top posting what now works for me: - changed calculation of memory size to loop through different memory nodes - added numa_node to banks - filter out banks that do not match the target mixup node (do not want to change ABI to add node information)
That's not satisfying overall but at least my code works with NUMA on Qemu.
*diff --git a/Kconfig b/Kconfig*
*index f8c1a77bed..4d3ab8cb49 100644*
*--- a/Kconfig*
*+++ b/Kconfig*
@@ -192,7 +192,9 @@ config NR_DRAM_BANKS
default 1 if ARCH_SUNXI || ARCH_OWL
default 4
help
- This defines the number of DRAM banks.
+ This defines the number of DRAM banks. For qemu with NUMA, you
+ may want to set this value to <slots> * <possible memdev>.
+ for instance, for a 2 slot with 4 memdevs set NR_DRAM_BANKS to 8.
config SYS_BOOT_GET_CMDLINE
bool "Enable kernel command line setup"
*diff --git a/arch/arm/lib/bootm-fdt.c b/arch/arm/lib/bootm-fdt.c*
*index 29020bd1c6..2b28ab8108 100644*
*--- a/arch/arm/lib/bootm-fdt.c*
*+++ b/arch/arm/lib/bootm-fdt.c*
@@ -42,12 +42,17 @@ int arch_fixup_fdt(void *blob)
u64 size[CONFIG_NR_DRAM_BANKS];
for (bank = 0; bank < CONFIG_NR_DRAM_BANKS; bank++) {
+ unsigned char node = bd->bi_dram[bank].numa_node;
start[bank] = bd->bi_dram[bank].start;
size[bank] = bd->bi_dram[bank].size;
#ifdef CONFIG_ARMV7_NONSEC
ret = armv7_apply_memory_carveout(&start[bank], &size[bank]);
if (ret)
return ret;
+#endif
+#ifdef CONFIG_OF_LIBFDT
+ /* add node info for the fdt_fixup_memory below */
+ start[bank] = (((phys_addr_t)node) << 56) | bd->bi_dram[bank].start;
#endif
}
*diff --git a/common/fdt_support.c b/common/fdt_support.c*
*index a9a32df1e7..3bca2ba888 100644*
*--- a/common/fdt_support.c*
*+++ b/common/fdt_support.c*
@@ -415,16 +415,29 @@ static int fdt_pack_reg(const void *fdt, void *buf, u64 *address, u64 *size,
return p - (char *)buf;
}
+static inline uint32_t fdt32_ld(const fdt32_t *p)
+{
+ const uint8_t *bp = (const uint8_t *)p;
+
+ return ((uint32_t)bp[0] << 24)
+ | ((uint32_t)bp[1] << 16)
+ | ((uint32_t)bp[2] << 8)
+ | bp[3];
+}
#if CONFIG_NR_DRAM_BANKS > 4
#define MEMORY_BANKS_MAX CONFIG_NR_DRAM_BANKS
#else
#define MEMORY_BANKS_MAX 4
#endif
+/* NUMA has yet to be properly handled
+ * This code appends memory to the first memory node that matches the NUMA node.
+ */
int fdt_fixup_memory_banks(void *blob, u64 start[], u64 size[], int banks)
{
int err, nodeoffset;
int len, i;
u8 tmp[MEMORY_BANKS_MAX * 16]; /* Up to 64-bit address + 64-bit size */
+ unsigned int numa_node;
if (banks > MEMORY_BANKS_MAX) {
printf("%s: num banks %d exceeds hardcoded limit %d."
