Message ID | 20230104183033.755668-7-pierre.gondois@arm.com (mailing list archive) |
---|---|
State | New, archived |
Headers | show |
Series | arch_topology: Build cacheinfo from primary CPU | expand |
Hi Pierre, On Wed, Jan 4, 2023 at 7:35 PM Pierre Gondois <pierre.gondois@arm.com> wrote: > commit 3fcbf1c77d08 ("arch_topology: Fix cache attributes detection > in the CPU hotplug path") > adds a call to detect_cache_attributes() to populate the cacheinfo > before updating the siblings mask. detect_cache_attributes() allocates > memory and can take the PPTT mutex (on ACPI platforms). On PREEMPT_RT > kernels, on secondary CPUs, this triggers a: > 'BUG: sleeping function called from invalid context' [1] > as the code is executed with preemption and interrupts disabled. > > The primary CPU was previously storing the cache information using > the now removed (struct cpu_topology).llc_id: > commit 5b8dc787ce4a ("arch_topology: Drop LLC identifier stash from > the CPU topology") > > allocate_cache_info() tries to build the cacheinfo from the primary > CPU prior secondary CPUs boot, if the DT/ACPI description > contains cache information. > If allocate_cache_info() fails, then fallback to the current state > for the cacheinfo allocation. [1] will be triggered in such case. > > When unplugging a CPU, the cacheinfo memory cannot be freed. If it > was, then the memory would be allocated early by the re-plugged > CPU and would trigger [1]. > > Note that populate_cache_leaves() might be called multiple times > due to populate_leaves being moved up. This is required since > detect_cache_attributes() might be called with per_cpu_cacheinfo(cpu) > being allocated but not populated. > > [1]: > [ 7.560791] BUG: sleeping function called from invalid context at kernel/locking/spinlock_rt.c:46 > [ 7.560794] in_atomic(): 1, irqs_disabled(): 128, non_block: 0, pid: 0, name: swapper/111 > [ 7.560796] preempt_count: 1, expected: 0 > [ 7.560797] RCU nest depth: 1, expected: 1 > [ 7.560799] 3 locks held by swapper/111/0: > [ 7.560800] #0: ffff403e406cae98 (&pcp->lock){+.+.}-{3:3}, at: get_page_from_freelist+0x218/0x12c8 > [ 7.560811] #1: ffffc5f8ed09f8e8 (rcu_read_lock){....}-{1:3}, at: rt_spin_trylock+0x48/0xf0 > [ 7.560820] #2: ffff403f400b4fd8 (&zone->lock){+.+.}-{3:3}, at: rmqueue_bulk+0x64/0xa80 > [ 7.560824] irq event stamp: 0 > [ 7.560825] hardirqs last enabled at (0): [<0000000000000000>] 0x0 > [ 7.560827] hardirqs last disabled at (0): [<ffffc5f8e9f7d594>] copy_process+0x5dc/0x1ab8 > [ 7.560830] softirqs last enabled at (0): [<ffffc5f8e9f7d594>] copy_process+0x5dc/0x1ab8 > [ 7.560833] softirqs last disabled at (0): [<0000000000000000>] 0x0 > [ 7.560834] Preemption disabled at: > [ 7.560835] [<ffffc5f8e9fd3c28>] migrate_enable+0x30/0x130 > [ 7.560838] CPU: 111 PID: 0 Comm: swapper/111 Tainted: G W 6.0.0-rc4-rt6-[...] > [ 7.560841] Call trace: > [...] > [ 7.560870] __kmalloc+0xbc/0x1e8 > [ 7.560873] detect_cache_attributes+0x2d4/0x5f0 > [ 7.560876] update_siblings_masks+0x30/0x368 > [ 7.560880] store_cpu_topology+0x78/0xb8 > [ 7.560883] secondary_start_kernel+0xd0/0x198 > [ 7.560885] __secondary_switched+0xb0/0xb4 > > Signed-off-by: Pierre Gondois <pierre.gondois@arm.com> > Reviewed-by: Sudeep Holla <sudeep.holla@arm.com> > Acked-by: Palmer Dabbelt <palmer@rivosinc.com> Thanks for your patch, which is now commit 5944ce092b97caed ("arch_topology: Build cacheinfo from primary CPU") in driver-core/driver-core-next. > --- a/drivers/base/arch_topology.c > +++ b/drivers/base/arch_topology.c > @@ -736,7 +736,7 @@ void update_siblings_masks(unsigned int cpuid) > > ret = detect_cache_attributes(cpuid); > if (ret && ret != -ENOENT) > - pr_info("Early cacheinfo failed, ret = %d\n", ret); > + pr_info("Early cacheinfo allocation failed, ret = %d\n", ret); > > /* update core and thread sibling masks */ > for_each_online_cpu(cpu) { > @@ -825,7 +825,7 @@ __weak int __init parse_acpi_topology(void) > #if defined(CONFIG_ARM64) || defined(CONFIG_RISCV) > void __init init_cpu_topology(void) > { > - int ret; > + int cpu, ret; > > reset_cpu_topology(); > ret = parse_acpi_topology(); > @@ -840,6 +840,14 @@ void __init