| Message ID | 20241031074618.3585491-2-guanjun@linux.alibaba.com (mailing list archive) |
|---|---|
| State | New |
| Headers | show |
| Series | Support for limiting the number of managed interrupts on every node per allocation. | expand |
On Thu, Oct 31 2024 at 15:46, guanjun@linux.alibaba.com wrote: > #ifdef CONFIG_SMP > > +static unsigned int __read_mostly managed_irqs_per_node; > +static struct cpumask managed_irqs_cpumsk[MAX_NUMNODES] __cacheline_aligned_in_smp = { > + [0 ... MAX_NUMNODES-1] = {CPU_BITS_ALL} > +}; > > +static void __group_prepare_affinity(struct cpumask *premask, > + cpumask_var_t *node_to_cpumask) > +{ > + nodemask_t nodemsk = NODE_MASK_NONE; > + unsigned int ncpus, n; > + > + get_nodes_in_cpumask(node_to_cpumask, premask, &nodemsk); > + > + for_each_node_mask(n, nodemsk) { > + cpumask_and(&managed_irqs_cpumsk[n], &managed_irqs_cpumsk[n], premask); > + cpumask_and(&managed_irqs_cpumsk[n], &managed_irqs_cpumsk[n], node_to_cpumask[n]); How is this managed_irqs_cpumsk array protected against concurrency? > + ncpus = cpumask_weight(&managed_irqs_cpumsk[n]); > + if (ncpus < managed_irqs_per_node) { > + /* Reset node n to current node cpumask */ > + cpumask_copy(&managed_irqs_cpumsk[n], node_to_cpumask[n]); This whole logic is incomprehensible and aside of the concurrency problem it's broken when CPUs are made present at run-time because these cpu masks are static and represent the stale state of the last invocation. Given the limitations of the x86 vector space, which is not going away anytime soon, there are only two options IMO to handle such a scenario. 1) Tell the nvme/block layer to disable queue affinity management 2) Restrict the devices and queues to the nodes they sit on Thanks, tglx
On Thu, Oct 31, 2024 at 6:35 PM Thomas Gleixner <tglx@linutronix.de> wrote: > > On Thu, Oct 31 2024 at 15:46, guanjun@linux.alibaba.com wrote: > > #ifdef CONFIG_SMP > > > > +static unsigned int __read_mostly managed_irqs_per_node; > > +static struct cpumask managed_irqs_cpumsk[MAX_NUMNODES] __cacheline_aligned_in_smp = { > > + [0 ... MAX_NUMNODES-1] = {CPU_BITS_ALL} > > +}; > > > > +static void __group_prepare_affinity(struct cpumask *premask, > > + cpumask_var_t *node_to_cpumask) > > +{ > > + nodemask_t nodemsk = NODE_MASK_NONE; > > + unsigned int ncpus, n; > > + > > + get_nodes_in_cpumask(node_to_cpumask, premask, &nodemsk); > > + > > + for_each_node_mask(n, nodemsk) { > > + cpumask_and(&managed_irqs_cpumsk[n], &managed_irqs_cpumsk[n], premask); > > + cpumask_and(&managed_irqs_cpumsk[n], &managed_irqs_cpumsk[n], node_to_cpumask[n]); > > How is this managed_irqs_cpumsk array protected against concurrency? > > > + ncpus = cpumask_weight(&managed_irqs_cpumsk[n]); > > + if (ncpus < managed_irqs_per_node) { > > + /* Reset node n to current node cpumask */ > > + cpumask_copy(&managed_irqs_cpumsk[n], node_to_cpumask[n]); > > This whole logic is incomprehensible and aside of the concurrency > problem it's broken when CPUs are made present at run-time because these > cpu masks are static and represent the stale state of the last > invocation. > > Given the limitations of the x86 vector space, which is not going away > anytime soon, there are only two options IMO to handle such a scenario. > > 1) Tell the nvme/block layer to disable queue affinity management +1 There are other use cases, such as cpu isolation, which can benefit from this way too. https://lore.kernel.org/linux-nvme/20240702104112.4123810-1-ming.lei@redhat.com/ Thanks,
