Skip to contentSkip to conversation

arm64: mm: Fix initialisation of DMA zones on non-NUMA systems

PATCH Applied Will Deacon 3 participants 4 messages Dec 3, 2019 – Dec 4, 2019 arm64

lore
Share & tools
lore

Will Deacon

author maintainer
#0 lore raw

PATCH 1/1 arm64: mm: Fix initialisation of DMA zones on non-NUMA systems

John reports that the recently merged commit 1a8e1cef7603 ("arm64: use both ZONE_DMA and ZONE_DMA32") breaks the boot on his DB845C board:

  | Booting Linux on physical CPU 0x0000000000 [0x517f803c]
  | Linux version 5.4.0-mainline-10675-g957a03b9e38f
  | Machine model: Thundercomm Dragonboard 845c
  | [...]
  | Built 1 zonelists, mobility grouping on.  Total pages: -188245
  | Kernel command line: earlycon
  | firmware_class.path=/vendor/firmware/ androidboot.hardware=db845c
  | init=/init androidboot.boot_devices=soc/1d84000.ufshc
  | printk.devkmsg=on buildvariant=userdebug root=/dev/sda2
  | androidboot.bootdevice=1d84000.ufshc androidboot.serialno=c4e1189c
  | androidboot.baseband=sda
  | msm_drm.dsi_display0=dsi_lt9611_1080_video_display:
  | androidboot.slot_suffix=_a skip_initramfs rootwait ro init=/init
  |
  | <hangs indefinitely here>

This is because, when CONFIG_NUMA=n, zone_sizes_init() fails to handle memblocks that fall entirely within the ZONE_DMA region and erroneously ends up trying to add a negatively-sized region into the following ZONE_DMA32, which is later interpreted as a large unsigned region by the core MM code.

Rework the non-NUMA implementation of zone_sizes_init() so that the start address of the memblock being processed is adjusted according to the end of the previous zone, which is then range-checked before updating the hole information of subsequent zones.

Cc Nicolas Saenz Julienne <[email protected]>
Cc Catalin Marinas <[email protected]>
Cc Christoph Hellwig <[email protected]>
Cc Bjorn Andersson <[email protected]>
Fixes 1a8e1cef7603 ("arm64: use both ZONE_DMA and ZONE_DMA32")
---
 arch/arm64/mm/init.c | 25 +++++++++++--------------
 1 file changed, 11 insertions(+), 14 deletions(-)

Compile-tested only.

arch/arm64/mm/init.c +11 -14
index be9481cdf3b9..b65dffdfb201 100644
@@ -214,15 +214,14 @@ static void __init zone_sizes_init(unsigned long min, unsigned long max)
 {
 	struct memblock_region *reg;
 	unsigned long zone_size[MAX_NR_ZONES], zhole_size[MAX_NR_ZONES];
-	unsigned long max_dma32 = min;
-	unsigned long __maybe_unused max_dma = min;
+	unsigned long __maybe_unused max_dma, max_dma32;
 
 	memset(zone_size, 0, sizeof(zone_size));
 
+	max_dma = max_dma32 = min;
 #ifdef CONFIG_ZONE_DMA
-	max_dma = PFN_DOWN(arm64_dma_phys_limit);
+	max_dma = max_dma32 = PFN_DOWN(arm64_dma_phys_limit);
 	zone_size[ZONE_DMA] = max_dma - min;
-	max_dma32 = max_dma;
 #endif
 #ifdef CONFIG_ZONE_DMA32
 	max_dma32 = PFN_DOWN(arm64_dma32_phys_limit);
@@ -236,25 +235,23 @@ static void __init zone_sizes_init(unsigned long min, unsigned long max)
 		unsigned long start = memblock_region_memory_base_pfn(reg);
 		unsigned long end = memblock_region_memory_end_pfn(reg);
 
