Add 4GB DMA32 zone
Subsystems (4)
PATCH 1/3 Add 4GB DMA32 zone
Add 4GB DMA32 zone
Add a new 4GB GFP_DMA32 zone between the GFP_DMA and GFP_NORMAL zones.
As a bit of historical background: when the x86-64 port was originally designed we had some discussion if we should use a 16MB DMA zone like i386 or a 4GB DMA zone like IA64 or both. Both was ruled out at this point because it was in early 2.4 when VM is still quite shakey and had bad troubles even dealing with one DMA zone. We settled on the 16MB DMA zone mainly because we worried about older soundcards and the floppy.
But this has always caused problems since then because device drivers had trouble getting enough DMA able memory. These days the VM works much better and the wide use of NUMA has proven it can deal with many zones successfully.
So this patch adds both zones.
This helps drivers who need a lot of memory below 4GB because their hardware is not accessing more (graphic drivers - proprietary and free ones, video frame buffer drivers, sound drivers etc.). Previously they could only use IOMMU+16MB GFP_DMA, which was not enough memory.
Another common problem is that hardware who has full memory addressing for >4GB misses it for some control structures in memory (like transmit rings or other metadata). They tended to allocate memory in the 16MB GFP_DMA or the IOMMU/swiotlb then using pci_alloc_consistent, but that can tie up a lot of precious 16MB GFPDMA/IOMMU/swiotlb memory (even on AMD systems the IOMMU tends to be quite small) especially if you have many devices. With the new zone pci_alloc_consistent can just put this stuff into memory below 4GB which works better.
One argument was still if the zone should be 4GB or 2GB. The main motivation for 2GB would be an unnamed not so unpopular hardware raid controller (mostly found in older machines from a particular four letter company) who has a strange 2GB restriction in firmware. But
that one works ok with swiotlb/IOMMU anyways, so it doesn't really need GFP_DMA32. I chose 4GB to be compatible with IA64 and because it seems to be the most common restriction.
The new zone is so far added only for x86-64.
For other architectures who don't set up this
new zone nothing changes. Architectures can set a compatibility define in Kconfig CONFIG_DMA_IS_DMA32 that will define GFP_DMA32 as GFP_DMA. Otherwise it's a nop because on 32bit architectures it's normally not needed because GFP_NORMAL (=0) is DMA able enough.
One problem is still that GFP_DMA means different things on different architectures. e.g. some drivers used to have #ifdef ia64 use GFP_DMA (trusting it to be 4GB) #elif __x86_64__ (use other hacks like the swiotlb because 16MB is not enough) ... . This was quite ugly and is now obsolete.
These should be now converted to use GFP_DMA32 unconditionally. I haven't done this yet. Or best only use pci_alloc_consistent/dma_alloc_coherent which will use GFP_DMA32 transparently.
arch/x86_64/mm/init.c +42 -23
@@ -318,32 +318,51 @@ void zap_low_mappings(void)flush_tlb_all (); } +/* Compute zone sizes for the DMA and DMA32 zones in a node. */ +__init void +size_zones (unsigned long *z,unsigned long *h, +unsigned long start_pfn,unsigned long end_pfn) +{ +int i; +unsigned long w; + +for (i =0 ; i <MAX_NR_ZONES ; i++) + z[i] =0 ; + +if (start_pfn <MAX_DMA_PFN ) + z[ZONE_DMA ] =MAX_DMA_PFN - start_pfn; +if (start_pfn <MAX_DMA32_PFN ) { +unsigned long dma32_pfn =MAX_DMA32_PFN ; +if (dma32_pfn > end_pfn) + dma32_pfn = end_pfn; + z[ZONE_DMA32 ] = dma32_pfn - start_pfn; + } + z[ZONE_NORMAL ] = end_pfn - start_pfn; + +/* Remove lower zones from higher ones. */ + w =0 ; +for (i =0 ; i <MAX_NR_ZONES ; i++) { +if (z[i]) + z[i] -= w; + w += z[i]; + } + +/* Compute holes */ + w =0 ; +for (i =0 ; i <MAX_NR_ZONES ; i++) { +unsigned long s = w; + w += z[i]; + h[i] =e820_hole_size (s, w); + } +} +#ifndef CONFIG_NUMA void __init paging_init (void ) { - { -unsigned long zones_size[MAX_NR_ZONES ]; -unsigned long holes[MAX_NR_ZONES ]; -unsigned int max_dma; - -memset (zones_size,0 ,sizeof (zones_size)); -memset (holes,0 ,sizeof (holes)); - - max_dma =virt_to_phys ((char *)MAX_DMA_ADDRESS ) >>PAGE_SHIFT ; - -if (end_pfn < max_dma) { - zones_size[ZONE_DMA ] = end_pfn; - holes[ZONE_DMA ] =e820_hole_size (0 , end_pfn); - }else { - zones_size[ZONE_DMA ] = max_dma; - holes[ZONE_DMA ] =e820_hole_size (0 , max_dma); - zones_size[ZONE_NORMAL ] = end_pfn - max_dma; - holes[ZONE_NORMAL ] =e820_hole_size (max_dma, end_pfn); - } -free_area_init_node (0 ,NODE_DATA (0 ), zones_size, -__pa (PAGE_OFFSET ) >>PAGE_SHIFT , holes); - } -return ; +unsigned long zones[MAX_NR_ZONES ], holes[MAX_NR_ZONES ]; +size_zones (zones, holes,0 , end_pfn); +free_area_init_node (0 ,NODE_DATA (0 ), zones, +__pa (PAGE_OFFSET ) >>PAGE_SHIFT , holes); }#endif
arch/x86_64/mm/numa.c +5 -20
@@ -132,29 +132,14 @@ void __init setup_node_zones(int nodeid)unsigned long start_pfn, end_pfn;unsigned long zones[MAX_NR_ZONES ];unsigned long holes[MAX_NR_ZONES ]; -unsigned long dma_end_pfn; -memset (zones,0 ,sizeof (unsigned long ) *MAX_NR_ZONES ); -memset (holes,0 ,sizeof (unsigned long ) *MAX_NR_ZONES ); + start_pfn =node_start_pfn (nodeid); + end_pfn =node_end_pfn (nodeid); - start_pfn =node_start_pfn (nodeid); - end_pfn =node_end_pfn (nodeid); +Dprintk (KERN_INFO "setting up node %d %lx-%lx\n" , + nodeid, start_pfn, end_pfn); -Dprintk (KERN_INFO "setting up node %d %lx-%lx\n" , nodeid, start_pfn, end_pfn); - -/* All nodes > 0 have a zero length zone DMA */ - dma_end_pfn =__pa (MAX_DMA_ADDRESS ) >>PAGE_SHIFT ; -if (start_pfn < dma_end_pfn) { - zones[ZONE_DMA ] = dma_end_pfn - start_pfn; - holes[ZONE_DMA ] =e820_hole_size (start_pfn, dma_end_pfn); - zones[ZONE_NORMAL ] = end_pfn - dma_end_pfn; - holes[ZONE_NORMAL ] =e820_hole_size (dma_end_pfn, end_pfn); - - }else { - zones[ZONE_NORMAL ] = end_pfn - start_pfn; - holes[ZONE_NORMAL ] =e820_hole_size (start_pfn, end_pfn); - } - +size_zones (zones, holes, start_pfn, end_pfn);free_area_init_node (nodeid,NODE_DATA (nodeid), zones, start_pfn, holes); }