@@ -444,6 +457,12 @@ int fdt_fixup_memory_banks(void *blob, u64 start[], u64 size[], int banks)
if (nodeoffset < 0)
return nodeoffset;
+ const __be32* numa_node_prop = fdt_getprop(blob, nodeoffset, "numa-node-id", &len);
+ if (numa_node_prop != NULL && len == sizeof(__be32)) {
+ numa_node = fdt32_ld(numa_node_prop);
+ }
+ else numa_node = 0;
+
err = fdt_setprop(blob, nodeoffset, "device_type", "memory",
sizeof("memory"));
if (err < 0) {
@@ -453,8 +472,27 @@ int fdt_fixup_memory_banks(void *blob, u64 start[], u64 size[], int banks)
}
for (i = 0; i < banks; i++) {
- if (start[i] == 0 && size[i] == 0)
+ /* clear node information */
+ unsigned int node;
+recheck:
+ if (start[i] == 0 && size[i] == 0)
break;
+ node = (start[i] >> 56) & 0xFF;
+ start[i] = start[i] & 0x00FFFFFFFFFFFFFF;
+ /* for the moment, just ignore the banks that are not in
+ * memory NUMA node */
+ if (node != numa_node) {
+ /* remove the bank from the list */
+ int j;
+ for (j=i; j < banks-1; j++) {
+ start[j] = start[j+1];
+ size[j] = size[j+1];
+ }
+ start[j]=0;
+ size[j]=0;
+ banks--;
+ goto recheck;
+ }
}
banks = i;
@@ -470,6 +508,7 @@ int fdt_fixup_memory_banks(void *blob, u64 start[], u64 size[], int banks)
"reg", fdt_strerror(err));
return err;
}
+
return 0;
}
*diff --git a/configs/qemu_arm64_defconfig b/configs/qemu_arm64_defconfig*
*index f6e586627a..0fdc22d71d 100644*
*--- a/configs/qemu_arm64_defconfig*
*+++ b/configs/qemu_arm64_defconfig*
@@ -1,7 +1,7 @@
CONFIG_ARM=y
CONFIG_POSITION_INDEPENDENT=y
CONFIG_ARCH_QEMU=y
-CONFIG_NR_DRAM_BANKS=1
+CONFIG_NR_DRAM_BANKS=32
CONFIG_ENV_SIZE=0x40000
CONFIG_ENV_SECT_SIZE=0x40000
CONFIG_AHCI=y
*diff --git a/include/asm-generic/u-boot.h b/include/asm-generic/u-boot.h*
*index 637de0c455..3cf45124ec 100644*
*--- a/include/asm-generic/u-boot.h*
*+++ b/include/asm-generic/u-boot.h*
@@ -71,6 +71,8 @@ struct bd_info {
struct { /* RAM configuration */
phys_addr_t start;
phys_size_t size;
+ unsigned numa_node;
+ unsigned pad; /* just to make sure we do not cause alignment */
} bi_dram[CONFIG_NR_DRAM_BANKS];
};
*diff --git a/lib/fdtdec.c b/lib/fdtdec.c*
*index 4b097fb588..b1934edc8d 100644*
*--- a/lib/fdtdec.c*
*+++ b/lib/fdtdec.c*
@@ -1064,77 +1064,101 @@ int fdtdec_decode_display_timing(const void *blob, int parent, int index,
return ret;
}
+ofnode get_next_memory_node(ofnode mem)
+{
+ do {
+ mem = ofnode_by_prop_value(mem, "device_type", "memory", 7);
+ } while (!ofnode_is_available(mem));
+
+ return mem;
+}
+
int fdtdec_setup_mem_size_base(void)
{
int ret;
+ int reg;
ofnode mem;
struct resource res;
+ phys_addr_t base = ~0;
+ phys_size_t size = 0;;
- mem = ofnode_path("/memory");
- if (!ofnode_valid(mem)) {
- debug("%s: Missing /memory node\n", __func__);
- return -EINVAL;
- }
+ for (mem = get_next_memory_node(mem); ofnode_valid(mem); mem = get_next_memory_node(mem)) {
- ret = ofnode_read_resource(mem, 0, &res);
- if (ret != 0) {
- debug("%s: Unable to decode first memory bank\n", __func__);
- return -EINVAL;
+ for(reg = 0, ret = 0; ret == 0 ; reg++) {
