Skip to contentSkip to conversation

memblock, arm: fixes for freeing of the memory map

PATCH Superseded Mike Rapoport 4 participants 15 messages May 19, 2021 – Jun 30, 2021

Subsystems (4)

v2 of 3 Superseded Latest is v3 →

lore
Share & tools
lore

Mike Rapoport

author maintainer
#0 lore raw

v2 cover memblock, arm: fixes for freeing of the memory map

From: Mike Rapoport <[email protected]>

Hi,

The coordination between freeing of unused memory map, pfn_valid() and core mm assumptions about validity of the memory map in various ranges was not designed for complex layouts of the physical memory with a lot of holes all over the place.

Kefen Wang reported crashes in move_freepages() on a system with the following memory layout:

  node   0: [mem 0x0000000080a00000-0x00000000855fffff] node   0: [mem 0x0000000086a00000-0x0000000087dfffff] node   0: [mem 0x000000008bd00000-0x000000008c4fffff] node   0: [mem 0x000000008e300000-0x000000008ecfffff] node   0: [mem 0x0000000090d00000-0x00000000bfffffff] node   0: [mem 0x00000000cc000000-0x00000000dc9fffff] node   0: [mem 0x00000000de700000-0x00000000de9fffff] node   0: [mem 0x00000000e0800000-0x00000000e0bfffff] node   0: [mem 0x00000000f4b00000-0x00000000f6ffffff] node   0: [mem 0x00000000fda00000-0x00000000ffffefff]

The crashes can be mitigated by enabling CONFIG_HOLES_IN_ZONE and essentially turning pfn_valid_within() to pfn_valid() instead of having it hardwired to 1.

Alternatively, we can update ARM's implementation of pfn_valid() to take into accounting rounding of the freed memory map to pageblock boundaries and make sure it returns true for PFNs that have memory map entries even if there is no physical memory.

I can take the entire series via memblock tree.

@Kefen, I didn't add your Tested-by yet because the patch is slightly different from the version you've tested.

v2:
* Use single memblock_overlaps_region() instead of several memblock_is_map_memory() lookups. This makes this series depend on update of MEMBLOCK_NOMAP handling in the memory map [2]

v1: Link: https://lore.kernel.org/lkml/[email protected]

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

Mike Rapoport (3):
memblock: free_unused_memmap: use pageblock units instead of MAX_ORDER
memblock: align freed memory map on pageblock boundaries with SPARSEMEM
arm: extend pfn_valid to take into accound freed memory map alignment

 arch/arm/mm/init.c | 13 ++++++++++++-
 mm/memblock.c      | 23 ++++++++++++-----------
 2 files changed, 24 insertions(+), 12 deletions(-)

base-commit: d07f6ca923ea0927a1024dfccafc5b53b61cfecc

Signature · 1 line
2.28.0
author maintainer
#1 lore raw

PATCH v2 1/3 memblock: free_unused_memmap: use pageblock units instead of MAX_ORDER

From: Mike Rapoport <[email protected]>

The code that frees unused memory map uses rounds start and end of the holes that are freed to MAX_ORDER_NR_PAGES to preserve continuity of the memory map for MAX_ORDER regions.

Lots of core memory management functionality relies on homogeneity of the memory map within each pageblock which size may differ from MAX_ORDER in certain configurations.

Although currently, for the architectures that use free_unused_memmap(), pageblock_order and MAX_ORDER are equivalent, it is cleaner to have common notation thought mm code.

Replace MAX_ORDER_NR_PAGES with pageblock_nr_pages and update the comments to make it more clear why the alignment to pageblock boundaries is required.

---
 mm/memblock.c | 16 ++++++++--------
 1 file changed, 8 insertions(+), 8 deletions(-)
mm/memblock.c +8 -8
index afaefa8fc6ab..97fa87541b5f 100644
@@ -1943,11 +1943,11 @@ static void __init free_unused_memmap(void)
 		start = min(start, ALIGN(prev_end, PAGES_PER_SECTION));
 #else
 		/*
-		 * Align down here since the VM subsystem insists that the
-		 * memmap entries are valid from the bank start aligned to
-		 * MAX_ORDER_NR_PAGES.
+		 * Align down here since many operations in VM subsystem
+		 * presume that there are no holes in the memory map inside
+		 * a pageblock
 		 */
-		start = round_down(start, MAX_ORDER_NR_PAGES);
+		start = round_down(start, pageblock_nr_pages);
 #endif
 
 		/*
@@ -1958,11 +1958,11 @@ static void __init free_unused_memmap(void)
 			free_memmap(prev_end, start);
 
 		/*
-		 * Align up here since the VM subsystem insists that the
-		 * memmap entries are valid from the bank end aligned to
-		 * MAX_ORDER_NR_PAGES.
+		 * Align up here since many operations in VM subsystem
+		 * presume that there are no holes in the memory map inside
+		 * a pageblock
 		 */
-		prev_end = ALIGN(end, MAX_ORDER_NR_PAGES);
+		prev_end = ALIGN(end, pageblock_nr_pages);
 	}
 
 #ifdef CONFIG_SPARSEMEM
Signature · 1 line
2.28.0
author maintainer
#2 lore raw

PATCH v2 2/3 memblock: align freed memory map on pageblock boundaries with SPARSEMEM

In reply to Mike Rapoport

From: Mike Rapoport <[email protected]>

When CONFIG_SPARSEMEM=y the ranges of the memory map that are freed are not aligned to the pageblock boundaries which breaks assumptions about homogeneity of the memory map throughout core mm code.

