[PATCH 3/4] drm/amdkfd: Add CWSR support
Liu, Shaoyun
Shaoyun.Liu at amd.com
Mon Nov 20 16:30:26 UTC 2017
The save/restore memory is allocated per queue in user mode and the address /size passed in the create_ioctl. The memory you noticed that allocated at kfd_process_init_cwsr is the memory for CWSR shader code itself. This shader code is executed by the shader in the user's address space(GPU point of view), they are similar as signal handler code in CPU side . The CWSR shader code has a fix size , one page plus another page for extra usage for the CWSR ( ex . parameter for second level handler ). So we can easily managed it in kernel . The save/restore data is quite big ( around 22M for vega10) , so we don't want to allocate them in kernel .
Other comments in line .
Regards
Shaoyun.liu
-----Original Message-----
From: Oded Gabbay [mailto:oded.gabbay at gmail.com]
Sent: Sunday, November 19, 2017 8:38 AM
To: Kuehling, Felix
Cc: amd-gfx list; Liu, Shaoyun; Zhao, Yong
Subject: Re: [PATCH 3/4] drm/amdkfd: Add CWSR support
On Tue, Nov 14, 2017 at 11:41 PM, Felix Kuehling <Felix.Kuehling at amd.com> wrote:
> This hardware feature allows the GPU to preempt shader execution in
> the middle of a compute wave, save the state and restore it later to
> resume execution.
>
> Memory for saving the state is allocated per queue in user mode and
> the address and size passed to the create_queue ioctl. The size
Is this a correct description?
It seems to me the memory is allocated at kfd_process_init_cwsr() and the address is saved internally and not passed in the create_ioctl.
Which begs the question, why indeed it is not allocated by the user and then passed through the create_ioctl function ?
> depends on the number of waves that can be in flight simultaneously on
> a given ASIC.
>
> Signed-off-by: Shaoyun.liu <shaoyun.liu at amd.com>
> Signed-off-by: Yong Zhao <yong.zhao at amd.com>
> Signed-off-by: Felix Kuehling <Felix.Kuehling at amd.com>
> ---
> drivers/gpu/drm/amd/amdkfd/kfd_chardev.c | 7 +-
> drivers/gpu/drm/amd/amdkfd/kfd_device.c | 20 ++++-
> .../gpu/drm/amd/amdkfd/kfd_device_queue_manager.c | 6 ++
> drivers/gpu/drm/amd/amdkfd/kfd_module.c | 4 +
> drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_vi.c | 27 +++++++
> drivers/gpu/drm/amd/amdkfd/kfd_priv.h | 31 +++++++-
> drivers/gpu/drm/amd/amdkfd/kfd_process.c | 87 +++++++++++++++++++++-
> include/uapi/linux/kfd_ioctl.h | 3 +-
> 8 files changed, 179 insertions(+), 6 deletions(-)
>
> diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c
> b/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c
> index 505d391..2a4612d 100644
> --- a/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c
> +++ b/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c
> @@ -117,7 +117,7 @@ static int kfd_open(struct inode *inode, struct file *filep)
> return -EPERM;
> }
>
> - process = kfd_create_process(current);
> + process = kfd_create_process(filep);
> if (IS_ERR(process))
> return PTR_ERR(process);
>
> @@ -206,6 +206,7 @@ static int set_queue_properties_from_user(struct queue_properties *q_properties,
> q_properties->ctx_save_restore_area_address =
> args->ctx_save_restore_address;
> q_properties->ctx_save_restore_area_size =
> args->ctx_save_restore_size;
> + q_properties->ctl_stack_size = args->ctl_stack_size;
> if (args->queue_type == KFD_IOC_QUEUE_TYPE_COMPUTE ||
> args->queue_type == KFD_IOC_QUEUE_TYPE_COMPUTE_AQL)
> q_properties->type = KFD_QUEUE_TYPE_COMPUTE; @@
> -1088,6 +1089,10 @@ static int kfd_mmap(struct file *filp, struct vm_area_struct *vma)
> KFD_MMAP_EVENTS_MASK) {
> vma->vm_pgoff = vma->vm_pgoff ^ KFD_MMAP_EVENTS_MASK;
> return kfd_event_mmap(process, vma);
> + } else if ((vma->vm_pgoff & KFD_MMAP_RESERVED_MEM_MASK) ==
> + KFD_MMAP_RESERVED_MEM_MASK) {
> + vma->vm_pgoff = vma->vm_pgoff ^ KFD_MMAP_RESERVED_MEM_MASK;
> + return kfd_reserved_mem_mmap(process, vma);
> }
>
> return -EFAULT;
> diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_device.c
> b/drivers/gpu/drm/amd/amdkfd/kfd_device.c
> index 621a3b5..4f05eac 100644
> --- a/drivers/gpu/drm/amd/amdkfd/kfd_device.c
> +++ b/drivers/gpu/drm/amd/amdkfd/kfd_device.c
> @@ -27,6 +27,7 @@
