On 12/01/2015 10:02 PM, [email protected] wrote:
> From: Wen Xiong <[email protected]>
>
> Enable MSIX vector mapping to CPU affinity. This feature is
> enabled by setting ipr_cpu_map=1 when loading ipr module.
>
> Signed-off-by: Wen Xiong <[email protected]>
> ---
> drivers/scsi/ipr.c | 220
> +++++++++++++++++++++++++++++++++++++++++++++++++++-
> drivers/scsi/ipr.h | 16 ++++
> 2 files changed, 235 insertions(+), 1 deletions(-)
>
> diff --git a/drivers/scsi/ipr.c b/drivers/scsi/ipr.c
> index 9f30887..494c0fc 100644
> --- a/drivers/scsi/ipr.c
> +++ b/drivers/scsi/ipr.c
> @@ -100,6 +100,11 @@ static unsigned int ipr_max_devs =
> IPR_DEFAULT_SIS64_DEVS;
> static unsigned int ipr_dual_ioa_raid = 1;
> static unsigned int ipr_number_of_msix = 16;
> static unsigned int ipr_fast_reboot;
> +static unsigned int ipr_cpu_map;
> +
> +static unsigned int *ipr_used_cpu;
How about just defining this like:
static unsigned int ipr_used_cpu[NR_CPUS];
Then we can eliminate the kmalloc / kfree of this and simplify this bit of the
code slightly.
> +static unsigned int ipr_possible_cpu_cnt;
> +
> static DEFINE_SPINLOCK(ipr_driver_lock);
>
> /* This table describes the differences between DMA controller chips */
> @@ -224,6 +229,8 @@ module_param_named(number_of_msix, ipr_number_of_msix,
> int, 0);
> MODULE_PARM_DESC(number_of_msix, "Specify the number of MSIX interrupts to
> use on capable adapters (1 - 16). (default:16)");
> module_param_named(fast_reboot, ipr_fast_reboot, int, S_IRUGO | S_IWUSR);
> MODULE_PARM_DESC(fast_reboot, "Skip adapter shutdown during reboot. Set to 1
> to enable. (default: 0)");
> +module_param_named(cpu_map, ipr_cpu_map, int, 0);
> +MODULE_PARM_DESC(cpu_map, "Enable CPU affinity per adapter. (default:0)");
Can we enable this by default? Would be nice for the default settings to
provide the best performance
under normal situations.
> MODULE_LICENSE("GPL");
> MODULE_VERSION(IPR_DRIVER_VERSION);
>
> @@ -1053,6 +1060,17 @@ static void ipr_send_blocking_cmd(struct ipr_cmnd
> *ipr_cmd,
> static int ipr_get_hrrq_index(struct ipr_ioa_cfg *ioa_cfg)
> {
> unsigned int hrrq;
> + struct ipr_vector_map_info *map_per_cpu;
> + u32 cpu;
> +
> + if (ioa_cfg->cpu_map && ioa_cfg->hrrq_num > 1) {
> + cpu = raw_smp_processor_id();
Not sure this is what we want here. I think what we want to do is to set the
irq affinity to be any thread on
the core, so you'd find the first available CPU, then use cpu_sibling_mask to
figure out all the CPUs on the core
which is what you'd use to set the affinity. Then in here, we'd just use
smp_processor_id() and find the HRRQ
that maps to.
I'm OK leaving this as a follow on optimization though since I'd like to get
this patch moving.