Make sure that the freed memory map is always aligned on pageblock boundaries regardless of the memory model selection.

---
 mm/memblock.c | 7 ++++---
 1 file changed, 4 insertions(+), 3 deletions(-)
mm/memblock.c +4 -3
index 97fa87541b5f..2e25d69739e0 100644
@@ -1941,14 +1941,13 @@ static void __init free_unused_memmap(void)
 		 * due to SPARSEMEM sections which aren't present.
 		 */
 		start = min(start, ALIGN(prev_end, PAGES_PER_SECTION));
-#else
+#endif
 		/*
 		 * Align down here since many operations in VM subsystem
 		 * presume that there are no holes in the memory map inside
 		 * a pageblock
 		 */
 		start = round_down(start, pageblock_nr_pages);
-#endif
 
 		/*
 		 * If we had a previous bank, and there is a space
@@ -1966,8 +1965,10 @@ static void __init free_unused_memmap(void)
 	}
 
 #ifdef CONFIG_SPARSEMEM
-	if (!IS_ALIGNED(prev_end, PAGES_PER_SECTION))
+	if (!IS_ALIGNED(prev_end, PAGES_PER_SECTION)) {
+		prev_end = ALIGN(end, pageblock_nr_pages);
 		free_memmap(prev_end, ALIGN(prev_end, PAGES_PER_SECTION));
+	}
 #endif
 }
 
Signature · 1 line
2.28.0
author maintainer
#3 lore raw

PATCH v2 3/3 arm: extend pfn_valid to take into accound freed memory map alignment

In reply to Mike Rapoport

From: Mike Rapoport <[email protected]>

When unused memory map is freed the preserved part of the memory map is extended to match pageblock boundaries because lots of core mm functionality relies on homogeneity of the memory map within pageblock boundaries.

Since pfn_valid() is used to check whether there is a valid memory map entry for a PFN, make it return true also for PFNs that have memory map entries even if there is no actual memory populated there.

---
 arch/arm/mm/init.c | 13 ++++++++++++-
 1 file changed, 12 insertions(+), 1 deletion(-)
arch/arm/mm/init.c +12 -1
index 9d4744a632c6..6162a070a410 100644
@@ -125,11 +125,22 @@ static void __init zone_sizes_init(unsigned long min, unsigned long max_low,
 int pfn_valid(unsigned long pfn)
 {
 	phys_addr_t addr = __pfn_to_phys(pfn);
+	unsigned long pageblock_size = PAGE_SIZE * pageblock_nr_pages;
 
 	if (__phys_to_pfn(addr) != pfn)
 		return 0;
 
-	return memblock_is_map_memory(addr);
+	/*
+	 * If address less than pageblock_size bytes away from a present
+	 * memory chunk there still will be a memory map entry for it
+	 * because we round freed memory map to the pageblock boundaries.
+	 */
+	if (memblock_overlaps_region(&memblock.memory,
+				     ALIGN_DOWN(addr, pageblock_size),
+				     pageblock_size))
+		return 1;
+
+	return 0;
 }
 EXPORT_SYMBOL(pfn_valid);
 #endif
Signature · 1 line
2.28.0
#4 lore raw

Re: [PATCH v2 3/3] arm: extend pfn_valid to take into accound freed memory map alignment

Hi,

Looks like this patch causes a boot regression at least for Cortex-A15. That's commit 990e6d0e1de8 ("arm: extend pfn_valid to take into accound freed memory map alignment") in Linux next.

Most of the time I see the following on beagle-x15 right after init starts:

Kernel panic - not syncing: Attempted to kill init! exitcode=0x0000000b
CPU0: stopping
CPU: 0 PID: 0 Comm: swapper/0 Not tainted 5.13.0-rc7-next-20210625 #100
Hardware name: Generic DRA74X (Flattened Device Tree)
[<c0110c54>] (unwind_backtrace) from [<c010b408>] (show_stack+0x10/0x14)
[<c010b408>] (show_stack) from [<c09fae04>] (dump_stack_lvl+0x40/0x4c)
[<c09fae04>] (dump_stack_lvl) from [<c010e768>] (do_handle_IPI+0x2c8/0x334)
[<c010e768>] (do_handle_IPI) from [<c010e7e8>] (ipi_handler+0x14/0x20)
[<c010e7e8>] (ipi_handler) from [<c01a5f14>] (handle_percpu_devid_irq+0xa8/0x22c)
[<c01a5f14>] (handle_percpu_devid_irq) from [<c019fc78>] (handle_domain_irq+0x64/0xa4)
[<c019fc78>] (handle_domain_irq) from [<c05b9bdc>] (gic_handle_irq+0x88/0xb0)
[<c05b9bdc>] (gic_handle_irq) from [<c0100b6c>] (__irq_svc+0x6c/0x90)
Exception stack(0xc0f01f08 to 0xc0f01f50)