+ ret = ofnode_read_resource(mem, reg, &res);
+ if (ret != 0)
+ break;
+ if ((phys_addr_t)res.start < base)
+ base = (phys_addr_t)res.start;
+ size += (phys_size_t)(res.end - res.start + 1);
+ }
}
+
+ gd->ram_base = (unsigned long)base;
+ gd->ram_size = (phys_size_t)size;
- gd->ram_size = (phys_size_t)(res.end - res.start + 1);
- gd->ram_base = (unsigned long)res.start;
+ debug("%s: Initial DRAM base %llx\n", __func__,
+ (unsigned long long)gd->ram_base);
debug("%s: Initial DRAM size %llx\n", __func__,
(unsigned long long)gd->ram_size);
return 0;
}
-ofnode get_next_memory_node(ofnode mem)
-{
- do {
- mem = ofnode_by_prop_value(mem, "device_type", "memory", 7);
- } while (!ofnode_is_available(mem));
-
- return mem;
-}
-
int fdtdec_setup_memory_banksize(void)
{
int bank, ret, reg = 0;
struct resource res;
ofnode mem = ofnode_null();
+ const __be32* numa_node_prop = NULL;
+ int len;
+ int numa_node = -1;
+ int count = 0;
+
+ for (mem = get_next_memory_node(mem); ofnode_valid(mem); mem = get_next_memory_node(mem)) {
- mem = get_next_memory_node(mem);
- if (!ofnode_valid(mem)) {
- debug("%s: Missing /memory node\n", __func__);
- return -EINVAL;
+ count++;
+
+ numa_node_prop = ofnode_get_property(mem, "numa-node-id", &len);
+ if (numa_node_prop != NULL && len == sizeof(__be32)) {
+ numa_node = of_read_number(numa_node_prop, 1);
}
+ else numa_node = 0;
- for (bank = 0; bank < CONFIG_NR_DRAM_BANKS; bank++) {
- ret = ofnode_read_resource(mem, reg++, &res);
- if (ret < 0) {
- reg = 0;
- mem = get_next_memory_node(mem);
- if (!ofnode_valid(mem))
- break;
+ debug("Found memory for node %d\n", numa_node);
- ret = ofnode_read_resource(mem, reg++, &res);
- if (ret < 0)
- break;
- }
+ ret = 0;
+ for(reg = 0; ret == 0 && bank < CONFIG_NR_DRAM_BANKS; reg++) {
+ ret = ofnode_read_resource(mem, reg, &res);
if (ret != 0)
- return -EINVAL;
+ break;
gd->bd->bi_dram[bank].start = (phys_addr_t)res.start;
gd->bd->bi_dram[bank].size =
(phys_size_t)(res.end - res.start + 1);
+ gd->bd->bi_dram[bank].numa_node = numa_node;
- debug("%s: DRAM Bank #%d: start = 0x%llx, size = 0x%llx\n",
+ debug("%s: DRAM Bank #%d: start = 0x%llx, size = 0x%llx"
+ " name_node = %d\n",
__func__, bank,
(unsigned long long)gd->bd->bi_dram[bank].start,
- (unsigned long long)gd->bd->bi_dram[bank].size);
+ (unsigned long long)gd->bd->bi_dram[bank].size,
+ gd->bd->bi_dram[bank].numa_node);
+
+ bank++;
+ }
+
+
+ }
+
+ if (count == 0) {
+ debug("%s: Missing /memory node\n", __func__);
+ return -EINVAL;
+ }
+ if (bank >= CONFIG_NR_DRAM_BANKS) {
+ printf("Too many DT memory nodes for CONFIG_NR_DRAM_BANKS=%d\n",
+ CONFIG_NR_DRAM_BANKS);
}
return 0;
On Wed, 7 Jul 2021 at 13:00, François Ozog francois.ozog@linaro.org wrote:
On Wed, 7 Jul 2021 at 12:16, AKASHI Takahiro takahiro.akashi@linaro.org wrote:
On Wed, Jul 07, 2021 at 11:37:19AM +0200, Fran??ois Ozog wrote:
On Wed, 7 Jul 2021 at 09:40, François Ozog francois.ozog@linaro.org
wrote:
On Wed, 7 Jul 2021 at 05:59, Heinrich Schuchardt xypron.glpk@gmx.de wrote:
Am 7. Juli 2021 05:18:20 MESZ schrieb Heinrich Schuchardt <
xypron.glpk@gmx.de>:
Am 7. Juli 2021 03:44:35 MESZ schrieb AKASHI Takahiro takahiro.akashi@linaro.org: >François, > >On Tue, Jul 06, 2021 at 08:10:08PM +0200, Heinrich Schuchardt
wrote:
>> On 7/6/21 6:13 PM, François Ozog wrote: >> > Hi Heinrich, U-Boot 2021-07rc5 does not take into account
memory
>> > description when using Qemu 5.2 NUMA configuration to adapt
memory
>map >> > (kernel_addr_r...): >> > >> > -smp 4 \ >> > -m 8G,slots=2,maxmem=16G \ >> > -object memory-backend-ram,size=4G,id=m0 \ >> > -object memory-backend-ram,size=4G,id=m1 \ >> > -numa node,cpus=0-1,nodeid=0,memdev=m0 \ >> > -numa node,cpus=2-3,nodeid=1,memdev=m1 >> > >> > kernel_addr_r is still 0x4040000 and thus you can't use it to >bootefi. >> > >> > fdt addr 0x13ede6de0; fdt print >> > >> > Displays fdt while I think it should not. >> > >> > If I load the kernel at dram.start, the load works but not
boot
>> > >> > U-Boot 2021.07 (Jul 06 2021 - 13:26:43 +0000) >> > >> > >> > DRAM:4 GiB >> > >> > Flash: 64 MiB >> > >> > Loading Environment from Flash... OK >> > >> > In:pl011@9000000 >> > >> > Out: pl011@9000000 >> > >> > Err: pl011@9000000 >> > >> > Net: eth0: virtio-net#32 >> > >> > Hit any key to stop autoboot:0 >> > >> > => >> > >> > => bdinfo >> > >> > boot_params = 0x0000000000000000 >> > >> > DRAM bank = 0x0000000000000000 >> > >> > -> start= 0x0000000140000000 >> > >> > -> size = 0x0000000100000000 >> > >> > flashstart= 0x0000000000000000 >> > >> > flashsize = 0x0000000004000000 >> > >> > flashoffset = 0x00000000000bc990 >> > >> > baudrate= 115200 bps >> > >> > relocaddr = 0x000000013ff27000 >> > >> > reloc off = 0x000000013ff27000 >> > >> > Build = 64-bit >> > >> > current eth = virtio-net#32 >> > >> > ethaddr = 52:52:52:52:52:52 >> > >> > IP addr = <NULL> >> > >> > fdt_blob= 0x000000013ede6de0 >> > >> > new_fdt = 0x000000013ede6de0 >> > >> > fdt_size= 0x0000000000100000 >> > >> > lmb_dump_all: >> > >> > memory.cnt= 0x1 >> > >> > memory.reg[0x0].base = 0x140000000 >> > >> > .size = 0x100000000 >> > >> > >> > reserved.cnt= 0x0 >> > >> > arch_number = 0x0000000000000000 >> > >> > TLB addr= 0x000000013fff0000 >> > >> > irq_sp= 0x000000013ede6dd0 >> > >> > sp start= 0x000000013ede6dd0 >> > >> > Early malloc usage: 3a8 / 2000 >> > >> > => load virtio 0:1 0x140000000 /oskit.efi >> > >> > 853424 bytes read in 1 ms (813.9 MiB/s) >> > >> > => bootefi0x140000000 0x13ede6dd0 >> > >> > ERROR: Failed to register WaitForKey event >> > >> > Setting OsIndications failed >> > >> > Error: Cannot initialize UEFI sub-system, r = 9 >> > >> > >> > I think there is a need to calculate memory map based on
previous