> 2024年10月31日 18:35，Thomas Gleixner <tglx@linutronix.de> 写道： > > On Thu, Oct 31 2024 at 15:46, guanjun@linux.alibaba.com wrote: >> #ifdef CONFIG_SMP >> >> +static unsigned int __read_mostly managed_irqs_per_node; >> +static struct cpumask managed_irqs_cpumsk[MAX_NUMNODES] __cacheline_aligned_in_smp = { >> + [0 ... MAX_NUMNODES-1] = {CPU_BITS_ALL} >> +}; >> >> +static void __group_prepare_affinity(struct cpumask *premask, >> + cpumask_var_t *node_to_cpumask) >> +{ >> + nodemask_t nodemsk = NODE_MASK_NONE; >> + unsigned int ncpus, n; >> + >> + get_nodes_in_cpumask(node_to_cpumask, premask, &nodemsk); >> + >> + for_each_node_mask(n, nodemsk) { >> + cpumask_and(&managed_irqs_cpumsk[n], &managed_irqs_cpumsk[n], premask); >> + cpumask_and(&managed_irqs_cpumsk[n], &managed_irqs_cpumsk[n], node_to_cpumask[n]); > > How is this managed_irqs_cpumsk array protected against concurrency? My intention was to allocate up to `managed_irq_per_node` cpu bits from `managed_irqs_cpumask[n]`, even if another task modifies some of the bits in the `managed_irqs_cpumask[n]` at the same time. > >> + ncpus = cpumask_weight(&managed_irqs_cpumsk[n]); >> + if (ncpus < managed_irqs_per_node) { >> + /* Reset node n to current node cpumask */ >> + cpumask_copy(&managed_irqs_cpumsk[n], node_to_cpumask[n]); > > This whole logic is incomprehensible and aside of the concurrency > problem it's broken when CPUs are made present at run-time because these > cpu masks are static and represent the stale state of the last > invocation. Sorry, I realize there is indeed a logic issue here (caused by developing on 5.10 LTS and rebase to the latest linux-next). > > Given the limitations of the x86 vector space, which is not going away > anytime soon, there are only two options IMO to handle such a scenario. > > 1) Tell the nvme/block layer to disable queue affinity management > > 2) Restrict the devices and queues to the nodes they sit on I have tried fixing this issue through nvme driver, but later discovered that the same issue exists with virtio net. Therefore, I want to address this with a more general solution. Thanks, Guanjun > > Thanks, > > tglx
On Fri, Nov 1, 2024 at 11:12 AM mapicccy <guanjun@linux.alibaba.com> wrote: > > > > 2024年10月31日 18:50，Ming Lei <ming.lei@redhat.com> 写道： > > On Thu, Oct 31, 2024 at 6:35 PM Thomas Gleixner <tglx@linutronix.de> wrote: > > > On Thu, Oct 31 2024 at 15:46, guanjun@linux.alibaba.com wrote: > > #ifdef CONFIG_SMP > > +static unsigned int __read_mostly managed_irqs_per_node; > +static struct cpumask managed_irqs_cpumsk[MAX_NUMNODES] __cacheline_aligned_in_smp = { > + [0 ... MAX_NUMNODES-1] = {CPU_BITS_ALL} > +}; > > +static void __group_prepare_affinity(struct cpumask *premask, > + cpumask_var_t *node_to_cpumask) > +{ > + nodemask_t nodemsk = NODE_MASK_NONE; > + unsigned int ncpus, n; > + > + get_nodes_in_cpumask(node_to_cpumask, premask, &nodemsk); > + > + for_each_node_mask(n, nodemsk) { > + cpumask_and(&managed_irqs_cpumsk[n], &managed_irqs_cpumsk[n], premask); > + cpumask_and(&managed_irqs_cpumsk[n], &managed_irqs_cpumsk[n], node_to_cpumask[n]); > > > How is this managed_irqs_cpumsk array protected against concurrency? > > + ncpus = cpumask_weight(&managed_irqs_cpumsk[n]); > + if (ncpus < managed_irqs_per_node) { > + /* Reset node n to