From c1f5204efcbcced83f67f12fa8f1a7f5f244fb87 Mon Sep 17 00:00:00 2001 From: Yury Norov Date: Sun, 28 Jan 2024 22:21:04 -0800 Subject: [PATCH 1/4] cpumask: add cpumask_weight_andnot() Similarly to cpumask_weight_and(), cpumask_weight_andnot() is a handy helper that may help to avoid creating an intermediate mask just to calculate number of bits that set in a 1st given mask, and clear in 2nd one. Signed-off-by: Yury Norov Reviewed-by: Jacob Keller Signed-off-by: Paolo Abeni --- include/linux/bitmap.h | 12 ++++++++++++ include/linux/cpumask.h | 13 +++++++++++++ lib/bitmap.c | 7 +++++++ 3 files changed, 32 insertions(+) diff --git a/include/linux/bitmap.h b/include/linux/bitmap.h index 99451431e4d6..5814e9ee40ba 100644 --- a/include/linux/bitmap.h +++ b/include/linux/bitmap.h @@ -54,6 +54,7 @@ struct device; * bitmap_full(src, nbits) Are all bits set in *src? * bitmap_weight(src, nbits) Hamming Weight: number set bits * bitmap_weight_and(src1, src2, nbits) Hamming Weight of and'ed bitmap + * bitmap_weight_andnot(src1, src2, nbits) Hamming Weight of andnot'ed bitmap * bitmap_set(dst, pos, nbits) Set specified bit area * bitmap_clear(dst, pos, nbits) Clear specified bit area * bitmap_find_next_zero_area(buf, len, pos, n, mask) Find bit free area @@ -169,6 +170,8 @@ bool __bitmap_subset(const unsigned long *bitmap1, unsigned int __bitmap_weight(const unsigned long *bitmap, unsigned int nbits); unsigned int __bitmap_weight_and(const unsigned long *bitmap1, const unsigned long *bitmap2, unsigned int nbits); +unsigned int __bitmap_weight_andnot(const unsigned long *bitmap1, + const unsigned long *bitmap2, unsigned int nbits); void __bitmap_set(unsigned long *map, unsigned int start, int len); void __bitmap_clear(unsigned long *map, unsigned int start, int len); @@ -425,6 +428,15 @@ unsigned long bitmap_weight_and(const unsigned long *src1, return __bitmap_weight_and(src1, src2, nbits); } +static __always_inline +unsigned long bitmap_weight_andnot(const unsigned long *src1, + const unsigned long *src2, unsigned int nbits) +{ + if (small_const_nbits(nbits)) + return hweight_long(*src1 & ~(*src2) & BITMAP_LAST_WORD_MASK(nbits)); + return __bitmap_weight_andnot(src1, src2, nbits); +} + static __always_inline void bitmap_set(unsigned long *map, unsigned int start, unsigned int nbits) { diff --git a/include/linux/cpumask.h b/include/linux/cpumask.h index cfb545841a2c..228c23eb36d2 100644 --- a/include/linux/cpumask.h +++ b/include/linux/cpumask.h @@ -719,6 +719,19 @@ static inline unsigned int cpumask_weight_and(const struct cpumask *srcp1, return bitmap_weight_and(cpumask_bits(srcp1), cpumask_bits(srcp2), small_cpumask_bits); } +/** + * cpumask_weight_andnot - Count of bits in (*srcp1 & ~*srcp2) + * @srcp1: the cpumask to count bits (< nr_cpu_ids) in. + * @srcp2: the cpumask to count bits (< nr_cpu_ids) in. + * + * Return: count of bits set in both *srcp1 and *srcp2 + */ +static inline unsigned int cpumask_weight_andnot(const struct cpumask *srcp1, + const struct cpumask *srcp2) +{ + return bitmap_weight_andnot(cpumask_bits(srcp1), cpumask_bits(srcp2), small_cpumask_bits); +} + /** * cpumask_shift_right - *dstp = *srcp >> n * @dstp: the cpumask result diff --git a/lib/bitmap.c b/lib/bitmap.c index 09522af227f1..b97692854966 100644 --- a/lib/bitmap.c +++ b/lib/bitmap.c @@ -348,6 +348,13 @@ unsigned int __bitmap_weight_and(const unsigned long *bitmap1, } EXPORT_SYMBOL(__bitmap_weight_and); +unsigned int __bitmap_weight_andnot(const unsigned long *bitmap1, + const unsigned long *bitmap2, unsigned int bits) +{ + return BITMAP_WEIGHT(bitmap1[idx] & ~bitmap2[idx], bits); +} +EXPORT_SYMBOL(__bitmap_weight_andnot); + void __bitmap_set(unsigned long *map, unsigned int start, int len) { unsigned long *p = map + BIT_WORD(start); From dcee228078c34b63089c4b589d4bddf08019d0f6 Mon Sep 17 00:00:00 2001 From: Yury Norov Date: Sun, 28 Jan 2024 22:21:05 -0800 Subject: [PATCH 2/4] cpumask: define cleanup function for cpumasks Now we can simplify code that allocates cpumasks for local needs. Signed-off-by: Yury Norov Signed-off-by: Paolo Abeni --- include/linux/cpumask.h | 3 +++ 1 file changed, 3 insertions(+) diff --git a/include/linux/cpumask.h b/include/linux/cpumask.h index 228c23eb36d2..1c29947db848 100644 --- a/include/linux/cpumask.h +++ b/include/linux/cpumask.h @@ -7,6 +7,7 @@ * set of CPUs in a system, one bit position per CPU number. In general, * only nr_cpu_ids (<= NR_CPUS) bits are valid. */ +#include #include #include #include @@ -990,6 +991,8 @@ static inline bool cpumask_available(cpumask_var_t mask) } #endif /* CONFIG_CPUMASK_OFFSTACK */ +DEFINE_FREE(free_cpumask_var, struct cpumask *, if (_T) free_cpumask_var(_T)); + /* It's common to want to use cpu_all_mask in struct member initializers, * so it has to refer to an address rather than a pointer. */ extern const DECLARE_BITMAP(cpu_all_bits, NR_CPUS); From 91bfe210e196780667c94c2795103ca6013233fc Mon Sep 17 00:00:00 2001 From: Yury Norov Date: Sun, 28 Jan 2024 22:21:06 -0800 Subject: [PATCH 3/4] net: mana: add a function to spread IRQs per CPUs Souradeep investigated that the driver performs faster if IRQs are spread on CPUs with the following heuristics: 1. No more than one IRQ per CPU, if possible; 2. NUMA locality is the second priority; 3. Sibling dislocality is the last priority. Let's consider this topology: Node 0 1 Core 0 1 2 3 CPU 0 1 2 3 4 5 6 7 The most performant IRQ distribution based on the above topology and heuristics may look like this: IRQ Nodes Cores CPUs 0 1 0 0-1 1 1 1 2-3 2 1 0 0-1 3 1 1 2-3 4 2 2 4-5 5 2 3 6-7 6 2 2 4-5 7 2 3 6-7 The irq_setup() routine introduced in this patch leverages the for_each_numa_hop_mask() iterator and assigns IRQs to sibling groups as described above. According to [1], for NUMA-aware but sibling-ignorant IRQ distribution based on cpumask_local_spread() performance test results look like this: ./ntttcp -r -m 16 NTTTCP for Linux 1.4.0 --------------------------------------------------------- 08:05:20 INFO: 17 threads created 08:05:28 INFO: Network activity progressing... 08:06:28 INFO: Test run completed. 08:06:28 INFO: Test cycle finished. 08:06:28 INFO: ##### Totals: ##### 08:06:28 INFO: test duration :60.00 seconds 08:06:28 INFO: total bytes :630292053310 08:06:28 INFO: throughput :84.04Gbps 08:06:28 INFO: retrans segs :4 08:06:28 INFO: cpu cores :192 08:06:28 INFO: cpu speed :3799.725MHz 08:06:28 INFO: user :0.05% 08:06:28 INFO: system :1.60% 08:06:28 INFO: idle :96.41% 08:06:28 INFO: iowait :0.00% 08:06:28 INFO: softirq :1.94% 08:06:28 INFO: cycles/byte :2.50 08:06:28 INFO: cpu busy (all) :534.41% For NUMA- and sibling-aware IRQ distribution, the same test works 15% faster: ./ntttcp -r -m 16 NTTTCP for Linux 1.4.0 --------------------------------------------------------- 08:08:51 INFO: 17 threads created 08:08:56 INFO: Network activity progressing... 