From c4bb19d46e1c9121a0948fa506098cbf2f64e2a6 Mon Sep 17 00:00:00 2001 From: David Nieto Date: Fri, 7 Oct 2016 16:25:04 -0700 Subject: nvgpu: gpu: arbiter for vf switch management JIRA DNVGPU-143 The arbiter is charged with selecting the proper frequencies when multiple applications submit simultaneously clock change requests On the current implementation, the arbiter guarantees that the selected frequency will be always higher or equal to the request, as long as the request is in range. The current code is not yet realtime friendly, as requests are not pre-allocated. Summary of changes: (1) pstate/vf switch no longer selects boot frequency (2) changed mclk code change to accept input freq (3) added arbiter (4) now a single session can submit concurrent requests the last request is the one that applies for that session (5) modified locking mechanism to reduce lock contention (6) Added callback to notify the arbiter that the VF table has changed and is no longer valid (PMU/Thermals must call this when VF table is invalid) (7) changed internal API to work with MHz (8) added debugfs for stats Change-Id: I6a7b05c9447761e8536f84ef86b5ab0793164d63 Signed-off-by: David Nieto Reviewed-on: http://git-master/r/1239461 Reviewed-by: Thomas Fleury GVS: Gerrit_Virtual_Submit Reviewed-by: Terje Bergstrom Reviewed-on: http://git-master/r/1267120 Reviewed-by: Automatic_Commit_Validation_User --- drivers/gpu/nvgpu/clk/clk.c | 61 +-- drivers/gpu/nvgpu/clk/clk.h | 1 - drivers/gpu/nvgpu/clk/clk_arb.c | 765 ++++++++++++++++++++++++++++---- drivers/gpu/nvgpu/clk/clk_arb.h | 26 +- drivers/gpu/nvgpu/clk/clk_mclk.c | 19 +- drivers/gpu/nvgpu/clk/clk_mclk.h | 10 +- drivers/gpu/nvgpu/gp106/clk_arb_gp106.c | 10 +- drivers/gpu/nvgpu/pstate/pstate.c | 4 - 8 files changed, 703 insertions(+), 193 deletions(-) (limited to 'drivers/gpu/nvgpu') diff --git a/drivers/gpu/nvgpu/clk/clk.c b/drivers/gpu/nvgpu/clk/clk.c index ef0834f4..bec5fad1 100644 --- a/drivers/gpu/nvgpu/clk/clk.c +++ b/drivers/gpu/nvgpu/clk/clk.c @@ -255,7 +255,7 @@ static int get_regime_id(struct gk20a *g, u32 domain, u32 *regimeid) return -EINVAL; } -int clk_program_fllclks(struct gk20a *g, struct change_fll_clk *fllclk) +int clk_program_fll_clks(struct gk20a *g, struct change_fll_clk *fllclk) { int status = -EINVAL; struct clk_domain *pdomain; @@ -277,8 +277,6 @@ int clk_program_fllclks(struct gk20a *g, struct change_fll_clk *fllclk) if (fllclk->clkmhz == 0) return -EINVAL; - mutex_lock(&pclk->changeclkmutex); - setfllclk.voltuv = fllclk->voltuv; setfllclk.gpc2clkmhz = fllclk->clkmhz; @@ -376,63 +374,6 @@ int clk_program_fllclks(struct gk20a *g, struct change_fll_clk *fllclk) if (status) goto done; done: - mutex_unlock(&pclk->changeclkmutex); - return status; -} - -int clk_set_boot_fll_clk(struct gk20a *g) -{ - int status; - struct change_fll_clk bootfllclk; - u16 gpc2clk_clkmhz = BOOT_GPC2CLK_MHZ; - u32 gpc2clk_voltuv = 0; - u32 gpc2clk_voltuv_sram = 0; - u16 mclk_clkmhz = BOOT_MCLK_MHZ; - u32 mclk_voltuv = 0; - u32 mclk_voltuv_sram = 0; - u32 voltuv = 0; - u32 voltuv_sram = 0; - - mutex_init(&g->clk_pmu.changeclkmutex); - status = clk_domain_get_f_or_v(g, CTRL_CLK_DOMAIN_GPC2CLK, - &gpc2clk_clkmhz, &gpc2clk_voltuv, CTRL_VOLT_DOMAIN_LOGIC); - if (status) - return status; - status = clk_domain_get_f_or_v(g, CTRL_CLK_DOMAIN_GPC2CLK, - &gpc2clk_clkmhz, &gpc2clk_voltuv_sram, CTRL_VOLT_DOMAIN_SRAM); - if (status) - return status; - status = clk_domain_get_f_or_v(g, CTRL_CLK_DOMAIN_MCLK, - &mclk_clkmhz, &mclk_voltuv, CTRL_VOLT_DOMAIN_LOGIC); - if (status) - return status; - status = clk_domain_get_f_or_v(g, CTRL_CLK_DOMAIN_MCLK, - &mclk_clkmhz, &mclk_voltuv_sram, CTRL_VOLT_DOMAIN_SRAM); - if (status) - return status; - - voltuv = ((gpc2clk_voltuv) > (mclk_voltuv)) ? (gpc2clk_voltuv) - : (mclk_voltuv); - - voltuv_sram = ((gpc2clk_voltuv_sram) > (mclk_voltuv_sram)) ? - (gpc2clk_voltuv_sram) : (mclk_voltuv_sram); - - status = volt_set_voltage(g, voltuv, voltuv_sram); - if (status) - gk20a_err(dev_from_gk20a(g), - "attempt to set boot voltage failed %d %d", - voltuv, voltuv_sram); - - bootfllclk.api_clk_domain = CTRL_CLK_DOMAIN_GPC2CLK; - bootfllclk.clkmhz = gpc2clk_clkmhz; - bootfllclk.voltuv = voltuv; - status = clk_program_fllclks(g, &bootfllclk); - if (status) - gk20a_err(dev_from_gk20a(g), "attempt to set boot gpc2clk failed"); - status = g->clk_pmu.clk_mclk.change(g, DEFAULT_BOOT_MCLK_SPEED); - if (status) - gk20a_err(dev_from_gk20a(g), "attempt to set boot mclk failed"); - return status; } diff --git a/drivers/gpu/nvgpu/clk/clk.h b/drivers/gpu/nvgpu/clk/clk.h index a0b88dcb..d0e82173 100644 --- a/drivers/gpu/nvgpu/clk/clk.h +++ b/drivers/gpu/nvgpu/clk/clk.h @@ -119,6 +119,5 @@ u32 clk_domain_get_f_points( u32 *fpointscount, u16 *freqpointsinmhz ); -int clk_set_boot_fll_clk(struct gk20a *g); int clk_program_fll_clks(struct gk20a *g, struct change_fll_clk *fllclk); #endif diff --git a/drivers/gpu/nvgpu/clk/clk_arb.c b/drivers/gpu/nvgpu/clk/clk_arb.c index 98b7cb5f..f868100b 100644 --- a/drivers/gpu/nvgpu/clk/clk_arb.c +++ b/drivers/gpu/nvgpu/clk/clk_arb.c @@ -18,9 +18,17 @@ #include #include #include +#include #include "clk/clk_arb.h" + +#define MAX_F_POINTS 127 + +#ifdef CONFIG_DEBUG_FS +static int nvgpu_clk_arb_debugfs_init(struct gk20a *g); +#endif + static int nvgpu_clk_arb_release_event_dev(struct inode *inode, struct file *filp); static int nvgpu_clk_arb_release_completion_dev(struct inode *inode, @@ -28,21 +36,57 @@ static int nvgpu_clk_arb_release_completion_dev(struct inode *inode, static unsigned int nvgpu_clk_arb_poll_dev(struct file *filp, poll_table *wait); static void nvgpu_clk_arb_run_arbiter_cb(struct work_struct *work); +static void nvgpu_clk_arb_run_vftable_cb(struct work_struct *work); +static int nvgpu_clk_arb_update_vftable(struct nvgpu_clk_arb *); + +struct nvgpu_clk_vf_point { + u16 mhz; + u32 uvolt; + u32 uvolt_sram; +}; struct nvgpu_clk_arb { - struct mutex req_lock; - struct mutex users_lock; + spinlock_t sessions_lock; + spinlock_t users_lock; + spinlock_t req_lock; + struct list_head users; + struct list_head sessions; struct list_head requests; - u64 gpc2clk_current_hz; - u64 gpc2clk_target_hz; - u64 gpc2clk_default_hz; - u64 mclk_current_hz; - u64 mclk_target_hz; - u64 mclk_default_hz; - atomic_t usercount; + struct gk20a *g; + spinlock_t data_lock; + spinlock_t vf_lock; + + u16 gpc2clk_actual_mhz; + u16 gpc2clk_default_mhz; + + u16 mclk_actual_mhz; + u16 mclk_default_mhz; + u32 voltuv_actual; + struct work_struct update_fn_work; + struct work_struct vftable_fn_work; + wait_queue_head_t vftable_wq; + + u16 *mclk_f_points; + bool vftable_set; + + struct nvgpu_clk_vf_point *mclk_vf_points; + u32 mclk_f_numpoints; + u16 *gpc2clk_f_points; + u32 gpc2clk_f_numpoints; + struct nvgpu_clk_vf_point *gpc2clk_vf_points; + +#ifdef CONFIG_DEBUG_FS + struct mutex debug_lock; + s64 switch_max; + s64 switch_min; + u64 switch_num; + s64 switch_avg; + s64 switch_std; + bool debugfs_set; +#endif }; @@ -51,15 +95,20 @@ struct nvgpu_clk_dev { struct list_head link; wait_queue_head_t readout_wq; atomic_t poll_mask; + u16 gpc2clk_target_mhz; + u16 mclk_target_mhz; }; struct nvgpu_clk_session { bool zombie; struct gk20a *g; struct kref refcount; + struct list_head link; + struct list_head targets; - u64 gpc2clk_target_hz; - u64 mclk_target_hz; + spinlock_t target_lock; + u16 gpc2clk_target_mhz; + u16 mclk_target_mhz; }; static const struct file_operations completion_dev_ops = { @@ -77,7 +126,7 @@ static const struct file_operations event_dev_ops = { int nvgpu_clk_arb_init_arbiter(struct gk20a *g) { struct nvgpu_clk_arb *arb; - u64 default_hz; + u16 default_mhz; int err; gk20a_dbg_fn(""); @@ -86,39 +135,104 @@ int nvgpu_clk_arb_init_arbiter(struct gk20a *g) return 0; arb = kzalloc(sizeof(struct nvgpu_clk_arb), GFP_KERNEL); - if (!arb) - return -ENOMEM; + if (!arb) { + err = -ENOMEM; + goto init_fail; + } + + arb->gpc2clk_f_numpoints = MAX_F_POINTS; + arb->mclk_f_numpoints = MAX_F_POINTS; + + arb->gpc2clk_f_points = kcalloc(MAX_F_POINTS, sizeof(u16), GFP_KERNEL); + if (!arb->gpc2clk_f_points) { + err = -ENOMEM; + goto init_fail; + } + + arb->mclk_f_points = kcalloc(MAX_F_POINTS, sizeof(u16), GFP_KERNEL); + if (!arb->mclk_f_points) { + err = -ENOMEM; + goto init_fail; + } + + arb->gpc2clk_vf_points = kcalloc(MAX_F_POINTS, + sizeof(struct nvgpu_clk_vf_point), GFP_KERNEL); + if (!arb->gpc2clk_vf_points) { + err = -ENOMEM; + goto init_fail; + } + + arb->mclk_vf_points = kcalloc(MAX_F_POINTS, + sizeof(struct nvgpu_clk_vf_point), GFP_KERNEL); + if (!arb->mclk_vf_points) { + err = -ENOMEM; + goto init_fail; + } g->clk_arb = arb; + arb->g = g; - mutex_init(&arb->req_lock); - mutex_init(&arb->users_lock); + spin_lock_init(&arb->sessions_lock); + spin_lock_init(&arb->users_lock); + spin_lock_init(&arb->req_lock); + spin_lock_init(&arb->data_lock); + spin_lock_init(&arb->vf_lock); err = g->ops.clk_arb.get_arbiter_clk_default(g, - NVGPU_GPU_CLK_DOMAIN_MCLK, &default_hz); - if (err) - return -EINVAL; + NVGPU_GPU_CLK_DOMAIN_MCLK, &default_mhz); + if (err) { + err = -EINVAL; + goto init_fail; + } - arb->mclk_target_hz = default_hz; - arb->mclk_current_hz = default_hz; - arb->mclk_default_hz = default_hz; + arb->mclk_default_mhz = default_mhz; err = g->ops.clk_arb.get_arbiter_clk_default(g, - NVGPU_GPU_CLK_DOMAIN_GPC2CLK, &default_hz); - if (err) - return -EINVAL; - - arb->gpc2clk_target_hz = default_hz; - arb->gpc2clk_current_hz = default_hz; - arb->gpc2clk_default_hz = default_hz; + NVGPU_GPU_CLK_DOMAIN_GPC2CLK, &default_mhz); + if (err) { + err = -EINVAL; + goto init_fail; + } - atomic_set(&arb->usercount, 0); + arb->gpc2clk_default_mhz = default_mhz; INIT_LIST_HEAD(&arb->users); + INIT_LIST_HEAD(&arb->sessions); INIT_LIST_HEAD(&arb->requests); + + init_waitqueue_head(&arb->vftable_wq); + + INIT_WORK(&arb->vftable_fn_work, nvgpu_clk_arb_run_vftable_cb); + INIT_WORK(&arb->update_fn_work, nvgpu_clk_arb_run_arbiter_cb); +#ifdef CONFIG_DEBUG_FS + mutex_init(&arb->debug_lock); + if (!arb->debugfs_set) { + if (nvgpu_clk_arb_debugfs_init(g)) + arb->debugfs_set = true; + } +#endif + err = nvgpu_clk_arb_update_vftable(arb); + if (err < 0) + goto init_fail; + + /* Schedule first run */ + schedule_work(&arb->update_fn_work); + return 0; + +init_fail: + + kfree(arb->gpc2clk_f_points); + kfree(arb->gpc2clk_vf_points); + + kfree(arb->mclk_f_points); + kfree(arb->mclk_vf_points); + + kfree(arb); + + return err; } void nvgpu_clk_arb_cleanup_arbiter(struct gk20a *g) @@ -170,6 +284,7 @@ static int nvgpu_clk_arb_install_fd(struct gk20a *g, fail: kfree(dev); put_unused_fd(fd); + return err; } @@ -190,12 +305,16 @@ int nvgpu_clk_arb_init_session(struct gk20a *g, session->g = g; kref_init(&session->refcount); - - atomic_inc(&arb->usercount); + spin_lock_init(&session->target_lock); session->zombie = false; - session->mclk_target_hz = arb->mclk_default_hz; - session->gpc2clk_target_hz = arb->gpc2clk_default_hz; + session->mclk_target_mhz = arb->mclk_default_mhz; + session->gpc2clk_target_mhz = arb->gpc2clk_default_mhz; + INIT_LIST_HEAD(&session->targets); + + spin_lock(&arb->sessions_lock); + list_add_tail(&session->link, &arb->sessions); + spin_unlock(&arb->sessions_lock); *_session = session; @@ -206,8 +325,15 @@ void nvgpu_clk_arb_free_session(struct kref *refcount) { struct nvgpu_clk_session *session = container_of(refcount, struct nvgpu_clk_session, refcount); + struct nvgpu_clk_arb *arb = session->g->clk_arb; + gk20a_dbg_fn(""); + + spin_lock(&arb->sessions_lock); + list_del(&session->link); + spin_unlock(&arb->sessions_lock); kfree(session); +; } void nvgpu_clk_arb_release_session(struct gk20a *g, @@ -215,12 +341,12 @@ void nvgpu_clk_arb_release_session(struct gk20a *g, { struct nvgpu_clk_arb *arb = g->clk_arb; + gk20a_dbg_fn(""); + session->zombie = true; kref_put(&session->refcount, nvgpu_clk_arb_free_session); - /* schedule arbiter if no more user */ - if (!atomic_dec_and_test(&arb->usercount)) - schedule_work(&arb->update_fn_work); + schedule_work(&arb->update_fn_work); } int nvgpu_clk_arb_install_event_fd(struct gk20a *g, @@ -230,19 +356,155 @@ int nvgpu_clk_arb_install_event_fd(struct gk20a *g, struct nvgpu_clk_dev *dev; int fd; + gk20a_dbg_fn(""); + fd = nvgpu_clk_arb_install_fd(g, session, &event_dev_ops, &dev); if (fd < 0) return fd; - mutex_lock(&arb->users_lock); + spin_lock(&arb->users_lock); list_add_tail(&dev->link, &arb->users); - mutex_unlock(&arb->users_lock); + spin_unlock(&arb->users_lock); *event_fd = fd; return 0; } +int nvgpu_clk_arb_install_request_fd(struct gk20a *g, + struct nvgpu_clk_session *session, int *request_fd) +{ + struct nvgpu_clk_dev *dev; + int fd; + + gk20a_dbg_fn(""); + + fd = nvgpu_clk_arb_install_fd(g, session, &completion_dev_ops, &dev); + if (fd < 0) + return fd; + + *request_fd = fd; + + return 0; +} + +static int nvgpu_clk_arb_update_vftable(struct nvgpu_clk_arb *arb) +{ + struct gk20a *g = arb->g; + + int i; + int status = 0; + u32 gpc2clk_voltuv = 0, mclk_voltuv = 0; + u32 gpc2clk_voltuv_sram = 0, mclk_voltuv_sram = 0; + + /* the flag must be visible in all threads */ + mb(); + ACCESS_ONCE(arb->vftable_set) = false; + + spin_lock(&arb->vf_lock); + + if (!clk_domain_get_f_points(arb->g, NVGPU_GPU_CLK_DOMAIN_GPC2CLK, + &arb->gpc2clk_f_numpoints, arb->gpc2clk_f_points) < 0) { + gk20a_err(dev_from_gk20a(g), + "failed to fetch GPC2CLK frequency points"); + goto exit_vftable; + } + if (clk_domain_get_f_points(arb->g, NVGPU_GPU_CLK_DOMAIN_MCLK, + &arb->mclk_f_numpoints, arb->mclk_f_points) < 0) { + gk20a_err(dev_from_gk20a(g), + "failed to fetch MCLK frequency points"); + goto exit_vftable; + } + + + memset(arb->mclk_vf_points, 0, + arb->mclk_f_numpoints*sizeof(struct nvgpu_clk_vf_point)); + memset(arb->gpc2clk_vf_points, 0, + arb->gpc2clk_f_numpoints*sizeof(struct nvgpu_clk_vf_point)); + + for (i = 0 ; i < arb->mclk_f_numpoints; i++) { + arb->mclk_vf_points[i].mhz = arb->mclk_f_points[i]; + mclk_voltuv = mclk_voltuv_sram = 0; + + status = clk_domain_get_f_or_v(g, CTRL_CLK_DOMAIN_MCLK, + &arb->mclk_vf_points[i].mhz, &mclk_voltuv, + CTRL_VOLT_DOMAIN_LOGIC); + if (status < 0) { + gk20a_err(dev_from_gk20a(g), + "failed to get MCLK LOGIC voltage"); + goto exit_vftable; + } + status = clk_domain_get_f_or_v(g, CTRL_CLK_DOMAIN_MCLK, + &arb->mclk_vf_points[i].mhz, &mclk_voltuv_sram, + CTRL_VOLT_DOMAIN_SRAM); + if (status < 0) { + gk20a_err(dev_from_gk20a(g), + "failed to get MCLK SRAM voltage"); + goto exit_vftable; + } + + arb->mclk_vf_points[i].uvolt = mclk_voltuv; + arb->mclk_vf_points[i].uvolt_sram = mclk_voltuv_sram; + } + + for (i = 0 ; i < arb->gpc2clk_f_numpoints; i++) { + arb->gpc2clk_vf_points[i].mhz = arb->gpc2clk_f_points[i]; + gpc2clk_voltuv = gpc2clk_voltuv_sram = 0; + + status = clk_domain_get_f_or_v(g, CTRL_CLK_DOMAIN_GPC2CLK, + &arb->gpc2clk_vf_points[i].mhz, &gpc2clk_voltuv, + CTRL_VOLT_DOMAIN_LOGIC); + if (status < 0) { + gk20a_err(dev_from_gk20a(g), + "failed