include/asm-x86_64/dma.h +9 -2
@@ -72,8 +72,15 @@#define MAX_DMA_CHANNELS 8 -/* The maximum address that we can perform a DMA transfer to on this platform */ -#define MAX_DMA_ADDRESS (PAGE_OFFSET+0x1000000) + +/* 16MB ISA DMA zone */ +#define MAX_DMA_PFN ((16*1024*1024) >> PAGE_SHIFT) + +/* 4GB broken PCI/AGP hardware bus master zone */ +#define MAX_DMA32_PFN ((4UL*1024*1024*1024) >> PAGE_SHIFT) + +/* Compat define for old dma zone */ +#define MAX_DMA_ADDRESS ((unsigned long)__va(MAX_DMA_PFN << PAGE_SHIFT))/* 8237 DMA controllers */ #define IO_DMA1_BASE 0x00/* 8 bit slave DMA, channels 0..3 */
include/asm-x86_64/proto.h +2 -0
@@ -23,6 +23,8 @@ extern void mtrr_bp_init(void);#define mtrr_bp_init () do {} while (0)#endif extern void init_memory_mapping (unsigned long start,unsigned long end); +extern void size_zones (unsigned long *z,unsigned long *h, +unsigned long start_pfn,unsigned long end_pfn);extern void system_call (void );extern int kernel_syscall (void );
include/linux/gfp.h +9 -0
@@ -14,6 +14,13 @@ struct vm_area_struct;/* Zone modifiers in GFP_ZONEMASK (see linux/mmzone.h - low two bits) */ #define __GFP_DMA 0x01u#define __GFP_HIGHMEM 0x02u +#ifdef CONFIG_DMA_IS_DMA32 +#define __GFP_DMA32 0x01/* ZONE_DMA is ZONE_DMA32 */ +#elif BITS_PER_LONG <64 +#define __GFP_DMA32 0x00/* ZONE_NORMAL is ZONE_DMA32 */ +#else +#define __GFP_DMA32 0x04/* Has own ZONE_DMA32 */ +#endif /* * Action modifiers - doesn't change the zoning
@@ -64,6 +71,8 @@ struct vm_area_struct;#define GFP_DMA __GFP_DMA +/* 4GB DMA on some platforms */ +#define GFP_DMA32 __GFP_DMA32/* * There is only one page-allocator function, and two main namespaces to
include/linux/mmzone.h +10 -8
@@ -70,11 +70,12 @@ struct per_cpu_pageset {#endif #define ZONE_DMA 0 -#define ZONE_NORMAL 1 -#define ZONE_HIGHMEM 2 +#define ZONE_DMA32 1 +#define ZONE_NORMAL 2 +#define ZONE_HIGHMEM 3 -#define MAX_NR_ZONES 3/* Sync this with ZONES_SHIFT */ -#define ZONES_SHIFT 2/* ceil(log2(MAX_NR_ZONES)) */ +#define MAX_NR_ZONES 4/* Sync this with ZONES_SHIFT */ +#define ZONES_SHIFT 3/* ceil(log2(MAX_NR_ZONES)) */ /*
@@ -90,7 +91,7 @@ struct per_cpu_pageset {* be 8 (2 ** 3) zonelists. GFP_ZONETYPES defines the number of possible * combinations of zone modifiers in "zone modifier space". */ -#define GFP_ZONEMASK 0x03 +#define GFP_ZONEMASK 0x07/* * As an optimisation any zone modifier bits which are only valid when * no other zone modifier bits are set (loners) should be placed in
@@ -110,6 +111,7 @@ struct per_cpu_pageset {* into multiple physical zones. On a PC we have 3 zones: * * ZONE_DMA < 16 MB ISA DMA capable memory +* ZONE_DMA32 0 MB Empty * ZONE_NORMAL 16-896 MB direct mapped by the kernel * ZONE_HIGHMEM > 896 MB only page cache and user processes */
@@ -428,10 +430,10 @@ extern struct pglist_data contig_page_da#if BITS_PER_LONG ==32 || defined(ARCH_HAS_ATOMIC_UNSIGNED )/* -* with 32 bit page->flags field, we reserve 8 bits for node/zone info. -* there are 3 zones (2 bits) and this leaves 8-2=6 bits for nodes. +* with 32 bit page->flags field, we reserve 9 bits for node/zone info. +* there are 4 zones (3 bits) and this leaves 9-3=6 bits for nodes. */ -#define FLAGS_RESERVED 8 +#define FLAGS_RESERVED 9#elif BITS_PER_LONG ==64 /*
mm/page_alloc.c +15 -4
@@ -58,8 +58,11 @@ long nr_swap_pages;* NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA +* +* TBD: should special case ZONE_DMA32 machines here - in those we normally +* don't need any ZONE_NORMAL reservation */ -int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES -1 ] = {256 ,32 }; +int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES -1 ] = {256 ,256 ,32 };EXPORT_SYMBOL (totalram_pages);EXPORT_SYMBOL (nr_swap_pages);
@@ -71,7 +74,7 @@ EXPORT_SYMBOL(nr_swap_pages);struct zone *zone_table[1 <<ZONETABLE_SHIFT ]__read_mostly ;EXPORT_SYMBOL (zone_table); -static char *zone_names[MAX_NR_ZONES ] = {"DMA" ,"Normal" ,"HighMem" }; +static char *zone_names[MAX_NR_ZONES ] = {"DMA" ,"DMA32" ,"Normal" ,"HighMem" };int min_free_kbytes =1024 ;unsigned long __initdata nr_kernel_pages;
@@ -1418,6 +1421,10 @@ static int __init build_zonelists_node(p zone = pgdat->node_zones +ZONE_NORMAL ;if (zone->present_pages) zonelist->zones[j++] = zone; +case ZONE_DMA32 : + zone = pgdat->node_zones +ZONE_DMA32 ; +if (zone->present_pages) + zonelist->zones[j++] = zone;case ZONE_DMA : zone = pgdat->node_zones +ZONE_DMA ;if (zone->present_pages)
@@ -1526,6 +1533,8 @@ static void __init build_zonelists(pg_da k =ZONE_NORMAL ;if (i & __GFP_HIGHMEM) k =ZONE_HIGHMEM ; +if (i & __GFP_DMA32) + k =ZONE_DMA32 ;if (i & __GFP_DMA) k =ZONE_DMA ;
@@ -1550,7 +1559,9 @@ static void __init build_zonelists(pg_da j =0 ; k =ZONE_NORMAL ;if (i & __GFP_HIGHMEM) - k =ZONE_HIGHMEM ; + k =ZONE_HIGHMEM ; +if (i & __GFP_DMA32) + k =ZONE_DMA32 ;if (i & __GFP_DMA) k =ZONE_DMA ;
@@ -1895,7 +1906,7 @@ static void __init free_area_init_core(sif (zholes_size) realsize -= zholes_size[j]; -if (j ==ZONE_DMA || j ==ZONE_NORMAL ) +if (j <ZONE_HIGHMEM ) nr_kernel_pages += realsize; nr_all_pages += realsize;
It seems a little strange to add individual zones one by one. I remember from an OS project I previously worked on that at some time our driver developers ran into one or more devices that were able to consume 31-bit physical addresses (but not 32-bit ones, and don't ask me for details on what exact devices these were, I never knew). I thus wonder whether it wouldn't make more sense to generalize the logic and allow drivers to specify to the allocator how many physical address bits they can deal with.
