Skip to contentSkip to conversation

memblocks: Move late alloc warning down to phys alloc

PATCH Superseded James Gowans 2 participants 4 messages Jun 14, 2024 – Jun 19, 2024

Subsystems (2)

v1 of 2 Superseded Latest is v2 →

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 memblocks: Move late alloc warning down to phys alloc

If a driver/subsystem tries to do an allocation after memblocks have been freed and the memory handed to the buddy allocator, it will not actually be legal to use that allocation - the buddy allocator owns the memory. This is handled by the memblocks function which does allocations and returns virtual addresses by printing a warning and doing a kmalloc instead. However, the physical allocation function does not to do this check - callers of the physical alloc function are unprotected against mis-use.

Improve the error catching here by moving the check into the physical allocation function which is used by the virtual addr allocation function.

Cc Mike Rapoport <[email protected]>
Cc Andrew Morton <[email protected]>
Cc Alex Graf <[email protected]>
---
 mm/memblock.c | 18 +++++++++++-------
 1 file changed, 11 insertions(+), 7 deletions(-)
mm/memblock.c +11 -7
index d09136e040d3..dd4f237dc1fc 100644
@@ -1457,6 +1457,17 @@ phys_addr_t __init memblock_alloc_range_nid(phys_addr_t size,
 		align = SMP_CACHE_BYTES;
 	}
 
+	/*
+	 * Detect any accidental use of these APIs after slab is ready, as at
+	 * this moment memblock may be deinitialized already and its
+	 * internal data may be destroyed (after execution of memblock_free_all)
+	 */
+	if (WARN_ON_ONCE(slab_is_available())) {
+		void *vaddr = kzalloc_node(size, GFP_NOWAIT, nid);
+
+		return vaddr ? virt_to_phys(vaddr) : 0;
+	}
+
 again:
 	found = memblock_find_in_range_node(size, align, start, end, nid,
 					    flags);
@@ -1576,13 +1587,6 @@ static void * __init memblock_alloc_internal(
 {
 	phys_addr_t alloc;
 
-	/*
-	 * Detect any accidental use of these APIs after slab is ready, as at
-	 * this moment memblock may be deinitialized already and its
-	 * internal data may be destroyed (after execution of memblock_free_all)
-	 */
-	if (WARN_ON_ONCE(slab_is_available()))
-		return kzalloc_node(size, GFP_NOWAIT, nid);
 
 	if (max_addr > memblock.current_limit)
 		max_addr = memblock.current_limit;
Signature · 1 line
2.34.1
maintainer
#1 lore raw
Quoted 1 line · James Gowans

Subject: [PATCH] memblocks: Move late alloc warning down to phys alloc

Nit: memblock

Quoted 8 lines

If a driver/subsystem tries to do an allocation after memblocks have been freed and the memory handed to the buddy allocator, it will not actually be legal to use that allocation - the buddy allocator owns the memory. This is handled by the memblocks function which does allocations and returns virtual addresses by printing a warning and doing a kmalloc instead. However, the physical allocation function does not to do this check - callers of the physical alloc function are unprotected against mis-use.

Did you see such misuse or this is a theoretical issue?

Quoted 20 lines

Improve the error catching here by moving the check into the physical allocation function which is used by the virtual addr allocation function.

Cc Mike Rapoport <[email protected]>
Cc Andrew Morton <[email protected]>
Cc Alex Graf <[email protected]>
---
 mm/memblock.c | 18 +++++++++++-------
 1 file changed, 11 insertions(+), 7 deletions(-)
mm/memblock.c +0 -0
index d09136e040d3..dd4f237dc1fc 100644
@@ -1457,6 +1457,17 @@ phys_addr_t __init memblock_alloc_range_nid(phys_addr_t size,
 		align = SMP_CACHE_BYTES;
 	}
 

mm/memblock.c phys_addr_t __init memblock_alloc_range_nid(phys_addr_t size, · line 1469

Quoted 10 lines
mm/memblock.c +10 -0
+	/*
+	 * Detect any accidental use of these APIs after slab is ready, as at
+	 * this moment memblock may be deinitialized already and its
+	 * internal data may be destroyed (after execution of memblock_free_all)
+	 */
+	if (WARN_ON_ONCE(slab_is_available())) {
+		void *vaddr = kzalloc_node(size, GFP_NOWAIT, nid);
+
+		return vaddr ? virt_to_phys(vaddr) : 0;
+	}

I'd move this before alignment check.

Quoted 21 lines
mm/memblock.c +1 -7
+
 again:
 	found = memblock_find_in_range_node(size, align, start, end, nid,
 					    flags);
@@ -1576,13 +1587,6 @@ static void * __init memblock_alloc_internal(
 {
 	phys_addr_t alloc;
 
-	/*
-	 * Detect any accidental use of these APIs after slab is ready, as at
-	 * this moment memblock may be deinitialized already and its
-	 * internal data may be destroyed (after execution of memblock_free_all)
-	 */
-	if (WARN_ON_ONCE(slab_is_available()))
-		return kzalloc_node(size, GFP_NOWAIT, nid);
 
 	if (max_addr > memblock.current_limit)
 		max_addr = memblock.current_limit;
Signature · 1 line
2.34.1
Signature · 2 lines
Sincerely yours,
Mike.
#2 lore raw
Quoted 4 lines · Mike Rapoport
Quoted 1 line · James Gowans

Subject: [PATCH] memblocks: Move late alloc warning down to phys alloc

Nit: memblock

Ack!

