UbiquitousOS
/
OpenCloudOS-Kernel
mirror of https://gitee.com/OpenCloudOS/OpenCloudOS-Kernel.git
11
0
Fork 0
Code Issues Projects Releases Wiki Activity

[SCSI] lpfc 8.3.39: Fixed driver vector mapping to CPU affinity 

Signed-off-by: James Smart <james.smart@emulex.com>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
This commit is contained in:
James Smart 2013-04-17 20:19:16 -04:00 committed by James Bottomley
parent 6a485eb9a9
commit 7bb03bbf1e
6 changed files with 488 additions and 15 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

1
drivers/scsi/lpfc/lpfc.h
View File

@ -709,6 +709,7 @@ struct lpfc_hba {
uint32_t cfg_poll_tmo;
uint32_t cfg_use_msi;
uint32_t cfg_fcp_imax;
uint32_t cfg_fcp_cpu_map;
uint32_t cfg_fcp_wq_count;
uint32_t cfg_fcp_eq_count;
uint32_t cfg_fcp_io_channel;

137
drivers/scsi/lpfc/lpfc_attr.c
View File

@ -3799,6 +3799,141 @@ lpfc_fcp_imax_init(struct lpfc_hba *phba, int val)
static DEVICE_ATTR(lpfc_fcp_imax, S_IRUGO | S_IWUSR,
lpfc_fcp_imax_show, lpfc_fcp_imax_store);
/**
* lpfc_state_show - Display current driver CPU affinity
* @dev: class converted to a Scsi_host structure.
* @attr: device attribute, not used.
* @buf: on return contains text describing the state of the link.
*
* Returns: size of formatted string.
**/
static ssize_t
lpfc_fcp_cpu_map_show(struct device *dev, struct device_attribute *attr,
char *buf)
{
struct Scsi_Host *shost = class_to_shost(dev);
struct lpfc_vport *vport = (struct lpfc_vport *)shost->hostdata;
struct lpfc_hba *phba = vport->phba;
struct lpfc_vector_map_info *cpup;
int idx, len = 0;
if ((phba->sli_rev != LPFC_SLI_REV4) ||
(phba->intr_type != MSIX))
return len;
switch (phba->cfg_fcp_cpu_map) {
case 0:
len += snprintf(buf + len, PAGE_SIZE-len,
"fcp_cpu_map: No mapping (%d)\n",
phba->cfg_fcp_cpu_map);
return len;
case 1:
len += snprintf(buf + len, PAGE_SIZE-len,
"fcp_cpu_map: HBA centric mapping (%d): "
"%d online CPUs\n",
phba->cfg_fcp_cpu_map,
phba->sli4_hba.num_online_cpu);
break;
case 2:
len += snprintf(buf + len, PAGE_SIZE-len,
"fcp_cpu_map: Driver centric mapping (%d): "
"%d online CPUs\n",
phba->cfg_fcp_cpu_map,
phba->sli4_hba.num_online_cpu);
break;
}
cpup = phba->sli4_hba.cpu_map;
for (idx = 0; idx < phba->sli4_hba.num_present_cpu; idx++) {
if (cpup->irq == LPFC_VECTOR_MAP_EMPTY)
len += snprintf(buf + len, PAGE_SIZE-len,
"CPU %02d io_chan %02d "
"physid %d coreid %d\n",
idx, cpup->channel_id, cpup->phys_id,
cpup->core_id);
else
len += snprintf(buf + len, PAGE_SIZE-len,
"CPU %02d io_chan %02d "
"physid %d coreid %d IRQ %d\n",
idx, cpup->channel_id, cpup->phys_id,
cpup->core_id, cpup->irq);
cpup++;
}
return len;
}
/**
* lpfc_fcp_cpu_map_store - Change CPU affinity of driver vectors
* @dev: class device that is converted into a Scsi_host.
