Some improvements for NVMe
cover Some improvements for NVMe
Hi,
The fisrt patch adds Arbitration Burst support, and other 2 patches are small improvements. Please help to review. Thanks.
Baolin Wang (3):
nvme: Add Arbitration Burst support
nvme-pci: Add controller memory buffer supported macro
nvme: Use USEC_PER_SEC instead of magic numbers
drivers/nvme/host/core.c | 25 ++++++++++++++++++++++++- drivers/nvme/host/nvme.h | 2 ++ drivers/nvme/host/pci.c | 4 +++- include/linux/nvme.h | 1 + 4 files changed, 30 insertions(+), 2 deletions(-)
Signature · 1 line
1.8.3.1
PATCH 1/3 nvme: Add Arbitration Burst support
From the NVMe spec, "In order to make efficient use of the non-volatile memory, it is often advantageous to execute multiple commands from a Submission Queue in parallel. For Submission Queues that are using weighted round robin with urgent priority class or round robin arbitration, host software may configure an Arbitration Burst setting". Thus add Arbitration Burst setting support.
--- drivers/nvme/host/core.c | 23 +++++++++++++++++++++++ drivers/nvme/host/nvme.h | 2 ++ drivers/nvme/host/pci.c | 2 ++ 3 files changed, 27 insertions(+)
drivers/nvme/host/core.c +23 -0
@@ -1241,6 +1241,28 @@ int nvme_set_queue_count(struct nvme_ctrl *ctrl, int *count) }EXPORT_SYMBOL_GPL (nvme_set_queue_count); +void nvme_set_arbitration_burst (struct nvme_ctrl *ctrl) +{ +u32 result; +int status; + +if (!ctrl->rab) +return ; + +/* +* The Arbitration Burst setting indicates the maximum number of +* commands that the controller may launch at one time from a +* particular Submission Queue. It is recommended that host software +* configure the Arbitration Burst setting as close to the recommended +* value by the controller as possible. +*/ + status =nvme_set_features (ctrl,NVME_FEAT_ARBITRATION , ctrl->rab, +NULL ,0 , &result); +if (status) +dev_warn (ctrl->device,"Failed to set Arbitration Burst\n" ); +} +EXPORT_SYMBOL_GPL (nvme_set_arbitration_burst); +#define NVME_AEN_SUPPORTED \ (NVME_AEN_CFG_NS_ATTR | NVME_AEN_CFG_FW_ACT | \ NVME_AEN_CFG_ANA_CHANGE | NVME_AEN_CFG_DISC_CHANGE)
@@ -2953,6 +2975,7 @@ int nvme_init_identify(struct nvme_ctrl *ctrl) }else ctrl->shutdown_timeout = shutdown_timeout; + ctrl->rab = id->rab; ctrl->npss = id->npss; ctrl->apsta = id->apsta; prev_apst_enabled = ctrl->apst_enabled;
drivers/nvme/host/nvme.h +2 -0
@@ -246,6 +246,7 @@ struct nvme_ctrl {u16 kas;u8 npss;u8 apsta; +u8 rab;u16 wctemp;u16 cctemp;u32 oaes;
@@ -550,6 +551,7 @@ int nvme_get_features(struct nvme_ctrl *dev, unsigned int fid, unsigned int dword11, void *buffer, size_t buflen, u32 *result);int nvme_set_queue_count (struct nvme_ctrl *ctrl,int *count); +void nvme_set_arbitration_burst (struct nvme_ctrl *ctrl);void nvme_stop_keep_alive (struct nvme_ctrl *ctrl);int nvme_reset_ctrl (struct nvme_ctrl *ctrl);int nvme_reset_ctrl_sync (struct nvme_ctrl *ctrl);
drivers/nvme/host/pci.c +2 -0
@@ -2111,6 +2111,8 @@ static int nvme_setup_io_queues(struct nvme_dev *dev)clear_bit (NVMEQ_ENABLED , &adminq->flags); +nvme_set_arbitration_burst (&dev->ctrl); +if (dev->cmb_use_sqes) { result =nvme_cmb_qdepth (dev, nr_io_queues,sizeof (struct nvme_command ));
Signature · 1 line
1.8.3.1
Keith Busch
maintainerRe: [PATCH 1/3] nvme: Add Arbitration Burst support
Quoted 6 lines · Baolin Wang
From the NVMe spec, "In order to make efficient use of the non-volatile memory, it is often advantageous to execute multiple commands from a Submission Queue in parallel. For Submission Queues that are using weighted round robin with urgent priority class or round robin arbitration, host software may configure an Arbitration Burst setting". Thus add Arbitration Burst setting support.