Jan
Quoted 1 line
Quoted 1 line
Quoted 1 line
"Andi Kleen" <[email protected]> 11.09.05 18:59:19 >>>
Add 4GB DMA32 zone
Add a new 4GB GFP_DMA32 zone between the GFP_DMA and GFP_NORMAL zones.
As a bit of historical background: when the x86-64 port was originally designed we had some discussion if we should use a 16MB DMA zone like i386 or a 4GB DMA zone like IA64 or both. Both was ruled out at this point because it was in early 2.4 when VM is still quite shakey and had bad troubles even dealing with one DMA zone. We settled on the 16MB DMA zone mainly because we worried about older soundcards and the floppy.
But this has always caused problems since then because device drivers had trouble getting enough DMA able memory. These days the VM works much better and the wide use of NUMA has proven it can deal with many zones successfully.
So this patch adds both zones.
This helps drivers who need a lot of memory below 4GB because their hardware is not accessing more (graphic drivers - proprietary and free ones, video frame buffer drivers, sound drivers etc.). Previously they could only use IOMMU+16MB GFP_DMA, which was not enough memory.
Another common problem is that hardware who has full memory addressing for >4GB misses it for some control structures in memory (like transmit rings or other metadata). They tended to allocate memory
in the 16MB GFP_DMA or the IOMMU/swiotlb then using pci_alloc_consistent,
but that can tie up a lot of precious 16MB GFPDMA/IOMMU/swiotlb memory
(even on AMD systems the IOMMU tends to be quite small) especially if you have
many devices. With the new zone pci_alloc_consistent can just put this stuff into memory below 4GB which works better.
One argument was still if the zone should be 4GB or 2GB. The main motivation for 2GB would be an unnamed not so unpopular hardware raid controller (mostly found in older machines from a particular four letter
company) who has a strange 2GB restriction in firmware. But that one works ok with swiotlb/IOMMU anyways, so it doesn't really need GFP_DMA32. I chose 4GB to be compatible with IA64 and because it seems to be the most common restriction.
The new zone is so far added only for x86-64.
For other architectures who don't set up this
new zone nothing changes. Architectures can set a compatibility define in Kconfig CONFIG_DMA_IS_DMA32 that will define GFP_DMA32 as GFP_DMA. Otherwise it's a nop because on 32bit architectures it's normally not needed because GFP_NORMAL (=0) is DMA able enough.
One problem is still that GFP_DMA means different things on different architectures. e.g. some drivers used to have #ifdef ia64 use GFP_DMA (trusting it to be 4GB) #elif __x86_64__ (use other hacks like the swiotlb because 16MB is not enough) ... . This was quite ugly and is now obsolete.
These should be now converted to use GFP_DMA32 unconditionally. I haven't done
this yet. Or best only use pci_alloc_consistent/dma_alloc_coherent which will use GFP_DMA32 transparently.
arch/x86_64/mm/init.c +38 -9
@@ -318,32 +318,51 @@ void zap_low_mappings(void)flush_tlb_all (); } +/* Compute zone sizes for the DMA and DMA32 zones in a node. */ +__init void +size_zones (unsigned long *z,unsigned long *h, +unsigned long start_pfn,unsigned long end_pfn) +{ +int i; +unsigned long w; + +for (i =0 ; i <MAX_NR_ZONES ; i++) + z[i] =0 ; + +if (start_pfn <MAX_DMA_PFN ) + z[ZONE_DMA ] =MAX_DMA_PFN - start_pfn; +if (start_pfn <MAX_DMA32_PFN ) { +unsigned long dma32_pfn =MAX_DMA32_PFN ; +if (dma32_pfn > end_pfn) + dma32_pfn = end_pfn; + z[ZONE_DMA32 ] = dma32_pfn - start_pfn; + } + z[ZONE_NORMAL ] = end_pfn - start_pfn; + +/* Remove lower zones from higher ones. */ + w =0 ; +for (i =0 ; i <MAX_NR_ZONES ; i++) { +if (z[i]) + z[i] -= w; + w += z[i]; + } + +/* Compute holes */ + w =0 ; +for (i =0 ; i <MAX_NR_ZONES ; i++) { +unsigned long s = w; + w += z[i]; + h[i] =e820_hole_size (s, w); + } +} +#ifndef CONFIG_NUMA void __init paging_init (void ) { - { -unsigned long zones_size[MAX_NR_ZONES ]; -unsigned long holes[MAX_NR_ZONES ]; -unsigned int max_dma; - -memset (zones_size,0 ,sizeof (zones_size)); -memset (holes,0 ,sizeof (holes)); - - max_dma = virt_to_phys((char *)MAX_DMA_ADDRESS) >>
PAGE_SHIFT;
-
- if (end_pfn < max_dma) {
- zones_size[ZONE_DMA] = end_pfn;
- holes[ZONE_DMA] = e820_hole_size(0, end_pfn);
- } else {
- zones_size[ZONE_DMA] = max_dma;
- holes[ZONE_DMA] = e820_hole_size(0, max_dma);
- zones_size[ZONE_NORMAL] = end_pfn - max_dma;
- holes[ZONE_NORMAL] = e820_hole_size(max_dma,end_pfn);
- }
- free_area_init_node(0, NODE_DATA(0), zones_size,
- __pa(PAGE_OFFSET) >> PAGE_SHIFT, holes);
- }
- return;
+ unsigned long zones[MAX_NR_ZONES], holes[MAX_NR_ZONES];
+ size_zones(zones, holes, 0, end_pfn);
+ free_area_init_node(0, NODE_DATA(0), zones,
+ __pa(PAGE_OFFSET) >> PAGE_SHIFT, holes);
}
#endif
arch/x86_64/mm/numa.c +4 -6
@@ -132,29 +132,14 @@ void __init setup_node_zones(int nodeid)unsigned long start_pfn, end_pfn;unsigned long zones[MAX_NR_ZONES ];unsigned long holes[MAX_NR_ZONES ]; -unsigned long dma_end_pfn; -memset (zones,0 ,sizeof (unsigned long ) *MAX_NR_ZONES ); -memset (holes,0 ,sizeof (unsigned long ) *MAX_NR_ZONES ); + start_pfn =node_start_pfn (nodeid); + end_pfn =node_end_pfn (nodeid); - start_pfn =node_start_pfn (nodeid); - end_pfn =node_end_pfn (nodeid); +Dprintk (KERN_INFO "setting up node %d %lx-%lx\n" , + nodeid, start_pfn, end_pfn); - Dprintk(KERN_INFO "setting up node %d %lx-%lx\n", nodeid,
start_pfn, end_pfn);
-
- /* All nodes > 0 have a zero length zone DMA */
- dma_end_pfn = __pa(MAX_DMA_ADDRESS) >> PAGE_SHIFT;
- if (start_pfn < dma_end_pfn) {
- zones[ZONE_DMA] = dma_end_pfn - start_pfn;
- holes[ZONE_DMA] = e820_hole_size(start_pfn,dma_end_pfn);
- zones[ZONE_NORMAL] = end_pfn - dma_end_pfn;
- holes[ZONE_NORMAL] = e820_hole_size(dma_end_pfn,end_pfn);
-
- } else {
- zones[ZONE_NORMAL] = end_pfn - start_pfn;
- holes[ZONE_NORMAL] = e820_hole_size(start_pfn,end_pfn);