1f00: 00000f38 00000f37 00000000 fe600000 c0ff90c0 00000000
1f20: c0f0520c c0f05260 00000000 c0f00000 00000000 c0e788f0 00000000 c0f01f58 1f40: c0126aa0 c0107dc4 60000013 ffffffff
[<c0100b6c>] (__irq_svc) from [<c0107dc4>] (arch_cpu_idle+0x1c/0x3c) [<c0107dc4>] (arch_cpu_idle) from [<c0a098d8>] (default_idle_call+0x38/0xe0) [<c0a098d8>] (default_idle_call) from [<c0172860>] (do_idle+0x214/0x2cc) [<c0172860>] (do_idle) from [<c0172c0c>] (cpu_startup_entry+0x18/0x1c) [<c0172c0c>] (cpu_startup_entry) from [<c0e00ef8>] (start_kernel+0x5cc/0x6c4)

Sometimes the system boots to console, but maybe only about 20% of the time. Reverting 990e6d0e1de8 makes Linux next boot again for me.

Regards,

Tony

#regzb introduced: 990e6d0e1de8 ("arm: extend pfn_valid to take into accound freed memory map alignment")

Quoted 47 lines · Mike Rapoport

From: Mike Rapoport <[email protected]>

When unused memory map is freed the preserved part of the memory map is extended to match pageblock boundaries because lots of core mm functionality relies on homogeneity of the memory map within pageblock boundaries.

Since pfn_valid() is used to check whether there is a valid memory map entry for a PFN, make it return true also for PFNs that have memory map entries even if there is no actual memory populated there.

---
 arch/arm/mm/init.c | 13 ++++++++++++-
 1 file changed, 12 insertions(+), 1 deletion(-)
arch/arm/mm/init.c +12 -1
index 9d4744a632c6..6162a070a410 100644
@@ -125,11 +125,22 @@ static void __init zone_sizes_init(unsigned long min, unsigned long max_low,
 int pfn_valid(unsigned long pfn)
 {
 	phys_addr_t addr = __pfn_to_phys(pfn);
+	unsigned long pageblock_size = PAGE_SIZE * pageblock_nr_pages;
 
 	if (__phys_to_pfn(addr) != pfn)
 		return 0;
 
-	return memblock_is_map_memory(addr);
+	/*
+	 * If address less than pageblock_size bytes away from a present
+	 * memory chunk there still will be a memory map entry for it
+	 * because we round freed memory map to the pageblock boundaries.
+	 */
+	if (memblock_overlaps_region(&memblock.memory,
+				     ALIGN_DOWN(addr, pageblock_size),
+				     pageblock_size))
+		return 1;
+
+	return 0;
 }
 EXPORT_SYMBOL(pfn_valid);
 #endif
Signature · 1 line
2.28.0
#5 lore raw

Re: [PATCH v2 3/3] arm: extend pfn_valid to take into accound freed memory map alignment

Hi Tony,

Quoted 30 lines · Tony Lindgren

Hi,

Looks like this patch causes a boot regression at least for Cortex-A15. That's commit 990e6d0e1de8 ("arm: extend pfn_valid to take into accound freed memory map alignment") in Linux next.

Most of the time I see the following on beagle-x15 right after init starts:

Kernel panic - not syncing: Attempted to kill init! exitcode=0x0000000b
CPU0: stopping
CPU: 0 PID: 0 Comm: swapper/0 Not tainted 5.13.0-rc7-next-20210625 #100
Hardware name: Generic DRA74X (Flattened Device Tree)
[<c0110c54>] (unwind_backtrace) from [<c010b408>] (show_stack+0x10/0x14)
[<c010b408>] (show_stack) from [<c09fae04>] (dump_stack_lvl+0x40/0x4c)
[<c09fae04>] (dump_stack_lvl) from [<c010e768>] (do_handle_IPI+0x2c8/0x334)
[<c010e768>] (do_handle_IPI) from [<c010e7e8>] (ipi_handler+0x14/0x20)
[<c010e7e8>] (ipi_handler) from [<c01a5f14>] (handle_percpu_devid_irq+0xa8/0x22c)
[<c01a5f14>] (handle_percpu_devid_irq) from [<c019fc78>] (handle_domain_irq+0x64/0xa4)
[<c019fc78>] (handle_domain_irq) from [<c05b9bdc>] (gic_handle_irq+0x88/0xb0)
[<c05b9bdc>] (gic_handle_irq) from [<c0100b6c>] (__irq_svc+0x6c/0x90)
Exception stack(0xc0f01f08 to 0xc0f01f50)