But if the user changed it to something else that better matches how they want to use queues, the driver is just going to undo that setting on the next reset.
Sagi Grimberg
maintainerRe: [PATCH 1/3] nvme: Add Arbitration Burst support
Quoted 11 lines
Quoted 7 lines
From the NVMe spec, "In order to make efficient use of the non-volatile
memory, it is often advantageous to execute multiple commands from a Submission Queue in parallel. For Submission Queues that are using weighted round robin with urgent priority class or round robin arbitration, host software may configure an Arbitration Burst setting". Thus add Arbitration Burst setting support.But if the user changed it to something else that better matches how they want to use queues, the driver is just going to undo that setting on the next reset.
Where do we do priority class arbitration? no one sets it
Keith Busch
maintainerRe: [PATCH 1/3] nvme: Add Arbitration Burst support
Quoted 8 lines · Sagi Grimberg
Quoted 8 lines
Quoted 7 lines
From the NVMe spec, "In order to make efficient use of the non-volatile
memory, it is often advantageous to execute multiple commands from a Submission Queue in parallel. For Submission Queues that are using weighted round robin with urgent priority class or round robin arbitration, host software may configure an Arbitration Burst setting". Thus add Arbitration Burst setting support.
Quoted 5 lines
Quoted 3 lines
But if the user changed it to something else that better matches how they want to use queues, the driver is just going to undo that setting on the next reset.
Where do we do priority class arbitration? no one sets it
Not the priority class, we don't use WRR in this driver. The RR arbitration can be set from user space and saved across controller resets, like:
# nvme set-feature -f 1 -v 3 --save
This patch would undo the saved feature value.
Re: [PATCH 1/3] nvme: Add Arbitration Burst support
Quoted 15 lines · Keith Busch
Quoted 8 lines · Sagi Grimberg
Quoted 8 lines
Quoted 7 lines
From the NVMe spec, "In order to make efficient use of the non-volatile
memory, it is often advantageous to execute multiple commands from a Submission Queue in parallel. For Submission Queues that are using weighted round robin with urgent priority class or round robin arbitration, host software may configure an Arbitration Burst setting". Thus add Arbitration Burst setting support.Quoted 5 lines
Quoted 3 lines
But if the user changed it to something else that better matches how they want to use queues, the driver is just going to undo that setting on the next reset.
Where do we do priority class arbitration? no one sets it
Quoted 7 lines
Not the priority class, we don't use WRR in this driver. The RR arbitration can be set from user space and saved across controller resets, like:
# nvme set-feature -f 1 -v 3 --saveThis patch would undo the saved feature value.
OK, I understaood your concern. Now we will select the RR arbitration as default in nvme_enable_ctrl(), but for some cases, we will not set the arbitration burst values from userspace, and we still want to use the defaut arbitration burst that recommended by the controller, taking into consideration any latency requirements.
So I'd like to add a parameter to decide if we can use the default arbitration burst like below, how do yo think? Thanks.
static bool use_default_arb;module_param(use_default_arb, bool, 0444); MODULE_PARM_DESC(use_default_arb, "use default arbitration burst that recommended by the controller");
Keith Busch
maintainerRe: [PATCH 1/3] nvme: Add Arbitration Burst support
Quoted 7 lines · Baolin Wang
OK, I understaood your concern. Now we will select the RR arbitration as default in
nvme_enable_ctrl(), but for some cases, we will not set the arbitration burst values from userspace, and we still want to use the defaut arbitration burst that recommended by the controller, taking into consideration any latency requirements.So I'd like to add a parameter to decide if we can use the default arbitration burst like below, how do yo think? Thanks.
