Skip to contentSkip to conversation

powerpc/memtrace: Don't offline memory blocks via offline_pages()

RFC PATCH In review David Hildenbrand 1 participant 5 messages Dec 17, 2019

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

cover powerpc/memtrace: Don't offline memory blocks via offline_pages()

This RFC is based on linux-next and
- 2 patches from "PATCH RFC v4 00/13] virtio-mem: paravirtualized memory" -> "mm: Allow to offline unmovable PageOffline() pages via

     MEM_GOING_OFFLINE" [1]

-> "mm/memory_hotplug: Introduce offline_and_remove_memory()" [2]
- "mm/memory_hotplug: Don't free usage map when removing a re-added early section" [3]

A branch with all patches (kept updated) is available at:
https://github.com/davidhildenbrand/linux.git memtrace

Stop using offline_pages() to offline memory blocks. Allocate the memory blocks using alloc_contig_pages() first and offline+remove the allcoated memory blocks using a clean MM interface. Offlining of allocated memory is made possible by using PageOffline() in combination with a memory notifier (similar to virto-mem).

Note: In the future, we might want to switch to only removing/readding the page tables of the allocated memory (while still marking it PageOffline()). However, that might have other implications, and requires work from PPC people (IOW, I won't fiddle with that :) ).

[1] https://lkml.kernel.org/r/[email protected]
[2] https://lkml.kernel.org/r/[email protected]
[3] https://lkml.kernel.org/r/[email protected]

David Hildenbrand (3):
powerpc/memtrace: Enforce power of 2 for memory buffer size powerpc/memtrace: Factor out readding memory into memtrace_free_node() powerpc/memtrace: Don't offline memory blocks via offline_pages()

 arch/powerpc/platforms/powernv/Kconfig    |   1 +
 arch/powerpc/platforms/powernv/memtrace.c | 217 ++++++++++++++--------
 2 files changed, 136 insertions(+), 82 deletions(-)
Signature · 1 line
2.23.0
#1 lore raw

RFC PATCH 1/3 powerpc/memtrace: Enforce power of 2 for memory buffer size

The code mentions "Trace memory needs to be aligned to the size", and e.g., round_up() is documented to work on power of 2 only. Also, the whole search is not optimized e.g., for being aligned to memory block size only while allocating multiple memory blocks.

Let's just limit to powers of 2 that are at least the size of memory blocks - the granularity we are using for alloc/offline/unplug.

Cc Benjamin Herrenschmidt <[email protected]>
Cc Paul Mackerras <[email protected]>
Cc Michael Ellerman <[email protected]>
Cc Andrew Morton <[email protected]>
Cc David Hildenbrand <[email protected]>
Cc Allison Randal <[email protected]>
Cc Anshuman Khandual <[email protected]>
Cc Balbir Singh <[email protected]>
Cc Rashmica Gupta <[email protected]>
---
 arch/powerpc/platforms/powernv/memtrace.c | 14 +++++---------
 1 file changed, 5 insertions(+), 9 deletions(-)
arch/powerpc/platforms/powernv/memtrace.c +5 -9
index eb2e75dac369..0c4c54d2e3c4 100644
@@ -268,15 +268,11 @@ static int memtrace_online(void)
 
 static int memtrace_enable_set(void *data, u64 val)
 {
-	u64 bytes;
-
-	/*
-	 * Don't attempt to do anything if size isn't aligned to a memory
-	 * block or equal to zero.
-	 */
-	bytes = memory_block_size_bytes();
-	if (val & (bytes - 1)) {
-		pr_err("Value must be aligned with 0x%llx\n", bytes);
+	const unsigned long bytes = memory_block_size_bytes();
+
+	if (val && (!is_power_of_2(val) || val < bytes)) {
+		pr_err("Value must be 0 or a power of 2 (at least 0x%lx)\n",
+		       bytes);
 		return -EINVAL;
 	}
 
Signature · 1 line
2.23.0
#2 lore raw

RFC PATCH 2/3 powerpc/memtrace: Factor out readding memory into memtrace_free_node()

In reply to David Hildenbrand

While at it, move it, we want to reuse it soon.