1f00: 00000f38 00000f37 00000000 fe600000 c0ff90c0 00000000
1f20: c0f0520c c0f05260 00000000 c0f00000 00000000 c0e788f0 00000000 c0f01f58 1f40: c0126aa0 c0107dc4 60000013 ffffffff
[<c0100b6c>] (__irq_svc) from [<c0107dc4>] (arch_cpu_idle+0x1c/0x3c) [<c0107dc4>] (arch_cpu_idle) from [<c0a098d8>] (default_idle_call+0x38/0xe0) [<c0a098d8>] (default_idle_call) from [<c0172860>] (do_idle+0x214/0x2cc) [<c0172860>] (do_idle) from [<c0172c0c>] (cpu_startup_entry+0x18/0x1c) [<c0172c0c>] (cpu_startup_entry) from [<c0e00ef8>] (start_kernel+0x5cc/0x6c4)

Quoted 2 lines

Sometimes the system boots to console, but maybe only about 20% of the time. Reverting 990e6d0e1de8 makes Linux next boot again for me.

Can you please send log with 'memblock=debug' added to the command line?

Quoted 55 lines

Regards,

Tony

#regzb introduced: 990e6d0e1de8 ("arm: extend pfn_valid to take into accound freed memory map alignment")

Quoted 47 lines · Mike Rapoport

From: Mike Rapoport <[email protected]>

When unused memory map is freed the preserved part of the memory map is extended to match pageblock boundaries because lots of core mm functionality relies on homogeneity of the memory map within pageblock boundaries.

Since pfn_valid() is used to check whether there is a valid memory map entry for a PFN, make it return true also for PFNs that have memory map entries even if there is no actual memory populated there.

---
 arch/arm/mm/init.c | 13 ++++++++++++-
 1 file changed, 12 insertions(+), 1 deletion(-)
arch/arm/mm/init.c +12 -1
index 9d4744a632c6..6162a070a410 100644
@@ -125,11 +125,22 @@ static void __init zone_sizes_init(unsigned long min, unsigned long max_low,
 int pfn_valid(unsigned long pfn)
 {
 	phys_addr_t addr = __pfn_to_phys(pfn);
+	unsigned long pageblock_size = PAGE_SIZE * pageblock_nr_pages;
 
 	if (__phys_to_pfn(addr) != pfn)
 		return 0;
 
-	return memblock_is_map_memory(addr);
+	/*
+	 * If address less than pageblock_size bytes away from a present
+	 * memory chunk there still will be a memory map entry for it
+	 * because we round freed memory map to the pageblock boundaries.
+	 */
+	if (memblock_overlaps_region(&memblock.memory,
+				     ALIGN_DOWN(addr, pageblock_size),
+				     pageblock_size))
+		return 1;
+
+	return 0;
 }
 EXPORT_SYMBOL(pfn_valid);
 #endif
Signature · 1 line
2.28.0
Signature · 2 lines
Sincerely yours,
Mike.
#6 lore raw

Re: [PATCH v2 3/3] arm: extend pfn_valid to take into accound freed memory map alignment

Quoted 1 line · Mike Rapoport

Can you please send log with 'memblock=debug' added to the command line?

Sure, log now available at:

http://muru.com/beagle-x15.txt

Regards,

Tony

#7 lore raw

Re: [PATCH v2 3/3] arm: extend pfn_valid to take into accound freed memory map alignment

Quoted 6 lines · Tony Lindgren
Quoted 1 line · Mike Rapoport

Can you please send log with 'memblock=debug' added to the command line?

Sure, log now available at:

http://muru.com/beagle-x15.txt

Hmm, no clues yet :(

Do you have CONFIG_DEBUG_VM, CONFIG_DEBUG_VM_PGFLAGS and CONFIG_PAGE_POISONING enabled in your config?
If not, can you please enable them and see if any of VM_BUG_* triggers?

Do you use FLATMEM or SPARSEMEM in your config?

Let's try seeing what PFNs get false results from pfn_valid, maybe this will give a better lead.

arch/arm/mm/init.c +6 -0
index 6162a070a410..66985fc3e730 100644
@@ -126,10 +126,16 @@ int pfn_valid(unsigned long pfn)
 {
 	phys_addr_t addr = __pfn_to_phys(pfn);
 	unsigned long pageblock_size = PAGE_SIZE * pageblock_nr_pages;
+	bool overlaps = memblock_overlaps_region(&memblock.memory,
+				     ALIGN_DOWN(addr, pageblock_size),
+				     pageblock_size);
 
 	if (__phys_to_pfn(addr) != pfn)
 		return 0;
 
+	if (memblock_is_map_memory(addr) != overlaps)
+		pr_info("%s(%pS): pfn: %lx: is_map: %d overlaps: %d\n", __func__, (void *)_RET_IP_, pfn, memblock_is_map_memory(addr), overlaps);
+
 	/*
 	 * If address less than pageblock_size bytes away from a present
 	 * memory chunk there still will be a memory map entry for it
Signature · 2 lines
Sincerely yours,
Mike.
#8 lore raw

Re: [PATCH v2 3/3] arm: extend pfn_valid to take into accound freed memory map alignment

Hi,

Quoted 6 lines · Mike Rapoport
Quoted 5 lines · Tony Lindgren
Quoted 1 line · Mike Rapoport

Can you please send log with 'memblock=debug' added to the command line?

