| Message ID | 20240319120026.2246389-1-rrichter@amd.com |
|---|---|
| Headers | show |
| Series | SRAT/CEDT fixes and updates | expand |
Robert Richter wrote: > For configurations that have the kconfig option NUMA_KEEP_MEMINFO > disabled, the SRAT lookup done with numa_fill_memblks() fails > returning NUMA_NO_MEMBLK (-1). An existing SRAT memory range cannot be > found for a CFMWS address range. This causes the addition of a > duplicate numa_memblk with a different node id and a subsequent page > fault and kernel crash during boot. > > numa_fill_memblks() is implemented and used in the init section only. > The option NUMA_KEEP_MEMINFO is only for the case when NUMA data will > be used outside of init. So fix the SRAT lookup by moving > numa_fill_memblks() out of the NUMA_KEEP_MEMINFO block to make it > always available in the init section. > > Note that the issue was initially introduced with [1]. But since > phys_to_target_node() was originally used that returned the valid node > 0, an additional numa_memblk was not added. Though, the node id was > wrong too. > > [1] fd49f99c1809 ("ACPI: NUMA: Add a node and memblk for each CFMWS not in SRAT") > > Fixes: 8f1004679987 ("ACPI/NUMA: Apply SRAT proximity domain to entire CFMWS window") > Cc: Derick Marks <derick.w.marks@intel.com> > Cc: Dan Williams <dan.j.williams@intel.com> > Cc: Alison Schofield <alison.schofield@intel.com> > Signed-off-by: Robert Richter <rrichter@amd.com> > --- > arch/x86/mm/numa.c | 4 ++-- > 1 file changed, 2 insertions(+), 2 deletions(-) > > diff --git a/arch/x86/mm/numa.c b/arch/x86/mm/numa.c > index 65e9a6e391c0..ce84ba86e69e 100644 > --- a/arch/x86/mm/numa.c > +++ b/arch/x86/mm/numa.c > @@ -929,6 +929,8 @@ int memory_add_physaddr_to_nid(u64 start) > } > EXPORT_SYMBOL_GPL(memory_add_physaddr_to_nid); > > +#endif > + > static int __init cmp_memblk(const void *a, const void *b) > { > const struct numa_memblk *ma = *(const struct numa_memblk **)a; > @@ -1001,5 +1003,3 @@ int __init numa_fill_memblks(u64 start, u64 end) > } > return 0; > } > - > -#endif Does this achieve the goal without an additional hunk like this? diff --git a/arch/x86/include/asm/sparsemem.h b/arch/x86/include/asm/sparsemem.h index 1be13b2dfe8b..1aaa447ef24b 100644 --- a/arch/x86/include/asm/sparsemem.h +++ b/arch/x86/include/asm/sparsemem.h @@ -37,9 +37,9 @@ extern int phys_to_target_node(phys_addr_t start); #define phys_to_target_node phys_to_target_node extern int memory_add_physaddr_to_nid(u64 start); #define memory_add_physaddr_to_nid memory_add_physaddr_to_nid +#endif extern int numa_fill_memblks(u64 start, u64 end); #define numa_fill_memblks numa_fill_memblks -#endif #endif /* __ASSEMBLY__ */ #endif /* _ASM_X86_SPARSEMEM_H */
Robert Richter wrote: > With the removal of the Itanium architecture [1] the last architecture > dependent functions: > > acpi_numa_slit_init(), acpi_numa_memory_affinity_init() > > were removed. Remove its remainings in the header files too an make > them static. It would have helped to clarify that this is a refactoring and absorbing logic into the helpers to validate the correctness of the conversion approach. I almost asked for it to be split to make that clearer, but figured it out eventually. > [1] cf8e8658100d arch: Remove Itanium (IA-64) architecture Checkpatch does not like this format and has some other things to say about this legacy code being touched. > > Signed-off-by: Robert Richter <rrichter@amd.com> > --- > drivers/acpi/numa/srat.c | 68 +++++++++++++--------------------------- > include/linux/acpi.h | 5 --- > 2 files changed, 21 insertions(+), 52 deletions(-) > > diff --git a/drivers/acpi/numa/srat.c b/drivers/acpi/numa/srat.c > index 50ae8557e8d1..910609a9754b 100644 > --- a/drivers/acpi/numa/srat.c > +++ b/drivers/acpi/numa/srat.c > @@ -208,16 +208,21 @@ int __init srat_disabled(void) > return acpi_numa < 0; > } > > -#if defined(CONFIG_X86) || defined(CONFIG_ARM64) || defined(CONFIG_LOONGARCH) > /* > * Callback for SLIT parsing. pxm_to_node() returns NUMA_NO_NODE for > * I/O localities since SRAT does not list them. I/O localities are > * not supported at this point. > */ > -void __init acpi_numa_slit_init(struct acpi_table_slit *slit) > +static int __init acpi_parse_slit(struct acpi_table_header *table) > { > + struct acpi_table_slit *slit = (struct acpi_table_slit *)table; > int i, j; > > + if (!slit_valid(slit)) { > + pr_info("SLIT table looks invalid. Not used.