* @attr: device attribute, not used.
* @buf: one or more lpfc_polling_flags values.
* @count: not used.
*
* Returns:
* -EINVAL - Not implemented yet.
**/
static ssize_t
lpfc_fcp_cpu_map_store(struct device *dev, struct device_attribute *attr,
const char *buf, size_t count)
{
int status = -EINVAL;
return status;
}
/*
# lpfc_fcp_cpu_map: Defines how to map CPUs to IRQ vectors
# for the HBA.
#
# Value range is [0 to 2]. Default value is LPFC_DRIVER_CPU_MAP (2).
# 0 - Do not affinitze IRQ vectors
# 1 - Affintize HBA vectors with respect to each HBA
# (start with CPU0 for each HBA)
# 2 - Affintize HBA vectors with respect to the entire driver
# (round robin thru all CPUs across all HBAs)
*/
static int lpfc_fcp_cpu_map = LPFC_DRIVER_CPU_MAP;
module_param(lpfc_fcp_cpu_map, int, S_IRUGO|S_IWUSR);
MODULE_PARM_DESC(lpfc_fcp_cpu_map,
"Defines how to map CPUs to IRQ vectors per HBA");
/**
* lpfc_fcp_cpu_map_init - Set the initial sr-iov virtual function enable
* @phba: lpfc_hba pointer.
* @val: link speed value.
*
* Description:
* If val is in a valid range [0-2], then affinitze the adapter's
* MSIX vectors.
*
* Returns:
* zero if val saved.
* -EINVAL val out of range
**/
static int
lpfc_fcp_cpu_map_init(struct lpfc_hba *phba, int val)
{
if (phba->sli_rev != LPFC_SLI_REV4) {
phba->cfg_fcp_cpu_map = 0;
return 0;
}
if (val >= LPFC_MIN_CPU_MAP && val <= LPFC_MAX_CPU_MAP) {
phba->cfg_fcp_cpu_map = val;
return 0;
}
lpfc_printf_log(phba, KERN_ERR, LOG_INIT,
"3326 fcp_cpu_map: %d out of range, using default\n",
val);
phba->cfg_fcp_cpu_map = LPFC_DRIVER_CPU_MAP;
return 0;
}
static DEVICE_ATTR(lpfc_fcp_cpu_map, S_IRUGO | S_IWUSR,
lpfc_fcp_cpu_map_show, lpfc_fcp_cpu_map_store);
/*
# lpfc_fcp_class: Determines FC class to use for the FCP protocol.
# Value range is [2,3]. Default value is 3.
@ -4154,6 +4289,7 @@ struct device_attribute *lpfc_hba_attrs[] = {
&dev_attr_lpfc_poll_tmo,
&dev_attr_lpfc_use_msi,
&dev_attr_lpfc_fcp_imax,
&dev_attr_lpfc_fcp_cpu_map,
&dev_attr_lpfc_fcp_wq_count,
&dev_attr_lpfc_fcp_eq_count,
&dev_attr_lpfc_fcp_io_channel,
@ -5136,6 +5272,7 @@ lpfc_get_cfgparam(struct lpfc_hba *phba)
lpfc_enable_rrq_init(phba, lpfc_enable_rrq);
lpfc_use_msi_init(phba, lpfc_use_msi);
lpfc_fcp_imax_init(phba, lpfc_fcp_imax);
lpfc_fcp_cpu_map_init(phba, lpfc_fcp_cpu_map);
lpfc_fcp_wq_count_init(phba, lpfc_fcp_wq_count);
lpfc_fcp_eq_count_init(phba, lpfc_fcp_eq_count);
lpfc_fcp_io_channel_init(phba, lpfc_fcp_io_channel);

5
drivers/scsi/lpfc/lpfc_hw4.h
View File

@ -200,6 +200,11 @@ struct lpfc_sli_intf {
#define LPFC_MAX_IMAX 5000000
#define LPFC_DEF_IMAX 50000
#define LPFC_MIN_CPU_MAP 0
#define LPFC_MAX_CPU_MAP 2
#define LPFC_HBA_CPU_MAP 1
#define LPFC_DRIVER_CPU_MAP 2 /* Default */
/* PORT_CAPABILITIES constants. */
#define LPFC_MAX_SUPPORTED_PAGES 8