Sure, log now available at:

Quoted 7 lines
Quoted 1 line

http://muru.com/beagle-x15.txt

Hmm, no clues yet :(

Do you have CONFIG_DEBUG_VM, CONFIG_DEBUG_VM_PGFLAGS and CONFIG_PAGE_POISONING enabled in your config?
If not, can you please enable them and see if any of VM_BUG_* triggers?

OK enabled, and no errors or warnings are triggered.

Quoted 1 line

Do you use FLATMEM or SPARSEMEM in your config?

Looks like make omap2plus_defconfig enables FLATMEM:

$ grep -e SPARSEMEM -e FLATMEM .config
CONFIG_ARCH_FLATMEM_ENABLE=y
CONFIG_ARCH_SPARSEMEM_ENABLE=y
CONFIG_FLATMEM_MANUAL=y
# CONFIG_SPARSEMEM_MANUAL is not set
CONFIG_FLATMEM=y
Quoted 2 lines

Let's try seeing what PFNs get false results from pfn_valid, maybe this will give a better lead.

With your patch below, system boots with lots of the following:

[   13.058654] Freeing unused kernel image (initmem) memory: 1024K
...
[   13.129211] pfn_valid(__sync_icache_dcache+0x2c/0x138): pfn: fffb6: is_map: 1 overlaps: 0
[   13.137481] pfn_valid(__sync_icache_dcache+0x2c/0x138): pfn: fffb7: is_map: 1 overlaps: 0
[   13.145751] pfn_valid(__sync_icache_dcache+0x2c/0x138): pfn: fffb8: is_map: 1 overlaps: 0
[   13.153991] pfn_valid(__sync_icache_dcache+0x2c/0x138): pfn: fffb9: is_map: 1 overlaps: 0
[   13.162200] pfn_valid(__sync_icache_dcache+0x2c/0x138): pfn: fffba: is_map: 1 overlaps: 0
[   13.170440] pfn_valid(__sync_icache_dcache+0x2c/0x138): pfn: fffbb: is_map: 1 overlaps: 0
[   13.178680] pfn_valid(__sync_icache_dcache+0x2c/0x138): pfn: fffbc: is_map: 1 overlaps: 0
[   13.186920] pfn_valid(__sync_icache_dcache+0x2c/0x138): pfn: fffbd: is_map: 1 overlaps: 0
[   13.195159] pfn_valid(__sync_icache_dcache+0x2c/0x138): pfn: fffbe: is_map: 1 overlaps: 0
[   13.203399] pfn_valid(__sync_icache_dcache+0x2c/0x138): pfn: fffbf: is_map: 1 overlaps: 0
[   13.211639] pfn_valid(__sync_icache_dcache+0x2c/0x138): pfn: fff00: is_map: 1 overlaps: 0

...

Then changing console loglevel to 0 boots system to login prompt. But I'm seeing some init processes segfaulting during start-up.

Regards,

Tony

Quoted 25 lines
arch/arm/mm/init.c +6 -0
index 6162a070a410..66985fc3e730 100644
@@ -126,10 +126,16 @@ int pfn_valid(unsigned long pfn)
 {
 	phys_addr_t addr = __pfn_to_phys(pfn);
 	unsigned long pageblock_size = PAGE_SIZE * pageblock_nr_pages;
+	bool overlaps = memblock_overlaps_region(&memblock.memory,
+				     ALIGN_DOWN(addr, pageblock_size),
+				     pageblock_size);
 
 	if (__phys_to_pfn(addr) != pfn)
 		return 0;
 
+	if (memblock_is_map_memory(addr) != overlaps)
+		pr_info("%s(%pS): pfn: %lx: is_map: %d overlaps: %d\n", __func__, (void *)_RET_IP_, pfn, memblock_is_map_memory(addr), overlaps);
+
 	/*
 	 * If address less than pageblock_size bytes away from a present
 	 * memory chunk there still will be a memory map entry for it
Signature · 3 lines
-- 
Sincerely yours,
Mike.
#9 lore raw

Re: [PATCH v2 3/3] arm: extend pfn_valid to take into accound freed memory map alignment

Hi,

Quoted 39 lines · Tony Lindgren

Hi,

Quoted 6 lines · Mike Rapoport
Quoted 5 lines · Tony Lindgren
Quoted 1 line · Mike Rapoport

Can you please send log with 'memblock=debug' added to the command line?

Sure, log now available at:

Quoted 7 lines
Quoted 1 line

http://muru.com/beagle-x15.txt

Hmm, no clues yet :(

Do you have CONFIG_DEBUG_VM, CONFIG_DEBUG_VM_PGFLAGS and CONFIG_PAGE_POISONING enabled in your config?
If not, can you please enable them and see if any of VM_BUG_* triggers?

OK enabled, and no errors or warnings are triggered.