08:09:56 INFO: Test run completed. 08:09:56 INFO: Test cycle finished. 08:09:56 INFO: ##### Totals: ##### 08:09:56 INFO: test duration :60.00 seconds 08:09:56 INFO: total bytes :741966608384 08:09:56 INFO: throughput :98.93Gbps 08:09:56 INFO: retrans segs :6 08:09:56 INFO: cpu cores :192 08:09:56 INFO: cpu speed :3799.791MHz 08:09:56 INFO: user :0.06% 08:09:56 INFO: system :1.81% 08:09:56 INFO: idle :96.18% 08:09:56 INFO: iowait :0.00% 08:09:56 INFO: softirq :1.95% 08:09:56 INFO: cycles/byte :2.25 08:09:56 INFO: cpu busy (all) :569.22% [1] https://lore.kernel.org/all/20231211063726.GA4977@linuxonhyperv3.guj3yctzbm1etfxqx2vob5hsef.xx.internal.cloudapp.net/ Signed-off-by: Yury Norov Co-developed-by: Souradeep Chakrabarti Signed-off-by: Paolo Abeni --- .../net/ethernet/microsoft/mana/gdma_main.c | 29 +++++++++++++++++++ 1 file changed, 29 insertions(+) diff --git a/drivers/net/ethernet/microsoft/mana/gdma_main.c b/drivers/net/ethernet/microsoft/mana/gdma_main.c index d33b27214539..05a0ac054823 100644 --- a/drivers/net/ethernet/microsoft/mana/gdma_main.c +++ b/drivers/net/ethernet/microsoft/mana/gdma_main.c @@ -1249,6 +1249,35 @@ void mana_gd_free_res_map(struct gdma_resource *r) r->size = 0; } +static __maybe_unused int irq_setup(unsigned int *irqs, unsigned int len, int node) +{ + const struct cpumask *next, *prev = cpu_none_mask; + cpumask_var_t cpus __free(free_cpumask_var); + int cpu, weight; + + if (!alloc_cpumask_var(&cpus, GFP_KERNEL)) + return -ENOMEM; + + rcu_read_lock(); + for_each_numa_hop_mask(next, node) { + weight = cpumask_weight_andnot(next, prev); + while (weight > 0) { + cpumask_andnot(cpus, next, prev); + for_each_cpu(cpu, cpus) { + if (len-- == 0) + goto done; + irq_set_affinity_and_hint(*irqs++, topology_sibling_cpumask(cpu)); + cpumask_andnot(cpus, cpus, topology_sibling_cpumask(cpu)); + --weight; + } + } + prev = next; + } +done: + rcu_read_unlock(); + return 0; +} + static int mana_gd_setup_irqs(struct pci_dev *pdev) { unsigned int max_queues_per_port = num_online_cpus(); From 8afefc361209b726aa5886833384d048412b159e Mon Sep 17 00:00:00 2001 From: Souradeep Chakrabarti Date: Sun, 28 Jan 2024 22:21:07 -0800 Subject: [PATCH 4/4] net: mana: Assigning IRQ affinity on HT cores Existing MANA design assigns IRQ to every CPU, including sibling hyper-threads. This may cause multiple IRQs to be active simultaneously in the same core and may reduce the network performance. Improve the performance by assigning IRQ to non sibling CPUs in local NUMA node. The performance improvement we are getting using ntttcp with following patch is around 15 percent against existing design and approximately 11 percent, when trying to assign one IRQ in each core across NUMA nodes, if enough cores are present. The change will improve the performance for the system with high number of CPU, where number of CPUs in a node is more than 64 CPUs. Nodes with 64 CPUs or less than 64 CPUs will not be affected by this change. The performance study was done using ntttcp tool in Azure. The node had 2 nodes with 32 cores each, total 128 vCPU and number of channels were 32 for 32 RX rings. The below table shows a comparison between existing design and new design: IRQ node-num core-num CPU performance(%) 1 0 | 0 0 | 0 0 | 0-1 0 2 0 | 0 0 | 1 1 | 2-3 3 3 0 | 0 1 | 2 2 | 4-5 10 4 0 | 0 1 | 3 3 | 6-7 15 5 0 | 0 2 | 4 4 | 8-9 15 ... ... 