The intermediate result of the old term (4UL * 1024 * 1024 * 1024) is 4 294 967 296 or 0x100000000 which is no problem on 64 bit systems. The patch does not change the later overall result of 0x100000 for MAX_DMA32_PFN. The new calculation yields the same result, but does not require 64 bit arithmetic.
On 32 bit systems the old calculation suffers from an arithmetic overflow in that intermediate term in braces: 4UL aka unsigned long int is 4 byte wide and an arithmetic overflow happens (the 0x100000000 does not fit in 4 bytes), the in braces result is truncated to zero, the following right shift does not alter that, so MAX_DMA32_PFN evaluates to 0 on 32 bit systems.
That wrong value is a problem in a comparision against MAX_DMA32_PFN in the init code for swiotlb in 'pci_swiotlb_detect_4gb()' to decide if swiotlb should be active. That comparison yields the opposite result, when compiling on 32 bit systems.
This was not possible before 1b7e03ef7570 ("x86, NUMA: Enable emulation on 32bit too") when that MAX_DMA32_PFN was first made visible to x86_32 (and which landed in v3.0).
In practice this wasn't a problem, unless you activated CONFIG_SWIOTLB on x86 (32 bit).
However for ARCH=x86 (32 bit) and if you have set CONFIG_IOMMU_INTEL, since c5a5dc4cbbf4 ("iommu/vt-d: Don't switch off swiotlb if bounce page is used") there's a dependency on CONFIG_SWIOTLB, which was not necessarily active before. That landed in v5.4, where we noticed it in the fli4l Linux distribution. We have CONFIG_IOMMU_INTEL active on both 32 and 64 bit kernel configs there (I could not find out why, so let's just say historical reasons).
The effect is at boot time 64 MiB (default size) were allocated for bounce buffers now, which is a noticeable amount of memory on small systems like pcengines ALIX 2D3 with 256 MiB memory, which are still frequently used as home routers.
We noticed this effect when migrating from kernel v4.19 (LTS) to v5.4 (LTS) in fli4l and got that kernel messages for example:
Linux version 5.4.22 (buildroot@buildroot) (gcc version 7.3.0 (Buildroot 2018.02.8)) #1 SMP Mon Nov 26 23:40:00 CET 2018 … Memory: 183484K/261756K available (4594K kernel code, 393K rwdata, 1660K rodata, 536K init, 456K bss , 78272K reserved, 0K cma-reserved, 0K highmem) … PCI-DMA: Using software bounce buffering for IO (SWIOTLB) software IO TLB: mapped [mem 0x0bb78000-0x0fb78000] (64MB)
The initial analysis and the suggested fix was done by user 'sourcejedi' at stackoverflow and explicitly marked as GPLv2 for inclusion in the Linux kernel:
https://unix.stackexchange.com/a/520525/50007
The new calculation, which does not suffer from that overflow, is the same as for arch/mips now as suggested by Robin Murphy.
The fix was tested by fli4l users on round about two dozen different systems, including both 32 and 64 bit archs, bare metal and virtualized machines.
Fixes: 1b7e03ef7570 ("x86, NUMA: Enable emulation on 32bit too") Fixes: https://web.nettworks.org/bugs/browse/FFL-2560 Fixes: https://unix.stackexchange.com/q/520065/50007 Reported-by: Alan Jenkins alan.christopher.jenkins@gmail.com Suggested-by: Robin Murphy robin.murphy@arm.com Signed-off-by: Alexander Dahl post@lespocky.de Cc: stable@vger.kernel.org ---
Notes: v3: - rewritten commit message to better explain that arithmetic overflow and added Fixes tag (Greg Kroah-Hartman) - rebased on v5.7-rc7
v2: - use the same calculation as with arch/mips (Robin Murphy)
arch/x86/include/asm/dma.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/arch/x86/include/asm/dma.h b/arch/x86/include/asm/dma.h index 00f7cf45e699..8e95aa4b0d17 100644 --- a/arch/x86/include/asm/dma.h +++ b/arch/x86/include/asm/dma.h @@ -74,7 +74,7 @@ #define MAX_DMA_PFN ((16UL * 1024 * 1024) >> PAGE_SHIFT)
/* 4GB broken PCI/AGP hardware bus master zone */ -#define MAX_DMA32_PFN ((4UL * 1024 * 1024 * 1024) >> PAGE_SHIFT) +#define MAX_DMA32_PFN (1UL << (32 - PAGE_SHIFT))
#ifdef CONFIG_X86_32 /* The maximum address that we can perform a DMA transfer to on this platform */