Quoted 1 line

Do you use FLATMEM or SPARSEMEM in your config?

Looks like make omap2plus_defconfig enables FLATMEM:

$ grep -e SPARSEMEM -e FLATMEM .config
CONFIG_ARCH_FLATMEM_ENABLE=y
CONFIG_ARCH_SPARSEMEM_ENABLE=y
CONFIG_FLATMEM_MANUAL=y
# CONFIG_SPARSEMEM_MANUAL is not set
CONFIG_FLATMEM=y
Quoted 2 lines

Let's try seeing what PFNs get false results from pfn_valid, maybe this will give a better lead.

With your patch below, system boots with lots of the following:

[   13.058654] Freeing unused kernel image (initmem) memory: 1024K
...
[   13.129211] pfn_valid(__sync_icache_dcache+0x2c/0x138): pfn: fffb6: is_map: 1 overlaps: 0
[   13.137481] pfn_valid(__sync_icache_dcache+0x2c/0x138): pfn: fffb7: is_map: 1 overlaps: 0

...

Quoted 2 lines

Then changing console loglevel to 0 boots system to login prompt. But I'm seeing some init processes segfaulting during start-up.

As it seems, the new version of pfn_valid() decides that last pages are not valid because of the overflow in memblock_overlaps_region(). As the result, __sync_icache_dcache() skips flushing these pages.

The patch below should fix this. I've left the prints for now, hopefully they will not appear anymore.

arch/arm/mm/init.c +7 -1
index 6162a070a410..7ba22d23eca4 100644
@@ -126,10 +126,16 @@ int pfn_valid(unsigned long pfn)
 {
 	phys_addr_t addr = __pfn_to_phys(pfn);
 	unsigned long pageblock_size = PAGE_SIZE * pageblock_nr_pages;
+	bool overlaps = memblock_overlaps_region(&memblock.memory,
+				     ALIGN_DOWN(addr, pageblock_size),
+				     pageblock_size - 1);
 
 	if (__phys_to_pfn(addr) != pfn)
 		return 0;
 
+	if (memblock_is_map_memory(addr) != overlaps)
+		pr_info("%s(%pS): pfn: %lx: is_map: %d overlaps: %d\n", __func__, (void *)_RET_IP_, pfn, memblock_is_map_memory(addr), overlaps);
+
 	/*
 	 * If address less than pageblock_size bytes away from a present
 	 * memory chunk there still will be a memory map entry for it
@@ -137,7 +143,7 @@ int pfn_valid(unsigned long pfn)
 	 */
 	if (memblock_overlaps_region(&memblock.memory,
 				     ALIGN_DOWN(addr, pageblock_size),
-				     pageblock_size))
+				     pageblock_size - 1))
 		return 1;
 
 	return 0;
Signature · 2 lines
Sincerely yours,
Mike.
#10 lore raw

Re: [PATCH v2 3/3] arm: extend pfn_valid to take into accound freed memory map alignment

Quoted 6 lines · Mike Rapoport

As it seems, the new version of pfn_valid() decides that last pages are not valid because of the overflow in memblock_overlaps_region(). As the result, __sync_icache_dcache() skips flushing these pages.

The patch below should fix this. I've left the prints for now, hopefully they will not appear anymore.

Yes this allows the system to boot for me :)

I'm still seeing these three prints though:

...
smp: Brought up 1 node, 2 CPUs
SMP: Total of 2 processors activated (3994.41 BogoMIPS).
CPU: All CPU(s) started in SVC mode. pfn_valid(__pageblock_pfn_to_page+0x14/0xa8): pfn: afe00: is_map: 0 overlaps: 1 pfn_valid(__pageblock_pfn_to_page+0x28/0xa8): pfn: affff: is_map: 0 overlaps: 1 pfn_valid(__pageblock_pfn_to_page+0x38/0xa8): pfn: afe00: is_map: 0 overlaps: 1
devtmpfs: initialized
...

Regards,

Tony

Quoted 34 lines
arch/arm/mm/init.c +7 -1
index 6162a070a410..7ba22d23eca4 100644
@@ -126,10 +126,16 @@ int pfn_valid(unsigned long pfn)
 {
 	phys_addr_t addr = __pfn_to_phys(pfn);
 	unsigned long pageblock_size = PAGE_SIZE * pageblock_nr_pages;
+	bool overlaps = memblock_overlaps_region(&memblock.memory,
+				     ALIGN_DOWN(addr, pageblock_size),
+				     pageblock_size - 1);
 
 	if (__phys_to_pfn(addr) != pfn)
 		return 0;
 
+	if (memblock_is_map_memory(addr) != overlaps)
+		pr_info("%s(%pS): pfn: %lx: is_map: %d overlaps: %d\n", __func__, (void *)_RET_IP_, pfn, memblock_is_map_memory(addr), overlaps);
+
 	/*
 	 * If address less than pageblock_size bytes away from a present
 	 * memory chunk there still will be a memory map entry for it
@@ -137,7 +143,7 @@ int pfn_valid(unsigned long pfn)
 	 */
 	if (memblock_overlaps_region(&memblock.memory,
 				     ALIGN_DOWN(addr, pageblock_size),
-				     pageblock_size))
+				     pageblock_size - 1))
 		return 1;
 
 	return 0;
Signature · 3 lines
-- 
Sincerely yours,
Mike.
#11 lore raw

Re: [PATCH v2 3/3] arm: extend pfn_valid to take into accound freed memory map alignment

Quoted 12 lines · Tony Lindgren
Quoted 6 lines · Mike Rapoport

As it seems, the new version of pfn_valid() decides that last pages are not valid because of the overflow in memblock_overlaps_region(). As the result, __sync_icache_dcache() skips flushing these pages.

