible to fall out of the for() loop traversing the lmb entries
and not find the requested drc_index.

Adding a check for this situation after the loop would do the same thing
and probably make this situation more clear.
 

quoted

+	if (hp_elog->id_type == PSERIES_HP_ELOG_ID_DRC_COUNT) {
+		lmbs_to_add = be32_to_cpu(hp_elog->_drc_u.drc_count);
+		pr_info("Attempting to hot-add %d LMB(s)\n", lmbs_to_add);
+	} else {
+		lmbs_to_add = 1;
+		pr_info("Attempting to hot-add LMB, drc index %x\n",
+			be32_to_cpu(hp_elog->_drc_u.drc_index));
+	}
+
+	dn = of_find_node_by_path("/ibm,dynamic-reconfiguration-memory");
+	if (!dn)
+		return -EINVAL;
+
+	prop = dlpar_clone_drconf_property(dn);
+	if (!prop) {
+		of_node_put(dn);
+		return -EINVAL;
+	}
+
+	p = prop->value;
+	entries = be32_to_cpu(*p++);
+	lmb = (struct of_drconf_cell *)p;

So if I'm reading this right the hp_elog either contains an index or a count of
LMBs to add. But it doesn't contain anything about which address ranges to add
or any of those details. That is all in the ibm,dynamic-memory property - but
how did it get in there?

The ibm,dynamic-memory property of the device tree is passed to the kernel
by phyp/qemu. The property is an array that associates each LMB with a
starting physical address and associativity.

quoted

+
+	for (i = 0; i < entries; i++, lmb++) {
+		u32 drc_index = be32_to_cpu(lmb->drc_index);
+
+		if (lmbs_to_add == lmbs_added)
+			break;
+
+		if (be32_to_cpu(lmb->flags) & DRCONF_MEM_ASSIGNED)
+			continue;
+
+		if (hp_elog->id_type == PSERIES_HP_ELOG_ID_DRC_INDEX
+		    && lmb->drc_index != hp_elog->_drc_u.drc_index)
+			continue;
+
+		rc = dlpar_acquire_drc(drc_index);
+		if (rc)
+			continue;
+
+		rc = dlpar_add_one_lmb(lmb);
+		if (rc) {
+			dlpar_release_drc(drc_index);
+			continue;
+		}

In both the above error cases you just move along. That means we potentially
hotplugged some memory but not everything that we were asked to. That seems
like a bad idea, we should either do everything or nothing.

That can be done, though will require some additional tracking.

The current hotplug handling for PowerVM make a best effort and tries to
add/remove as much of the requested memory as possible. I was going with
that same approach here, but have no problem moving to an all or nothing
approach.

We will need to keep track of the LMBs added/removed during a request so
we can return to the original state if the request cannot be satisfied.
 

quoted

+
+		lmb->flags |= cpu_to_be32(DRCONF_MEM_ASSIGNED);
+		lmbs_added++;
+		pr_info("Memory at %llx (drc index %x) has been hot-added\n",
+			be64_to_cpu(lmb->base_addr), drc_index);
+	}
+
+	if (lmbs_added)
+		rc = of_update_property(dn, prop);
+	else
+		dlpar_free_drconf_property(prop);

The value of rc here is not clear. It could be EINVAL or it could be the result
of the last dlpar_add_one_lmb(lmb). gcc would have told you that if you hadn't
initialised it.

quoted

+
+	of_node_put(dn);
+	return rc ? rc : lmbs_added;

This looks wrong.

Doesn't the rc eventually go back to dlpar_write(), which expects 0 for success?

That should show up as the write failing in userspace.

Based on previous comments I think the handling of rc will be updated so
we either return success or failure.

quoted

 int dlpar_memory(struct pseries_hp_errorlog *hp_elog)
 {
 	int rc = 0;

Don't initialise to zero, that way gcc can tell you if there's a path where you
forget to initialise it. It also means you can't accidentally return success.

quoted

+	if (hp_elog->id_type != PSERIES_HP_ELOG_ID_DRC_COUNT
+	    && hp_elog->id_type != PSERIES_HP_ELOG_ID_DRC_INDEX)
+		return -EINVAL;

This would look nicer as a switch I think.

can do.

-Nathan