- }
-
+ size_zones(zones, holes, start_pfn, end_pfn);
free_area_init_node(nodeid, NODE_DATA(nodeid), zones,
start_pfn, holes);
}
include/asm-x86_64/dma.h +0 -1
@@ -72,8 +72,15 @@#define MAX_DMA_CHANNELS 8 -/* The maximum address that we can perform a DMA transfer to on this
platform */
-#define MAX_DMA_ADDRESS (PAGE_OFFSET+0x1000000)
+
+/* 16MB ISA DMA zone */
+#define MAX_DMA_PFN ((16*1024*1024) >> PAGE_SHIFT)
+
+/* 4GB broken PCI/AGP hardware bus master zone */ +#define MAX_DMA32_PFN ((4UL*1024*1024*1024) >> PAGE_SHIFT) +
+/* Compat define for old dma zone */
+#define MAX_DMA_ADDRESS ((unsigned long)__va(MAX_DMA_PFN << PAGE_SHIFT))
/* 8237 DMA controllers */
#define IO_DMA1_BASE 0x00 /* 8 bit slave DMA, channels 0..3 */
include/asm-x86_64/proto.h +0 -0
@@ -23,6 +23,8 @@ extern void mtrr_bp_init(void);#define mtrr_bp_init () do {} while (0)#endif extern void init_memory_mapping(unsigned long start, unsigned long
end);
+extern void size_zones(unsigned long *z, unsigned long *h,
+ unsigned long start_pfn, unsigned longend_pfn);
extern void system_call(void);
extern int kernel_syscall(void);
include/linux/gfp.h +0 -0
@@ -14,6 +14,13 @@ struct vm_area_struct;/* Zone modifiers in GFP_ZONEMASK (see linux/mmzone.h - low two bits)
*/
#define __GFP_DMA 0x01u
#define __GFP_HIGHMEM 0x02u
+#ifdef CONFIG_DMA_IS_DMA32
+#define __GFP_DMA32 0x01 /* ZONE_DMA is ZONE_DMA32 */
+#elif BITS_PER_LONG < 64
+#define __GFP_DMA32 0x00 /* ZONE_NORMAL is ZONE_DMA32 */
+#else
+#define __GFP_DMA32 0x04 /* Has own ZONE_DMA32 */
+#endif
/*
* Action modifiers - doesn't change the zoning
@@ -64,6 +71,8 @@ struct vm_area_struct;
#define GFP_DMA __GFP_DMA
+/* 4GB DMA on some platforms */
+#define GFP_DMA32 __GFP_DMA32
/*
* There is only one page-allocator function, and two main namespaces
to
include/linux/mmzone.h +3 -3
@@ -70,11 +70,12 @@ struct per_cpu_pageset {#endif #define ZONE_DMA 0 -#define ZONE_NORMAL 1 -#define ZONE_HIGHMEM 2 +#define ZONE_DMA32 1 +#define ZONE_NORMAL 2 +#define ZONE_HIGHMEM 3 -#define MAX_NR_ZONES 3/* Sync this with ZONES_SHIFT
*/
-#define ZONES_SHIFT 2 /* ceil(log2(MAX_NR_ZONES)) */
+#define MAX_NR_ZONES 4 /* Sync this with ZONES_SHIFT
*/
+#define ZONES_SHIFT 3 /* ceil(log2(MAX_NR_ZONES)) */
/*
@@ -90,7 +91,7 @@ struct per_cpu_pageset {
* be 8 (2 ** 3) zonelists. GFP_ZONETYPES defines the number of
possible
* combinations of zone modifiers in "zone modifier space". */
-#define GFP_ZONEMASK 0x03
+#define GFP_ZONEMASK 0x07
/*
* As an optimisation any zone modifier bits which are only valid
when
* no other zone modifier bits are set (loners) should be placed in
@@ -110,6 +111,7 @@ struct per_cpu_pageset {
* into multiple physical zones. On a PC we have 3 zones:
*
* ZONE_DMA < 16 MB ISA DMA capable memory
+ * ZONE_DMA32 0 MB Empty
* ZONE_NORMAL 16-896 MB direct mapped by the kernel
* ZONE_HIGHMEM > 896 MB only page cache and user
processes
*/
@@ -428,10 +430,10 @@ extern struct pglist_data contig_page_da
#if BITS_PER_LONG == 32 || defined(ARCH_HAS_ATOMIC_UNSIGNED) /*
- * with 32 bit page->flags field, we reserve 8 bits for node/zone info.
- * there are 3 zones (2 bits) and this leaves 8-2=6 bits for nodes.
+ * with 32 bit page->flags field, we reserve 9 bits for node/zone info.
+ * there are 4 zones (3 bits) and this leaves 9-3=6 bits for nodes. */
-#define FLAGS_RESERVED 8
+#define FLAGS_RESERVED 9
#elif BITS_PER_LONG == 64
/*
mm/page_alloc.c +0 -0
@@ -58,8 +58,11 @@ long nr_swap_pages;* NORMAL allocation will leave 784M/256 of ram reserved in the
ZONE_DMA
* HIGHMEM allocation will leave 224M/32 of ram reserved inZONE_NORMAL
* HIGHMEM allocation will (224M+784M)/256 of ram reserved inZONE_DMA
+ *
+ * TBD: should special case ZONE_DMA32 machines here - in those we normally
+ * don't need any ZONE_NORMAL reservation
*/
-int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = { 256, 32 }; +int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = { 256, 256, 32 };
EXPORT_SYMBOL(totalram_pages); EXPORT_SYMBOL(nr_swap_pages);
@@ -71,7 +74,7 @@ EXPORT_SYMBOL(nr_swap_pages);
struct zone *zone_table[1 << ZONETABLE_SHIFT] __read_mostly; EXPORT_SYMBOL(zone_table);
-static char *zone_names[MAX_NR_ZONES] = { "DMA", "Normal", "HighMem" };
+static char *zone_names[MAX_NR_ZONES] = { "DMA", "DMA32", "Normal", "HighMem" };
int min_free_kbytes = 1024;
unsigned long __initdata nr_kernel_pages;
@@ -1418,6 +1421,10 @@ static int __init build_zonelists_node(p
zone = pgdat->node_zones + ZONE_NORMAL; if (zone->present_pages)
zonelist->zones[j++] = zone;
+ case ZONE_DMA32:
+ zone = pgdat->node_zones + ZONE_DMA32;
+ if (zone->present_pages)
+ zonelist->zones[j++] = zone;
case ZONE_DMA:
zone = pgdat->node_zones + ZONE_DMA; if (zone->present_pages)
@@ -1526,6 +1533,8 @@ static void __init build_zonelists(pg_da
k = ZONE_NORMAL;
if (i & __GFP_HIGHMEM)
k = ZONE_HIGHMEM;
+ if (i & __GFP_DMA32)
+ k = ZONE_DMA32;
if (i & __GFP_DMA)
k = ZONE_DMA;
@@ -1550,7 +1559,9 @@ static void __init build_zonelists(pg_da
j = 0;
k = ZONE_NORMAL;
if (i & __GFP_HIGHMEM)
- k = ZONE_HIGHMEM;
+ k = ZONE_HIGHMEM;
+ if (i & __GFP_DMA32)
+ k = ZONE_DMA32;
if (i & __GFP_DMA)
k = ZONE_DMA;
@@ -1895,7 +1906,7 @@ static void __init free_area_init_core(s
if (zholes_size)
realsize -= zholes_size[j];
- if (j == ZONE_DMA || j == ZONE_NORMAL)
+ if (j < ZONE_HIGHMEM)
nr_kernel_pages += realsize;
nr_all_pages += realsize;
Quoted 4 lines · Jan Beulich
It seems a little strange to add individual zones one by one. I remember from an OS project I previously worked on that at some time our driver developers ran into one or more devices that were able to consume 31-bit physical addresses
That's likely the unnamed RAID controller with the broken firmware refered to below (they might actually have fixed the firmware now) But for block devices it's not really needed anyways. From my experience and those of other folks (IA64) the 4GB zone + a small fallback zone is a good compromise.
