Re: [PATCH 1/6 v5] Kernel DLPAR Infrastructure

[Nathan Fontenot <nfont@austin.ibm.com>]

Benjamin Herrenschmidt wrote:
> On Wed, 2009-10-28 at 15:53 -0500, Nathan Fontenot wrote:
>> This patch provides the kernel DLPAR infrastructure in a new filed named
>> dlpar.c.  The functionality provided is for acquiring and releasing a resource
>> from firmware and the parsing of information returned from the
>> ibm,configure-connector rtas call.  Additionally this exports the pSeries
>> reconfiguration notifier chain so that it can be invoked when device tree 
>> updates are made.
>>
>> Signed-off-by: Nathan Fontenot <nfont at austin.ibm.com> 
>> ---
> 
> Hi Nathan !
> 
> Finally I get to review this stuff :-)
> 

Thanks!

>> +#define CFG_CONN_WORK_SIZE	4096
>> +static char workarea[CFG_CONN_WORK_SIZE];
>> +static DEFINE_SPINLOCK(workarea_lock);
> 
> So I'm not a huge fan of this workarea static. First a static is in
> effect a global name (as far as System.map etc... are concerned) so it
> would warrant a better name. Then, do we really want that 4K of BSS
> taken even on platforms that don't do dlpar ? Any reason why you don't
> just pop a free page with __get_free_page() inside of
> configure_connector() ?
> 

I'm not either, having a static buffer and a lock feels like overkill
for this.  I tried kmalloc, but that didn't work.  I'll try using
__get_free_page.

>> +struct cc_workarea {
>> +	u32	drc_index;
>> +	u32	zero;
>> +	u32	name_offset;
>> +	u32	prop_length;
>> +	u32	prop_offset;
>> +};
>> +
>> +static struct property *parse_cc_property(char *workarea)
>> +{
>> +	struct property *prop;
>> +	struct cc_workarea *ccwa;
>> +	char *name;
>> +	char *value;
>> +
>> +	prop = kzalloc(sizeof(*prop), GFP_KERNEL);
>> +	if (!prop)
>> +		return NULL;
>> +
>> +	ccwa = (struct cc_workarea *)workarea;
>> +	name = workarea + ccwa->name_offset;
>> +	prop->name = kzalloc(strlen(name) + 1, GFP_KERNEL);
>> +	if (!prop->name) {
>> +		kfree(prop);
>> +		return NULL;
>> +	}
>> +
>> +	strcpy(prop->name, name);
>> +
>> +	prop->length = ccwa->prop_length;
>> +	value = workarea + ccwa->prop_offset;
>> +	prop->value = kzalloc(prop->length, GFP_KERNEL);
>> +	if (!prop->value) {
>> +		kfree(prop->name);
>> +		kfree(prop);
>> +		return NULL;
>> +	}
>> +
>> +	memcpy(prop->value, value, prop->length);
>> +	return prop;
>> +}
>> +
>> +static void free_property(struct property *prop)
>> +{
>> +	kfree(prop->name);
>> +	kfree(prop->value);
>> +	kfree(prop);
>> +}
>> +
>> +static struct device_node *parse_cc_node(char *work_area)
>> +{
> 
> const char* maybe ?

sure.

> 
>> +	struct device_node *dn;
>> +	struct cc_workarea *ccwa;
>> +	char *name;
>> +
>> +	dn = kzalloc(sizeof(*dn), GFP_KERNEL);
>> +	if (!dn)
>> +		return NULL;
>> +
>> +	ccwa = (struct cc_workarea *)work_area;
>> +	name = work_area + ccwa->name_offset;
> 
> I'm wondering whether work_area should be a struct cc_workarea * in the
> first place with a char data[] at the end, but that would mean probably
> tweaking the offsets... no big deal, up to you.
>

I'll look onto that.  Anything that makes this easier to understand is good.

 
>> +	dn->full_name = kzalloc(strlen(name) + 1, GFP_KERNEL);
>> +	if (!dn->full_name) {
>> +		kfree(dn);
>> +		return NULL;
>> +	}
>> +
>> +	strcpy(dn->full_name, name);
> 
> kstrdup ?

yep, should have used kstrdup.

> 
>  .../...
> 
>> +#define NEXT_SIBLING    1
>> +#define NEXT_CHILD      2
>> +#define NEXT_PROPERTY   3
>> +#define PREV_PARENT     4
>> +#define MORE_MEMORY     5
>> +#define CALL_AGAIN	-2
>> +#define ERR_CFG_USE     -9003
>> +
>> +struct device_node *configure_connector(u32 drc_index)
>> +{
> 
> It's a global exported function, I'd rather you call it
> dlpar_configure_connector()
>

ok.
 
>> +	struct device_node *dn;
>> +	struct device_node *first_dn = NULL;
>> +	struct device_node *last_dn = NULL;
>> +	struct property *property;
>> +	struct property *last_property = NULL;
>> +	struct cc_workarea *ccwa;
>> +	int cc_token;
>> +	int rc;
>> +
>> +	cc_token = rtas_token("ibm,configure-connector");
>> +	if (cc_token == RTAS_UNKNOWN_SERVICE)
>> +		return NULL;
>> +
>> +	spin_lock(&workarea_lock);
>> +
>> +	ccwa = (struct cc_workarea *)&workarea[0];
>> +	ccwa->drc_index = drc_index;
>> +	ccwa->zero = 0;
> 
> Popping a free page with gfp (or just kmalloc'ing 4K) would avoid the
> need for the lock too.

yes, see comment at beginning.