320
drivers/scsi/lpfc/lpfc_init.c
View File

@ -33,6 +33,7 @@
#include <linux/slab.h>
#include <linux/firmware.h>
#include <linux/miscdevice.h>
#include <linux/percpu.h>
#include <scsi/scsi.h>
#include <scsi/scsi_device.h>
@ -58,6 +59,9 @@ char *_dump_buf_dif;
unsigned long _dump_buf_dif_order;
spinlock_t _dump_buf_lock;
/* Used when mapping IRQ vectors in a driver centric manner */
uint16_t lpfc_used_cpu[LPFC_MAX_CPU];
static void lpfc_get_hba_model_desc(struct lpfc_hba *, uint8_t *, uint8_t *);
static int lpfc_post_rcv_buf(struct lpfc_hba *);
static int lpfc_sli4_queue_verify(struct lpfc_hba *);
@ -4861,6 +4865,7 @@ lpfc_sli_driver_resource_unset(struct lpfc_hba *phba)
static int
lpfc_sli4_driver_resource_setup(struct lpfc_hba *phba)
{
struct lpfc_vector_map_info *cpup;
struct lpfc_sli *psli;
LPFC_MBOXQ_t *mboxq;
int rc, i, hbq_count, max_buf_size;
@ -5198,6 +5203,26 @@ lpfc_sli4_driver_resource_setup(struct lpfc_hba *phba)
goto out_free_fcp_eq_hdl;
}
phba->sli4_hba.cpu_map = kzalloc((sizeof(struct lpfc_vector_map_info) *
phba->sli4_hba.num_present_cpu),
GFP_KERNEL);
if (!phba->sli4_hba.cpu_map) {
lpfc_printf_log(phba, KERN_ERR, LOG_INIT,
"3327 Failed allocate memory for msi-x "
"interrupt vector mapping\n");
rc = -ENOMEM;
goto out_free_msix;
}
/* Initialize io channels for round robin */
cpup = phba->sli4_hba.cpu_map;
rc = 0;
for (i = 0; i < phba->sli4_hba.num_present_cpu; i++) {
cpup->channel_id = rc;
rc++;
if (rc >= phba->cfg_fcp_io_channel)
rc = 0;
}
/*
* Enable sr-iov virtual functions if supported and configured
* through the module parameter.
@ -5217,6 +5242,8 @@ lpfc_sli4_driver_resource_setup(struct lpfc_hba *phba)
return 0;
out_free_msix:
kfree(phba->sli4_hba.msix_entries);
out_free_fcp_eq_hdl:
kfree(phba->sli4_hba.fcp_eq_hdl);
out_free_fcf_rr_bmask:
@ -5246,6 +5273,11 @@ lpfc_sli4_driver_resource_unset(struct lpfc_hba *phba)
{
struct lpfc_fcf_conn_entry *conn_entry, *next_conn_entry;
/* Free memory allocated for msi-x interrupt vector to CPU mapping */
kfree(phba->sli4_hba.cpu_map);
phba->sli4_hba.num_present_cpu = 0;
phba->sli4_hba.num_online_cpu = 0;
/* Free memory allocated for msi-x interrupt vector entries */
kfree(phba->sli4_hba.msix_entries);
@ -6792,6 +6824,7 @@ lpfc_sli4_queue_verify(struct lpfc_hba *phba)
int cfg_fcp_io_channel;
uint32_t cpu;
uint32_t i = 0;
uint32_t j = 0;
/*
@ -6802,15 +6835,21 @@ lpfc_sli4_queue_verify(struct lpfc_hba *phba)
/* Sanity check on HBA EQ parameters */
cfg_fcp_io_channel = phba->cfg_fcp_io_channel;
/* It doesn't make sense to have more io channels then CPUs */
for_each_online_cpu(cpu) {
/* It doesn't make sense to have more io channels then online CPUs */
for_each_present_cpu(cpu) {
if (cpu_online(cpu))
i++;
j++;
}
phba->sli4_hba.num_online_cpu = i;
phba->sli4_hba.num_present_cpu = j;
if (i < cfg_fcp_io_channel) {
lpfc_printf_log(phba,
KERN_ERR, LOG_INIT,
"3188 Reducing IO channels to match number of "
"CPUs: from %d to %d\n", cfg_fcp_io_channel, i);
"online CPUs: from %d to %d\n",
cfg_fcp_io_channel, i);
cfg_fcp_io_channel = i;
}
@ -8309,6 +8348,269 @@ lpfc_sli_disable_intr(struct lpfc_hba *phba)
return;
}
/**
* lpfc_find_next_cpu - Find next available CPU that matches the phys_id
* @phba: pointer to lpfc hba data structure.
*
* Find next available CPU to use for IRQ to CPU affinity.
*/
static int
lpfc_find_next_cpu(struct lpfc_hba *phba, uint32_t phys_id)
{
struct lpfc_vector_map_info *cpup;
int cpu;
cpup = phba->sli4_hba.cpu_map;
for (cpu = 0; cpu < phba->sli4_hba.num_present_cpu; cpu++) {