Quoted 5 lines
(but not 32-bit ones, and don't ask me for details on
what exact devices these were, I never knew). I thus wonder whether it wouldn't make more sense to generalize the logic and allow drivers to specify to the allocator how many physical address bits they can deal with.
Because that would likely either impact the page allocation fast path by having unsuited data structures for the normal case or make the code to allocate pages with arbitary boundaries really slow because the allocation wouldn't be O(1). Didn't seem like a good tradeoff.
-Andi
Quoted 4 lines
One argument was still if the zone should be 4GB or 2GB. The main motivation for 2GB would be an unnamed not so unpopular hardware raid controller (mostly found in older machines from a particular four letter company) who has a strange 2GB restriction in firmware. But
Adaptec AACRAID is one offender
Quoted 1 line
that one works ok with swiotlb/IOMMU anyways, so it doesn't really
Old aacraid actually cannot use IOMMU. It isn't alone in that limitation. Most hardware that has a 30/31bit limit can't go via the IOMMU because IOMMU space appears on the bus above 2GB so is itself invisible to the hardware.
Other devices with similar limits include the broadcomm b44
Quoted 6 lines · Alan Cox
Quoted 4 lines
One argument was still if the zone should be 4GB or 2GB. The main motivation for 2GB would be an unnamed not so unpopular hardware raid controller (mostly found in older machines from a particular four letter company) who has a strange 2GB restriction in firmware. But
Adaptec AACRAID is one offender
Yes, that one was considered. Believe me we went over all the broken hardware cases quite a lot before comming up with this patch.
I refuse to add unnecessary limits to everybody else just because of that single broken firmware. 4GB limit is really common and the oddballs like these have to use the same workarounds (custom bounce buffer in low GFP_DMA memory) they always did on machines with enough memory.
Also the aacraid is not really an big issue on x86-64, because afaik nobody shipped EM64T or AMD64 machines with these beasts. (most aacraid is older Xeons without 64bit capability). There are a few users who bought plugin cards later, but near all of these ran into other problems because they didn't have enough memory (I cannot in fact remember a report of someone running especially into this problem) And the cards seem to be essentially dead in the market now. So it's really more a theoretical problem than a practical one. [Proof of it: the current sources don't seem to handle it, so it cannot be that bad ;-]
And the probability of someone using a b44 in a machine with >2GB of memory is small at best. Usually they are in really lowend boxes where you couldn't even plug in more memory than that. That is why I essentially ignored the b44. AFAIK the driver has a GFP_DMA bounce workaround anyways, so it would work anyways.
Yes I know some soundcards have similar limits, but for all these we still have GFP_DMA and they always have been quite happy with that.
Basically this a solution to make a lot of common hardware happy and the oddballs and really broken cases are not worse than they have been before.
Quoted 6 lines
Quoted 1 line
that one works ok with swiotlb/IOMMU anyways, so it doesn't really
Old aacraid actually cannot use IOMMU. It isn't alone in that limitation. Most hardware that has a 30/31bit limit can't go via the IOMMU because IOMMU space appears on the bus above 2GB so is itself invisible to the hardware.
Yes, true. Use GFP_DMA then.
Actually swiotlb would work in theory because it tends to be pretty low, but that is not enabled on all machines and the code doesn't attempt to handle it (and I don't plan to do it)
Hopefully the patch can go into 2.6.13.
-Andi
Quoted 5 lines · Andi Kleen
(I cannot in fact remember a report of someone running especially into this problem) And the cards seem to be essentially dead in the market now. So it's really more a theoretical problem than a practical one. [Proof of it: the current sources don't seem to handle it, so it cannot be that bad ;-]
Current sources don't handle DMA32, so that can't be bad either. Can we stick to sensible discussion, its quicker.
There have been various reports over time, quite a few early on when we had a long series of on list discussions trying to debug what appeared to be an iommu bug but was in fact a kernel bug.
You hit it on any size AMD64 with iommu, but since most aacraid users are intel boxes it doesn't hurt too many, and the rest all know about turning the iommu off on the box (but that hurts the rest), or just run something else because Linux "doesn't work".
Quoted 3 lines
That is why I essentially ignored the b44. AFAIK the driver has a
GFP_DMAbounce workaround anyways, so it would work anyways.
Usually - the DMA zone at 16MB is too small so allocations sometimes fail. It btw would want 1GB limits.
Quoted 3 lines
Yes I know some soundcards have similar limits, but for all these we still have
GFP_DMAand they always have been quite happy with that.
No current shipping card, also those that need it typically need small amounts (they'll live with 8K)
Quoted 6 lines
Quoted 4 lines
Old aacraid actually cannot use IOMMU. It isn't alone in that limitation. Most hardware that has a 30/31bit limit can't go via the IOMMU because IOMMU space appears on the bus above 2GB so is itself invisible to the hardware.
Yes, true. Use
GFP_DMAthen.
Doesn't work. The DMA area is way too small with all the users already and the aacraid can and does want a lot of outstanding I/O. Using a 1GB or 2GB boundary line hurts nobody doing "allocate me some memory below 4Gb" because nobody asks for .5GB chunks. A 4GB zone means we need to either increase the 16MB zone or add yet another one.
Alan
Quoted 9 lines · Alan Cox
Quoted 5 lines · Andi Kleen
(I cannot in fact remember a report of someone running especially into this problem) And the cards seem to be essentially dead in the market now. So it's really more a theoretical problem than a practical one. [Proof of it: the current sources don't seem to handle it, so it cannot be that bad ;-]
Current sources don't handle DMA32, so that can't be bad either. Can we stick to sensible discussion, its quicker.
Well discussing broken hardware that is rarely used on 64bit systems with enough memory doesn't seem very sensible to me.
But ok:
Even a 2GB DMA32 wouldn't magically fix it anyways.
Quoted 9 lines
There have been various reports over time, quite a few early on when we had a long series of on list discussions trying to debug what appeared to be an iommu bug but was in fact a kernel bug.
You hit it on any size AMD64 with iommu, but since most aacraid users are intel boxes it doesn't hurt too many, and the rest all know about turning the iommu off on the box (but that hurts the rest), or just run something else because Linux "doesn't work".