>> > firmware (TFA, QEMU can be considered as previous frimware) >information >> > (DT or blob_list). >> > >> > What do you think ? >> > >> > Cheers >> > >> > FF >> > >> > -- >> > >> > François-Frédéric Ozog | /Director Business Development/ >> > T: +33.67221.6485 >> > francois.ozog@linaro.org mailto:francois.ozog@linaro.org >| Skype: ffozog >> > >> > >> >> The kernel load address is hard coded here: >> include/configs/qemu-arm.h:41: "kernel_addr_r=0x40400000\0" \ >> >> bdinfo shows: >> DRAM start = 0x140000000 >> DRAM size = 0x100000000 >> >> fdt addr $fdt_addr >> fdt printf >> >> shows two memory areas. One at 40000000, one at 140000000. > >(This shows that U-Boot receives a correct memory map via dtb.) > >Is this a NUMA machine, isn't it? Why should we care of which >memory region be used here? Please note that this is a virtual machine, >there is no practical difference between two regions. > >The root problem is that U-Boot did not recognize there were two >memory regions. We can fix this issue in either way: > >1) >diff --git a/configs/qemu_arm64_defconfig >b/configs/qemu_arm64_defconfig >index f6e586627a8e..b70ffae8bf6e 100644 >--- a/configs/qemu_arm64_defconfig >+++ b/configs/qemu_arm64_defconfig >@@ -1,7 +1,7 @@ > CONFIG_ARM=y > CONFIG_POSITION_INDEPENDENT=y > CONFIG_ARCH_QEMU=y >-CONFIG_NR_DRAM_BANKS=1 >+CONFIG_NR_DRAM_BANKS=2 > CONFIG_ENV_SIZE=0x40000 > CONFIG_ENV_SECT_SIZE=0x40000 > CONFIG_AHCI=y > >2) >diff --git a/lib/fdtdec.c b/lib/fdtdec.c >index 4b097fb588ed..4067ea2dead6 100644 >--- a/lib/fdtdec.c >+++ b/lib/fdtdec.c >@@ -1111,7 +1111,7 @@ int fdtdec_setup_memory_banksize(void) > return -EINVAL; > } > >- for (bank = 0; bank < CONFIG_NR_DRAM_BANKS; bank++) { >+ for (bank = 0; ; bank++) { > ret = ofnode_read_resource(mem, reg++, &res); > if (ret < 0) { > reg = 0; > > (fdtdec_setup_memory_banksize() is called in
dram_init_banksize().)
> > >(2) seems much better, but I don't know why we had to use >CONFIG_NR_DRAM_BANKS here. >
- alone does not work as other places in the code refer to
CONFIG_NR_DRAM_BANKS. Setting ...BANKS to 32 makes my code work and bdinfo seems now correct:
=> bdinfo
boot_params = 0x0000000000000000 DRAM bank = 0x0000000000000000 -> start = 0x0000000140000000 -> size = 0x0000000100000000 DRAM bank = 0x0000000000000001 -> start = 0x0000000040000000 -> size = 0x0000000100000000 flashstart = 0x0000000000000000 flashsize = 0x0000000004000000 flashoffset = 0x00000000000bcb88 baudrate = 115200 bps relocaddr = 0x000000013ff27000 reloc off = 0x000000013ff27000 Build = 64-bit current eth = virtio-net#32 ethaddr = 52:52:52:52:52:52 IP addr = <NULL> fdt_blob = 0x000000013ede6cf0 new_fdt = 0x000000013ede6cf0 fdt_size = 0x0000000000100000 lmb_dump_all: memory.cnt = 0x1 memory.reg[0x0].base = 0x40000000 .size = 0x200000000 reserved.cnt = 0x1 reserved.reg[0x0].base = 0x13ede58f0 .size = 0x121a710 arch_number = 0x0000000000000000 TLB addr = 0x000000013fff0000 irq_sp = 0x000000013ede6ce0 sp start = 0x000000013ede6ce0 Early malloc usage: 3a8 / 2000
May I suggest you propose a combined patch Akashi-san? If we assume NUMA systems to be tested up to 8 nodes to mimic real existing enterprise hardware and up to 4 memory slots (say for memory hot plugging tests) what about a default value of 32? Alternatively, we could set this value to a much higher one if the costs are negligible.