I wouldn't call this the 'default' arbitration since it usually is not the default. This is the 'recommended' arbitration value.
Quoted 4 lines
static bool
use_default_arb;module_param(use_default_arb, bool, 0444);MODULE_PARM_DESC(use_default_arb, "use default arbitration burst that recommended by the controller");
"use controller's recommended arbitration burst"
Re: [PATCH 1/3] nvme: Add Arbitration Burst support
Quoted 11 lines · Keith Busch
Quoted 7 lines · Baolin Wang
OK, I understaood your concern. Now we will select the RR arbitration as default in
nvme_enable_ctrl(), but for some cases, we will not set the arbitration burst values from userspace, and we still want to use the defaut arbitration burst that recommended by the controller, taking into consideration any latency requirements.So I'd like to add a parameter to decide if we can use the default arbitration burst like below, how do yo think? Thanks.
I wouldn't call this the 'default' arbitration since it usually is not the default. This is the 'recommended' arbitration value.
OK. Change to rename the variable as:
static bool use_recommended_arb;
Quoted 7 lines
Quoted 5 lines
static bool
use_default_arb;module_param(use_default_arb, bool, 0444);MODULE_PARM_DESC(use_default_arb, "use default arbitration burst that recommended by the controller");"use controller's recommended arbitration burst"
Sure. Thanks.
Keith Busch
maintainerRe: [PATCH 1/3] nvme: Add Arbitration Burst support
In reply to Baolin WangQuoted 6 lines · Baolin Wang
drivers/nvme/host/core.c +6
-0+void nvme_set_arbitration_burst (struct nvme_ctrl *ctrl) +{ +u32 result; +int status; + +if (!ctrl->rab)
drivers/nvme/host/core.c int nvme_set_queue_count(struct nvme_ctrl *ctrl, int *count) · line 1259
Quoted 10 lines
drivers/nvme/host/core.c +10
-0+return ; + +/* +* The Arbitration Burst setting indicates the maximum number of +* commands that the controller may launch at one time from a +* particular Submission Queue. It is recommended that host software +* configure the Arbitration Burst setting as close to the recommended +* value by the controller as possible. +*/ + status = nvme_set_features(ctrl, NVME_FEAT_ARBITRATION, ctrl->rab,
Since 'rab' is an 8-bit field, but the feature's AB value is only 3 bits, we should add a validity check.
drivers/nvme/host/core.c int nvme_set_queue_count(struct nvme_ctrl *ctrl, int *count) · line 1264
Quoted 2 lines
drivers/nvme/host/core.c +2
-0+} +EXPORT_SYMBOL_GPL (nvme_set_arbitration_burst);
I don't see any particular reason to export this function just for the pci transport to use. Just make it a local static function an call it from nvme_init_identify().
Re: [PATCH 1/3] nvme: Add Arbitration Burst support
Quoted 12 lines · Keith Busch
Quoted 11 lines · Baolin Wang
drivers/nvme/host/core.c +11
-0+void nvme_set_arbitration_burst (struct nvme_ctrl *ctrl) +{ +u32 result; +int status; + +if (!ctrl->rab) +return ; + +/* +* The Arbitration Burst setting indicates the maximum number of +* commands that the controller may launch at one time from a
Quoted 8 lines
drivers/nvme/host/core.c int nvme_set_queue_count(struct nvme_ctrl *ctrl, int *count) · line 1259
Quoted 5 lines
drivers/nvme/host/core.c +5
-0+* particular Submission Queue. It is recommended that host software +* configure the Arbitration Burst setting as close to the recommended +* value by the controller as possible. +*/ + status = nvme_set_features(ctrl, NVME_FEAT_ARBITRATION, ctrl->rab,Since 'rab' is an 8-bit field, but the feature's AB value is only 3 bits, we should add a validity check.