to get GPC2CLK LOGIC voltage"); + goto exit_vftable; + } + status = clk_domain_get_f_or_v(g, CTRL_CLK_DOMAIN_GPC2CLK, + &arb->gpc2clk_vf_points[i].mhz, &gpc2clk_voltuv_sram, + CTRL_VOLT_DOMAIN_SRAM); + if (status < 0) { + gk20a_err(dev_from_gk20a(g), + "failed to get GPC2CLK SRAM voltage"); + goto exit_vftable; + } + + arb->gpc2clk_vf_points[i].uvolt = gpc2clk_voltuv; + arb->gpc2clk_vf_points[i].uvolt_sram = gpc2clk_voltuv_sram; + + } + + /* make flag visible when all data has resolved in the tables */ + wmb(); + ACCESS_ONCE(arb->vftable_set) = true; + + wake_up(&arb->vftable_wq); +exit_vftable: + + spin_unlock(&arb->vf_lock); + + return status; +} + +void nvgpu_clk_arb_schedule_vftable_update(struct gk20a *g) +{ + struct nvgpu_clk_arb *arb = g->clk_arb; + + ACCESS_ONCE(arb->vftable_set) = false; + /* Disable the flag in case arbiter gets scheduled first */ + mb(); + + schedule_work(&arb->vftable_fn_work); + schedule_work(&arb->update_fn_work); +} + +static void nvgpu_clk_arb_run_vftable_cb(struct work_struct *work) +{ + struct nvgpu_clk_arb *arb = + container_of(work, struct nvgpu_clk_arb, update_fn_work); + + nvgpu_clk_arb_update_vftable(arb); +} + static void nvgpu_clk_arb_run_arbiter_cb(struct work_struct *work) { struct nvgpu_clk_arb *arb = @@ -250,67 +512,270 @@ static void nvgpu_clk_arb_run_arbiter_cb(struct work_struct *work) struct nvgpu_clk_session *session; struct nvgpu_clk_dev *dev; struct nvgpu_clk_dev *tmp; + struct gk20a *g = arb->g; + + struct change_fll_clk fllclk; + u32 gpc2clk_voltuv = 0, mclk_voltuv = 0; + u32 gpc2clk_voltuv_sram = 0, mclk_voltuv_sram = 0; + + u32 voltuv, voltuv_sram; + + int status; + + /* Temporary variables for checking target frequency */ + u16 gpc2clk_target, mclk_target; - mutex_lock(&arb->req_lock); + /* iteration index */ + u32 index; - arb->mclk_target_hz = arb->mclk_default_hz; - arb->gpc2clk_target_hz = arb->gpc2clk_default_hz; +#ifdef CONFIG_DEBUG_FS + u64 t0, t1; +#endif - list_for_each_entry(dev, &arb->requests, link) { - session = dev->session; + gk20a_dbg_fn(""); + +#ifdef CONFIG_DEBUG_FS + g->ops.read_ptimer(g, &t0); +#endif + + /* Only one arbiter should be running */ + gpc2clk_target = 0; + mclk_target = 0; + + spin_lock(&arb->sessions_lock); + list_for_each_entry(session, &arb->sessions, link) { if (!session->zombie) { - /* TODO: arbiter policy. For now last request wins */ + spin_lock(&arb->req_lock); + spin_lock(&session->target_lock); + + mclk_target = mclk_target > session->mclk_target_mhz ? + mclk_target : session->mclk_target_mhz; + + gpc2clk_target = + gpc2clk_target > session->gpc2clk_target_mhz ? + gpc2clk_target : session->gpc2clk_target_mhz; + /* Move processed requests to notification list*/ + list_for_each_entry_safe(dev, tmp, &session->targets, + link) { + list_del_init(&dev->link); + list_add_tail(&dev->link, &arb->requests); + } + spin_unlock(&session->target_lock); + spin_unlock(&arb->req_lock); + + } + } + spin_unlock(&arb->sessions_lock); + + gpc2clk_target = (gpc2clk_target > 0) ? gpc2clk_target : + arb->gpc2clk_actual_mhz ? gpc2clk_target : + arb->gpc2clk_default_mhz; - arb->mclk_target_hz = session->mclk_target_hz; - arb->gpc2clk_target_hz = session->gpc2clk_target_hz; + mclk_target = (mclk_target > 0) ? mclk_target : + arb->mclk_actual_mhz ? mclk_target : + arb->mclk_default_mhz; + + if (!gpc2clk_target && !mclk_target) { + mclk_target = arb->mclk_default_mhz; + gpc2clk_target = arb->gpc2clk_default_mhz; + } + + if (!gpc2clk_target) + gpc2clk_target = arb->gpc2clk_actual_mhz; + + do { + /* Check that the table is set */ + mb(); + wait_event(arb->vftable_wq, arb->vftable_set); + } while (!ACCESS_ONCE(arb->vftable_set)); + + spin_lock(&arb->vf_lock); + /* round up the freq