current node cpumask */ > + cpumask_copy(&managed_irqs_cpumsk[n], node_to_cpumask[n]); > > > This whole logic is incomprehensible and aside of the concurrency > problem it's broken when CPUs are made present at run-time because these > cpu masks are static and represent the stale state of the last > invocation. > > Given the limitations of the x86 vector space, which is not going away > anytime soon, there are only two options IMO to handle such a scenario. > > 1) Tell the nvme/block layer to disable queue affinity management > > > +1 > > There are other use cases, such as cpu isolation, which can benefit from > this way too. > > https://lore.kernel.org/linux-nvme/20240702104112.4123810-1-ming.lei@redhat.com/ > I wonder if we need to do the same for virtio-blk. > > Thanks for your reminder. However, in this link only modified the NVMe driver, > but there is the same issue in the virtio net driver as well. I guess you meant virtio-blk actually? > > Guanjun > > > Thanks, > Thanks
Hi, On 31. 10. 24, 8:46, 'Guanjun' wrote: > From: Guanjun <guanjun@linux.alibaba.com> > > Commit c410abbbacb9 (genirq/affinity: Add is_managed to struct irq_affinity_desc) > introduced is_managed bit to struct irq_affinity_desc. Due to queue interrupts > treated as managed interrupts, in scenarios where a large number of > devices are present (using massive msix queue interrupts), an excessive number > of IRQ matrix bits (about num_online_cpus() * nvecs) are reserved during > interrupt allocation. This sequently leads to the situation where interrupts > for some devices cannot be properly allocated. > > Support for limiting the number of managed interrupts on every node per allocation. > > Signed-off-by: Guanjun <guanjun@linux.alibaba.com> > --- > .../admin-guide/kernel-parameters.txt | 9 +++ > block/blk-mq-cpumap.c | 2 +- > drivers/virtio/virtio_vdpa.c | 2 +- > fs/fuse/virtio_fs.c | 2 +- > include/linux/group_cpus.h | 2 +- > kernel/irq/affinity.c | 11 ++-- > lib/group_cpus.c | 55 ++++++++++++++++++- > 7 files changed, 73 insertions(+), 10 deletions(-) > > diff --git a/Documentation/admin-guide/kernel-parameters.txt b/Documentation/admin-guide/kernel-parameters.txt > index 9b61097a6448..ac80f35d04c9 100644 > --- a/Documentation/admin-guide/kernel-parameters.txt > +++ b/Documentation/admin-guide/kernel-parameters.txt > @@ -3238,6 +3238,15 @@ > different yeeloong laptops. > Example: machtype=lemote-yeeloong-2f-7inch > > + managed_irqs_per_node= > + [KNL,SMP] Support for limiting the number of managed > + interrupts on every node to prevent the case that > + interrupts cannot be properly allocated where a large > + number of devices are present. The default number is 0, > + that means no limit to the number of managed irqs. > + Format: integer between 0 and num_possible_cpus() / num_possible_nodes() > + Default: 0 Kernel parameters suck. Esp. here you have to guess to even properly boot. Could this be auto-tuned instead? > --- a/lib/group_cpus.c > +++ b/lib/group_cpus.c > @@ -11,6 +11,30 @@ > > #ifdef CONFIG_SMP > > +static unsigned int __read_mostly managed_irqs_per_node; > +static struct cpumask managed_irqs_cpumsk[MAX_NUMNODES] __cacheline_aligned_in_smp = { This is quite excessive. On SUSE configs, this is 8192 cpu bits * 1024 nodes = 1 M. For everyone. You have to allocate this dynamically instead. See e.g. setup_node_to_cpumask_map(). > + [0 ... MAX_NUMNODES-1] = {CPU_BITS_ALL} > +}; > + > +static int __init irq_managed_setup(char *str) > +{ > + int ret; > + > + ret = kstrtouint(str, 10, &managed_irqs_per_node); > + if (ret < 0) { > + pr_warn("managed_irqs_per_node= cannot parse, ignored\n"); could not be parsed > + return 0; > + } > + > + if (managed_irqs_per_node * num_possible_nodes() > num_possible_cpus()) { > + managed_irqs_per_node = num_possible_cpus() / num_possible_nodes(); > + pr_warn("managed_irqs_per_node= cannot be larger than %u\n", > + managed_irqs_per_node); > + } > + return 1; > +} > +__setup("managed_irqs_per_node=", irq_managed_setup); > + > static void grp_spread_init_one(struct cpumask *irqmsk, struct cpumask *nmsk, > unsigned int cpus_per_grp) > { ... > @@ -332,6 +380,7 @@ static int __group_cpus_evenly(unsigned int startgrp, unsigned int numgrps, > /** > * group_cpus_evenly - Group all CPUs evenly per NUMA/CPU locality > * @numgrps: number of groups > + * @is_managed: if these groups managed by kernel are managed by the kernel > * > * Return: cpumask array if successful, NULL otherwise. And each element > * includes CPUs assigned to this group thanks,
On Fri, Nov 01 2024 at 11:03, mapicccy wrote: >> 2024年10月31日 18:35，Thomas Gleixner <tglx@linutronix.de> 写道： >>> + get_nodes_in_cpumask(node_to_cpumask, premask, &nodemsk); >>> + >>> + for_each_node_mask(n, nodemsk) { >>> + cpumask_and(&managed_irqs_cpumsk[n], &managed_irqs_cpumsk[n], premask); >>> + cpumask_and(&managed_irqs_cpumsk[n], &managed_irqs_cpumsk[n], node_to_cpumask[n]); >> >> How is this managed_irqs_cpumsk array protected against concurrency? > > My intention was to allocate up to `managed_irq_per_node` cpu bits from `managed_irqs_cpumask[n]`, > even if another task modifies some of the bits in the `managed_irqs_cpumask[n]` at the same time. That may have been your intention, but how is this even remotely correct? Aside of that. If it's intentional and you think it's correct then you should have documented that in the code and also annotated it to not trigger santiziers. >> Given the limitations of the x86 vector space, which is not going away >> anytime soon, there are only two options IMO to handle such a scenario. >> >> 1) Tell the nvme/block layer to disable queue affinity management >> >> 2) Restrict the devices and queues to the nodes they sit on > > I have tried fixing this issue through nvme driver, but later > discovered that the same issue exists with virtio net. Therefore, I > want to address this with a more general solution. I understand, but a general solution for this problem won't exist ever. It's very reasonable to restrict this for one particular device type or subsystem while maintaining the strict managed property for others, no? General solutions are definitely preferred, but not for the price that they break existing completely correct and working setups. Which is what your 2/2 patch does for sure. Thanks, tglx
