Skip to contentSkip to conversation

drivers/base/memory.c: memory subsys init: skip search for missing blocks

PATCH In review Scott Cheloha 3 participants 7 messages Nov 1, 2019 – Nov 2, 2019

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 drivers/base/memory.c: memory subsys init: skip search for missing blocks

Add a flag to init_memory_block() to enable/disable searching for a matching block before creating a device for the block and adding it to the memory subsystem's bus.

When the memory subsystem is being initialized there is no need to check if a given block has already been added to its bus. The bus is new, so the block in question cannot yet have a corresponding device.

The search for a missing block is O(n) so this saves substantial time at boot if there are many such blocks to add.

---
 drivers/base/memory.c | 18 +++++++++++-------
 1 file changed, 11 insertions(+), 7 deletions(-)
drivers/base/memory.c +11 -7
index 55907c27075b..1160df4a8feb 100644
@@ -643,17 +643,21 @@ int register_memory(struct memory_block *memory)
 }
 
 static int init_memory_block(struct memory_block **memory,
-			     unsigned long block_id, unsigned long state)
+			     unsigned long block_id, unsigned long state,
+			     bool may_exist)
 {
 	struct memory_block *mem;
 	unsigned long start_pfn;
 	int ret = 0;
 
-	mem = find_memory_block_by_id(block_id);
-	if (mem) {
-		put_device(&mem->dev);
-		return -EEXIST;
+	if (may_exist) {
+		mem = find_memory_block_by_id(block_id);
+		if (mem) {
+			put_device(&mem->dev);
+			return -EEXIST;
+		}
 	}
+
 	mem = kzalloc(sizeof(*mem), GFP_KERNEL);
 	if (!mem)
 		return -ENOMEM;
@@ -684,7 +688,7 @@ static int add_memory_block(unsigned long base_section_nr)
 	if (section_count == 0)
 		return 0;
 	ret = init_memory_block(&mem, base_memory_block_id(base_section_nr),
-				MEM_ONLINE);
+				MEM_ONLINE, false);
 	if (ret)
 		return ret;
 	mem->section_count = section_count;
@@ -720,7 +724,7 @@ int create_memory_block_devices(unsigned long start, unsigned long size)
 
 	mutex_lock(&mem_sysfs_mutex);
 	for (block_id = start_block_id; block_id != end_block_id; block_id++) {
-		ret = init_memory_block(&mem, block_id, MEM_OFFLINE);
+		ret = init_memory_block(&mem, block_id, MEM_OFFLINE, true);
 		if (ret)
 			break;
 		mem->section_count = sections_per_block;
Signature · 1 line
2.24.0.rc1
#1 lore raw
Quoted 32 lines · Scott Cheloha

Add a flag to init_memory_block() to enable/disable searching for a matching block before creating a device for the block and adding it to the memory subsystem's bus.

When the memory subsystem is being initialized there is no need to check if a given block has already been added to its bus. The bus is new, so the block in question cannot yet have a corresponding device.

The search for a missing block is O(n) so this saves substantial time at boot if there are many such blocks to add.

---
  drivers/base/memory.c | 18 +++++++++++-------
  1 file changed, 11 insertions(+), 7 deletions(-)
drivers/base/memory.c +2 -1
index 55907c27075b..1160df4a8feb 100644
@@ -643,17 +643,21 @@ int register_memory(struct memory_block *memory)
  }
 
  static int init_memory_block(struct memory_block **memory,
-			     unsigned long block_id, unsigned long state)
+			     unsigned long block_id, unsigned long state,
+			     bool may_exist)
  {
  	struct memory_block *mem;
  	unsigned long start_pfn;
  	int ret = 0;
 

drivers/base/memory.c int register_memory(struct memory_block *memory) · line 658

Quoted 10 lines
drivers/base/memory.c +6 -4
-	mem = find_memory_block_by_id(block_id);
-	if (mem) {
-		put_device(&mem->dev);
-		return -EEXIST;
+	if (may_exist) {
+		mem = find_memory_block_by_id(block_id);
+		if (mem) {
+			put_device(&mem->dev);
+			return -EEXIST;
+		}

No, I don't really like that. Can we please speed up the lookup via a radix tree as noted in the comment of "find_memory_block()".

/*
* For now, we have a linear search to go find the appropriate
* memory_block corresponding to a particular phys_index. If
* this gets to be a real problem, we can always use a radix
* tree or something here.
*
* This could be made generic for all device subsystems. */

This will speed up all users of walk_memory_blocks() similarly.
Especially, it will also speed up link_mem_sections() used during boot, which relies on walk_memory_blocks().

Signature · 3 lines
Thanks,

David / dhildenb
#2 lore raw
Quoted 56 lines · David Hildenbrand
Quoted 32 lines · Scott Cheloha

Add a flag to init_memory_block() to enable/disable searching for a matching block before creating a device for the block and adding it to the memory subsystem's bus.

When the memory subsystem is being initialized there is no need to check if a given block has already been added to its bus. The bus is new, so the block in question cannot yet have a corresponding device.

The search for a missing block is O(n) so this saves substantial time at boot if there are many such blocks to add.