init_cpu_topology(void) > reset_cpu_topology(); > return; > } > + > + for_each_possible_cpu(cpu) { > + ret = fetch_cache_info(cpu); > + if (ret) { > + pr_err("Early cacheinfo failed, ret = %d\n", ret); This triggers on all my RV64 platforms (K210, Icicle, Starlight, RZ/Five). This seems to be a respin of https://lore.kernel.org/all/CAMuHMdUBZ791fxCPkKQ6HCwLE4GJB2S35QC=SQ+X8w5Q4C_70g@mail.gmail.com which had the same issue. > + break; > + } > + } > } > Gr{oetje,eeting}s, Geert -- Geert Uytterhoeven -- There's lots of Linux beyond ia32 -- geert@linux-m68k.org In personal conversations with technical people, I call myself a hacker. But when I'm talking to journalists I just say "programmer" or something like that. -- Linus Torvalds
On Tue, Jan 24, 2023 at 02:50:16PM +0100, Geert Uytterhoeven wrote: > Hi Pierre, > [...] > > @@ -840,6 +840,14 @@ void __init init_cpu_topology(void) > > reset_cpu_topology(); > > return; > > } > > + > > + for_each_possible_cpu(cpu) { > > + ret = fetch_cache_info(cpu); > > + if (ret) { > > + pr_err("Early cacheinfo failed, ret = %d\n", ret); > > This triggers on all my RV64 platforms (K210, Icicle, Starlight, > RZ/Five). > > This seems to be a respin of > https://lore.kernel.org/all/CAMuHMdUBZ791fxCPkKQ6HCwLE4GJB2S35QC=SQ+X8w5Q4C_70g@mail.gmail.com > which had the same issue. > I need to recollect my memories reading all the thread, but even after the fixes there were few platforms that failed with so early allocation but were fine with initcalls. Are these such platforms or am I mixing up things here ? Do you still see all the cacheinfo in the sysfs with initcalls that happen later in the boot ? Conor might help me remember the details. -- Regards, Sudeep
On Tue, Jan 24, 2023 at 02:04:20PM +0000, Sudeep Holla wrote: > Conor might help me remember the details. And I can't shirk either since you know I just replied to Pierre! > On Tue, Jan 24, 2023 at 02:50:16PM +0100, Geert Uytterhoeven wrote: > > > @@ -840,6 +840,14 @@ void __init init_cpu_topology(void) > > > reset_cpu_topology(); > > > return; > > > } > > > + > > > + for_each_possible_cpu(cpu) { > > > + ret = fetch_cache_info(cpu); > > > + if (ret) { > > > + pr_err("Early cacheinfo failed, ret = %d\n", ret); > > > > This triggers on all my RV64 platforms (K210, Icicle, Starlight, > > RZ/Five). > > > > This seems to be a respin of > > https://lore.kernel.org/all/CAMuHMdUBZ791fxCPkKQ6HCwLE4GJB2S35QC=SQ+X8w5Q4C_70g@mail.gmail.com > > which had the same issue. > > > > I need to recollect my memories reading all the thread, but even after the > fixes there were few platforms that failed with so early allocation but were > fine with initcalls. Are these such platforms or am I mixing up things here ? > Do you still see all the cacheinfo in the sysfs with initcalls that happen > later in the boot ? IIRC that stuff was failing back then because riscv calls init_cpu_topology() far sooner in boot than arm64 does, and therefore caused allocation failures. You made that warning go away in the below patch by moving detect_cache_attributes() to update_siblings_masks(), which both arches call later during boot IIRC: https://lore.kernel.org/all/20220713133344.1201247-1-sudeep.holla@arm.com Pierre's patch has added fetch_cache_info() to the problematic init_cpu_topology() which is called before we can actually do any allocation in smp_prepare_boot_cpu() or something like that. That's what I get for only reviewing the patch that was specifically for riscv, and not the rest of the series... D'oh. This actually came up a few weeks ago, although I kinda considered the reason it was triggered to be a bit bogus there, since that dmips property is not (yet?) a valid property on RISC-V. The patch for that is here: https://patchwork.kernel.org/project/linux-riscv/patch/20230105033705.3946130-1-leyfoon.tan@starfivetech.com/ I tried it on a PolarFire SoC (unfortunately not an Icicle, I just went and bricked mine an hour ago) & it should be a fix for this problem too. My suggested commit message for that is somewhat prophetic now that I look back at it: https://lore.kernel.org/all/Y7V4byskevAWKM3G@spud/ I'll ping Palmer about that patch I guess... Cheers, Conor.