Well, lets not rush as there are other twists:
the 4G bank in node 1 is marked BootServicesData in the UEFI
GetMemoryMap
which I assume is not the case. EDK2 reports it as ConventionalMemory.
The root cause seem to be gd->ramtop not being setup properly.
Further analysis shows that the DT passed to the booted EFI payload does not seem to be correct:
DT fragment passed to U-Boot
memory@140000000 { numa-node-id = <0x00000001>; reg = <0x00000001 0x40000000 0x00000001 0x00000000>; device_type = "memory"; }; memory@40000000 { numa-node-id = <0x00000000>; reg = <0x00000000 0x40000000 0x00000001 0x00000000>; device_type = "memory"; };
DT passed to payload (as per my debug code):
memory@140000000: memory
numa-node-id 1 reg (len= 32) 140000000 100000000 40000000 100000000memory@40000000: memory
numa-node-id 0 reg (len= 16) 40000000 100000000I am investigating this further...
You should check the logic of fdt_fixup_memory_banks() which is called this way: efi_dt_fixup() image_setup_libfdt() arch_fixup_fdt() fdt_fixup_memory_banks()
What it does is to put *all* the memory regions unconditionally as a single "reg" array into the *first-detected* "memory" node, which is "memory@140000000" in this case. It means that this function doesn't respect NUMA configuration.
Thanks.
tweaking ram_top to be the correct in https://elixir.bootlin.com/u-boot/latest/source/lib/efi_loader/efi_memory.c#... results in memory on node 1 to be considered as ConventionalMemory but U-Boot places all code and data at the end of node 1 while it should position it on the current node. That said it is an acceptable work around for my test case.
Bottom line, we need to introduce NUMA node management in memory management all over the place. It is unclear if there is a business case for that. I'll ask LEDGE members...
-Takahiro Akashi
In this case, other occurrences of CONFIG_NR_DRAM_BANKS in this file >should be replaced with a variable for it. > >> Your use case is well beyond the typical U-Boot usage. So I
guess it
>> will be up to Linaro to provide the necessary patches: >> >> * determine the active CPU >> * determine the RAM assigned to the active CPU according >> to the numa-node-id in the device-tree >> * make sure that U-Boot only uses the memory of the active CPU >> internally >> * make sure that the UEFI memory map contains a compliant description >> * possibly, dynamically set up the environment variables >> >> +CC Tuomas Tynkkynen (maintainer for qemu_arm64_defconfig) > >For (1), we'd better have a different config, or increase >the value of CONFIG_NR_DRAM_BANKS to a bigger number?
Is the system configured such that each CPU can access the others
CPU's
RAM when entering U-Boot?
Best regards
Heinrich
At least the comments for this patch sound as if on a physical
system
cross NUMA node memory access is only available after full SMP initialization:
https://patchwork.kernel.org/project/linux-acpi/patch/20180625130552.5636-1-...
QEMU may be less restrictive.
QEMU allows the node distance to be 255 indicating that cross node
access is infeasible.
Best regards
Heinrich
> >-Takahiro Akashi > > >> Best regards >> >> Heinrich
-- François-Frédéric Ozog | Director Business Development T: +33.67221.6485 francois.ozog@linaro.org | Skype: ffozog
-- François-Frédéric Ozog | *Director Business Development* T: +33.67221.6485 francois.ozog@linaro.org | Skype: ffozog
-- François-Frédéric Ozog | *Director Business Development* T: +33.67221.6485 francois.ozog@linaro.org | Skype: ffozog