-		if (start >= max)
-			continue;
 #ifdef CONFIG_ZONE_DMA
-		if (start < max_dma) {
-			unsigned long dma_end = min_not_zero(end, max_dma);
+		if (start >= min && start < max_dma) {
+			unsigned long dma_end = min(end, max_dma);
 			zhole_size[ZONE_DMA] -= dma_end - start;
+			start = dma_end;
 		}
 #endif
 #ifdef CONFIG_ZONE_DMA32
-		if (start < max_dma32) {
+		if (start >= max_dma && start < max_dma32) {
 			unsigned long dma32_end = min(end, max_dma32);
-			unsigned long dma32_start = max(start, max_dma);
-			zhole_size[ZONE_DMA32] -= dma32_end - dma32_start;
+			zhole_size[ZONE_DMA32] -= dma32_end - start;
+			start = dma32_end;
 		}
 #endif
-		if (end > max_dma32) {
+		if (start >= max_dma32 && start < max) {
 			unsigned long normal_end = min(end, max);
-			unsigned long normal_start = max(start, max_dma32);
-			zhole_size[ZONE_NORMAL] -= normal_end - normal_start;
+			zhole_size[ZONE_NORMAL] -= normal_end - start;
 		}
 	}
 
Signature · 1 line
2.17.1
Tested-by on v1 1/1
#1 lore raw
Quoted 3 lines · Will Deacon

John reports that the recently merged commit 1a8e1cef7603 ("arm64: use both ZONE_DMA and ZONE_DMA32") breaks the boot on his DB845C board:

I've also tripped on the same issue with the original HiKey board, and can validate that this patch resolves the issue there.

thanks
-john

#2 lore raw
Quoted 10 lines · John Stultz
Quoted 3 lines · Will Deacon

John reports that the recently merged commit 1a8e1cef7603 ("arm64: use both ZONE_DMA and ZONE_DMA32") breaks the boot on his DB845C board:

I've also tripped on the same issue with the original HiKey board, and can validate that this patch resolves the issue there.

Also validated on db845c.
Thanks so much for the quick fix!
-john

maintainer Applied
#3 lore raw
In reply to Will Deacon
Quoted 29 lines · Will Deacon

John reports that the recently merged commit 1a8e1cef7603 ("arm64: use both ZONE_DMA and ZONE_DMA32") breaks the boot on his DB845C board:

  | Booting Linux on physical CPU 0x0000000000 [0x517f803c]
  | Linux version 5.4.0-mainline-10675-g957a03b9e38f
  | Machine model: Thundercomm Dragonboard 845c
  | [...]
  | Built 1 zonelists, mobility grouping on.  Total pages: -188245
  | Kernel command line: earlycon
  | firmware_class.path=/vendor/firmware/ androidboot.hardware=db845c
  | init=/init androidboot.boot_devices=soc/1d84000.ufshc
  | printk.devkmsg=on buildvariant=userdebug root=/dev/sda2
  | androidboot.bootdevice=1d84000.ufshc androidboot.serialno=c4e1189c
  | androidboot.baseband=sda
  | msm_drm.dsi_display0=dsi_lt9611_1080_video_display:
  | androidboot.slot_suffix=_a skip_initramfs rootwait ro init=/init
  |
  | <hangs indefinitely here>

This is because, when CONFIG_NUMA=n, zone_sizes_init() fails to handle memblocks that fall entirely within the ZONE_DMA region and erroneously ends up trying to add a negatively-sized region into the following ZONE_DMA32, which is later interpreted as a large unsigned region by the core MM code.

Rework the non-NUMA implementation of zone_sizes_init() so that the start address of the memblock being processed is adjusted according to the end of the previous zone, which is then range-checked before updating the hole information of subsequent zones.

Quoted 8 lines
Cc Nicolas Saenz Julienne <[email protected]>
Cc Catalin Marinas <[email protected]>
Cc Christoph Hellwig <[email protected]>
Cc Bjorn Andersson <[email protected]>
Fixes 1a8e1cef7603 ("arm64: use both ZONE_DMA and ZONE_DMA32")

Queued for 5.5-rc1. Thanks.

Signature · 1 line
Catalin