diff options
author | Dick Kennedy <dick.kennedy@broadcom.com> | 2020-05-02 00:43:06 +0300 |
---|---|---|
committer | Martin K. Petersen <martin.petersen@oracle.com> | 2020-05-08 05:47:24 +0300 |
commit | 3048e3e805e36a61cf02f185b02b4144f46d8ff3 (patch) | |
tree | 4e4ad1788e1a7b8260076869438851532330b0c6 /drivers/scsi/lpfc/lpfc_attr.c | |
parent | f809da6db68a8be49e317f0ccfbced1af9258839 (diff) | |
download | linux-3048e3e805e36a61cf02f185b02b4144f46d8ff3.tar.xz |
scsi: lpfc: Change default queue allocation for reduced memory consumption
By default, the driver attempts to allocate a hdwq per logical cpu in order
to provide good cpu affinity. Some systems have extremely high cpu counts
and this can significantly raise memory consumption.
In testing on x86 platforms (non-AMD) it is found that sharing of a hdwq by
a physical cpu and its HT cpu can occur with little performance
degredation. By sharing, the hdwq count can be halved, significantly
reducing the memory overhead.
Change the default behavior of the driver on non-AMD x86 platforms to
share a hdwq by the cpu and its HT cpu.
Link: https://lore.kernel.org/r/20200501214310.91713-6-jsmart2021@gmail.com
Reviewed-by: Hannes Reinecke <hare@suse.de>
Signed-off-by: Dick Kennedy <dick.kennedy@broadcom.com>
Signed-off-by: James Smart <jsmart2021@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
Diffstat (limited to 'drivers/scsi/lpfc/lpfc_attr.c')
-rw-r--r-- | drivers/scsi/lpfc/lpfc_attr.c | 106 |
1 files changed, 84 insertions, 22 deletions
diff --git a/drivers/scsi/lpfc/lpfc_attr.c b/drivers/scsi/lpfc/lpfc_attr.c index 1354c141d614..2791efa770af 100644 --- a/drivers/scsi/lpfc/lpfc_attr.c +++ b/drivers/scsi/lpfc/lpfc_attr.c @@ -5704,17 +5704,69 @@ LPFC_ATTR_R(hdw_queue, LPFC_HBA_HDWQ_MIN, LPFC_HBA_HDWQ_MAX, "Set the number of I/O Hardware Queues"); -static inline void -lpfc_assign_default_irq_numa(struct lpfc_hba *phba) +#if IS_ENABLED(CONFIG_X86) +/** + * lpfc_cpumask_irq_mode_init - initalizes cpumask of phba based on + * irq_chann_mode + * @phba: Pointer to HBA context object. + **/ +static void +lpfc_cpumask_irq_mode_init(struct lpfc_hba *phba) +{ + unsigned int cpu, first_cpu, numa_node = NUMA_NO_NODE; + const struct cpumask *sibling_mask; + struct cpumask *aff_mask = &phba->sli4_hba.irq_aff_mask; + + cpumask_clear(aff_mask); + + if (phba->irq_chann_mode == NUMA_MODE) { + /* Check if we're a NUMA architecture */ + numa_node = dev_to_node(&phba->pcidev->dev); + if (numa_node == NUMA_NO_NODE) { + phba->irq_chann_mode = NORMAL_MODE; + return; + } + } + + for_each_possible_cpu(cpu) { + switch (phba->irq_chann_mode) { + case NUMA_MODE: + if (cpu_to_node(cpu) == numa_node) + cpumask_set_cpu(cpu, aff_mask); + break; + case NHT_MODE: + sibling_mask = topology_sibling_cpumask(cpu); + first_cpu = cpumask_first(sibling_mask); + if (first_cpu < nr_cpu_ids) + cpumask_set_cpu(first_cpu, aff_mask); + break; + default: + break; + } + } +} +#endif + +static void +lpfc_assign_default_irq_chann(struct lpfc_hba *phba) { #if