\n"); > + return -EINVAL; > + } > + > for (i = 0; i < slit->locality_count; i++) { > const int from_node = pxm_to_node(i); > > @@ -234,19 +239,25 @@ void __init acpi_numa_slit_init(struct acpi_table_slit *slit) > slit->entry[slit->locality_count * i + j]); > } > } > + > + return 0; > } > > -/* > - * Default callback for parsing of the Proximity Domain <-> Memory > - * Area mappings > - */ > -int __init > -acpi_numa_memory_affinity_init(struct acpi_srat_mem_affinity *ma) > +static int __initdata parsed_numa_memblks; Checkpatch wants __initdata moved after the symbol name. Otherwise this conversion looks correct. > +static int __init > +acpi_parse_memory_affinity(union acpi_subtable_headers * header, > + const unsigned long table_end) > { > + struct acpi_srat_mem_affinity *ma; > u64 start, end; > u32 hotpluggable; > int node, pxm; > > + ma = (struct acpi_srat_mem_affinity *)header; > + > + acpi_table_print_srat_entry(&header->common); > + > if (srat_disabled()) > goto out_err; > if (ma->header.length < sizeof(struct acpi_srat_mem_affinity)) { > @@ -293,6 +304,8 @@ acpi_numa_memory_affinity_init(struct acpi_srat_mem_affinity *ma) > > max_possible_pfn = max(max_possible_pfn, PFN_UP(end - 1)); > > + parsed_numa_memblks++; > + > return 0; > out_err_bad_srat: > bad_srat(); > @@ -448,27 +461,6 @@ static int __init acpi_parse_cfmws(union acpi_subtable_headers *header, > (*fake_pxm)++; > return 0; > } > -#else > -static inline void acpi_table_print_cedt(void) {} Would be nice to move this patch before the printing patch otherwise feels icky to add a line in patch2 that gets deleted in patch3.
Dan, On 19.03.24 13:15:48, Dan Williams wrote: > Does this achieve the goal without an additional hunk like this? > > diff --git a/arch/x86/include/asm/sparsemem.h b/arch/x86/include/asm/sparsemem.h > index 1be13b2dfe8b..1aaa447ef24b 100644 > --- a/arch/x86/include/asm/sparsemem.h > +++ b/arch/x86/include/asm/sparsemem.h > @@ -37,9 +37,9 @@ extern int phys_to_target_node(phys_addr_t start); > #define phys_to_target_node phys_to_target_node > extern int memory_add_physaddr_to_nid(u64 start); > #define memory_add_physaddr_to_nid memory_add_physaddr_to_nid > +#endif > extern int numa_fill_memblks(u64 start, u64 end); > #define numa_fill_memblks numa_fill_memblks > -#endif > #endif /* __ASSEMBLY__ */ > > #endif /* _ASM_X86_SPARSEMEM_H */ right, there are some build issues that need to be fixed in this patch. Sorry for that. Thanks, -Robert
On Thu, Mar 21, 2024 at 05:55:57PM +0100, Robert Richter wrote: > Alison, > > On 20.03.24 10:46:07, Alison Schofield wrote: > > On Tue, Mar 19, 2024 at 01:00:23PM +0100, Robert Richter wrote: > > > For configurations that have the kconfig option NUMA_KEEP_MEMINFO > > > disabled, the SRAT lookup done with numa_fill_memblks() fails > > > returning NUMA_NO_MEMBLK (-1). An existing SRAT memory range cannot be > > > found for a CFMWS address range. This causes the addition of a > > > duplicate numa_memblk with a different node id and a subsequent page > > > fault and kernel crash during boot. > > > > > > numa_fill_memblks() is implemented and used in the init section only. > > > The option NUMA_KEEP_MEMINFO is only for the case when NUMA data will > > > be used outside of init. So fix the SRAT lookup by moving > > > numa_fill_memblks() out of the NUMA_KEEP_MEMINFO block to make it > > > always available in the init section. > > > > > > Note that the issue was initially introduced with [1]. But since > > > phys_to_target_node() was originally used that returned the valid node > > > 0, an additional numa_memblk was not