---
   drivers/base/memory.c | 18 +++++++++++-------
   1 file changed, 11 insertions(+), 7 deletions(-)
drivers/base/memory.c +2 -1
index 55907c27075b..1160df4a8feb 100644
@@ -643,17 +643,21 @@ int register_memory(struct memory_block *memory)
   }
 
   static int init_memory_block(struct memory_block **memory,
-			     unsigned long block_id, unsigned long state)
+			     unsigned long block_id, unsigned long state,
+			     bool may_exist)
   {
   	struct memory_block *mem;
   	unsigned long start_pfn;
   	int ret = 0;
 

drivers/base/memory.c int register_memory(struct memory_block *memory) · line 658

Quoted 10 lines
drivers/base/memory.c +6 -4
-	mem = find_memory_block_by_id(block_id);
-	if (mem) {
-		put_device(&mem->dev);
-		return -EEXIST;
+	if (may_exist) {
+		mem = find_memory_block_by_id(block_id);
+		if (mem) {
+			put_device(&mem->dev);
+			return -EEXIST;
+		}

No, I don't really like that. Can we please speed up the lookup via a radix tree as noted in the comment of "find_memory_block()".

/*
* For now, we have a linear search to go find the appropriate
* memory_block corresponding to a particular phys_index. If
* this gets to be a real problem, we can always use a radix
* tree or something here.
*
* This could be made generic for all device subsystems. */

Quoted 4 lines

This will speed up all users of walk_memory_blocks() similarly. Especially, it will also speed up link_mem_sections() used during boot, which relies on walk_memory_blocks().

Also, I think this check that I added back than should not happen in the wild. If we care, we could turn that into a VM_BUG_ON(find_memory_block_by_id(block_id)).

Signature · 3 lines
Thanks,

David / dhildenb
#3 lore raw
In reply to David Hildenbrand
Quoted 2 lines · David Hildenbrand

No, I don't really like that. Can we please speed up the lookup via a radix tree as noted in the comment of "find_memory_block()".

I agree with the general sentiment that a redesign is the correct long term action - it has been for some time now - but implementing a new storage and retrieval method and verifying that it introduces no new problems itself is non-trivial. There's a reason it remains a comment.

I don't see any issues with the patch itself. Do we really want to forego the short term, low-hanging, low risk fruit in favor of waiting indefinitely for that well-tested long-term solution?

Rick

#4 lore raw
Quoted 10 lines

Am 01.11.2019 um 23:32 schrieb Rick Lindsley <[email protected]>:

Quoted 1 line · David Hildenbrand

No, I don't really like that. Can we please speed up the lookup via a radix tree as noted in the comment of "find_memory_block()".

I agree with the general sentiment that a redesign is the correct long term action - it has been for some time now - but implementing a new storage and retrieval method and verifying that it introduces no new problems itself is non-trivial. There's a reason it remains a comment.

I don't see any issues with the patch itself. Do we really want to forego the short term, low-hanging, low risk fruit in favor of waiting indefinitely for that well-tested long-term solution?

The low hanging fruit for me is to convert it to a simple VM_BUG_ON(). As I said, this should never really happen with current code.

Also, I don‘t think adding a radix tree here is rocket science and takes indefinitely ;) feel free to prove me wrong.

Quoted 2 lines

Rick

#5 lore raw
Quoted 6 lines · David Hildenbrand
Quoted 6 lines

Am 01.11.2019 um 23:32 schrieb Rick Lindsley <[email protected]>:

Quoted 1 line · David Hildenbrand

No, I don't really like that. Can we please speed up the lookup via a radix tree as noted in the comment of "find_memory_block()".

Quoted 7 lines
Quoted 3 lines

I agree with the general sentiment that a redesign is the correct long term action - it has been for some time now - but implementing a new storage and retrieval method and verifying that it introduces no new problems itself is non-trivial. There's a reason it remains a comment.

I don't see any issues with the patch itself. Do we really want to forego the short term, low-hanging, low risk fruit in favor of waiting indefinitely for that well-tested long-term solution?

The low hanging fruit for me is to convert it to a simple VM_BUG_ON(). As I said, this should never really happen with current code.

Also, I don‘t think adding a radix tree here is rocket science and takes indefinitely ;) feel free to prove me wrong.

To clarify the goal here, "adding a radix tree" means changing subsys_private's klist_devices member from a klist to a radix tree or xarray, right?

-Scott

#6 lore raw
Quoted 15 lines

Am 02.11.2019 um 20:43 schrieb Scott Cheloha <[email protected]>:

Quoted 10 lines · David Hildenbrand
Quoted 9 lines
Quoted 2 lines

Am 01.11.2019 um 23:32 schrieb Rick Lindsley <[email protected]>:

Quoted 1 line · David Hildenbrand

No, I don't really like that. Can we please speed up the lookup via a radix tree as noted in the comment of "find_memory_block()".

I agree with the general sentiment that a redesign is the correct long term action - it has been for some time now - but implementing a new storage and retrieval method and verifying that it introduces no new problems itself is non-trivial. There's a reason it remains a comment.

I don't see any issues with the patch itself. Do we really want to forego the short term, low-hanging, low risk fruit in favor of waiting indefinitely for that well-tested long-term solution?

Quoted 7 lines
Quoted 3 lines

The low hanging fruit for me is to convert it to a simple VM_BUG_ON(). As I said, this should never really happen with current code.

Also, I don‘t think adding a radix tree here is rocket science and takes indefinitely ;) feel free to prove me wrong.

To clarify the goal here, "adding a radix tree" means changing subsys_private's klist_devices member from a klist to a radix tree or xarray, right?

I wouldn‘t go that far and only use a subsystem local data structure as a fast lookup cache. The memory subsystem is one of the rare subsystems that deals with such a big number of devices (AFAIK). Most other subsystems don‘t really need that.

I do agree that converting the klist to a radix tree would be more involved, but at least I think we can keep this subsystem-local, at least for now. Introducing a local cache should be simple.

Cheers!

Quoted 2 lines

-Scott