> 
>> +	rc = rtas_call(cc_token, 2, 1, NULL, workarea, NULL);
>> +	while (rc) {
>> +		switch (rc) {
>> +		case NEXT_SIBLING:
>> +			dn = parse_cc_node(workarea);
>> +			if (!dn)
>> +				goto cc_error;
>> +
>> +			dn->parent = last_dn->parent;
>> +			last_dn->sibling = dn;
>> +			last_dn = dn;
>> +			break;
>> +
>> +		case NEXT_CHILD:
>> +			dn = parse_cc_node(workarea);
>> +			if (!dn)
>> +				goto cc_error;
>> +
>> +			if (!first_dn)
>> +				first_dn = dn;
>> +			else {
>> +				dn->parent = last_dn;
>> +				if (last_dn)
>> +					last_dn->child = dn;
>> +			}
>> +
>> +			last_dn = dn;
>> +			break;
>> +
>> +		case NEXT_PROPERTY:
>> +			property = parse_cc_property(workarea);
>> +			if (!property)
>> +				goto cc_error;
>> +
>> +			if (!last_dn->properties)
>> +				last_dn->properties = property;
>> +			else
>> +				last_property->next = property;
>> +
>> +			last_property = property;
>> +			break;
>> +
>> +		case PREV_PARENT:
>> +			last_dn = last_dn->parent;
>> +			break;
>> +
>> +		case CALL_AGAIN:
>> +			break;
>> +
>> +		case MORE_MEMORY:
>> +		case ERR_CFG_USE:
>> +		default:
>> +			printk(KERN_ERR "Unexpected Error (%d) "
>> +			       "returned from configure-connector\n", rc);
>> +			goto cc_error;
>> +		}
>> +
>> +		rc = rtas_call(cc_token, 2, 1, NULL, workarea, NULL);
>> +	}
>> +
>> +	spin_unlock(&workarea_lock);
>> +	return first_dn;
>> +
>> +cc_error:
>> +	spin_unlock(&workarea_lock);
>> +
>> +	if (first_dn)
>> +		free_cc_nodes(first_dn);
>> +
>> +	return NULL;
>> +}
>> +
>> +static struct device_node *derive_parent(const char *path)
>> +{
>> +	struct device_node *parent;
>> +	char parent_path[128];
>> +	int parent_path_len;
>> +
>> +	parent_path_len = strrchr(path, '/') - path + 1;
>> +	strlcpy(parent_path, path, parent_path_len);
>> +
>> +	parent = of_find_node_by_path(parent_path);
>> +
>> +	return parent;
>> +}
> 
> This ...
> 
>> +static int add_one_node(struct device_node *dn)
>> +{
>> +	struct proc_dir_entry *ent;
>> +	int rc;
>> +
>> +	of_node_set_flag(dn, OF_DYNAMIC);
>> +	kref_init(&dn->kref);
>> +	dn->parent = derive_parent(dn->full_name);
>> +
>> +	rc = blocking_notifier_call_chain(&pSeries_reconfig_chain,
>> +					  PSERIES_RECONFIG_ADD, dn);
>> +	if (rc == NOTIFY_BAD) {
>> +		printk(KERN_ERR "Failed to add device node %s\n",
>> +		       dn->full_name);
>> +		return -ENOMEM; /* For now, safe to assume kmalloc failure */
>> +	}
>> +
>> +	of_attach_node(dn);
>> +
>> +#ifdef CONFIG_PROC_DEVICETREE
>> +	ent = proc_mkdir(strrchr(dn->full_name, '/') + 1, dn->parent->pde);
>> +	if (ent)
>> +		proc_device_tree_add_node(dn, ent);
>> +#endif
>> +
>> +	of_node_put(dn->parent);
>> +	return 0;
>> +}
> 
>  ... and this ...
> 
>> +int add_device_tree_nodes(struct device_node *dn)
>> +{
>> +	struct device_node *child = dn->child;
>> +	struct device_node *sibling = dn->sibling;
>> +	int rc;
>> +
>> +	dn->child = NULL;
>> +	dn->sibling = NULL;
>> +	dn->parent = NULL;
>> +
>> +	rc = add_one_node(dn);
>> +	if (rc)
>> +		return rc;
>> +
>> +	if (child) {
>> +		rc = add_device_tree_nodes(child);
>> +		if (rc)
>> +			return rc;
>> +	}
>> +
>> +	if (sibling)
>> +		rc = add_device_tree_nodes(sibling);
>> +
>> +	return rc;
>> +}
> 
>  ... and this ...
> 
>> +static int remove_one_node(struct device_node *dn)
>> +{
>> +	struct device_node *parent = dn->parent;
>> +	struct property *prop = dn->properties;
>> +
>> +#ifdef CONFIG_PROC_DEVICETREE
>> +	while (prop) {
>> +		remove_proc_entry(prop->name, dn->pde);
>> +		prop = prop->next;
>> +	}
>> +
>> +	if (dn->pde)
>> +		remove_proc_entry(dn->pde->name, parent->pde);
>> +#endif
>> +
>> +	blocking_notifier_call_chain(&pSeries_reconfig_chain,
>> +			    PSERIES_RECONFIG_REMOVE, dn);
>> +	of_detach_node(dn);
>> +	of_node_put(dn); /* Must decrement the refcount */
>> +
>> +	return 0;
>> +}
> 
>  ... and this ...
> 
>> +static int _remove_device_tree_nodes(struct device_node *dn)
>> +{
>> +	int rc;
>> +
>> +	if (dn->child) {
>> +		rc = _remove_device_tree_nodes(dn->child);
>> +		if (rc)
>> +			return rc;
>> +	}
>> +
>> +	if (dn->sibling) {
>> +		rc = _remove_device_tree_nodes(dn->sibling);
>> +		if (rc)
>> +			return rc;
>> +	}
>> +
>> +	rc = remove_one_node(dn);
>> +	return rc;
>> +}
> 
>  ... repeat myself ...
> 
>> +int remove_device_tree_nodes(struct device_node *dn)
>> +{
>> +	int rc;
>> +
>> +	if (dn->child) {
>> +		rc = _remove_device_tree_nodes(dn->child);
>> +		if (rc)
>> +			return rc;
>> +	}
>> +
>> +	rc = remove_one_node(dn);
>> +	return rc;
>> +}
> 
>  ... should probably all go to something like drivers/of/dynamic.c or at
> least for now arch/powerpc/kernel/of_dynamic.c along with everything
> related to dynamically adding and removing nodes. I see that potentially
> useful for more than just DLPAR (though DLPAR is the only user right
> now) and should also all be prefixed with of_*