You need essentially the same code to fix it with GFP_DMA32 as with GFP_DMA. Some bounce code that allocates low memory and does the necessary bounces.
The only difference with GFP_DMA is that the bounce pool is smaller
Limiting everybody else just to get a bigger bounce pool on a single broken device that isn't even shipping anymore doesn't seem like sensible design approach to me.
Quoted 7 lines
Quoted 4 lines
That is why I essentially ignored the b44. AFAIK the driver has a
GFP_DMAbounce workaround anyways, so it would work anyways.Usually - the DMA zone at 16MB is too small so allocations sometimes fail. It btw would want 1GB limits.
mempool - sleep on a waitqueue until someone frees. In fact the kernel's normal allocator is already quite good at that so you might not even need that.
Quoted 6 lines
Doesn't work. The DMA area is way too small with all the users already and the aacraid can and does want a lot of outstanding I/O. Using a 1GB or 2GB boundary line hurts nobody doing "allocate me some memory below 4Gb" because nobody asks for .5GB chunks. A 4GB zone means we need to either increase the 16MB zone or add yet another one.
This is really only a problem if a significant fraction of your memory is beyond the limit (otherwise most buffers can go directly).
And with the mempool sleep approach they will just get small queues. Yes that will be slower, but if you want performance on boxes with a lot of memory you should not buy broken hardware.
Basically aacraid was always broken and it is not more or not less broken than it was before with DMA32.
-Andi
Quoted 3 lines · Andi Kleen
And with the mempool sleep approach they will just get small queues. Yes that will be slower, but if you want performance on boxes with a lot of memory you should not buy broken hardware.
Ironically its broken hardware it works best with. AMD64 is problematic but Intel with the swiotlb works ;)
Ok - points made anyway, if you think the 4GB one is the best way to do it even considering these then I've no problem with that.
Quoted 7 lines · Alan Cox
Quoted 3 lines · Andi Kleen
And with the mempool sleep approach they will just get small queues. Yes that will be slower, but if you want performance on boxes with a lot of memory you should not buy broken hardware.
Ironically its broken hardware it works best with. AMD64 is problematic but Intel with the swiotlb works ;)
Actually the swiotlb code currently doesn't attempt to handle dma masks <4GB even when the bounce pool happens to be located lower - it will just fail or use GFP_DMA. It could be fixed in theory, but it would be pretty unreliable and sometimes work on one system and sometimes not, so I would be reluctant to go down that path.
Also BTW on many systems which don't allocate the IOMMU aperture in BIOS and Linux has to allocate it over memory it tends to be as low (or high) as the swiotlb pool - it is bootmem allocated at roughly the same place in boot. But again the code doesn't attempt to handle that, it just uses hardcoded 0xffffffff masks.
-Andi
Quoted 8 lines · Alan Cox
Quoted 3 lines · Andi Kleen
Yes I know some soundcards have similar limits, but for all these we still have
GFP_DMAand they always have been quite happy with that.No current shipping card, also those that need it typically need small amounts (they'll live with 8K)
[...just because I love broken hardware, not because I've been following this thread...]
RealTek's ALS4000 PCI card is a SoundBlaster ISA clone chip glued onto a PCI bus. Its DMA mask is 24-bit, IIRC. :)
Jeff
Quoted 15 lines · Jeff Garzik
Quoted 10 lines · Alan Cox
Quoted 3 lines · Andi Kleen
Yes I know some soundcards have similar limits, but for all these we still have
GFP_DMAand they always have been quite happy with that.No current shipping card, also those that need it typically need small amounts (they'll live with 8K)
Quoted 6 lines
[...just because I love broken hardware, not because I've been following this thread...]
RealTek's ALS4000 PCI card is a SoundBlaster ISA clone chip glued onto a PCI bus. Its DMA mask is 24-bit, IIRC. :)
Jeff
Yep. You're absolutely right about the card. Google doesn't find anyone still selling them, though... Apart from ebay ;-)
(Wrote the driver and got rid of the d*mn thing 2 weeks later...)
Cheers,
Bart
Quoted 4 lines · Bart Hartgers
Yep. You're absolutely right about the card. Google doesn't find anyone still selling them, though... Apart from ebay ;-)
(Wrote the driver and got rid of the d*mn thing 2 weeks later...)
Just to avoid any misconceptions - it should just work great on x86-64 because exactly for such hardware we kept the 16MB DMA zone. (unless of course the driver has other problems than just DMAing)
-Andi
Quoted 4 lines · Andi Kleen
And the probability of someone using a b44 in a machine with >2GB of memory is small at best. Usually they are in really lowend boxes where you couldn't even plug in more memory than that.
Data point:
My current model Dell notebook as b44 and 2GB RAM. The 2GB is the limit only until >1GB SODIMMs become available.
Cheers
Hi,
include/linux/mmzone.h struct per_cpu_pageset { · line 78
Quoted 4 lines · Andi Kleen
include/linux/mmzone.h +2
-2-#define MAX_NR_ZONES 3/* Sync this with ZONES_SHIFT */ -#define ZONES_SHIFT 2/* ceil(log2(MAX_NR_ZONES)) */ +#define MAX_NR_ZONES 4/* Sync this with ZONES_SHIFT */ +#define ZONES_SHIFT 3/* ceil(log2(MAX_NR_ZONES)) */
Why needs ZONES_SHIFT to be increased?
include/linux/mmzone.h extern struct pglist_data contig_page_da · line 436
Quoted 2 lines
include/linux/mmzone.h +1
-1-#define FLAGS_RESERVED 8 +#define FLAGS_RESERVED 9
I would prefer to keep this at 8.
bye, Roman
Quoted 10 lines · Roman Zippel
Hi,
include/linux/mmzone.h struct per_cpu_pageset { · line 78
Quoted 4 lines · Andi Kleen
include/linux/mmzone.h +2
-2-#define MAX_NR_ZONES 3/* Sync this with ZONES_SHIFT */ -#define ZONES_SHIFT 2/* ceil(log2(MAX_NR_ZONES)) */ +#define MAX_NR_ZONES 4/* Sync this with ZONES_SHIFT */ +#define ZONES_SHIFT 3/* ceil(log2(MAX_NR_ZONES)) */ Why needs
ZONES_SHIFTto be increased?
Quoted 4 lines
include/linux/mmzone.h extern struct pglist_data contig_page_da · line 436
Quoted 2 lines
include/linux/mmzone.h +1
-1-#define FLAGS_RESERVED 8 +#define FLAGS_RESERVED 9I would prefer to keep this at 8.
sparsemem needs these two.
-Andi
Hi,
Quoted 7 lines · Andi Kleen
Quoted 7 lines
include/linux/mmzone.h struct per_cpu_pageset { · line 78
Quoted 4 lines
include/linux/mmzone.h +2
-2-#define MAX_NR_ZONES 3/* Sync this with ZONES_SHIFT */ -#define ZONES_SHIFT 2/* ceil(log2(MAX_NR_ZONES)) */ +#define MAX_NR_ZONES 4/* Sync this with ZONES_SHIFT */ +#define ZONES_SHIFT 3/* ceil(log2(MAX_NR_ZONES)) */ Why needs
ZONES_SHIFTto be increased?