> #include "kfd_priv.h"
> #include "kfd_device_queue_manager.h"
> #include "kfd_pm4_headers_vi.h"
> +#include "cwsr_trap_handler_gfx8.asm"
>
> #define MQD_SIZE_ALIGNED 768
>
> @@ -38,7 +39,8 @@ static const struct kfd_device_info kaveri_device_info = {
> .ih_ring_entry_size = 4 * sizeof(uint32_t),
> .event_interrupt_class = &event_interrupt_class_cik,
> .num_of_watch_points = 4,
> - .mqd_size_aligned = MQD_SIZE_ALIGNED
> + .mqd_size_aligned = MQD_SIZE_ALIGNED,
> + .supports_cwsr = false,
> };
>
> static const struct kfd_device_info carrizo_device_info = { @@ -49,7
> +51,8 @@ static const struct kfd_device_info carrizo_device_info = {
> .ih_ring_entry_size = 4 * sizeof(uint32_t),
> .event_interrupt_class = &event_interrupt_class_cik,
> .num_of_watch_points = 4,
> - .mqd_size_aligned = MQD_SIZE_ALIGNED
> + .mqd_size_aligned = MQD_SIZE_ALIGNED,
> + .supports_cwsr = true,
> };
>
> struct kfd_deviceid {
> @@ -212,6 +215,17 @@ static int iommu_invalid_ppr_cb(struct pci_dev *pdev, int pasid,
> return AMD_IOMMU_INV_PRI_RSP_INVALID; }
>
> +static void kfd_cwsr_init(struct kfd_dev *kfd) {
> + if (cwsr_enable && kfd->device_info->supports_cwsr) {
> + BUILD_BUG_ON(sizeof(cwsr_trap_gfx8_hex) > PAGE_SIZE);
> +
> + kfd->cwsr_isa = cwsr_trap_gfx8_hex;
> + kfd->cwsr_isa_size = sizeof(cwsr_trap_gfx8_hex);
> + kfd->cwsr_enabled = true;
> + }
> +}
> +
> bool kgd2kfd_device_init(struct kfd_dev *kfd,
> const struct kgd2kfd_shared_resources
> *gpu_resources) { @@ -286,6 +300,8 @@ bool kgd2kfd_device_init(struct
> kfd_dev *kfd,
> goto device_iommu_pasid_error;
> }
>
> + kfd_cwsr_init(kfd);
> +
> if (kfd_resume(kfd))
> goto kfd_resume_error;
>
> diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c
> b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c
> index e202921..5c06502 100644
> --- a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c
> +++ b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c
> @@ -173,6 +173,9 @@ static int create_queue_nocpsch(struct device_queue_manager *dqm,
> *allocated_vmid = qpd->vmid;
> q->properties.vmid = qpd->vmid;
>
> + q->properties.tba_addr = qpd->tba_addr;
> + q->properties.tma_addr = qpd->tma_addr;
> +
> if (q->properties.type == KFD_QUEUE_TYPE_COMPUTE)
> retval = create_compute_queue_nocpsch(dqm, q, qpd);
> else if (q->properties.type == KFD_QUEUE_TYPE_SDMA) @@ -846,6
> +849,9 @@ static int create_queue_cpsch(struct device_queue_manager *dqm, struct queue *q,
> }
>
> dqm->asic_ops.init_sdma_vm(dqm, q, qpd);
> +
> + q->properties.tba_addr = qpd->tba_addr;
> + q->properties.tma_addr = qpd->tma_addr;
> retval = mqd->init_mqd(mqd, &q->mqd, &q->mqd_mem_obj,
> &q->gart_mqd_addr, &q->properties);
> if (retval)
> diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_module.c
> b/drivers/gpu/drm/amd/amdkfd/kfd_module.c
> index 6c5a9ca..4b2423b 100644
> --- a/drivers/gpu/drm/amd/amdkfd/kfd_module.c
> +++ b/drivers/gpu/drm/amd/amdkfd/kfd_module.c
> @@ -49,6 +49,10 @@ module_param(sched_policy, int, 0444);
> MODULE_PARM_DESC(sched_policy,
> "Scheduling policy (0 = HWS (Default), 1 = HWS without
> over-subscription, 2 = Non-HWS (Used for debugging only)");
>
> +int cwsr_enable = 1;
> +module_param(cwsr_enable, int, 0444); MODULE_PARM_DESC(cwsr_enable,
> +"CWSR enable (0 = Off, 1 = On (Default))");
> +
> int max_num_of_queues_per_device =
> KFD_MAX_NUM_OF_QUEUES_PER_DEVICE_DEFAULT;
> module_param(max_num_of_queues_per_device, int, 0444);
> MODULE_PARM_DESC(max_num_of_queues_per_device,
> diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_vi.c
> b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_vi.c
> index 2ba7cea..00e1f1a 100644
> --- a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_vi.c
> +++ b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_vi.c
> @@ -89,6 +89,28 @@ static int init_mqd(struct mqd_manager *mm, void **mqd,
> if (q->format == KFD_QUEUE_FORMAT_AQL)
> m->cp_hqd_iq_rptr = 1;
>
> + if (q->tba_addr) {
> + m->compute_tba_lo = lower_32_bits(q->tba_addr >> 8);
> + m->compute_tba_hi = upper_32_bits(q->tba_addr >> 8);
> + m->compute_tma_lo = lower_32_bits(q->tma_addr >> 8);
> + m->compute_tma_hi = upper_32_bits(q->tma_addr >> 8);
Why the >> 8 on the addresses ?