I agree, there should be at least a powerpc generic implementation of these
routines.  The reason I put them here is that I am doing some oddities with
the next, child, and sibling pointers since they point to items not yet in
the device tree.

I saw that Grant Likely is doing updates to all of the of_* stuff right now,
would it be ok to have these routines here, renamed as dlpar_*, and look
to merge them in with Grant's updates when he finishes?
  
> 
>> +#define DR_ENTITY_SENSE		9003
>> +#define DR_ENTITY_PRESENT	1
>> +#define DR_ENTITY_UNUSABLE	2
>> +#define ALLOCATION_STATE	9003
>> +#define ALLOC_UNUSABLE		0
>> +#define ALLOC_USABLE		1
>> +#define ISOLATION_STATE		9001
>> +#define ISOLATE			0
>> +#define UNISOLATE		1
>> +
>> +int acquire_drc(u32 drc_index)
>> +{
>> +	int dr_status, rc;
>> +
>> +	rc = rtas_call(rtas_token("get-sensor-state"), 2, 2, &dr_status,
>> +		       DR_ENTITY_SENSE, drc_index);
>> +	if (rc || dr_status != DR_ENTITY_UNUSABLE)
>> +		return -1;
>> +
>> +	rc = rtas_set_indicator(ALLOCATION_STATE, drc_index, ALLOC_USABLE);
>> +	if (rc)
>> +		return rc;
>> +
>> +	rc = rtas_set_indicator(ISOLATION_STATE, drc_index, UNISOLATE);
>> +	if (rc) {
>> +		rtas_set_indicator(ALLOCATION_STATE, drc_index, ALLOC_UNUSABLE);
>> +		return rc;
>> +	}
>> +
>> +	return 0;
>> +}
>> +
>> +int release_drc(u32 drc_index)
>> +{
>> +	int dr_status, rc;
>> +
>> +	rc = rtas_call(rtas_token("get-sensor-state"), 2, 2, &dr_status,
>> +		       DR_ENTITY_SENSE, drc_index);
>> +	if (rc || dr_status != DR_ENTITY_PRESENT)
>> +		return -1;
>> +
>> +	rc = rtas_set_indicator(ISOLATION_STATE, drc_index, ISOLATE);
>> +	if (rc)
>> +		return rc;
>> +
>> +	rc = rtas_set_indicator(ALLOCATION_STATE, drc_index, ALLOC_UNUSABLE);
>> +	if (rc) {
>> +		rtas_set_indicator(ISOLATION_STATE, drc_index, UNISOLATE);
>> +		return rc;
>> +	}
>> +
>> +	return 0;
>> +}
> 
> Both above should have a dlpar_* prefix

will do.

> 
>> +static int pseries_dlpar_init(void)
>> +{
>> +	if (!machine_is(pseries))
>> +		return 0;
>> +
>> +	return 0;
>> +}
>> +device_initcall(pseries_dlpar_init);
> 
> What the point ? :-)

Yeah, its a bit odd looking but later patches actually add code to the init routine
to set up memory probe/release and cpu probe/release handlers.

I'll look to add ifdef's around the initcall for cases where no work is to be done.

-Nathan Fontenot

> 
> Cheers
> Ben.
>