Sure.
Quoted 7 lines
drivers/nvme/host/core.c int nvme_set_queue_count(struct nvme_ctrl *ctrl, int *count) · line 1264
Quoted 3 lines
drivers/nvme/host/core.c +2
-0+} +EXPORT_SYMBOL_GPL (nvme_set_arbitration_burst);I don't see any particular reason to export this function just for the pci transport to use. Just make it a local static function an call it from
nvme_init_identify().
OK. Thanks.
PATCH 2/3 nvme-pci: Add controller memory buffer supported macro
In reply to Baolin WangIntroduce a new capability macro to indicate if the controller supports the memory buffer or not, instead of reading the NVME_REG_CMBSZ register.
--- drivers/nvme/host/pci.c | 2 +- include/linux/nvme.h | 1 + 2 files changed, 2 insertions(+), 1 deletion(-)
drivers/nvme/host/pci.c +1 -1
@@ -1760,7 +1760,7 @@ static void nvme_map_cmb(struct nvme_dev *dev)struct pci_dev *pdev =to_pci_dev (dev->dev);int bar; -if (dev->cmb_size) +if (!NVME_CAP_CMBS (dev->ctrl.cap) || dev->cmb_size)return ; dev->cmbsz =readl (dev->bar +NVME_REG_CMBSZ );
include/linux/nvme.h +1 -0
@@ -134,6 +134,7 @@ enum {#define NVME_CAP_NSSRC (cap) (((cap) >> 36) & 0x1)#define NVME_CAP_MPSMIN (cap) (((cap) >> 48) & 0xf)#define NVME_CAP_MPSMAX (cap) (((cap) >> 52) & 0xf) +#define NVME_CAP_CMBS (cap) (((cap) >> 57) & 0x1)#define NVME_CMB_BIR (cmbloc) ((cmbloc) & 0x7)#define NVME_CMB_OFST (cmbloc) (((cmbloc) >> 12) & 0xfffff)
Signature · 1 line
1.8.3.1
Christoph Hellwig
maintainerRe: [PATCH 2/3] nvme-pci: Add controller memory buffer supported macro
Quoted 3 lines · Baolin Wang
Introduce a new capability macro to indicate if the controller supports the memory buffer or not, instead of reading the
NVME_REG_CMBSZregister.
This is a complex issue. The CMBS bit was only added in NVMe 1.4 as a backwards incompatible change, as the CMB addressing scheme can lead to data corruption. The CMBS was added as part of the horribe hack that also involves the CBA field, which we'll need to see before using it to work around the addressing issue. At the same time we should also continue supporting the legacy pre-1.4 CMB with a warning (and may reject it if we know we run in a VM).
Re: [PATCH 2/3] nvme-pci: Add controller memory buffer supported macro
Quoted 7 lines · Christoph Hellwig
Quoted 3 lines · Baolin Wang
Introduce a new capability macro to indicate if the controller supports the memory buffer or not, instead of reading the
NVME_REG_CMBSZregister.This is a complex issue. The CMBS bit was only added in NVMe 1.4 as a backwards incompatible change, as the CMB addressing scheme can lead
Ah, right, I think I should add another version condition:
if ((ctrl->vs >= NVME_VS(1, 4, 0) && !NVME_CAP_CMBS(dev->ctrl.cap)) ||
dev->cmb_size)
return;Quoted 5 lines
to data corruption. The CMBS was added as part of the horribe hack that also involves the CBA field, which we'll need to see before using it to work around the addressing issue. At the same time we should also continue supporting the legacy pre-1.4 CMB with a warning (and may reject it if we know we run in a VM).