/* CPU must be online */
if (cpu_online(cpu)) {
if ((cpup->irq == LPFC_VECTOR_MAP_EMPTY) &&
(lpfc_used_cpu[cpu] == LPFC_VECTOR_MAP_EMPTY) &&
(cpup->phys_id == phys_id)) {
return cpu;
}
}
cpup++;
}
/*
* If we get here, we have used ALL CPUs for the specific
* phys_id. Now we need to clear out lpfc_used_cpu and start
* reusing CPUs.
*/
for (cpu = 0; cpu < phba->sli4_hba.num_present_cpu; cpu++) {
if (lpfc_used_cpu[cpu] == phys_id)
lpfc_used_cpu[cpu] = LPFC_VECTOR_MAP_EMPTY;
}
cpup = phba->sli4_hba.cpu_map;
for (cpu = 0; cpu < phba->sli4_hba.num_present_cpu; cpu++) {
/* CPU must be online */
if (cpu_online(cpu)) {
if ((cpup->irq == LPFC_VECTOR_MAP_EMPTY) &&
(cpup->phys_id == phys_id)) {
return cpu;
}
}
cpup++;
}
return LPFC_VECTOR_MAP_EMPTY;
}
/**
* lpfc_sli4_set_affinity - Set affinity for HBA IRQ vectors
* @phba: pointer to lpfc hba data structure.
* @vectors: number of HBA vectors
*
* Affinitize MSIX IRQ vectors to CPUs. Try to equally spread vector
* affinization across multple physical CPUs (numa nodes).
* In addition, this routine will assign an IO channel for each CPU
* to use when issuing I/Os.
*/
static int
lpfc_sli4_set_affinity(struct lpfc_hba *phba, int vectors)
{
int i, idx, saved_chann, used_chann, cpu, phys_id;
int max_phys_id, num_io_channel, first_cpu;
struct lpfc_vector_map_info *cpup;
#ifdef CONFIG_X86
struct cpuinfo_x86 *cpuinfo;
#endif
struct cpumask *mask;
uint8_t chann[LPFC_FCP_IO_CHAN_MAX+1];
/* If there is no mapping, just return */
if (!phba->cfg_fcp_cpu_map)
return 1;
/* Init cpu_map array */
memset(phba->sli4_hba.cpu_map, 0xff,
(sizeof(struct lpfc_vector_map_info) *
phba->sli4_hba.num_present_cpu));
max_phys_id = 0;
phys_id = 0;
num_io_channel = 0;
first_cpu = LPFC_VECTOR_MAP_EMPTY;
/* Update CPU map with physical id and core id of each CPU */
cpup = phba->sli4_hba.cpu_map;
for (cpu = 0; cpu < phba->sli4_hba.num_present_cpu; cpu++) {
#ifdef CONFIG_X86
cpuinfo = &cpu_data(cpu);
cpup->phys_id = cpuinfo->phys_proc_id;
cpup->core_id = cpuinfo->cpu_core_id;
#else
/* No distinction between CPUs for other platforms */
cpup->phys_id = 0;
cpup->core_id = 0;
#endif
lpfc_printf_log(phba, KERN_INFO, LOG_INIT,
"3328 CPU physid %d coreid %d\n",
cpup->phys_id, cpup->core_id);
if (cpup->phys_id > max_phys_id)
max_phys_id = cpup->phys_id;
cpup++;
}
/* Now associate the HBA vectors with specific CPUs */
for (idx = 0; idx < vectors; idx++) {
cpup = phba->sli4_hba.cpu_map;
cpu = lpfc_find_next_cpu(phba, phys_id);
if (cpu == LPFC_VECTOR_MAP_EMPTY) {
/* Try for all phys_id's */
for (i = 1; i < max_phys_id; i++) {
phys_id++;
if (phys_id > max_phys_id)
phys_id = 0;
cpu = lpfc_find_next_cpu(phba, phys_id);
if (cpu == LPFC_VECTOR_MAP_EMPTY)
continue;
goto found;
}
lpfc_printf_log(phba, KERN_ERR, LOG_INIT,
"3329 Cannot set affinity:"
"Error mapping vector %d (%d)\n",
idx, vectors);
return 0;
}
found:
cpup += cpu;
if (phba->cfg_fcp_cpu_map == LPFC_DRIVER_CPU_MAP)
lpfc_used_cpu[cpu] = phys_id;
/* Associate vector with selected CPU */
cpup->irq = phba->sli4_hba.msix_entries[idx].vector;
/* Associate IO channel with selected CPU */
cpup->channel_id = idx;
num_io_channel++;
if (first_cpu == LPFC_VECTOR_MAP_EMPTY)
first_cpu = cpu;
/* Now affinitize to the selected CPU */
mask = &cpup->maskbits;
cpumask_clear(mask);
cpumask_set_cpu(cpu, mask);
i = irq_set_affinity_hint(phba->sli4_hba.msix_entries[idx].
vector, mask);
lpfc_printf_log(phba, KERN_INFO, LOG_INIT,
"3330 Set Affinity: CPU %d channel %d "
"irq %d (%x)\n",
cpu, cpup->channel_id,