On Tue, Jan 24, 2023 at 02:39:19PM +0000, Conor Dooley wrote: > On Tue, Jan 24, 2023 at 02:04:20PM +0000, Sudeep Holla wrote: > > Conor might help me remember the details. > > And I can't shirk either since you know I just replied to Pierre! > > > On Tue, Jan 24, 2023 at 02:50:16PM +0100, Geert Uytterhoeven wrote: > > > > > @@ -840,6 +840,14 @@ void __init init_cpu_topology(void) > > > > reset_cpu_topology(); > > > > return; > > > > } > > > > + > > > > + for_each_possible_cpu(cpu) { > > > > + ret = fetch_cache_info(cpu); > > > > + if (ret) { > > > > + pr_err("Early cacheinfo failed, ret = %d\n", ret); > > > > > > This triggers on all my RV64 platforms (K210, Icicle, Starlight, > > > RZ/Five). > > > > > > This seems to be a respin of > > > https://lore.kernel.org/all/CAMuHMdUBZ791fxCPkKQ6HCwLE4GJB2S35QC=SQ+X8w5Q4C_70g@mail.gmail.com > > > which had the same issue. > > > > > > > I need to recollect my memories reading all the thread, but even after the > > fixes there were few platforms that failed with so early allocation but were > > fine with initcalls. Are these such platforms or am I mixing up things here ? > > Do you still see all the cacheinfo in the sysfs with initcalls that happen > > later in the boot ? > > IIRC that stuff was failing back then because riscv calls > init_cpu_topology() far sooner in boot than arm64 does, and therefore > caused allocation failures. You made that warning go away in the below > patch by moving detect_cache_attributes() to update_siblings_masks(), > which both arches call later during boot IIRC: > https://lore.kernel.org/all/20220713133344.1201247-1-sudeep.holla@arm.com > > Pierre's patch has added fetch_cache_info() to the problematic > init_cpu_topology() which is called before we can actually do any > allocation in smp_prepare_boot_cpu() or something like that. > > That's what I get for only reviewing the patch that was specifically for > riscv, and not the rest of the series... D'oh. > > This actually came up a few weeks ago, although I kinda considered the > reason it was triggered to be a bit bogus there, since that dmips property > is not (yet?) a valid property on RISC-V. The patch for that is here: > https://patchwork.kernel.org/project/linux-riscv/patch/20230105033705.3946130-1-leyfoon.tan@starfivetech.com/ > I tried it on a PolarFire SoC (unfortunately not an Icicle, I just went > and bricked mine an hour ago) & it should be a fix for this problem too. > > My suggested commit message for that is somewhat prophetic now that I > look back at it: > https://lore.kernel.org/all/Y7V4byskevAWKM3G@spud/ > Ah, that thread, I remember that :). I still need to understand how this is related to memory allocation. Pierre was suggesting(in private) if we need to keep fetch_cache_info() arch specific but I really don't want to go down that patch until I understand and there is no other option. Thanks for your time. I will try to recall boot flow and see if I can gather the reasoning for the seen behaviour.
On Tue, Jan 24, 2023 at 02:48:39PM +0000, Sudeep Holla wrote: > > Ah, that thread, I remember that :). > > I still need to understand how this is related to memory allocation. > Pierre was suggesting(in private) if we need to keep fetch_cache_info() > arch specific but I really don't want to go down that patch until I > understand and there is no other option. > > Thanks for your time. I will try to recall boot flow and see if I can > gather the reasoning for the seen behaviour. > OK, I must have atleast taken a look at the code before I replied. smp_prepare_boot_cpu() is called quite early before page_alloc_init() and mm_init()(in init_main.c) while smp_prepare_cpus() get called quite late from kernel_init->kernel_init_freeable(). Geert, can you please try with the patch Conor pointed out and see if that helps to fix the allocation failures[1]
Hi Geert, On Tue, Jan 24, 2023 at 02:55:41PM +0000, Sudeep Holla wrote: > > Geert, can you please try with the patch Conor pointed out and see if > that helps to fix the allocation failures[1] > Sorry for the nag, but did you get the chance to test -next with [1] and see if it fixes the cacheinfo memory failure you were observing ?
