RFC [PATCH] x86/pci: reserve extra page to avoid error caused by P2P pref DMA reads
Diag guys, found one system when loading is high, will have gart wark error. root cause is P2P bridge try to prefetch for several intel e1000 under it. and that skb is near GART iommu area.
try to reserve page in the boundary at first.
last page near TOM2, and last page near MMIO
also gart first and last page.
need one better way for all arch support PCI and memory with a lot of holes etc.
--- arch/x86/kernel/pci-dma.c | 28 ++++++++++++++++++++++++++++ arch/x86/kernel/pci-gart_64.c | 6 ++++++ 2 files changed, 34 insertions(+)
arch/x86/kernel/pci-dma.c +28 -0
@@ -72,12 +72,40 @@ static int __init parse_dma32_size_opt(c }early_param ("dma32_size" , parse_dma32_size_opt); +static void __init reserve_last_page (unsigned long pfn) +{ +unsigned long phys; +void *ptr; + + phys = (pfn -1 )<<PAGE_SHIFT ; + ptr =__alloc_bootmem_nopanic (PAGE_SIZE ,PAGE_SIZE , phys); + +if (!ptr ||virt_to_phys (ptr) != phys) +printk (KERN_WARNING "Can not hold last page near %lx for workaround P2P pref DMA reads!\n" , phys); +else +printk (KERN_WARNING "Last page is reserved near %lx for workaround P2P pref DMA reads!\n" , phys); +}void __init dma32_reserve_bootmem (void ) {unsigned long size, align; + +/* +* try to reserve last page to workaround P2P bridge pref DMA reads +* normally don't need to reserve the page near mmio, +* because always has acpi etc sit there. +* but some system has that acpi in the middle of ram below 4g +* so just reserve it. +*/ +if (max_low_pfn_mapped < max_pfn_mapped) +reserve_last_page (max_low_pfn_mapped); + +/* less than 4G, don't need iommu */ if (max_pfn <=MAX_DMA32_PFN )return ; +/* try to reserve last page to workaround P2P bridge pref DMA reads */ +reserve_last_page (max_pfn); +/* * check aperture_64.c allocate_aperture() for reason about * using 512M as goal
arch/x86/kernel/pci-gart_64.c +6 -0
@@ -826,6 +826,9 @@ void __init gart_iommu_init(void)*/ set_bit_string (iommu_gart_bitmap,0 ,EMERGENCY_PAGES ); +/* reserve one page at tail, for P2P bridge pref DMA reads */ +set_bit_string (iommu_gart_bitmap, iommu_pages -1 ,1 ); + agp_memory_reserved = iommu_size; printk(KERN_INFO "PCI-DMA: Reserving %luMB of IOMMU area in the AGP aperture\n",
@@ -870,6 +873,9 @@ void __init gart_iommu_init(void)for (i =EMERGENCY_PAGES ; i < iommu_pages; i++) iommu_gatt_base[i] = gart_unmapped_entry; +/* we need set unmapped on head too, for P2P bridge pref DMA reads */ + iommu_gatt_base[0 ] = gart_unmapped_entry; +flush_gart (); dma_ops = &gart_dma_ops; }
Quoted 12 lines · Yinghai Lu
Diag guys, found one system when loading is high, will have gart wark error. root cause is P2P bridge try to prefetch for several intel e1000 under it. and that skb is near GART iommu area.
try to reserve page in the boundary at first. last page near TOM2, and last page near MMIO also gart first and last page.
need one better way for all arch support PCI and memory with a lot of holes etc.
void
__initdma32_reserve_bootmem(void)
{
arch/x86/kernel/pci-dma.c static int __init parse_dma32_size_opt(c · line 98
Quoted 9 lines
arch/x86/kernel/pci-dma.c +8
-0unsigned long size, align; + +/* +* try to reserve last page to workaround P2P bridge pref DMA reads +* normally don't need to reserve the page near mmio, +* because always has acpi etc sit there. +* but some system has that acpi in the middle of ram below 4g +* so just reserve it. +*/
Nice! Could this be the root cause of those skb corruptions and e1000 crashes you've been reporting? So the _usual_ setup accidentally protects us from these prefetch induced failures.
I think your patch is fine for the iommu bits, but the reserve_last_page() thing should be done in a cleaner way. Cannot we just reserve it all at the e820 stage, before passing that RAM to the bootmem allocator?
Also, what is the guarantee that 4K of a space is enough to stop all prefetching across that boundary?