Quoted 6 lines
Quoted 4 lines
include/linux/mmzone.h extern struct pglist_data contig_page_da · line 436
Quoted 2 lines
include/linux/mmzone.h +1
-1-#define FLAGS_RESERVED 8 +#define FLAGS_RESERVED 9I would prefer to keep this at 8.
sparsemem needs these two.
What two? What are you talking about???
bye, Roman
Quoted 10 lines · Roman Zippel
Hi,
Quoted 7 lines · Andi Kleen
Quoted 7 lines
include/linux/mmzone.h struct per_cpu_pageset { · line 78
Quoted 4 lines
include/linux/mmzone.h +2
-2-#define MAX_NR_ZONES 3/* Sync this with ZONES_SHIFT */ -#define ZONES_SHIFT 2/* ceil(log2(MAX_NR_ZONES)) */ +#define MAX_NR_ZONES 4/* Sync this with ZONES_SHIFT */ +#define ZONES_SHIFT 3/* ceil(log2(MAX_NR_ZONES)) */ Why needs
ZONES_SHIFTto be increased?
Quoted 8 lines
Quoted 6 lines
Quoted 4 lines
include/linux/mmzone.h extern struct pglist_data contig_page_da · line 436
Quoted 2 lines
include/linux/mmzone.h +1
-1-#define FLAGS_RESERVED 8 +#define FLAGS_RESERVED 9I would prefer to keep this at 8.
sparsemem needs these two.
What two? What are you talking about???
The two changes you quoted.
-Andi
Hi,
Quoted 13 lines · Andi Kleen
Quoted 13 lines · Andi Kleen
Quoted 12 lines
Quoted 11 lines
include/linux/mmzone.h struct per_cpu_pageset { · line 78
Quoted 4 lines
include/linux/mmzone.h +2
-2-#define MAX_NR_ZONES 3/* Sync this with ZONES_SHIFT */ -#define ZONES_SHIFT 2/* ceil(log2(MAX_NR_ZONES)) */ +#define MAX_NR_ZONES 4/* Sync this with ZONES_SHIFT */ +#define ZONES_SHIFT 3/* ceil(log2(MAX_NR_ZONES)) */ Why needs
ZONES_SHIFTto be increased?include/linux/mmzone.h extern struct pglist_data contig_page_da · line 436
Quoted 2 lines
include/linux/mmzone.h +1
-1-#define FLAGS_RESERVED 8 +#define FLAGS_RESERVED 9I would prefer to keep this at 8.
Quoted 5 lines
Quoted 3 lines
Quoted 1 line
sparsemem needs these two.
What two? What are you talking about???
The two changes you quoted.
Sorry, but I still don't see why. Could you _please_ be a little more verbose? Thanks.
bye, Roman
Hi,
Quoted 8 lines · Andi Kleen
Quoted 8 lines · Andi Kleen
include/linux/mmzone.h struct per_cpu_pageset { · line 78
Quoted 4 lines
include/linux/mmzone.h +2
-2-#define MAX_NR_ZONES 3/* Sync this with ZONES_SHIFT */ -#define ZONES_SHIFT 2/* ceil(log2(MAX_NR_ZONES)) */ +#define MAX_NR_ZONES 4/* Sync this with ZONES_SHIFT */ +#define ZONES_SHIFT 3/* ceil(log2(MAX_NR_ZONES)) */ Why needs
ZONES_SHIFTto be increased?
Quoted 6 lines
Quoted 4 lines
include/linux/mmzone.h extern struct pglist_data contig_page_da · line 436
Quoted 2 lines
include/linux/mmzone.h +1
-1-#define FLAGS_RESERVED 8 +#define FLAGS_RESERVED 9I would prefer to keep this at 8.
sparsemem needs these two.
Did I somehow offend you, that I don't deserve an answer? The reason for my question is rather simple (and I thought obvious), the four zone types fit into two bits, so what is sparsemem doing with this extra bit?
bye, Roman
Quoted 11 lines · Roman Zippel
Hi,
Quoted 8 lines · Andi Kleen
Quoted 8 lines · Andi Kleen
include/linux/mmzone.h struct per_cpu_pageset { · line 78
Quoted 4 lines
include/linux/mmzone.h +2
-2-#define MAX_NR_ZONES 3/* Sync this with ZONES_SHIFT */ -#define ZONES_SHIFT 2/* ceil(log2(MAX_NR_ZONES)) */ +#define MAX_NR_ZONES 4/* Sync this with ZONES_SHIFT */ +#define ZONES_SHIFT 3/* ceil(log2(MAX_NR_ZONES)) */ Why needs
ZONES_SHIFTto be increased?
Quoted 8 lines
Quoted 6 lines
Quoted 4 lines
include/linux/mmzone.h extern struct pglist_data contig_page_da · line 436
Quoted 2 lines
include/linux/mmzone.h +1
-1-#define FLAGS_RESERVED 8 +#define FLAGS_RESERVED 9I would prefer to keep this at 8.
sparsemem needs these two.
Did I somehow offend you, that I don't deserve an answer?
Well, your aggressive tone definitely doesn't encourage speedy answers.
Quoted 3 lines
The reason for my question is rather simple (and I thought obvious), the four zone types fit into two bits, so what is sparsemem doing with this extra bit?
iirc it was a patch that came in over Andrew. I can't find the email anymore unfortunately. The argument looked plausible and I think it fixed a boot problem for the submitter on some arch (probably IA64, on x86-64 it worked fine without it, but I've never tried sparsemem and the code was originally written before sparsemem). Andrew do you still have the patch with the description? It must have been between 2.6.13mm1 and 2.6.13mm2.
You're right that four zones should in theory fit into 2 bits, so I'm also not sure why it was needed.
-Andi
Quoted 3 lines · Andi Kleen
Andrew do you still have the patch with
the description? It must have been between 2.6.13mm1 and 2.6.13mm2.
From: KAMEZAWA Hiroyuki <[email protected]>
Because 2.6.13-mm2 adds new zone DMA32, ZONES_SHIFT becomes 3. So, flags bits reserved for (SECTION | NODE | ZONE) should be increase.
ZONE_SHIFT is increased, FLAGS_RESERVED should be.
Signed-off-by Kamezawa Hiroyuki <[email protected]>
---
include/linux/mmzone.h | 5 +++-- 1 files changed, 3 insertions(+), 2 deletions(-)
include/linux/mmzone.h renamed from include/linux/mmzone.h~x86_64-dma32-fix +3 -2
@@ -431,9 +431,10 @@ extern struct pglist_data contig_page_da#if BITS_PER_LONG ==32 || defined(ARCH_HAS_ATOMIC_UNSIGNED )/* * with 32 bit page->flags field, we reserve 8 bits for node/zone info. -* there are 3 zones (2 bits) and this leaves 8-2=6 bits for nodes. +* there are 4 zones (3 bits) and this leaves 8-2=6 bits for nodes. +* +6bits for sections if CONFIG_SPARSEMEM */ -#define FLAGS_RESERVED 8 +#define FLAGS_RESERVED 9#elif BITS_PER_LONG ==64 /*
_
Hmm ok description is not very enlightening. 4 zones should indeed still fit into 2 bits.
Kamezawa-san, can you please explain why exactly you did that change?