The patch below should fix this. I've left the prints for now, hopefully they will not appear anymore.

Yes this allows the system to boot for me :)

I'm still seeing these three prints though:

Quoted 7 lines

...
smp: Brought up 1 node, 2 CPUs
SMP: Total of 2 processors activated (3994.41 BogoMIPS).
CPU: All CPU(s) started in SVC mode. pfn_valid(__pageblock_pfn_to_page+0x14/0xa8): pfn: afe00: is_map: 0 overlaps: 1 pfn_valid(__pageblock_pfn_to_page+0x28/0xa8): pfn: affff: is_map: 0 overlaps: 1 pfn_valid(__pageblock_pfn_to_page+0x38/0xa8): pfn: afe00: is_map: 0 overlaps: 1

These pfns do have memory map despite they are stolen in arm_memblock_steal():

memblock_free: [0xaff00000-0xafffffff] arm_memblock_steal+0x50/0x70
memblock_remove: [0xaff00000-0xafffffff] arm_memblock_steal+0x5c/0x70
...
memblock_free: [0xafe00000-0xafefffff] arm_memblock_steal+0x50/0x70
memblock_remove: [0xafe00000-0xafefffff] arm_memblock_steal+0x5c/0x70

But the struct pages there are never initialized.

I'll resend v3 of the entire set with an addition patch to take care of that as well.

Quoted 42 lines

devtmpfs: initialized
...

Regards,

Tony

Quoted 34 lines
arch/arm/mm/init.c +7 -1
index 6162a070a410..7ba22d23eca4 100644
@@ -126,10 +126,16 @@ int pfn_valid(unsigned long pfn)
 {
 	phys_addr_t addr = __pfn_to_phys(pfn);
 	unsigned long pageblock_size = PAGE_SIZE * pageblock_nr_pages;
+	bool overlaps = memblock_overlaps_region(&memblock.memory,
+				     ALIGN_DOWN(addr, pageblock_size),
+				     pageblock_size - 1);
 
 	if (__phys_to_pfn(addr) != pfn)
 		return 0;
 
+	if (memblock_is_map_memory(addr) != overlaps)
+		pr_info("%s(%pS): pfn: %lx: is_map: %d overlaps: %d\n", __func__, (void *)_RET_IP_, pfn, memblock_is_map_memory(addr), overlaps);
+
 	/*
 	 * If address less than pageblock_size bytes away from a present
 	 * memory chunk there still will be a memory map entry for it
@@ -137,7 +143,7 @@ int pfn_valid(unsigned long pfn)
 	 */
 	if (memblock_overlaps_region(&memblock.memory,
 				     ALIGN_DOWN(addr, pageblock_size),
-				     pageblock_size))
+				     pageblock_size - 1))
 		return 1;
 
 	return 0;
Signature · 3 lines
-- 
Sincerely yours,
Mike.
Signature · 2 lines
Sincerely yours,
Mike.
#12 lore raw

Re: [PATCH v2 3/3] arm: extend pfn_valid to take into accound freed memory map alignment

Quoted 30 lines · Mike Rapoport
Quoted 12 lines · Tony Lindgren
Quoted 6 lines · Mike Rapoport

As it seems, the new version of pfn_valid() decides that last pages are not valid because of the overflow in memblock_overlaps_region(). As the result, __sync_icache_dcache() skips flushing these pages.

The patch below should fix this. I've left the prints for now, hopefully they will not appear anymore.

Yes this allows the system to boot for me :)

I'm still seeing these three prints though:

Quoted 7 lines

...
smp: Brought up 1 node, 2 CPUs
SMP: Total of 2 processors activated (3994.41 BogoMIPS).
CPU: All CPU(s) started in SVC mode. pfn_valid(__pageblock_pfn_to_page+0x14/0xa8): pfn: afe00: is_map: 0 overlaps: 1 pfn_valid(__pageblock_pfn_to_page+0x28/0xa8): pfn: affff: is_map: 0 overlaps: 1 pfn_valid(__pageblock_pfn_to_page+0x38/0xa8): pfn: afe00: is_map: 0 overlaps: 1

These pfns do have memory map despite they are stolen in arm_memblock_steal():

memblock_free: [0xaff00000-0xafffffff] arm_memblock_steal+0x50/0x70
memblock_remove: [0xaff00000-0xafffffff] arm_memblock_steal+0x5c/0x70
...
memblock_free: [0xafe00000-0xafefffff] arm_memblock_steal+0x50/0x70
memblock_remove: [0xafe00000-0xafefffff] arm_memblock_steal+0x5c/0x70

Quoted 4 lines

But the struct pages there are never initialized.