> + map_per_cpu = ioa_cfg->cpu_map_tbl;
> + if (cpu < ioa_cfg->possible_cpu_cnt) {
> + map_per_cpu += cpu;
> + return map_per_cpu->hrrq_id;
> + }
> + }
>
> if (ioa_cfg->hrrq_num == 1)
> hrrq = 0;
> @@ -4086,6 +4104,77 @@ static struct device_attribute ipr_ioa_fw_type_attr = {
> .show = ipr_show_fw_type
> };
>
> +static ssize_t
> +ipr_show_cpu_map_tbl(struct device *dev, struct device_attribute *attr,
> + char *buf)
> +{
> + struct Scsi_Host *shost = class_to_shost(dev);
> + struct ipr_ioa_cfg *ioa_cfg = (struct ipr_ioa_cfg *)shost->hostdata;
> + struct ipr_vector_map_info *map_per_cpu;
> + int len = 0, show_current_cpu = 0;
> +
> + if (ioa_cfg->intr_flag != IPR_USE_MSIX)
> + return len;
> +
> + switch (ioa_cfg->cpu_map) {
> + case 0:
> + len += snprintf(buf + len, PAGE_SIZE - len,
> + "ipr driver cpu map: No mapping (%d)\n",
> + ioa_cfg->cpu_map);
> + return len;
> + case 1:
> + len += snprintf(buf + len, PAGE_SIZE - len,
> + "ipr driver cpu map: IRQ vector to CPU mapping (%d): "
> + "%d online CPUs\n",
> + ioa_cfg->cpu_map,
> + num_online_cpus());
> + break;
> + }
> +
> + while (show_current_cpu < ioa_cfg->possible_cpu_cnt) {
> + map_per_cpu = &ioa_cfg->cpu_map_tbl[show_current_cpu];
> +
> + if (map_per_cpu->irq_id == IPR_VECTOR_MAP_EMPTY)
> + len += snprintf(buf + len, PAGE_SIZE - len,
> + "CPU %02d io_hrrq %02d\n",
> + show_current_cpu,
> + map_per_cpu->hrrq_id);
> + else
> + len += snprintf(buf + len, PAGE_SIZE - len,
> + "CPU %02d io_hrrq %02d "
> + " IRQ=%d\n",
> + show_current_cpu,
> + map_per_cpu->hrrq_id,
> map_per_cpu->irq_id);
> +
> + show_current_cpu++;
> +
> + if (show_current_cpu < ioa_cfg->possible_cpu_cnt
> + && (len >= (PAGE_SIZE - 64))) {
> + len += snprintf(buf + len, PAGE_SIZE - len,
> "more...\n");
> + break;
> + }
> + }
> +
> + return len;
> +}
> +
> +static ssize_t
> +ipr_store_cpu_map_tbl(struct device *dev, struct device_attribute *attr,
> + const char *buf, size_t count)
> +{
> + int status = -EINVAL;
> + return status;
> +}
> +
> +static struct device_attribute ipr_cpu_map_attr = {
> + .attr = {
> + .name = "cpu_map",
> + .mode = S_IRUGO | S_IWUSR,
Since we don't support writing the attribute, let's change the mode to just be
S_IRUGO
and then we don't need to have a ipr_store_cpu_map_tbl function.
> + },
> + .show = ipr_show_cpu_map_tbl,
> + .store = ipr_store_cpu_map_tbl
> +};
> +
> static struct device_attribute *ipr_ioa_attrs[] = {
> &ipr_fw_version_attr,
> &ipr_log_level_attr,
> @@ -4095,6 +4184,7 @@ static struct device_attribute *ipr_ioa_attrs[] = {
> &ipr_update_fw_attr,
> &ipr_ioa_fw_type_attr,
> &ipr_iopoll_weight_attr,
> + &ipr_cpu_map_attr,
> NULL,
> };
>
> @@ -9333,6 +9423,8 @@ static void ipr_free_mem(struct ipr_ioa_cfg *ioa_cfg)
>
> ipr_free_dump(ioa_cfg);
> kfree(ioa_cfg->trace);
> + if (ioa_cfg->cpu_map)
> + kfree(ioa_cfg->cpu_map_tbl);
Its valid to call kfree on a NULL pointer, so the if check isn't technically
required.
> }
>
> /**
> @@ -9352,9 +9444,13 @@ static void ipr_free_irqs(struct ipr_ioa_cfg *ioa_cfg)
> if (ioa_cfg->intr_flag == IPR_USE_MSI ||
> ioa_cfg->intr_flag == IPR_USE_MSIX) {
> int i;
> - for (i = 0; i < ioa_cfg->nvectors; i++)
> + for (i = 0; i < ioa_cfg->nvectors; i++) {
> + if (ioa_cfg->cpu_map)