[shaoyunl] The hardware register requirements , they need to be shift 8 bits.
> + m->compute_pgm_rsrc2 |=
> + (1 << COMPUTE_PGM_RSRC2__TRAP_PRESENT__SHIFT);
> + }
> +
> + if (mm->dev->cwsr_enabled && q->ctx_save_restore_area_address) {
> + m->cp_hqd_persistent_state |=
> + (1 << CP_HQD_PERSISTENT_STATE__QSWITCH_MODE__SHIFT);
> + m->cp_hqd_ctx_save_base_addr_lo =
> + lower_32_bits(q->ctx_save_restore_area_address);
> + m->cp_hqd_ctx_save_base_addr_hi =
> + upper_32_bits(q->ctx_save_restore_area_address);
> + m->cp_hqd_ctx_save_size = q->ctx_save_restore_area_size;
> + m->cp_hqd_cntl_stack_size = q->ctl_stack_size;
> + m->cp_hqd_cntl_stack_offset = q->ctl_stack_size;
> + m->cp_hqd_wg_state_offset = q->ctl_stack_size;
Just wanted to make sure the last two lines are not copy-paste error from the third line
[Shaoyunl] ye, that's requirement form hardware spec .
> + }
> +
> *mqd = m;
> if (gart_addr)
> *gart_addr = addr;
> @@ -167,6 +189,11 @@ static int __update_mqd(struct mqd_manager *mm, void *mqd,
> 2 << CP_HQD_PQ_CONTROL__SLOT_BASED_WPTR__SHIFT;
> }
>
> + if (mm->dev->cwsr_enabled && q->ctx_save_restore_area_address)
> + m->cp_hqd_ctx_save_control =
> + atc_bit << CP_HQD_CTX_SAVE_CONTROL__ATC__SHIFT |
> + mtype <<
> + CP_HQD_CTX_SAVE_CONTROL__MTYPE__SHIFT;
> +
> q->is_active = (q->queue_size > 0 &&
> q->queue_address != 0 &&
> q->queue_percent > 0); diff --git
> a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h
> b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h
> index 4750473..a668764 100644
> --- a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h
> +++ b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h
> @@ -41,6 +41,7 @@
>
> #define KFD_MMAP_DOORBELL_MASK 0x8000000000000 #define
> KFD_MMAP_EVENTS_MASK 0x4000000000000
> +#define KFD_MMAP_RESERVED_MEM_MASK 0x2000000000000
>
> /*
> * When working with cp scheduler we should assign the HIQ manually
> or via @@ -63,6 +64,15 @@ #define KFD_MAX_NUM_OF_QUEUES_PER_PROCESS
> 1024
>
> /*
> + * Size of the per-process TBA+TMA buffer: 2 pages
> + *
> + * The first page is the TBA used for the CWSR ISA code. The second
> + * page is used as TMA for daisy changing a user-mode trap handler.