Hi Sudeep, On Wed, Jan 25, 2023 at 3:54 PM Sudeep Holla <sudeep.holla@arm.com> wrote: > On Tue, Jan 24, 2023 at 02:55:41PM +0000, Sudeep Holla wrote: > > Geert, can you please try with the patch Conor pointed out and see if > > that helps to fix the allocation failures[1] > > > > Sorry for the nag, but did you get the chance to test -next with [1] > and see if it fixes the cacheinfo memory failure you were observing ? > [1] https://lore.kernel.org/all/20230103035316.3841303-1-leyfoon.tan@starfivetech.com/ After applying that patch, the issue is gone. Thanks, sending my Tb! Gr{oetje,eeting}s, Geert -- Geert Uytterhoeven -- There's lots of Linux beyond ia32 -- geert@linux-m68k.org In personal conversations with technical people, I call myself a hacker. But when I'm talking to journalists I just say "programmer" or something like that. -- Linus Torvalds
On Wed, Jan 25, 2023 at 04:28:04PM +0100, Geert Uytterhoeven wrote: > Hi Sudeep, > > On Wed, Jan 25, 2023 at 3:54 PM Sudeep Holla <sudeep.holla@arm.com> wrote: > > On Tue, Jan 24, 2023 at 02:55:41PM +0000, Sudeep Holla wrote: > > > Geert, can you please try with the patch Conor pointed out and see if > > > that helps to fix the allocation failures[1] > > > > > > > Sorry for the nag, but did you get the chance to test -next with [1] > > and see if it fixes the cacheinfo memory failure you were observing ? > > > [1] https://lore.kernel.org/all/20230103035316.3841303-1-leyfoon.tan@starfivetech.com/ > > After applying that patch, the issue is gone. > Thanks, sending my Tb! > Thanks for testing and sorry for wrong information in my PR. I had mixed up things and assumed it was tested. I usually leave it in -next for a week before sending PR but rushed things once I was back from the holidays, won't repeat hopefully
Hello Geert, Sudeep, On 1/25/23 16:28, Geert Uytterhoeven wrote: > Hi Sudeep, > > On Wed, Jan 25, 2023 at 3:54 PM Sudeep Holla <sudeep.holla@arm.com> wrote: >> On Tue, Jan 24, 2023 at 02:55:41PM +0000, Sudeep Holla wrote: >>> Geert, can you please try with the patch Conor pointed out and see if >>> that helps to fix the allocation failures[1] >>> >> >> Sorry for the nag, but did you get the chance to test -next with [1] >> and see if it fixes the cacheinfo memory failure you were observing ? > >> [1] https://lore.kernel.org/all/20230103035316.3841303-1-leyfoon.tan@starfivetech.com/ > > After applying that patch, the issue is gone. > Thanks, sending my Tb! > > Gr{oetje,eeting}s, > > Geert > > -- > Geert Uytterhoeven -- There's lots of Linux beyond ia32 -- geert@linux-m68k.org > > In personal conversations with technical people, I call myself a hacker. But > when I'm talking to journalists I just say "programmer" or something like that. > -- Linus Torvalds Happy to hear the issue disappears with the patch, thanks for solving the issue, Regards, Pierre
diff --git a/arch/riscv/kernel/cacheinfo.c b/arch/riscv/kernel/cacheinfo.c index 440a3df5944c..3a13113f1b29 100644 --- a/arch/riscv/kernel/cacheinfo.c +++ b/arch/riscv/kernel/cacheinfo.c @@ -113,11 +113,6 @@ static void fill_cacheinfo(struct cacheinfo **this_leaf, } } -int init_cache_level(unsigned int cpu) -{ - return init_of_cache_level(cpu); -} - int populate_cache_leaves(unsigned int cpu) { struct cpu_cacheinfo *this_cpu_ci = get_cpu_cacheinfo(cpu); diff --git a/drivers/base/arch_topology.c b/drivers/base/arch_topology.c index e7d6e6657ffa..b1c1dd38ab01 100644 --- a/drivers/base/arch_topology.c +++ b/drivers/base/arch_topology.c @@ -736,7 +736,7 @@ void update_siblings_masks(unsigned int cpuid) ret = detect_cache_attributes(cpuid); if (ret && ret != -ENOENT) - pr_info("Early cacheinfo failed, ret = %d\n", ret); + pr_info("Early cacheinfo allocation failed, ret = %d\n", ret); /* update core and thread sibling masks */ for_each_online_cpu(cpu) { @@ -825,7 +825,7 @@ __weak int __init