Cc Benjamin Herrenschmidt <[email protected]>
Cc Paul Mackerras <[email protected]>
Cc Michael Ellerman <[email protected]>
Cc Andrew Morton <[email protected]>
Cc David Hildenbrand <[email protected]>
Cc Allison Randal <[email protected]>
Cc Jens Axboe <[email protected]>
Cc Anshuman Khandual <[email protected]>
Cc Thomas Gleixner <[email protected]>
Cc Balbir Singh <[email protected]>
Cc Rashmica Gupta <[email protected]>
---
 arch/powerpc/platforms/powernv/memtrace.c | 44 ++++++++++++++---------
 1 file changed, 27 insertions(+), 17 deletions(-)
arch/powerpc/platforms/powernv/memtrace.c +27 -17
index 0c4c54d2e3c4..2d2a0a2acd60 100644
@@ -50,6 +50,32 @@ static const struct file_operations memtrace_fops = {
 	.open	= simple_open,
 };
 
+static int online_mem_block(struct memory_block *mem, void *arg)
+{
+	return device_online(&mem->dev);
+}
+
+static int memtrace_free_node(int nid, unsigned long start, unsigned long size)
+{
+	int ret;
+
+	ret = add_memory(nid, start, size);
+	if (!ret) {
+		/*
+		 * If the kernel isn't compiled with the auto online option, we
+		 * will try to online ourselves. We'll ignore any errors here -
+		 * user space can try to online itself later (after all, the
+		 * memory was added successfully).
+		 */
+		if (!memhp_auto_online) {
+			lock_device_hotplug();
+			walk_memory_blocks(start, size, NULL, online_mem_block);
+			unlock_device_hotplug();
+		}
+	}
+	return ret;
+}
+
 static int check_memblock_online(struct memory_block *mem, void *arg)
 {
 	if (mem->state != MEM_ONLINE)
@@ -202,11 +228,6 @@ static int memtrace_init_debugfs(void)
 	return ret;
 }
 
-static int online_mem_block(struct memory_block *mem, void *arg)
-{
-	return device_online(&mem->dev);
-}
-
 /*
  * Iterate through the chunks of memory we have removed from the kernel
  * and attempt to add them back to the kernel.
@@ -229,24 +250,13 @@ static int memtrace_online(void)
 			ent->mem = 0;
 		}
 
-		if (add_memory(ent->nid, ent->start, ent->size)) {
+		if (memtrace_free_node(ent->nid, ent->start, ent->size)) {
 			pr_err("Failed to add trace memory to node %d\n",
 				ent->nid);
 			ret += 1;
 			continue;
 		}
 
-		/*
-		 * If kernel isn't compiled with the auto online option
-		 * we need to online the memory ourselves.
-		 */
-		if (!memhp_auto_online) {
-			lock_device_hotplug();
-			walk_memory_blocks(ent->start, ent->size, NULL,
-					   online_mem_block);
-			unlock_device_hotplug();
-		}
-
 		/*
 		 * Memory was added successfully so clean up references to it
 		 * so on reentry we can tell that this chunk was added.
Signature · 1 line
2.23.0
#3 lore raw

RFC PATCH 3/3 powerpc/memtrace: Don't offline memory blocks via offline_pages()

In reply to David Hildenbrand

offline_pages() should not be called outside of the MM core. Especially, having to manually fiddle with the memory block states is a sign that this is not a good idea. To offline memory block devices cleanly, device_offline() should be used. This is the only remaining caller of offline_pages(), except the official device_offline() way.

E.g., when trying to allocate right now we trigger messages like

[   11.227817] page:c00c000000182000 refcount:1 mapcount:0 mapping:0000000000000000 index:0x0
[   11.228056] raw: 007ffff000000000 c000000001538860 c000000001538860 0000000000000000
[   11.228070] raw: 0000000000000000 0000000000000001 00000001ffffffff 0000000000000000
[   11.228097] page dumped because: unmovable page

and theoretically we might end up looping quite a long time trying to offline memory, which would have to be canceled by the user manually (CTRL-C).

Memtrace needs to identify+allocate multiple consecutive memory blocks. It also has to remove the memory blocks to remove all page tables (HW requirement).

Let's use alloc_contig_pages() to allocate memory that spans multiple memory block devices. We can then set all pages PageOffline() to allow these pages to get isolated. A temporary memory notifier can then make offlining of these pages succeed by dropping its reference to the pages on MEM_GOING_OFFLINE events(as documented in include/linux/page-flags.h for PageOffline() pages). Error handling is a bit tricky.