> + */
> +#define KFD_CWSR_TBA_TMA_SIZE (PAGE_SIZE * 2) #define
> +KFD_CWSR_TMA_OFFSET PAGE_SIZE
> +
> +/*
> * Kernel module parameter to specify maximum number of supported queues per
> * device
> */
> @@ -78,6 +88,8 @@ extern int max_num_of_queues_per_device;
> /* Kernel module parameter to specify the scheduling policy */
> extern int sched_policy;
>
> +extern int cwsr_enable;
> +
> /*
> * Kernel module parameter to specify whether to send sigterm to HSA process on
> * unhandled exception
> @@ -131,6 +143,7 @@ struct kfd_device_info {
> size_t ih_ring_entry_size;
> uint8_t num_of_watch_points;
> uint16_t mqd_size_aligned;
> + bool supports_cwsr;
> };
>
> struct kfd_mem_obj {
> @@ -200,6 +213,11 @@ struct kfd_dev {
>
> /* Debug manager */
> struct kfd_dbgmgr *dbgmgr;
> +
> + /* CWSR */
> + bool cwsr_enabled;
> + const void *cwsr_isa;
> + unsigned int cwsr_isa_size;
> };
>
> /* KGD2KFD callbacks */
> @@ -332,6 +350,9 @@ struct queue_properties {
> uint32_t eop_ring_buffer_size;
> uint64_t ctx_save_restore_area_address;
> uint32_t ctx_save_restore_area_size;
> + uint32_t ctl_stack_size;
> + uint64_t tba_addr;
> + uint64_t tma_addr;
> };
>
> /**
> @@ -439,6 +460,11 @@ struct qcm_process_device {
> uint32_t num_gws;
> uint32_t num_oac;
> uint32_t sh_hidden_private_base;
> +
> + /* CWSR memory */
> + void *cwsr_kaddr;
> + uint64_t tba_addr;
> + uint64_t tma_addr;
> };
>
>
> @@ -563,7 +589,7 @@ struct amdkfd_ioctl_desc {
>
> void kfd_process_create_wq(void);
> void kfd_process_destroy_wq(void);
> -struct kfd_process *kfd_create_process(const struct task_struct *);
> +struct kfd_process *kfd_create_process(struct file *filep);
> struct kfd_process *kfd_get_process(const struct task_struct *);
> struct kfd_process *kfd_lookup_process_by_pasid(unsigned int pasid);
>
> @@ -577,6 +603,9 @@ struct kfd_process_device
> *kfd_get_process_device_data(struct kfd_dev *dev, struct kfd_process_device *kfd_create_process_device_data(struct kfd_dev *dev,
> struct
> kfd_process *p);
>
> +int kfd_reserved_mem_mmap(struct kfd_process *process,
> + struct vm_area_struct *vma);
> +
> /* Process device data iterator */
> struct kfd_process_device *kfd_get_first_process_device_data(
> struct
> kfd_process *p); diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_process.c
> b/drivers/gpu/drm/amd/amdkfd/kfd_process.c
> index 1bb9b26..39f4c19 100644
> --- a/drivers/gpu/drm/amd/amdkfd/kfd_process.c
> +++ b/drivers/gpu/drm/amd/amdkfd/kfd_process.c
> @@ -28,6 +28,7 @@
> #include <linux/amd-iommu.h>
> #include <linux/notifier.h>
> #include <linux/compat.h>
> +#include <linux/mman.h>
>
> struct mm_struct;
>
> @@ -53,6 +54,8 @@ struct kfd_process_release_work {
>
> static struct kfd_process *find_process(const struct task_struct
> *thread); static struct kfd_process *create_process(const struct
> task_struct *thread);
> +static int kfd_process_init_cwsr(struct kfd_process *p, struct file
> +*filep);
> +
>
> void kfd_process_create_wq(void)
> {
> @@ -68,9 +71,10 @@ void kfd_process_destroy_wq(void)
> }
> }
>
> -struct kfd_process *kfd_create_process(const struct task_struct
> *thread)
> +struct kfd_process *kfd_create_process(struct file *filep)
> {
> struct kfd_process *process;
> + struct task_struct *thread = current;
>
> if (!thread->mm)
> return ERR_PTR(-EINVAL); @@ -101,6 +105,8 @@ struct
> kfd_process *kfd_create_process(const struct task_struct *thread)
>
> up_write(&thread->mm->mmap_sem);
>
> + kfd_process_init_cwsr(process, filep);
> +
> return process;
> }
>