phba->sli4_hba.msix_entries[idx].vector, i);
/* Spread vector mapping across multple physical CPU nodes */
phys_id++;
if (phys_id > max_phys_id)
phys_id = 0;
}
/*
* Finally fill in the IO channel for any remaining CPUs.
* At this point, all IO channels have been assigned to a specific
* MSIx vector, mapped to a specific CPU.
* Base the remaining IO channel assigned, to IO channels already
* assigned to other CPUs on the same phys_id.
*/
for (i = 0; i <= max_phys_id; i++) {
/*
* If there are no io channels already mapped to
* this phys_id, just round robin thru the io_channels.
* Setup chann[] for round robin.
*/
for (idx = 0; idx < phba->cfg_fcp_io_channel; idx++)
chann[idx] = idx;
saved_chann = 0;
used_chann = 0;
/*
* First build a list of IO channels already assigned
* to this phys_id before reassigning the same IO
* channels to the remaining CPUs.
*/
cpup = phba->sli4_hba.cpu_map;
cpu = first_cpu;
cpup += cpu;
for (idx = 0; idx < phba->sli4_hba.num_present_cpu;
idx++) {
if (cpup->phys_id == i) {
/*
* Save any IO channels that are
* already mapped to this phys_id.
*/
if (cpup->irq != LPFC_VECTOR_MAP_EMPTY) {
chann[saved_chann] =
cpup->channel_id;
saved_chann++;
goto out;
}
/* See if we are using round-robin */
if (saved_chann == 0)
saved_chann =
phba->cfg_fcp_io_channel;
/* Associate next IO channel with CPU */
cpup->channel_id = chann[used_chann];
num_io_channel++;
used_chann++;
if (used_chann == saved_chann)
used_chann = 0;
lpfc_printf_log(phba, KERN_INFO, LOG_INIT,
"3331 Set IO_CHANN "
"CPU %d channel %d\n",
idx, cpup->channel_id);
}
out:
cpu++;
if (cpu >= phba->sli4_hba.num_present_cpu) {
cpup = phba->sli4_hba.cpu_map;
cpu = 0;
} else {
cpup++;
}
}
}
if (phba->sli4_hba.num_online_cpu != phba->sli4_hba.num_present_cpu) {
cpup = phba->sli4_hba.cpu_map;
for (idx = 0; idx < phba->sli4_hba.num_present_cpu; idx++) {
if (cpup->channel_id == LPFC_VECTOR_MAP_EMPTY) {
cpup->channel_id = 0;
num_io_channel++;
lpfc_printf_log(phba, KERN_INFO, LOG_INIT,
"3332 Assign IO_CHANN "
"CPU %d channel %d\n",
idx, cpup->channel_id);
}
cpup++;
}
}
/* Sanity check */
if (num_io_channel != phba->sli4_hba.num_present_cpu)
lpfc_printf_log(phba, KERN_ERR, LOG_INIT,
"3333 Set affinity mismatch:"
"%d chann != %d cpus: %d vactors\n",
num_io_channel, phba->sli4_hba.num_present_cpu,
vectors);
phba->cfg_fcp_io_sched = LPFC_FCP_SCHED_BY_CPU;
return 1;
}
/**
* lpfc_sli4_enable_msix - Enable MSI-X interrupt mode to SLI-4 device
* @phba: pointer to lpfc hba data structure.
@ -8360,9 +8662,7 @@ enable_msix_vectors:
phba->sli4_hba.msix_entries[index].vector,
phba->sli4_hba.msix_entries[index].entry);
/*
* Assign MSI-X vectors to interrupt handlers
*/
/* Assign MSI-X vectors to interrupt handlers */
for (index = 0; index < vectors; index++) {
memset(&phba->sli4_hba.handler_name[index], 0, 16);
sprintf((char *)&phba->sli4_hba.handler_name[index],
@ -8390,6 +8690,8 @@ enable_msix_vectors:
phba->cfg_fcp_io_channel, vectors);
phba->cfg_fcp_io_channel = vectors;
}
lpfc_sli4_set_affinity(phba, vectors);
return rc;
cfg_fail_out:
@ -10667,6 +10969,7 @@ static struct miscdevice lpfc_mgmt_dev = {
static int __init
lpfc_init(void)
{
int cpu;
int error = 0;
printk(LPFC_MODULE_DESC "\n");
@ -10693,6 +10996,11 @@ lpfc_init(void)
return -ENOMEM;
}
}
/* Initialize in case vector mapping is needed */
for (cpu = 0; cpu < LPFC_MAX_CPU; cpu++)
lpfc_used_cpu[cpu] = LPFC_VECTOR_MAP_EMPTY;
error = pci_register_driver(&lpfc_driver);
if (error) {
fc_release_transport(lpfc_transport_template);