Ingo
Quoted 9 lines · Yinghai Lu
Diag guys, found one system when loading is high, will have gart wark error. root cause is P2P bridge try to prefetch for several intel e1000 under it. and that skb is near GART iommu area.
try to reserve page in the boundary at first.
last page near TOM2, and last page near MMIO
also gart first and last page.need one better way for all arch support PCI and memory with a lot of holes etc.
Sounds like a similar problem for which we don't unmap GART pages, just map them read-only and route them to a dummy page. In that case, the right approach is to write the last page of the GART to the dummy page as well and mark it occupied. That shouldn't involve actually reserving memory.
I'm not familiar with the x86 PCI DMA code but afaik, it's fairly similar to ours (powerpc) in that area.
Ben.
Quoted 79 lines
Signed-off-by Yinghai Lu <[email protected]>Cc Pavel Machek <[email protected]>Cc Benjamin Herrenschmidt <[email protected]>Cc Jesse Barnes<[email protected]>--- arch/x86/kernel/pci-dma.c | 28 ++++++++++++++++++++++++++++ arch/x86/kernel/pci-gart_64.c | 6 ++++++ 2 files changed, 34 insertions(+)Index:
linux-2.6/arch/x86/kernel/pci-dma.c
===================================================================arch/x86/kernel/pci-dma.c +28
-0@@ -72,12 +72,40 @@ static int __init parse_dma32_size_opt(c }early_param ("dma32_size" , parse_dma32_size_opt); +static void __init reserve_last_page (unsigned long pfn) +{ +unsigned long phys; +void *ptr; + + phys = (pfn -1 )<<PAGE_SHIFT ; + ptr =__alloc_bootmem_nopanic (PAGE_SIZE ,PAGE_SIZE , phys); + +if (!ptr ||virt_to_phys (ptr) != phys) +printk (KERN_WARNING "Can not hold last page near %lx for workaround P2P pref DMA reads!\n" , phys); +else +printk (KERN_WARNING "Last page is reserved near %lx for workaround P2P pref DMA reads!\n" , phys); +}void __init dma32_reserve_bootmem (void ) {unsigned long size, align; + +/* +* try to reserve last page to workaround P2P bridge pref DMA reads +* normally don't need to reserve the page near mmio, +* because always has acpi etc sit there. +* but some system has that acpi in the middle of ram below 4g +* so just reserve it. +*/ +if (max_low_pfn_mapped < max_pfn_mapped) +reserve_last_page (max_low_pfn_mapped); + +/* less than 4G, don't need iommu */ if (max_pfn <=MAX_DMA32_PFN )return ; +/* try to reserve last page to workaround P2P bridge pref DMA reads */ +reserve_last_page (max_pfn); +/* * check aperture_64.c allocate_aperture() for reason about * using 512M as goal Index:
linux-2.6/arch/x86/kernel/pci-gart_64.c
===================================================================arch/x86/kernel/pci-gart_64.c +6
-0@@ -826,6 +826,9 @@ void __init gart_iommu_init(void)*/ set_bit_string (iommu_gart_bitmap,0 ,EMERGENCY_PAGES ); +/* reserve one page at tail, for P2P bridge pref DMA reads */ +set_bit_string (iommu_gart_bitmap, iommu_pages -1 ,1 ); + agp_memory_reserved = iommu_size; printk(KERN_INFO "PCI-DMA: Reserving %luMB of IOMMU area in the AGP aperture\n",@@ -870,6 +873,9 @@ void __init gart_iommu_init(void)for (i =EMERGENCY_PAGES ; i < iommu_pages; i++) iommu_gatt_base[i] = gart_unmapped_entry; +/* we need set unmapped on head too, for P2P bridge pref DMA reads */ + iommu_gatt_base[0 ] = gart_unmapped_entry; +flush_gart (); dma_ops = &gart_dma_ops; }
Quoted 9 lines · Benjamin Herrenschmidt
Quoted 8 lines · Yinghai Lu
Diag guys, found one system when loading is high, will have gart wark error. root cause is P2P bridge try to prefetch for several intel e1000 under it. and that skb is near GART iommu area.
try to reserve page in the boundary at first.
last page near TOM2, and last page near MMIO
also gart first and last page.
Quoted 7 lines
Quoted 1 line
need one better way for all arch support PCI and memory with a lot of holes etc.
Sounds like a similar problem for which we don't unmap GART pages, just map them read-only and route them to a dummy page. In that case, the right approach is to write the last page of the GART to the dummy page as well and mark it occupied. That shouldn't involve actually reserving memory.
i should split it into two patches. that patch include two patch second part is doing map to scratch memory.
but memory near TOP of MEMORY and TOP of MEMORY2, could have the same problem.
YH