> @@ -168,6 +174,11 @@ static void kfd_process_wq_release(struct work_struct *work)
> amd_iommu_unbind_pasid(pdd->dev->pdev,
> p->pasid);
>
> list_del(&pdd->per_device_list);
> +
> + if (pdd->qpd.cwsr_kaddr)
> + free_pages((unsigned long)pdd->qpd.cwsr_kaddr,
> + get_order(KFD_CWSR_TBA_TMA_SIZE));
> +
> kfree(pdd);
> }
>
> @@ -260,6 +271,46 @@ static const struct mmu_notifier_ops kfd_process_mmu_notifier_ops = {
> .release = kfd_process_notifier_release, };
>
> +static int kfd_process_init_cwsr(struct kfd_process *p, struct file
> +*filep) {
> + int err = 0;
> + unsigned long offset;
> + struct kfd_process_device *temp, *pdd = NULL;
> + struct kfd_dev *dev = NULL;
> + struct qcm_process_device *qpd = NULL;
> +
> + mutex_lock(&p->mutex);
> + list_for_each_entry_safe(pdd, temp, &p->per_device_data,
> + per_device_list) {
> + dev = pdd->dev;
> + qpd = &pdd->qpd;
> + if (!dev->cwsr_enabled || qpd->cwsr_kaddr)
> + continue;
> + offset = (dev->id | KFD_MMAP_RESERVED_MEM_MASK) << PAGE_SHIFT;
> + qpd->tba_addr = (int64_t)vm_mmap(filep, 0,
> + KFD_CWSR_TBA_TMA_SIZE, PROT_READ | PROT_EXEC,
> + MAP_SHARED, offset);
> +
> + if (IS_ERR_VALUE(qpd->tba_addr)) {
> + pr_err("Failure to set tba address. error -%d.\n",
> + (int)qpd->tba_addr);
> + err = qpd->tba_addr;
> + qpd->tba_addr = 0;
> + qpd->cwsr_kaddr = NULL;
> + goto out;
> + }
> +
> + memcpy(qpd->cwsr_kaddr, dev->cwsr_isa,
> + dev->cwsr_isa_size);
> +
> + qpd->tma_addr = qpd->tba_addr + KFD_CWSR_TMA_OFFSET;
> + pr_debug("set tba :0x%llx, tma:0x%llx, cwsr_kaddr:%p for pqm.\n",
> + qpd->tba_addr, qpd->tma_addr, qpd->cwsr_kaddr);
> + }
> +out:
> + mutex_unlock(&p->mutex);
> + return err;
> +}
> +
> static struct kfd_process *create_process(const struct task_struct
> *thread) {
> struct kfd_process *process;
> @@ -535,3 +586,37 @@ struct kfd_process
> *kfd_lookup_process_by_pasid(unsigned int pasid)
>
> return p;
> }
> +
> +int kfd_reserved_mem_mmap(struct kfd_process *process,
> + struct vm_area_struct *vma) {
> + struct kfd_dev *dev = kfd_device_by_id(vma->vm_pgoff);
> + struct kfd_process_device *pdd;
> + struct qcm_process_device *qpd;
> +
> + if (!dev)
> + return -EINVAL;
> + if ((vma->vm_end - vma->vm_start) != KFD_CWSR_TBA_TMA_SIZE) {
> + pr_err("Incorrect CWSR mapping size.\n");
> + return -EINVAL;
> + }
> +
> + pdd = kfd_get_process_device_data(dev, process);
> + if (!pdd)
> + return -EINVAL;
> + qpd = &pdd->qpd;
> +
> + qpd->cwsr_kaddr = (void *)__get_free_pages(GFP_KERNEL | __GFP_ZERO,
> + get_order(KFD_CWSR_TBA_TMA_SIZE));
> + if (!qpd->cwsr_kaddr) {
> + pr_err("Error allocating per process CWSR buffer.\n");
> + return -ENOMEM;
> + }
> +
> + vma->vm_flags |= VM_IO | VM_DONTCOPY | VM_DONTEXPAND
> + | VM_NORESERVE | VM_DONTDUMP | VM_PFNMAP;
> + /* Mapping pages to user process */
> + return remap_pfn_range(vma, vma->vm_start,
> + PFN_DOWN(__pa(qpd->cwsr_kaddr)),
> + KFD_CWSR_TBA_TMA_SIZE,
> +vma->vm_page_prot); }
> diff --git a/include/uapi/linux/kfd_ioctl.h
> b/include/uapi/linux/kfd_ioctl.h index 731d0df..7039f16 100644
> --- a/include/uapi/linux/kfd_ioctl.h
> +++ b/include/uapi/linux/kfd_ioctl.h
> @@ -58,7 +58,8 @@ struct kfd_ioctl_create_queue_args {
> __u64 eop_buffer_address; /* to KFD */
> __u64 eop_buffer_size; /* to KFD */
> __u64 ctx_save_restore_address; /* to KFD */
> - __u64 ctx_save_restore_size; /* to KFD */
> + __u32 ctx_save_restore_size; /* to KFD */
> + __u32 ctl_stack_size; /* to KFD */
> };
>
> struct kfd_ioctl_destroy_queue_args {
> --
> 2.7.4
>
More information about the amd-gfx
mailing list