> + irq_set_affinity_hint(
> + ioa_cfg->vectors_info[i].vec, NULL);
> free_irq(ioa_cfg->vectors_info[i].vec,
> &ioa_cfg->hrrq[i]);
> + }
> } else
> free_irq(pdev->irq, &ioa_cfg->hrrq[0]);
>
> @@ -9585,11 +9681,35 @@ static int ipr_alloc_mem(struct ipr_ioa_cfg *ioa_cfg)
> if (!ioa_cfg->trace)
> goto out_free_hostrcb_dma;
>
> + if (ioa_cfg->cpu_map) {
> + ioa_cfg->cpu_map_tbl =
> + kzalloc(sizeof(struct ipr_vector_map_info) *
> + ioa_cfg->possible_cpu_cnt, GFP_KERNEL);
> +
> + if (!ioa_cfg->cpu_map_tbl)
> + goto out_free_trace;
> + }
> +
> + if (ipr_used_cpu == NULL) {
> + ipr_used_cpu = kzalloc((sizeof(uint16_t) *
> + ipr_possible_cpu_cnt), GFP_KERNEL);
> + if (!ipr_used_cpu)
> + goto out_free_cpu_map_tbl;
> +
> + for (i = 0; i < ipr_possible_cpu_cnt; i++)
> + ipr_used_cpu[i] = IPR_VECTOR_MAP_EMPTY;
> + }
> +
> rc = 0;
> out:
> LEAVE;
> return rc;
>
> +out_free_cpu_map_tbl:
> + if (ioa_cfg->cpu_map)
> + kfree(ioa_cfg->cpu_map_tbl);
> +out_free_trace:
> + kfree(ioa_cfg->trace);
> out_free_hostrcb_dma:
> while (i-- > 0) {
> dma_free_coherent(&pdev->dev, sizeof(struct ipr_hostrcb),
> @@ -9791,6 +9911,92 @@ static void ipr_wait_for_pci_err_recovery(struct
> ipr_ioa_cfg *ioa_cfg)
> }
> }
>
> +static int
> +ipr_find_next_free_cpu(struct ipr_ioa_cfg *ioa_cfg, u32 cpu_id)
> +{
> + struct ipr_vector_map_info *map_per_cpu;
> + int i;
> +
> + map_per_cpu = ioa_cfg->cpu_map_tbl;
> + for (i = 0; i < ioa_cfg->possible_cpu_cnt; i++) {
> + if (cpu_online(i)) {
> + if ((map_per_cpu->irq_id == IPR_VECTOR_MAP_EMPTY) &&
> + (ipr_used_cpu[i] == IPR_VECTOR_MAP_EMPTY) &&
> + (map_per_cpu->cpu_id == cpu_id))
> + return i;
> + }
> + map_per_cpu++;
> + }
> +
> + return IPR_VECTOR_MAP_EMPTY;
> +}
> +
> +static int ipr_set_cpu_affinity_for_hrrq(struct ipr_ioa_cfg *ioa_cfg,
> + int vectors)
> +{
> + struct ipr_vector_map_info *map_per_cpu;
> + int cpu_idx = 0, hrrq_num, next_cpu_id, i;
> +
> + memset(ioa_cfg->cpu_map_tbl, 0xff,
> + (sizeof(struct ipr_vector_map_info) *
> + ioa_cfg->possible_cpu_cnt));
> +
> + map_per_cpu = ioa_cfg->cpu_map_tbl;
> + for (i = 0; i < ioa_cfg->possible_cpu_cnt; i++, map_per_cpu++)
> + map_per_cpu->cpu_id = i;
> +
> + for (i = 0; i < ipr_possible_cpu_cnt; i++) {
> + if (ipr_used_cpu[i] == IPR_VECTOR_MAP_EMPTY) {
> + cpu_idx = i;
> + break;
> + }
> + }
> +
> + hrrq_num = 0;
> + for (i = 0; i < vectors; i++) {
> + map_per_cpu = ioa_cfg->cpu_map_tbl;
> + next_cpu_id = ipr_find_next_free_cpu(ioa_cfg, cpu_idx);
> +
> + if (next_cpu_id == IPR_VECTOR_MAP_EMPTY) {
> + int j;
> +
> + for (j = 0; j < i; j++)
> + irq_set_affinity_hint(
> + ioa_cfg->vectors_info[j].vec, NULL);
> + ioa_cfg->cpu_map = 0;
> + dev_info(&ioa_cfg->pdev->dev, "Cannot set cpu
> affinity\n");
> + return 0;
> + }
> +
> + map_per_cpu += next_cpu_id;
> + ipr_used_cpu[next_cpu_id] = next_cpu_id;
> +
> + map_per_cpu->irq_id = ioa_cfg->vectors_info[i].vec;
> + map_per_cpu->hrrq_id = hrrq_num;
> + irq_set_affinity_hint(ioa_cfg->vectors_info[i].vec,
> + get_cpu_mask(next_cpu_id));
Hmm. It looks like you've changed ipr_get_hrrq_index to use
raw_smp_processor_id(), which
should return
It would be interesting to see how performance is impacted here if we set the
affinity
hint to be every
> + cpu_idx++;
> + hrrq_num++;