25 0 | 0 12| 24 24| 48-49 12 ... 32 0 | 0 15| 31 31| 62-63 12 33 0 | 0 16| 0 32| 0-1 10 ... 64 0 | 0 31| 31 63| 62-63 0 Signed-off-by: Souradeep Chakrabarti Signed-off-by: Paolo Abeni --- .../net/ethernet/microsoft/mana/gdma_main.c | 59 +++++++++++++++---- 1 file changed, 49 insertions(+), 10 deletions(-) diff --git a/drivers/net/ethernet/microsoft/mana/gdma_main.c b/drivers/net/ethernet/microsoft/mana/gdma_main.c index 05a0ac054823..1332db9a08eb 100644 --- a/drivers/net/ethernet/microsoft/mana/gdma_main.c +++ b/drivers/net/ethernet/microsoft/mana/gdma_main.c @@ -1249,7 +1249,7 @@ void mana_gd_free_res_map(struct gdma_resource *r) r->size = 0; } -static __maybe_unused int irq_setup(unsigned int *irqs, unsigned int len, int node) +static int irq_setup(unsigned int *irqs, unsigned int len, int node) { const struct cpumask *next, *prev = cpu_none_mask; cpumask_var_t cpus __free(free_cpumask_var); @@ -1280,13 +1280,16 @@ done: static int mana_gd_setup_irqs(struct pci_dev *pdev) { - unsigned int max_queues_per_port = num_online_cpus(); struct gdma_context *gc = pci_get_drvdata(pdev); + unsigned int max_queues_per_port; struct gdma_irq_context *gic; unsigned int max_irqs, cpu; - int nvec, irq; + int start_irq_index = 1; + int nvec, *irqs, irq; int err, i = 0, j; + cpus_read_lock(); + max_queues_per_port = num_online_cpus(); if (max_queues_per_port > MANA_MAX_NUM_QUEUES) max_queues_per_port = MANA_MAX_NUM_QUEUES; @@ -1294,8 +1297,18 @@ static int mana_gd_setup_irqs(struct pci_dev *pdev) max_irqs = max_queues_per_port + 1; nvec = pci_alloc_irq_vectors(pdev, 2, max_irqs, PCI_IRQ_MSIX); - if (nvec < 0) + if (nvec < 0) { + cpus_read_unlock(); return nvec; + } + if (nvec <= num_online_cpus()) + start_irq_index = 0; + + irqs = kmalloc_array((nvec - start_irq_index), sizeof(int), GFP_KERNEL); + if (!irqs) { + err = -ENOMEM; + goto free_irq_vector; + } gc->irq_contexts = kcalloc(nvec, sizeof(struct gdma_irq_context), GFP_KERNEL); @@ -1323,17 +1336,41 @@ static int mana_gd_setup_irqs(struct pci_dev *pdev) goto free_irq; } - err = request_irq(irq, mana_gd_intr, 0, gic->name, gic); - if (err) - goto free_irq; + if (!i) { + err = request_irq(irq, mana_gd_intr, 0, gic->name, gic); + if (err) + goto free_irq; - cpu = cpumask_local_spread(i, gc->numa_node); - irq_set_affinity_and_hint(irq, cpumask_of(cpu)); + /* If number of IRQ is one extra than number of online CPUs, + * then we need to assign IRQ0 (hwc irq) and IRQ1 to + * same CPU. + * Else we will use different CPUs for IRQ0 and IRQ1. + * Also we are using cpumask_local_spread instead of + * cpumask_first for the node, because the node can be + * mem only. + */ + if (start_irq_index) { + cpu = cpumask_local_spread(i, gc->numa_node); + irq_set_affinity_and_hint(irq, cpumask_of(cpu)); + } else { + irqs[start_irq_index] = irq; + } + } else { + irqs[i - start_irq_index] = irq; + err = request_irq(irqs[i - start_irq_index], mana_gd_intr, 0, + gic->name, gic); + if (err) + goto free_irq; + } } + err = irq_setup(irqs, (nvec - start_irq_index), gc->numa_node); + if (err) + goto free_irq; + gc->max_num_msix = nvec; gc->num_msix_usable = nvec; - + cpus_read_unlock(); return 0; free_irq: @@ -1346,8 +1383,10 @@ free_irq: } kfree(gc->irq_contexts); + kfree(irqs); gc->irq_contexts = NULL; free_irq_vector: + cpus_read_unlock(); pci_free_irq_vectors(pdev); return err; }