Note1: ZONE_MOVABLE memory blocks won't be considered. Not sure if that was ever really relevant. (unmovable data would end up on these memory blocks for a tiny little time frame)

Note2: We don't have to care about online_page_callback_t, as we forbid re-onlining from our memory notifier.

Note3: I was told this feature is never used along with DIMM-based memory hotunplug - otherwise bad things will happen when a DIMM would try to remove "alread-removed" memory (that is still in use).

Tested under QEMU with powernv emulation (kernel + initramfs).

$ mount -t debugfs none /sys/kernel/debug/
$ cat /sys/devices/system/memory/block_size_bytes
10000000
$ echo 0x20000000 > /sys/kernel/debug/powerpc/memtrace/enable
[   19.809790] Offlined Pages 4096
[   19.835842] Offlined Pages 4096
[   19.853136] memtrace: Allocated trace memory on node 0 at 0x0000000040000000

Unfortunately, QEMU does not support NUMA for powernv yet, so I cannot test that.

Cc Benjamin Herrenschmidt <[email protected]>
Cc Paul Mackerras <[email protected]>
Cc Michael Ellerman <[email protected]>
Cc Andrew Morton <[email protected]>
Cc David Hildenbrand <[email protected]>
Cc Allison Randal <[email protected]>
Cc Jens Axboe <[email protected]>
Cc Anshuman Khandual <[email protected]>
Cc Thomas Gleixner <[email protected]>
Cc Michal Hocko <[email protected]>
Cc Oscar Salvador <[email protected]>
Cc Balbir Singh <[email protected]>
Cc Rashmica Gupta <[email protected]>
---
 arch/powerpc/platforms/powernv/Kconfig    |   1 +
 arch/powerpc/platforms/powernv/memtrace.c | 175 ++++++++++++++--------
 2 files changed, 112 insertions(+), 64 deletions(-)
arch/powerpc/platforms/powernv/Kconfig +1 -0
index 938803eab0ad..571a0fa9f055 100644
@@ -29,6 +29,7 @@ config OPAL_PRD
 config PPC_MEMTRACE
 	bool "Enable removal of RAM from kernel mappings for tracing"
 	depends on PPC_POWERNV && MEMORY_HOTREMOVE
+	select CONTIG_ALLOC
 	help
 	  Enabling this option allows for the removal of memory (RAM)
 	  from the kernel mappings to be used for hardware tracing.
arch/powerpc/platforms/powernv/memtrace.c +111 -64

194 lines not on this page · raw message

Signature · 1 line
2.23.0
#4 lore raw
In reply to David Hildenbrand
Quoted 27 lines · David Hildenbrand

This RFC is based on linux-next and
- 2 patches from "PATCH RFC v4 00/13] virtio-mem: paravirtualized memory" -> "mm: Allow to offline unmovable PageOffline() pages via

     MEM_GOING_OFFLINE" [1]

-> "mm/memory_hotplug: Introduce offline_and_remove_memory()" [2]
- "mm/memory_hotplug: Don't free usage map when removing a re-added early section" [3]

A branch with all patches (kept updated) is available at:
https://github.com/davidhildenbrand/linux.git memtrace

Stop using offline_pages() to offline memory blocks. Allocate the memory blocks using alloc_contig_pages() first and offline+remove the allcoated memory blocks using a clean MM interface. Offlining of allocated memory is made possible by using PageOffline() in combination with a memory notifier (similar to virto-mem).

Note: In the future, we might want to switch to only removing/readding the page tables of the allocated memory (while still marking it PageOffline()). However, that might have other implications, and requires work from PPC people (IOW, I won't fiddle with that :) ).

[1] https://lkml.kernel.org/r/[email protected]
[2] https://lkml.kernel.org/r/[email protected]
[3] https://lkml.kernel.org/r/[email protected]

Quoted 9 lines

David Hildenbrand (3):
powerpc/memtrace: Enforce power of 2 for memory buffer size powerpc/memtrace: Factor out readding memory into memtrace_free_node() powerpc/memtrace: Don't offline memory blocks via offline_pages()

 arch/powerpc/platforms/powernv/Kconfig    |   1 +
 arch/powerpc/platforms/powernv/memtrace.c | 217 ++++++++++++++--------
 2 files changed, 136 insertions(+), 82 deletions(-)

(CC linuxppc-dev on the cover letter, my fancy sendmail cc-cmd.sh script missed it, sorry)

Signature · 3 lines
Thanks,

David / dhildenb