Thanks,
-Andi
Quoted 36 lines · Andrew Morton
Quoted 2 lines · Andi Kleen
Andrew do you still have the patch with
the description? It must have been between 2.6.13mm1 and 2.6.13mm2.From: KAMEZAWA Hiroyuki <[email protected]>
Because 2.6.13-mm2 adds new zone DMA32,
ZONES_SHIFTbecomes 3. So, flags bits reserved for (SECTION | NODE | ZONE) should be increase.
ZONE_SHIFTis increased,FLAGS_RESERVEDshould be.Signed-off-by Kamezawa Hiroyuki <[email protected]>
Cc Andi Kleen <[email protected]>Signed-off-by Andrew Morton <[email protected]>---
include/linux/mmzone.h | 5 +++-- 1 files changed, 3 insertions(+), 2 deletions(-)diff -puN
include/linux/mmzone.h~x86_64-dma32-fixinclude/linux/mmzone.hinclude/linux/mmzone.h renamed from include/linux/mmzone.h~x86_64-dma32-fix +3
-2@@ -431,9 +431,10 @@ extern struct pglist_data contig_page_da#if BITS_PER_LONG ==32 || defined(ARCH_HAS_ATOMIC_UNSIGNED )/* * with 32 bit page->flags field, we reserve 8 bits for node/zone info. -* there are 3 zones (2 bits) and this leaves 8-2=6 bits for nodes. +* there are 4 zones (3 bits) and this leaves 8-2=6 bits for nodes. +* +6bits for sections if CONFIG_SPARSEMEM */ -#define FLAGS_RESERVED 8 +#define FLAGS_RESERVED 9#elif BITS_PER_LONG ==64 /* _
Hi,
Quoted 4 lines · Andi Kleen
Hmm ok description is not very enlightening. 4 zones should indeed still fit into 2 bits.
Kamezawa-san, can you please explain why exactly you did that change?
Probably because it triggers this check:
#if SECTIONS_WIDTH+NODES_WIDTH+ZONES_WIDTH > FLAGS_RESERVED #error SECTIONS_WIDTH+NODES_WIDTH+ZONES_WIDTH > FLAGS_RESERVED #endif
bye, Roman
Quoted 9 lines · Roman Zippel
Quoted 1 line
Kamezawa-san, can you please explain why exactly you did that change?
Probably because it triggers this check:
#if SECTIONS_WIDTH+NODES_WIDTH+ZONES_WIDTH >
FLAGS_RESERVED#error SECTIONS_WIDTH+NODES_WIDTH+ZONES_WIDTH >FLAGS_RESERVED#endif
Yes, it was for this.
If still ZONES_WIDTH = ZONES_SHIFT =2, I have no problem.
Thanks,
-- Kame
Quoted 2 lines
bye, Roman
Quoted 63 lines · Andi Kleen
Add 4GB DMA32 zone
Add a new 4GB
GFP_DMA32zone between theGFP_DMAandGFP_NORMALzones.As a bit of historical background: when the x86-64 port was originally designed we had some discussion if we should use a 16MB DMA zone like i386 or a 4GB DMA zone like IA64 or both. Both was ruled out at this point because it was in early 2.4 when VM is still quite shakey and had bad troubles even dealing with one DMA zone. We settled on the 16MB DMA zone mainly because we worried about older soundcards and the floppy.
But this has always caused problems since then because device drivers had trouble getting enough DMA able memory. These days the VM works much better and the wide use of NUMA has proven it can deal with many zones successfully.
So this patch adds both zones.
This helps drivers who need a lot of memory below 4GB because their hardware is not accessing more (graphic drivers - proprietary and free ones, video frame buffer drivers, sound drivers etc.). Previously they could only use IOMMU+16MB
GFP_DMA, which was not enough memory.Another common problem is that hardware who has full memory addressing for >4GB misses it for some control structures in memory (like transmit rings or other metadata). They tended to allocate memory in the 16MB
GFP_DMAor the IOMMU/swiotlb then usingpci_alloc_consistent, but that can tie up a lot of precious 16MB GFPDMA/IOMMU/swiotlb memory (even on AMD systems the IOMMU tends to be quite small) especially if you have many devices. With the new zonepci_alloc_consistentcan just put this stuff into memory below 4GB which works better.One argument was still if the zone should be 4GB or 2GB. The main motivation for 2GB would be an unnamed not so unpopular hardware raid controller (mostly found in older machines from a particular four letter company) who has a strange 2GB restriction in firmware. But
that one works ok with swiotlb/IOMMU anyways, so it doesn't really needGFP_DMA32. I chose 4GB to be compatible with IA64 and because it seems to be the most common restriction.The new zone is so far added only for x86-64.
For other architectures who don't set up this
new zone nothing changes. Architectures can set a compatibility define inKconfigCONFIG_DMA_IS_DMA32that will defineGFP_DMA32asGFP_DMA. Otherwise it's a nop because on 32bit architectures it's normally not needed becauseGFP_NORMAL(=0) is DMA able enough.One problem is still that
GFP_DMAmeans different things on different architectures. e.g. some drivers used to have #ifdef ia64 useGFP_DMA(trusting it to be 4GB) #elif__x86_64__(use other hacks like the swiotlb because 16MB is not enough) ... . This was quite ugly and is now obsolete.These should be now converted to use
GFP_DMA32unconditionally. I haven't done this yet. Or best only use pci_alloc_consistent/dma_alloc_coherent which will useGFP_DMA32transparently.Signed-off-by Andi Kleen <[email protected]>
<snip>
Quoted 31 lines
Index:
linux/include/linux/mmzone.h
===================================================================include/linux/mmzone.h +6
-5@@ -70,11 +70,12 @@ struct per_cpu_pageset {#endif #define ZONE_DMA 0 -#define ZONE_NORMAL 1 -#define ZONE_HIGHMEM 2 +#define ZONE_DMA32 1 +#define ZONE_NORMAL 2 +#define ZONE_HIGHMEM 3 -#define MAX_NR_ZONES 3/* Sync this with ZONES_SHIFT */ -#define ZONES_SHIFT 2/* ceil(log2(MAX_NR_ZONES)) */ +#define MAX_NR_ZONES 4/* Sync this with ZONES_SHIFT */ +#define ZONES_SHIFT 3/* ceil(log2(MAX_NR_ZONES)) */ /* @@ -90,7 +91,7 @@ struct per_cpu_pageset {* be 8 (2 ** 3) zonelists. GFP_ZONETYPES defines the number of possible * combinations of zone modifiers in "zone modifier space". */ -#define GFP_ZONEMASK 0x03 +#define GFP_ZONEMASK 0x07/* * As an optimisation any zone modifier bits which are only valid when * no other zone modifier bits are set (loners) should be placed in
include/linux/mmzone.h struct per_cpu_pageset { · line 111
Quoted 1 line
include/linux/mmzone.h +0
-0@@ -110,6 +111,7 @@ struct per_cpu_pageset {* into multiple physical zones. On a PC we have 3 zones:
Now 4 zones.
include/linux/mmzone.h struct per_cpu_pageset { · line 117
Quoted 6 lines
include/linux/mmzone.h +1
-0* * ZONE_DMA < 16 MB ISA DMA capable memory +* ZONE_DMA32 0 MB Empty * ZONE_NORMAL 16-896 MB direct mapped by the kernel * ZONE_HIGHMEM > 896 MB only page cache and user processes */
<snip>
--
Coywolf Qi Hunt
http://sosdg.org/~coywolf/