Quoted 11 lines
Quoted 9 lines

If a driver/subsystem tries to do an allocation after memblocks have been freed and the memory handed to the buddy allocator, it will not actually be legal to use that allocation - the buddy allocator owns the memory. This is handled by the memblocks function which does allocations and returns virtual addresses by printing a warning and doing a kmalloc instead. However, the physical allocation function does not to do this check - callers of the physical alloc function are unprotected against mis-use.

Did you see such misuse or this is a theoretical issue?

Yeah, I was driving the memblock allocator badly when prototyping something. Allocating a large contiguous block of memory for an in- memory filesystem and I was doing the allocation in an initcall, but by that point it was too late. The memblock allocator happily gave me a large chunk of memory, but it was already in use by the buddy allocator, so ended up with memory corruption. Oops! Getting this warning would have made the problem immediately obvious.

Quoted 21 lines
Quoted 21 lines

Improve the error catching here by moving the check into the physical allocation function which is used by the virtual addr allocation function.

Cc Mike Rapoport <[email protected]>
Cc Andrew Morton <[email protected]>
Cc Alex Graf <[email protected]>
---
 mm/memblock.c | 18 +++++++++++-------
 1 file changed, 11 insertions(+), 7 deletions(-)
mm/memblock.c +0 -0
index d09136e040d3..dd4f237dc1fc 100644
@@ -1457,6 +1457,17 @@ phys_addr_t __init memblock_alloc_range_nid(phys_addr_t size,
              align = SMP_CACHE_BYTES;
      }
 
Quoted 12 lines

mm/memblock.c phys_addr_t __init memblock_alloc_range_nid(phys_addr_t size, · line 1469

Quoted 10 lines
mm/memblock.c +10 -0
+     /*
+      * Detect any accidental use of these APIs after slab is ready, as at
+      * this moment memblock may be deinitialized already and its
+      * internal data may be destroyed (after execution of memblock_free_all)
+      */
+     if (WARN_ON_ONCE(slab_is_available())) {
+             void *vaddr = kzalloc_node(size, GFP_NOWAIT, nid);
+
+             return vaddr ? virt_to_phys(vaddr) : 0;
+     }

I'd move this before alignment check.

Ack, will do in V2.

Anything else or should I make the tweaks and post V2?

JG

maintainer
#3 lore raw
Quoted 10 lines · James
Quoted 4 lines · Mike Rapoport
Quoted 1 line · James Gowans

Subject: [PATCH] memblocks: Move late alloc warning down to phys alloc

Nit: memblock

Ack!

Quoted 2 lines
Quoted 2 lines

If a driver/subsystem tries to do an allocation after memblocks have

And another one here :) ^

Quoted 52 lines
Quoted 9 lines
Quoted 7 lines

been freed and the memory handed to the buddy allocator, it will not actually be legal to use that allocation - the buddy allocator owns the memory. This is handled by the memblocks function which does allocations and returns virtual addresses by printing a warning and doing a kmalloc instead. However, the physical allocation function does not to do this check - callers of the physical alloc function are unprotected against mis-use.

Did you see such misuse or this is a theoretical issue?

Yeah, I was driving the memblock allocator badly when prototyping something. Allocating a large contiguous block of memory for an in- memory filesystem and I was doing the allocation in an initcall, but by that point it was too late. The memblock allocator happily gave me a large chunk of memory, but it was already in use by the buddy allocator, so ended up with memory corruption. Oops! Getting this warning would have made the problem immediately obvious.

Quoted 21 lines
Quoted 21 lines

Improve the error catching here by moving the check into the physical allocation function which is used by the virtual addr allocation function.

Cc Mike Rapoport <[email protected]>
Cc Andrew Morton <[email protected]>
Cc Alex Graf <[email protected]>
---
 mm/memblock.c | 18 +++++++++++-------
 1 file changed, 11 insertions(+), 7 deletions(-)
mm/memblock.c +0 -0
index d09136e040d3..dd4f237dc1fc 100644
@@ -1457,6 +1457,17 @@ phys_addr_t __init memblock_alloc_range_nid(phys_addr_t size,
              align = SMP_CACHE_BYTES;
      }
 
Quoted 12 lines

mm/memblock.c phys_addr_t __init memblock_alloc_range_nid(phys_addr_t size, · line 1469

Quoted 10 lines
mm/memblock.c +10 -0
+     /*
+      * Detect any accidental use of these APIs after slab is ready, as at
+      * this moment memblock may be deinitialized already and its
+      * internal data may be destroyed (after execution of memblock_free_all)
+      */
+     if (WARN_ON_ONCE(slab_is_available())) {
+             void *vaddr = kzalloc_node(size, GFP_NOWAIT, nid);
+
+             return vaddr ? virt_to_phys(vaddr) : 0;
+     }

I'd move this before alignment check.

Quoted 3 lines

Ack, will do in V2.

Anything else or should I make the tweaks and post V2?

Looks ok to me.

Quoted 1 line

JG

Signature · 2 lines
Sincerely yours,
Mike.