> + }
> +
> + hrrq_num = 0;
> + if (vectors < ioa_cfg->possible_cpu_cnt) {
> + map_per_cpu = ioa_cfg->cpu_map_tbl;
> + for (i = 0; i < ioa_cfg->possible_cpu_cnt; i++) {
> + if (map_per_cpu->hrrq_id == IPR_VECTOR_MAP_EMPTY) {
> + map_per_cpu->hrrq_id = hrrq_num;
> + hrrq_num++;
> + }
> +
> + if (hrrq_num == ioa_cfg->hrrq_num)
> + hrrq_num = 0;
> + map_per_cpu++;
> + }
> + }
> +
> + return 1;
> +}
> +
> static int ipr_enable_msix(struct ipr_ioa_cfg *ioa_cfg)
> {
> struct msix_entry entries[IPR_MAX_MSIX_VECTORS];
> @@ -10045,6 +10251,7 @@ static int ipr_probe_ioa(struct pci_dev *pdev,
> ioa_cfg->hdw_dma_regs = ipr_regs;
> ioa_cfg->hdw_dma_regs_pci = ipr_regs_pci;
> ioa_cfg->ioa_mailbox = ioa_cfg->chip_cfg->mailbox + ipr_regs;
> + ioa_cfg->cpu_map = ipr_cpu_map;
>
> ipr_init_regs(ioa_cfg);
>
> @@ -10106,6 +10313,8 @@ static int ipr_probe_ioa(struct pci_dev *pdev,
> }
> }
>
> + ioa_cfg->possible_cpu_cnt = ipr_possible_cpu_cnt;
> +
> if (ioa_cfg->intr_flag == IPR_USE_MSI ||
> ioa_cfg->intr_flag == IPR_USE_MSIX) {
> rc = ipr_test_msi(ioa_cfg, pdev);
> @@ -10200,6 +10409,9 @@ static int ipr_probe_ioa(struct pci_dev *pdev,
> goto cleanup_nolog;
> }
>
> + if (ioa_cfg->cpu_map)
> + ipr_set_cpu_affinity_for_hrrq(ioa_cfg, ioa_cfg->nvectors);
> +
> if ((dev_id->driver_data & IPR_USE_PCI_WARM_RESET) ||
> (dev_id->device == PCI_DEVICE_ID_IBM_OBSIDIAN_E &&
> !ioa_cfg->revid)) {
> ioa_cfg->needs_warm_reset = 1;
> @@ -10644,9 +10856,14 @@ static struct notifier_block ipr_notifier = {
> **/
> static int __init ipr_init(void)
> {
> + int cpu;
> ipr_info("IBM Power RAID SCSI Device Driver version: %s %s\n",
> IPR_DRIVER_VERSION, IPR_DRIVER_DATE);
>
> + ipr_used_cpu = NULL;
> + for_each_present_cpu(cpu)
> + ipr_possible_cpu_cnt++;
I don't think this will handle CPU hotplug, but we can look at adding that
support
in a follow on patch.
Also, this could be simplified to just use num_present_cpus().
> +
> register_reboot_notifier(&ipr_notifier);
> return pci_register_driver(&ipr_driver);
> }
> @@ -10663,6 +10880,7 @@ static void __exit ipr_exit(void)
> {
> unregister_reboot_notifier(&ipr_notifier);
> pci_unregister_driver(&ipr_driver);
> + kfree(ipr_used_cpu);
> }
>
> module_init(ipr_init);
> diff --git a/drivers/scsi/ipr.h b/drivers/scsi/ipr.h
> index a34c7a5..a7ef83f 100644
> --- a/drivers/scsi/ipr.h
> +++ b/drivers/scsi/ipr.h
> @@ -1464,6 +1464,16 @@ enum ipr_sdt_state {
> DUMP_OBTAINED
> };
>
> +#define IPR_VECTOR_MAP_EMPTY 0xffff
> +#define IPR_SET_IO_CPU_AFFINITY 0x1
> +
> +/* IRQ vector to CPU mapping when set CPU affinity*/
> +struct ipr_vector_map_info {
> + u16 cpu_id;
> + u16 irq_id;
> + u16 hrrq_id;
> +};
> +
> /* Per-controller data */
> struct ipr_ioa_cfg {
> char eye_catcher[8];
> @@ -1595,6 +1605,12 @@ struct ipr_ioa_cfg {
>
> u32 iopoll_weight;
>
> + u32 cpu_map;
> + struct ipr_vector_map_info *cpu_map_tbl;
> + u16 possible_cpu_cnt;
> +
> + u32 use_blk_mq;
> +
> }; /* struct ipr_ioa_cfg */
>
> struct ipr_cmnd {
>
--
Brian King
Power Linux I/O
IBM Linux Technology Center
------------------------------------------------------------------------------
_______________________________________________
Iprdd-devel mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/iprdd-devel