diff --git a/Documentation/admin-guide/kernel-parameters.txt b/Documentation/admin-guide/kernel-parameters.txt index 9b61097a6448..ac80f35d04c9 100644 --- a/Documentation/admin-guide/kernel-parameters.txt +++ b/Documentation/admin-guide/kernel-parameters.txt @@ -3238,6 +3238,15 @@ different yeeloong laptops. Example: machtype=lemote-yeeloong-2f-7inch + managed_irqs_per_node= + [KNL,SMP] Support for limiting the number of managed + interrupts on every node to prevent the case that + interrupts cannot be properly allocated where a large + number of devices are present. The default number is 0, + that means no limit to the number of managed irqs. + Format: integer between 0 and num_possible_cpus() / num_possible_nodes() + Default: 0 + maxcpus= [SMP,EARLY] Maximum number of processors that an SMP kernel will bring up during bootup. maxcpus=n : n >= 0 limits the kernel to bring up 'n' processors. Surely after diff --git a/block/blk-mq-cpumap.c b/block/blk-mq-cpumap.c index 9638b25fd521..481c81318e00 100644 --- a/block/blk-mq-cpumap.c +++ b/block/blk-mq-cpumap.c @@ -20,7 +20,7 @@ void blk_mq_map_queues(struct blk_mq_queue_map *qmap) const struct cpumask *masks; unsigned int queue, cpu; - masks = group_cpus_evenly(qmap->nr_queues); + masks = group_cpus_evenly(qmap->nr_queues, true); if (!masks) { for_each_possible_cpu(cpu) qmap->mq_map[cpu] = qmap->queue_offset; diff --git a/drivers/virtio/virtio_vdpa.c b/drivers/virtio/virtio_vdpa.c index 7364bd53e38d..cd303ac64046 100644 --- a/drivers/virtio/virtio_vdpa.c +++ b/drivers/virtio/virtio_vdpa.c @@ -330,7 +330,7 @@ create_affinity_masks(unsigned int nvecs, struct irq_affinity *affd) for (i = 0, usedvecs = 0; i < affd->nr_sets; i++) { unsigned int this_vecs = affd->set_size[i]; int j; - struct cpumask *result = group_cpus_evenly(this_vecs); + struct cpumask *result = group_cpus_evenly(this_vecs, true); if (!result) { kfree(masks); diff --git a/fs/fuse/virtio_fs.c b/fs/fuse/virtio_fs.c index f68527891929..41b3bcc03f9c 100644 --- a/fs/fuse/virtio_fs.c +++ b/fs/fuse/virtio_fs.c @@ -881,7 +881,7 @@ static void virtio_fs_map_queues(struct virtio_device *vdev, struct virtio_fs *f return; fallback: /* Attempt to map evenly in groups over the CPUs */ - masks = group_cpus_evenly(fs->num_request_queues); + masks = group_cpus_evenly(fs->num_request_queues, true); /* If even this fails we default to all CPUs use queue zero */ if (!masks) { for_each_possible_cpu(cpu) diff --git a/include/linux/group_cpus.h b/include/linux/group_cpus.h index e42807ec61f6..10a12b9a7ed4 100644 --- a/include/linux/group_cpus.h +++ b/include/linux/group_cpus.h @@ -9,6 +9,6 @@ #include <linux/kernel.h> #include <linux/cpu.h> -struct cpumask *group_cpus_evenly(unsigned int numgrps); +struct cpumask *group_cpus_evenly(unsigned int numgrps, bool is_managed); #endif diff --git a/kernel/irq/affinity.c b/kernel/irq/affinity.c index 44a4eba80315..775ab8537ddc 100644 --- a/kernel/irq/affinity.c +++ b/kernel/irq/affinity.c @@ -64,6 +64,10 @@ irq_create_affinity_masks(unsigned int nvecs, struct irq_affinity *affd) for (curvec = 0; curvec < affd->pre_vectors; curvec++) cpumask_copy(&masks[curvec].mask, irq_default_affinity); + /* Mark the managed interrupts */ + for (i = curvec; i < nvecs - affd->post_vectors; i++) + masks[i].is_managed = 1; + /* * Spread on present CPUs starting from affd->pre_vectors. If we * have multiple sets, build each sets affinity mask separately. @@ -71,7 +75,8 @@ irq_create_affinity_masks(unsigned int nvecs, struct irq_affinity *affd) for (i = 0, usedvecs = 0; i < affd->nr_sets; i++) { unsigned int this_vecs = affd->set_size[i]; int j; - struct cpumask *result = group_cpus_evenly(this_vecs); + struct cpumask *result = group_cpus_evenly(this_vecs, + masks[curvec].is_managed); if (!result) { kfree(masks); @@ -94,10 +99,6 @@ irq_create_affinity_masks(unsigned int nvecs, struct irq_affinity *affd) for (; curvec < nvecs; curvec++) cpumask_copy(&masks[curvec].mask, irq_default_affinity); - /* Mark the managed interrupts */ - for (i = affd->pre_vectors; i < nvecs - affd->post_vectors; i++) - masks[i].is_managed = 1; - return masks; } diff --git a/lib/group_cpus.c b/lib/group_cpus.c index ee272c4cefcc..769a139491bc 100644 --- a/lib/group_cpus.c +++ b/lib/group_cpus.c @@ -11,6 +11,30 @@ #ifdef CONFIG_SMP +static unsigned int __read_mostly managed_irqs_per_node; +static struct cpumask managed_irqs_cpumsk[MAX_NUMNODES] __cacheline_aligned_in_smp = { + [0 ... MAX_NUMNODES-1] = {CPU_BITS_ALL} +}; + +static int __init irq_managed_setup(char *str) +{ + int ret; + + ret = kstrtouint(str, 10, &managed_irqs_per_node); + if (ret < 0) { + pr_warn("managed_irqs_per_node= cannot parse, ignored\n"); + return 0; + } + + if (managed_irqs_per_node * num_possible_nodes() > num_possible_cpus()) { + managed_irqs_per_node = num_possible_cpus() / num_possible_nodes(); + pr_warn("managed_irqs_per_node= cannot be larger than %u\n", + managed_irqs_per_node); + } + return 1; +} +__setup("managed_irqs_per_node=", irq_managed_setup); + static void grp_spread_init_one(struct cpumask *irqmsk, struct cpumask *nmsk, unsigned int cpus_per_grp) { @@ -246,6 +270,30 @@ static void alloc_nodes_groups(unsigned int numgrps, } } +static void __group_prepare_affinity(struct cpumask *premask, + cpumask_var_t *node_to_cpumask) +{ + nodemask_t nodemsk = NODE_MASK_NONE; + unsigned int ncpus, n; + + get_nodes_in_cpumask(node_to_cpumask, premask, &nodemsk); + + for_each_node_mask(n, nodemsk) { + cpumask_and(&managed_irqs_cpumsk[n], &managed_irqs_cpumsk[n], premask); + cpumask_and(&managed_irqs_cpumsk[n], &managed_irqs_cpumsk[n], node_to_cpumask[n]); + + ncpus = cpumask_weight(&managed_irqs_cpumsk[n]); + if (ncpus < managed_irqs_per_node) { + /* Reset node n to current node cpumask */ + cpumask_copy(&managed_irqs_cpumsk[n], node_to_cpumask[n]); + continue; + } + + grp_spread_init_one(premask, &managed_irqs_cpumsk[n], managed_irqs_per_node); + } +} + + static int __group_cpus_evenly(unsigned int startgrp, unsigned int numgrps, cpumask_var_t *node_to_cpumask, const struct cpumask *cpu_mask, @@ -332,6 +380,7 @@ static int __group_cpus_evenly(unsigned int startgrp, unsigned int numgrps, /** * group_cpus_evenly - Group all CPUs evenly per NUMA/CPU locality * @numgrps: number of groups + * @is_managed: if these groups managed by kernel * * Return: cpumask array if successful, NULL otherwise. And each element * includes CPUs assigned to this group @@ -344,7 +393,7 @@ static int __group_cpus_evenly(unsigned int startgrp, unsigned int numgrps, * We guarantee in the resulted grouping that all CPUs are covered, and * no same CPU is assigned to multiple groups */ -struct cpumask *group_cpus_evenly(unsigned int numgrps) +struct cpumask *group_cpus_evenly(unsigned int numgrps, bool is_managed) { unsigned int curgrp = 0, nr_present = 0, nr_others = 0; cpumask_var_t *node_to_cpumask; @@ -382,6 +431,10 @@ struct cpumask *group_cpus_evenly(unsigned int numgrps) */ cpumask_copy(npresmsk, data_race(cpu_present_mask)); + /* Limit the count of managed interrupts on every node */ + if (is_managed && managed_irqs_per_node) + __group_prepare_affinity(npresmsk, node_to_cpumask); + /* grouping present CPUs first */ ret = __group_cpus_evenly(curgrp, numgrps, node_to_cpumask, npresmsk, nmsk, masks);