I'll resend v3 of the entire set with an addition patch to take care of that as well.

OK sounds good to me :)

Thanks,

Tony

#13 lore raw

Re: [PATCH v2 3/3] arm: extend pfn_valid to take into accound freed memory map alignment

In reply to Mike Rapoport
Quoted 24 lines · Mike Rapoport
Quoted 12 lines · Tony Lindgren
Quoted 6 lines · Mike Rapoport

As it seems, the new version of pfn_valid() decides that last pages are not valid because of the overflow in memblock_overlaps_region(). As the result, __sync_icache_dcache() skips flushing these pages.

The patch below should fix this. I've left the prints for now, hopefully they will not appear anymore.

Yes this allows the system to boot for me :)

I'm still seeing these three prints though:

Quoted 7 lines

...
smp: Brought up 1 node, 2 CPUs
SMP: Total of 2 processors activated (3994.41 BogoMIPS).
CPU: All CPU(s) started in SVC mode. pfn_valid(__pageblock_pfn_to_page+0x14/0xa8): pfn: afe00: is_map: 0 overlaps: 1 pfn_valid(__pageblock_pfn_to_page+0x28/0xa8): pfn: affff: is_map: 0 overlaps: 1 pfn_valid(__pageblock_pfn_to_page+0x38/0xa8): pfn: afe00: is_map: 0 overlaps: 1

These pfns do have memory map despite they are stolen in arm_memblock_steal():

Quoted 7 lines

memblock_free: [0xaff00000-0xafffffff] arm_memblock_steal+0x50/0x70
memblock_remove: [0xaff00000-0xafffffff] arm_memblock_steal+0x5c/0x70
...
memblock_free: [0xafe00000-0xafefffff] arm_memblock_steal+0x50/0x70
memblock_remove: [0xafe00000-0xafefffff] arm_memblock_steal+0x5c/0x70

But the struct pages there are never initialized.

Actually, with FLATMEM these struct pages will be always set to 0 because we don't do memory map poisoning with FLATMEM.

I could not find a case where zeroed struct page would cause real trouble, so I'd say it is more theoretical issue and it can be addressed unrelated to these changes.

Signature · 2 lines
Sincerely yours,
Mike.
#14 lore raw
In reply to Mike Rapoport
Quoted 32 lines · Mike Rapoport

From: Mike Rapoport <[email protected]>

Hi,

The coordination between freeing of unused memory map, pfn_valid() and core mm assumptions about validity of the memory map in various ranges was not designed for complex layouts of the physical memory with a lot of holes all over the place.

Kefen Wang reported crashes in move_freepages() on a system with the following memory layout:

  node   0: [mem 0x0000000080a00000-0x00000000855fffff] node   0: [mem 0x0000000086a00000-0x0000000087dfffff] node   0: [mem 0x000000008bd00000-0x000000008c4fffff] node   0: [mem 0x000000008e300000-0x000000008ecfffff] node   0: [mem 0x0000000090d00000-0x00000000bfffffff] node   0: [mem 0x00000000cc000000-0x00000000dc9fffff] node   0: [mem 0x00000000de700000-0x00000000de9fffff] node   0: [mem 0x00000000e0800000-0x00000000e0bfffff] node   0: [mem 0x00000000f4b00000-0x00000000f6ffffff] node   0: [mem 0x00000000fda00000-0x00000000ffffefff]

The crashes can be mitigated by enabling CONFIG_HOLES_IN_ZONE and essentially turning pfn_valid_within() to pfn_valid() instead of having it hardwired to 1.

Alternatively, we can update ARM's implementation of pfn_valid() to take into accounting rounding of the freed memory map to pageblock boundaries and make sure it returns true for PFNs that have memory map entries even if there is no physical memory.

Quoted 4 lines

I can take the entire series via memblock tree.

@Kefen, I didn't add your Tested-by yet because the patch is slightly different from the version you've tested.

Backport this version(also with link[2]) and oom test could pass too.

Quoted 23 lines

v2:
* Use single memblock_overlaps_region() instead of several memblock_is_map_memory() lookups. This makes this series depend on update of MEMBLOCK_NOMAP handling in the memory map [2]

v1: Link: https://lore.kernel.org/lkml/[email protected]

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

Mike Rapoport (3):
memblock: free_unused_memmap: use pageblock units instead of MAX_ORDER
memblock: align freed memory map on pageblock boundaries with SPARSEMEM
arm: extend pfn_valid to take into accound freed memory map alignment

  arch/arm/mm/init.c | 13 ++++++++++++-
  mm/memblock.c      | 23 ++++++++++++-----------
  2 files changed, 24 insertions(+), 12 deletions(-)

base-commit: d07f6ca923ea0927a1024dfccafc5b53b61cfecc