IS_ENABLED(CONFIG_X86) - /* If AMD architecture, then default is LPFC_IRQ_CHANN_NUMA */ - if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD) - phba->cfg_irq_numa = 1; - else - phba->cfg_irq_numa = 0; + switch (boot_cpu_data.x86_vendor) { + case X86_VENDOR_AMD: + /* If AMD architecture, then default is NUMA_MODE */ + phba->irq_chann_mode = NUMA_MODE; + break; + case X86_VENDOR_INTEL: + /* If Intel architecture, then default is no hyperthread mode */ + phba->irq_chann_mode = NHT_MODE; + break; + default: + phba->irq_chann_mode = NORMAL_MODE; + break; + } + lpfc_cpumask_irq_mode_init(phba); #else - phba->cfg_irq_numa = 0; + phba->irq_chann_mode = NORMAL_MODE; #endif } @@ -5726,6 +5778,7 @@ lpfc_assign_default_irq_numa(struct lpfc_hba *phba) * * 0 = Configure number of IRQ Channels to: * if AMD architecture, number of CPUs on HBA's NUMA node + * if Intel architecture, number of physical CPUs. * otherwise, number of active CPUs. * [1,256] = Manually specify how many IRQ Channels to use. * @@ -5751,35 +5804,44 @@ MODULE_PARM_DESC(lpfc_irq_chann, "Set number of interrupt vectors to allocate"); static int lpfc_irq_chann_init(struct lpfc_hba *phba, uint32_t val) { - const struct cpumask *numa_mask; + const struct cpumask *aff_mask; if (phba->cfg_use_msi != 2) { lpfc_printf_log(phba, KERN_INFO, LOG_INIT, "8532 use_msi = %u ignoring cfg_irq_numa\n", phba->cfg_use_msi); - phba->cfg_irq_numa = 0; - phba->cfg_irq_chann = LPFC_IRQ_CHANN_MIN; + phba->irq_chann_mode = NORMAL_MODE; + phba->cfg_irq_chann = LPFC_IRQ_CHANN_DEF; return 0; } /* Check if default setting was passed */ if (val == LPFC_IRQ_CHANN_DEF) - lpfc_assign_default_irq_numa(phba); + lpfc_assign_default_irq_chann(phba); - if (phba->cfg_irq_numa) { - numa_mask = &phba->sli4_hba.numa_mask; + if (phba->irq_chann_mode != NORMAL_MODE) { + aff_mask = &phba->sli4_hba.irq_aff_mask; - if (cpumask_empty(numa_mask)) { + if (cpumask_empty(aff_mask)) { lpfc_printf_log(phba, KERN_INFO, LOG_INIT, - "8533 Could not identify NUMA node, " - "ignoring cfg_irq_numa\n"); - phba->cfg_irq_numa = 0; - phba->cfg_irq_chann = LPFC_IRQ_CHANN_MIN; + "8533 Could not identify CPUS for " + "mode %d, ignoring\n", + phba->irq_chann_mode); + phba->irq_chann_mode = NORMAL_MODE; + phba->cfg_irq_chann = LPFC_IRQ_CHANN_DEF; } else { - phba->cfg_irq_chann = cpumask_weight(numa_mask); + phba->cfg_irq_chann = cpumask_weight(aff_mask); + + /* If no hyperthread mode, then set hdwq count to + * aff_mask weight as well + */ + if (phba->irq_chann_mode == NHT_MODE) + phba->cfg_hdw_queue = phba->cfg_irq_chann; + lpfc_printf_log(phba, KERN_INFO, LOG_INIT, "8543 lpfc_irq_chann set to %u " - "(numa)\n", phba->cfg_irq_chann); + "(mode: %d)\n", phba->cfg_irq_chann, + phba->irq_chann_mode); } } else { if (val > LPFC_IRQ_CHANN_MAX) { @@ -5790,7 +5852,7 @@ lpfc_irq_chann_init(struct lpfc_hba *phba, uint32_t val) val, LPFC_IRQ_CHANN_MIN, LPFC_IRQ_CHANN_MAX); - phba->cfg_irq_chann = LPFC_IRQ_CHANN_MIN; + phba->cfg_irq_chann = LPFC_IRQ_CHANN_DEF; return -EINVAL; } phba->cfg_irq_chann = val; |