parse_acpi_topology(void) #if defined(CONFIG_ARM64) || defined(CONFIG_RISCV) void __init init_cpu_topology(void) { - int ret; + int cpu, ret; reset_cpu_topology(); ret = parse_acpi_topology(); @@ -840,6 +840,14 @@ void __init init_cpu_topology(void) reset_cpu_topology(); return; } + + for_each_possible_cpu(cpu) { + ret = fetch_cache_info(cpu); + if (ret) { + pr_err("Early cacheinfo failed, ret = %d\n", ret); + break; + } + } } void store_cpu_topology(unsigned int cpuid) diff --git a/drivers/base/cacheinfo.c b/drivers/base/cacheinfo.c index 339cd1e3d5af..b57fbd0d7114 100644 --- a/drivers/base/cacheinfo.c +++ b/drivers/base/cacheinfo.c @@ -386,10 +386,6 @@ static void free_cache_attributes(unsigned int cpu) return; cache_shared_cpu_map_remove(cpu); - - kfree(per_cpu_cacheinfo(cpu)); - per_cpu_cacheinfo(cpu) = NULL; - cache_leaves(cpu) = 0; } int __weak init_cache_level(unsigned int cpu) @@ -402,29 +398,71 @@ int __weak populate_cache_leaves(unsigned int cpu) return -ENOENT; } +static inline +int allocate_cache_info(int cpu) +{ + per_cpu_cacheinfo(cpu) = kcalloc(cache_leaves(cpu), + sizeof(struct cacheinfo), GFP_ATOMIC); + if (!per_cpu_cacheinfo(cpu)) { + cache_leaves(cpu) = 0; + return -ENOMEM; + } + + return 0; +} + +int fetch_cache_info(unsigned int cpu) +{ + struct cpu_cacheinfo *this_cpu_ci; + unsigned int levels, split_levels; + int ret; + + if (acpi_disabled) { + ret = init_of_cache_level(cpu); + if (ret < 0) + return ret; + } else { + ret = acpi_get_cache_info(cpu, &levels, &split_levels); + if (ret < 0) + return ret; + + this_cpu_ci = get_cpu_cacheinfo(cpu); + this_cpu_ci->num_levels = levels; + /* + * This assumes that: + * - there cannot be any split caches (data/instruction) + * above a unified cache + * - data/instruction caches come by pair + */ + this_cpu_ci->num_leaves = levels + split_levels; + } + if (!cache_leaves(cpu)) + return -ENOENT; + + return allocate_cache_info(cpu); +} + int detect_cache_attributes(unsigned int cpu) { int ret; - /* Since early detection of the cacheinfo is allowed via this - * function and this also gets called as CPU hotplug callbacks via - * cacheinfo_cpu_online, the initialisation can be skipped and only - * CPU maps can be updated as the CPU online status would be update - * if called via cacheinfo_cpu_online path. + /* Since early initialization/allocation of the cacheinfo is allowed + * via fetch_cache_info() and this also gets called as CPU hotplug + * callbacks via cacheinfo_cpu_online, the init/alloc can be skipped + * as it will happen only once (the cacheinfo memory is never freed). + * Just populate the cacheinfo. */ if (per_cpu_cacheinfo(cpu)) - goto update_cpu_map; + goto populate_leaves; if (init_cache_level(cpu) || !cache_leaves(cpu)) return -ENOENT; - per_cpu_cacheinfo(cpu) = kcalloc(cache_leaves(cpu), - sizeof(struct cacheinfo), GFP_ATOMIC); - if (per_cpu_cacheinfo(cpu) == NULL) { - cache_leaves(cpu) = 0; - return -ENOMEM; - } + ret = allocate_cache_info(cpu); + if (ret) + return ret; +populate_leaves: /* * populate_cache_leaves() may completely setup the cache leaves and * shared_cpu_map or it may leave it partially setup. @@ -433,7 +471,6 @@ int detect_cache_attributes(unsigned int cpu) if (ret) goto free_ci; -update_cpu_map: /* * For systems using DT for cache hierarchy, fw_token * and shared_cpu_map will be set up here only if they are diff --git a/include/linux/cacheinfo.h b/include/linux/cacheinfo.h index 00d8e7f9d1c6..dfef57077cd0 100644 --- a/include/linux/cacheinfo.h +++ b/include/linux/cacheinfo.h @@ -85,6 +85,7 @@ int populate_cache_leaves(unsigned int cpu); int cache_setup_acpi(unsigned int cpu); bool last_level_cache_is_valid(unsigned int cpu); bool last_level_cache_is_shared(unsigned int cpu_x, unsigned int cpu_y); +int fetch_cache_info(unsigned int cpu); int detect_cache_attributes(unsigned int cpu); #ifndef CONFIG_ACPI_PPTT /*