22
drivers/scsi/lpfc/lpfc_sli.c
View File

@ -7923,15 +7923,21 @@ lpfc_sli4_bpl2sgl(struct lpfc_hba *phba, struct lpfc_iocbq *piocbq,
static inline uint32_t
lpfc_sli4_scmd_to_wqidx_distr(struct lpfc_hba *phba)
{
int i;
struct lpfc_vector_map_info *cpup;
int chann, cpu;
if (phba->cfg_fcp_io_sched == LPFC_FCP_SCHED_BY_CPU)
i = smp_processor_id();
else
i = atomic_add_return(1, &phba->fcp_qidx);
i = (i % phba->cfg_fcp_io_channel);
return i;
if (phba->cfg_fcp_io_sched == LPFC_FCP_SCHED_BY_CPU) {
cpu = smp_processor_id();
if (cpu < phba->sli4_hba.num_present_cpu) {
cpup = phba->sli4_hba.cpu_map;
cpup += cpu;
return cpup->channel_id;
}
chann = cpu;
}
chann = atomic_add_return(1, &phba->fcp_qidx);
chann = (chann % phba->cfg_fcp_io_channel);
return chann;
}
/**

16
drivers/scsi/lpfc/lpfc_sli4.h
View File

@ -435,6 +435,17 @@ struct lpfc_sli4_lnk_info {
#define LPFC_SLI4_HANDLER_NAME_SZ 16
/* Used for IRQ vector to CPU mapping */
struct lpfc_vector_map_info {
uint16_t phys_id;
uint16_t core_id;
uint16_t irq;
uint16_t channel_id;
struct cpumask maskbits;
};
#define LPFC_VECTOR_MAP_EMPTY 0xffff
#define LPFC_MAX_CPU 256
/* SLI4 HBA data structure entries */
struct lpfc_sli4_hba {
void __iomem *conf_regs_memmap_p; /* Kernel memory mapped address for
@ -568,6 +579,11 @@ struct lpfc_sli4_hba {
struct lpfc_iov iov;
spinlock_t abts_scsi_buf_list_lock; /* list of aborted SCSI IOs */
spinlock_t abts_sgl_list_lock; /* list of aborted els IOs */
/* CPU to vector mapping information */
struct lpfc_vector_map_info *cpu_map;
uint16_t num_online_cpu;
uint16_t num_present_cpu;
};
enum lpfc_sge_type {