Thanks for your explanation. I saw we've added an CMB sysfs attribute to get the CMB location and size if we enabled CMB, so should we still add some information in nvme_map_cmb() let user know explicitly?
dev_info(dev->ctrl.device, "Registered the CMB, bar=0x%x, size=%lld\n", bar, size);
Re: [PATCH 2/3] nvme-pci: Add controller memory buffer supported macro
Quoted 6 lines · Christoph Hellwig
Quoted 5 lines · Baolin Wang
Quoted 3 lines
Introduce a new capability macro to indicate if the controller supports the memory buffer or not, instead of reading the
NVME_REG_CMBSZregister.
Quoted 7 lines
Quoted 2 lines
This is a complex issue. The CMBS bit was only added in NVMe 1.4 as a backwards incompatible change, as the CMB addressing scheme can lead
Ah, right, I think I should add another version condition:
if ((ctrl->vs>=NVME_VS(1, 4, 0)&& !NVME_CAP_CMBS(dev->ctrl.cap)) ||dev->cmb_size)return;
After more thinking, now we should read NVME_REG_VS register to get the controller version when using the CMBS bit in Capabilities register, there is no benefit, so I'll drop this patch. Thanks.
Quoted 5 lines
Quoted 5 lines
to data corruption. The CMBS was added as part of the horribe hack that also involves the CBA field, which we'll need to see before using it to work around the addressing issue. At the same time we should also continue supporting the legacy pre-1.4 CMB with a warning (and may reject it if we know we run in a VM).
Keith Busch
maintainerRe: [PATCH 2/3] nvme-pci: Add controller memory buffer supported macro
In reply to Christoph HellwigQuoted 5 lines · Christoph Hellwig
Quoted 3 lines · Baolin Wang
Introduce a new capability macro to indicate if the controller supports the memory buffer or not, instead of reading the
NVME_REG_CMBSZregister.
Quoted 7 lines
This is a complex issue. The CMBS bit was only added in NVMe 1.4 as a backwards incompatible change, as the CMB addressing scheme can lead to data corruption. The CMBS was added as part of the horribe hack that also involves the CBA field, which we'll need to see before using it to work around the addressing issue. At the same time we should also continue supporting the legacy pre-1.4 CMB with a warning (and may reject it if we know we run in a VM).
Well, a CMB from an emulated controller (like qemu's) can be used within a VM. It's only if you direct assign a PCI function that CMB usage breaks.
Christoph Hellwig
maintainerRe: [PATCH 2/3] nvme-pci: Add controller memory buffer supported macro
Quoted 14 lines · Keith Busch
Quoted 5 lines · Christoph Hellwig
Quoted 3 lines · Baolin Wang
Introduce a new capability macro to indicate if the controller supports the memory buffer or not, instead of reading the
NVME_REG_CMBSZregister.Quoted 7 lines
This is a complex issue. The CMBS bit was only added in NVMe 1.4 as a backwards incompatible change, as the CMB addressing scheme can lead to data corruption. The CMBS was added as part of the horribe hack that also involves the CBA field, which we'll need to see before using it to work around the addressing issue. At the same time we should also continue supporting the legacy pre-1.4 CMB with a warning (and may reject it if we know we run in a VM).
Quoted 3 lines
Well, a CMB from an emulated controller (like qemu's) can be used within a VM. It's only if you direct assign a PCI function that CMB usage breaks.
But we have no idea if a controller is assigned or emulated.
PATCH 3/3 nvme: Use USEC_PER_SEC instead of magic numbers
In reply to Baolin WangUse USEC_PER_SEC instead of magic numbers to make code more readable.
--- drivers/nvme/host/core.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
drivers/nvme/host/core.c +1 -1
@@ -2963,7 +2963,7 @@ int nvme_init_identify(struct nvme_ctrl *ctrl)if (id->rtd3e) {/* us -> s */ -u32 transition_time =le32_to_cpu (id->rtd3e) /1000000 ; +u32 transition_time =le32_to_cpu (id->rtd3e) /USEC_PER_SEC ; ctrl->shutdown_timeout = clamp_t(unsigned int, transition_time, shutdown_timeout,60 );
Signature · 1 line
1.8.3.1
Re: [PATCH 3/3] nvme: Use USEC_PER_SEC instead of magic numbers