added. Though, the node id was > > > wrong too. > > > > Hi Richard, > > > > I recall a bit of wrangling w #defines to make ARM64 and LOONGARCH build. > > I'm seeing an x86 build error today: > > > > >> arch/x86/mm/numa.c:957:12: error: redefinition of 'numa_fill_memblks' > > 957 | int __init numa_fill_memblks(u64 start, u64 end) > > > > include/linux/numa.h:40:26: note: previous definition of 'numa_fill_memblks' with type > > +'int(u64, u64)' {aka 'int(long long unsigned int, long long unsigned int)'} > > 40 | static inline int __init numa_fill_memblks(u64 start, u64 end) > > | ^~~~~~~~~~~~~~~~~ > > > > In addition to what you suggest, would something like this diff below be > > a useful safety measure to distinguish num_fill_memblks() success (rc:0) > > and possible non-existence (rc:-1). I don't think it hurts to take a > > second look using phys_to_target_node() (totall untested) > > > > diff --git a/drivers/acpi/numa/srat.c b/drivers/acpi/numa/srat.c > > index 070a52e4daa8..0c48fe32ced4 100644 > > --- a/drivers/acpi/numa/srat.c > > +++ b/drivers/acpi/numa/srat.c > > @@ -437,9 +437,16 @@ static int __init acpi_parse_cfmws(union acpi_subtable_headers *header, > > * found for any portion of the window to cover the entire > > * window. > > */ > > - if (!numa_fill_memblks(start, end)) > > + rc = numa_fill_memblks(start, end); > > + if (!rc) > > return 0; > > > > + if (rc == NUMA_NO_MEMBLK) { > > + node = phys_to_target_node(start); > > + if (node != NUMA_NO_NODE) > > + return 0; > > + } > > + > > for non-x86 the numa_add_memblk() function looks good in a way that it > is able to handle presumable overlapping regions. numa_fill_memblks() > would just fail then and numa_add_memblk() being called. For x86 we > need numa_fill_memblks() since x86 specific numa_add_memblk() cannot > handle the overlapping case. > > That said, we do not need the 2nd check. It looks to me that it > actually breaks non-x86 as the whole block may not be registered (if > it is larger than anything existing). > > For x86 the 2nd check may never happen if numa_fill_memblks() is > always enabled (which is this patch for). Hi Robert, (<-- got it right this time ;)) I wasn't thinking of x86, but rather archs that may not support numa_fill_memblks() and return NUMA_NO_MEMBLK (-1) per the #ifndef numa_fill_memblks in include/linux/numa.h In those cases, take a second look at phys_to_targe_node() before blindly adding another memblk. Is that the failure signature you reported here? I can wait and see your final patch and how the different archs will handle it. I'm worried that NUMA_NO_MEMBLK is overloaded and we need to diffentiate between archs that don't even look for a node, versus archs that look but don't find a node. --Alison > > So we should be good without your change. > > Thanks, > > -Robert > > > /* No SRAT description. Create a new node. */ > > > > --Alison > > > > > > > > [1] fd49f99c1809 ("ACPI: NUMA: Add a node and memblk for each CFMWS not in SRAT") > > > > > > Fixes: 8f1004679987 ("ACPI/NUMA: Apply SRAT proximity domain to entire CFMWS window") > > > Cc: Derick Marks <derick.w.marks@intel.com> > > > Cc: Dan Williams <dan.j.williams@intel.com> > > > Cc: Alison Schofield <alison.schofield@intel.com> > > > Signed-off-by: Robert Richter <rrichter@amd.com> > > > --- > > > arch/x86/mm/numa.c | 4 ++-- > > > 1 file changed, 2 insertions(+), 2 deletions(-) > > > > > > diff --git a/arch/x86/mm/numa.c b/arch/x86/mm/numa.c > > > index 65e9a6e391c0..ce84ba86e69e 100644 > > > --- a/arch/x86/mm/numa.c > > > +++ b/arch/x86/mm/numa.c > > > @@ -929,6 +929,8 @@ int memory_add_physaddr_to_nid(u64 start) > > > } > > > EXPORT_SYMBOL_GPL(memory_add_physaddr_to_nid); > > > > > > +#endif > > > + > > > static int __init cmp_memblk(const void *a, const void *b) > > > { > > > const struct numa_memblk *ma = *(const struct numa_memblk **)a; > > > @@ -1001,5 +1003,3 @@ int __init numa_fill_memblks(u64 start, u64 end) > > > } > > > return 0; > > > } > > > - > > > -#endif > > > -- > > > 2.39.2 > > >