requests */ + for (index = 0; index < arb->gpc2clk_f_numpoints; index++) { + if (arb->gpc2clk_vf_points[index].mhz >= gpc2clk_target) { + gpc2clk_target = arb->gpc2clk_vf_points[index].mhz; + gpc2clk_voltuv = arb->gpc2clk_vf_points[index].uvolt; + gpc2clk_voltuv_sram = + arb->gpc2clk_vf_points[index].uvolt_sram; + break; } } - /* TODO: loop up higher or equal VF points */ + if (index == arb->gpc2clk_f_numpoints) { + gpc2clk_target = arb->gpc2clk_vf_points[index].mhz; + gpc2clk_voltuv = arb->gpc2clk_vf_points[index].uvolt; + gpc2clk_voltuv_sram = + arb->gpc2clk_vf_points[index].uvolt_sram; + } + + if (!mclk_target) + mclk_target = arb->mclk_actual_mhz; + + for (index = 0; index < arb->mclk_f_numpoints; index++) { + if (arb->mclk_vf_points[index].mhz >= mclk_target) { + mclk_target = arb->mclk_vf_points[index].mhz; + mclk_voltuv = arb->mclk_vf_points[index].uvolt; + mclk_voltuv_sram = + arb->mclk_vf_points[index].uvolt_sram; + break; + } + } + if (index == arb->mclk_f_numpoints) { + mclk_target = arb->mclk_vf_points[index].mhz; + mclk_voltuv = arb->mclk_vf_points[index].uvolt; + mclk_voltuv_sram = + arb->mclk_vf_points[index].uvolt_sram; + } + spin_unlock(&arb->vf_lock); + + /* Program clocks */ + /* A change in both mclk of gpc2clk may require a change in voltage */ + if ((arb->gpc2clk_actual_mhz == gpc2clk_target) && + (arb->mclk_actual_mhz == mclk_target)) { + goto exit_arb; + } + + voltuv = gpc2clk_voltuv > mclk_voltuv ? gpc2clk_voltuv : mclk_voltuv; + voltuv_sram = gpc2clk_voltuv_sram > mclk_voltuv_sram ? + gpc2clk_voltuv_sram : mclk_voltuv_sram; + + /* if voltage ascends we do: + * (1) FLL change + * (2) Voltage change + * (3) MCLK change + * If it goes down + * (1) MCLK change + * (2) Voltage change + * (3) FLL change + */ + + /* descending */ + if (voltuv <= arb->voltuv_actual) { + status = g->clk_pmu.clk_mclk.change(g, mclk_target); + if (status < 0) + goto exit_arb; + + status = volt_set_voltage(g, voltuv, voltuv_sram); + if (status < 0) + goto exit_arb; + + fllclk.api_clk_domain = CTRL_CLK_DOMAIN_GPC2CLK; + fllclk.clkmhz = gpc2clk_target; + fllclk.voltuv = voltuv; + status = clk_program_fll_clks(g, &fllclk); + if (status < 0) + goto exit_arb; + } else { + fllclk.api_clk_domain = CTRL_CLK_DOMAIN_GPC2CLK; + fllclk.clkmhz = gpc2clk_target; + fllclk.voltuv = voltuv; + status = clk_program_fll_clks(g, &fllclk); + if (status < 0) + goto exit_arb; + + status = volt_set_voltage(g, voltuv, voltuv_sram); + if (status < 0) + goto exit_arb; + + status = g->clk_pmu.clk_mclk.change(g, mclk_target); + if (status < 0) + goto exit_arb; + } + + spin_lock(&arb->data_lock); + arb->gpc2clk_actual_mhz = gpc2clk_target; + arb->mclk_actual_mhz = mclk_target; + arb->voltuv_actual = voltuv; + /* Make changes visible to other threads */ + wmb(); + + spin_unlock(&arb->data_lock); + +#ifdef CONFIG_DEBUG_FS + g->ops.read_ptimer(g, &t1); + arb->switch_num++; + + mutex_lock(&arb->debug_lock); + if (arb->switch_num == 1) { + arb->switch_max = arb->switch_min = + arb->switch_avg = (t1-t0)/1000; + arb->switch_std = 0; + } else { + s64 prev_avg; + u64 curr = (t1-t0)/1000; + + arb->switch_max = curr > arb->switch_max ? + curr : arb->switch_max; + arb->switch_min = arb->switch_min ? + (curr < arb->switch_min ? + curr : arb->switch_min) : curr; + prev_avg = arb->switch_avg; + arb->switch_avg = (curr + + (arb->switch_avg * (arb->switch_num-1))) / + arb->switch_num; + arb->switch_std += + (curr - arb->switch_avg) * (curr - prev_avg); + } + mutex_unlock(&arb->debug_lock); - arb->mclk_current_hz = arb->mclk_target_hz; - arb->gpc2clk_current_hz = arb->gpc2clk_target_hz; +#endif - /* TODO: actually program the clocks */ +exit_arb: + spin_lock(&arb->req_lock); /* notify completion for all requests */ list_for_each_entry_safe(dev, tmp, &arb->requests, link) { atomic_set(&dev->poll_mask, POLLIN | POLLRDNORM); wake_up_interruptible(&dev->readout_wq); list_del_init(&dev->link); } - mutex_unlock(&arb->req_lock); + spin_unlock(&arb->req_lock); /* notify event for all users */ - mutex_lock(&arb->users_lock); + spin_lock(&arb->users_lock); list_for_each_entry(dev, &arb->users, link) { atomic_set(&dev->poll_mask, POLLIN | POLLRDNORM); wake_up_interruptible(&dev->readout_wq); } - mutex_unlock(&arb->users_lock); - + spin_unlock(&arb->users_lock); } -int nvgpu_clk_arb_apply_session_constraints(struct gk20a *g, - struct nvgpu_clk_session *session, int *completion_fd) +int nvgpu_clk_arb_commit_request_fd(struct gk20a *g, + struct nvgpu_clk_session *session, int request_fd) { struct nvgpu_clk_arb *arb = g->clk_arb; struct nvgpu_clk_dev *dev; - int fd; + struct fd fd; + int err = 0; - fd = nvgpu_clk_arb_install_fd(g, session, &completion_dev_ops, &dev); - if (fd < 0) - return fd; + gk20a_dbg_fn(""); + + fd = fdget(request_fd); + + if (!fd.file) + return -EINVAL; + + dev = (struct nvgpu_clk_dev *) fd.file->private_data; - *completion_fd = fd; + if (!dev || dev->session != session) { + err = -EINVAL; + goto fdput_fd; + } + spin_lock(&session->target_lock); + session->mclk_target_mhz = dev->mclk_target_mhz ? dev->mclk_target_mhz : + session->mclk_target_mhz; + session->gpc2clk_target_mhz = dev->gpc2clk_target_mhz ? + dev->gpc2clk_target_mhz : + session->gpc2clk_target_mhz; - mutex_lock(&arb->req_lock); - list_add_tail(&dev->link, &arb->requests); - mutex_unlock(&arb->req_lock); + list_add_tail(&dev->link, &session->targets); + spin_unlock(&session->target_lock); schedule_work(&arb->update_fn_work); - return 0; +fdput_fd: + fdput(fd); + return err; } static unsigned int nvgpu_clk_arb_poll_dev(struct file *filp, poll_table *wait) @@ -328,11 +793,22 @@ static int nvgpu_clk_arb_release_completion_dev(struct inode *inode, { struct nvgpu_clk_dev *dev = filp->private_data; struct nvgpu_clk_session *session = dev->session; + struct nvgpu_clk_arb *arb; + + arb = session->g->clk_arb; gk20a_dbg_fn(""); + spin_lock(&arb->req_lock); + spin_lock(&session->target_lock); + if (!list_empty(&dev->link)) + list_del_init(&dev->link); + spin_unlock(&session->target_lock); + spin_unlock(&arb->req_lock); + kref_put(&session->refcount, nvgpu_clk_arb_free_session); kfree(dev); + return 0; } @@ -341,94 +817,123 @@ static int nvgpu_clk_arb_release_event_dev(struct inode *inode, { struct nvgpu_clk_dev *dev = filp->private_data; struct nvgpu_clk_session *session = dev->session; - struct nvgpu_clk_arb *arb = session->g->clk_arb; + struct nvgpu_clk_arb *arb; + + arb = session->g->clk_arb; gk20a_dbg_fn(""); - mutex_lock(&arb->users_lock); - list_del_init(&dev->link); - mutex_unlock(&arb->users_lock); + spin_lock(&arb->users_lock); + list_del(&dev->link); + spin_unlock(&arb->users_lock); kref_put(&session->refcount, nvgpu_clk_arb_free_session); kfree(dev); + return 0; } -int nvgpu_clk_arb_set_session_target_hz(struct nvgpu_clk_session *session, - u32 api_domain, u64 target_hz) +int nvgpu_clk_arb_set_session_target_mhz(struct nvgpu_clk_session *session, + int request_fd, u32 api_domain, u16 target_mhz) { + struct nvgpu_clk_dev *dev; + struct fd fd; + int err = 0; - gk20a_dbg_fn("domain=0x%08x target_hz=%llu", api_domain, target_hz); + gk20a_dbg_fn("domain=0x%08x target_mhz=%u", api_domain, target_mhz); + + fd = fdget(request_fd); + + if (!fd.file) + return -EINVAL; + + dev = fd.file->private_data; + if (!dev || dev->session != session) { + err = -EINVAL; + goto fdput_fd; + } switch (api_domain) { case NVGPU_GPU_CLK_DOMAIN_MCLK: - session->mclk_target_hz = target_hz; - return 0; + dev->mclk_target_mhz = target_mhz; + break; case NVGPU_GPU_CLK_DOMAIN_GPC2CLK: - session->gpc2clk_target_hz = target_hz; - return 0; + dev->gpc2clk_target_mhz = target_mhz; + break; default: - return -EINVAL; + err = -EINVAL; } + +fdput_fd: + fdput(fd); + return err; } -int nvgpu_clk_arb_get_session_target_hz(struct nvgpu_clk_session *session, - u32 api_domain, u64 *freq_hz) +int nvgpu_clk_arb_get_session_target_mhz(struct nvgpu_clk_session *session, + u32 api_domain, u16 *freq_mhz) { + int err = 0; + + spin_lock(&session->target_lock); + switch (api_domain) { case NVGPU_GPU_CLK_DOMAIN_MCLK: - *freq_hz = session->mclk_target_hz; - return 0; + *freq_mhz = session->mclk_target_mhz; + break; case NVGPU_GPU_CLK_DOMAIN_GPC2CLK: - *freq_hz = session->gpc2clk_target_hz; - return 0; + *freq_mhz = session->gpc2clk_target_mhz; + break; default: - *freq_hz = 0; - return -EINVAL; + *freq_mhz = 0; + err = -EINVAL; } + + spin_unlock(&session->target_lock); + return err; } -int nvgpu_clk_arb_get_arbiter_actual_hz(struct gk20a *g, - u32 api_domain, u64 *freq_hz) +int nvgpu_clk_arb_get_arbiter_actual_mhz(struct gk20a *g, + u32 api_domain, u16 *freq_mhz) { struct nvgpu_clk_arb *arb = g->clk_arb; int err = 0; - mutex_lock(&arb->req_lock); + spin_lock(&arb->data_lock); + switch (api_domain) { case NVGPU_GPU_CLK_DOMAIN_MCLK: - *freq_hz = arb->mclk_current_hz; + *freq_mhz = arb->mclk_actual_mhz; break; case NVGPU_GPU_CLK_DOMAIN_GPC2CLK: - *freq_hz = arb->gpc2clk_current_hz; + *freq_mhz = arb->gpc2clk_actual_mhz; break; default: - *freq_hz = 0; + *freq_mhz = 0; err = -EINVAL; } - mutex_unlock(&arb->req_lock); + spin_unlock(&arb->data_lock); return err; } -int nvgpu_clk_arb_get_arbiter_effective_hz(struct gk20a *g, - u32 api_domain, u64 *freq_hz) +int nvgpu_clk_arb_get_arbiter_effective_mhz(struct gk20a *g, + u32 api_domain, u16 *freq_mhz) { /* TODO: measure clocks from counters */ - return nvgpu_clk_arb_get_arbiter_actual_hz(g, api_domain, freq_hz); + return nvgpu_clk_arb_get_arbiter_actual_mhz(g, api_domain, freq_mhz); } int nvgpu_clk_arb_get_arbiter_clk_range(struct gk20a *g, u32 api_domain, - u64 *min_hz, u64 *max_hz) + u16 *min_mhz, u16 *max_mhz) { return g->ops.clk_arb.get_arbiter_clk_range(g, api_domain, - min_hz, max_hz); + min_mhz, max_mhz); } u32 nvgpu_clk_arb_get_arbiter_clk_domains(struct gk20a *g) @@ -441,3 +946,67 @@ int nvgpu_clk_arb_get_arbiter_clk_f_points(struct gk20a *g, { return (int)clk_domain_get_f_points(g, api_domain, max_points, fpoints); } + +#ifdef CONFIG_DEBUG_FS +static int nvgpu_clk_arb_stats_show(struct seq_file *s, void *unused) +{ + struct gk20a *g = s->private; + struct nvgpu_clk_arb *arb = g->clk_arb; + u64 num; + s64 tmp, avg, std, max, min; + + /* Make copy of structure to reduce time with lock held */ + mutex_lock(&arb->debug_lock); + std = arb->switch_std; + avg = arb->switch_avg; + max = arb->switch_max; + min = arb->switch_min; + num = arb->switch_num; + mutex_unlock(&arb->debug_lock); + + tmp = std; + do_div(tmp, num); + seq_printf(s, "Number of transitions: %lld\n", + num); + seq_printf(s, "max / min : %lld / %lld usec\n", + max, min); + seq_printf(s, "avg / std : %lld / %ld usec\n", + avg, int_sqrt(tmp)); + + return 0; +} + +static int nvgpu_clk_arb_stats_open(struct inode *inode, struct file *file) +{ + return single_open(file, nvgpu_clk_arb_stats_show, inode->i_private); +} + +static const struct file_operations nvgpu_clk_arb_stats_fops = { + .open = nvgpu_clk_arb_stats_open, + .read = seq_read, + .llseek = seq_lseek, + .release = single_release, +}; + + +static int nvgpu_clk_arb_debugfs_init(struct gk20a *g) +{ + struct gk20a_platform *platform = dev_get_drvdata(g->dev); + + struct dentry *gpu_root = platform->debugfs; + struct dentry *d; + + gk20a_dbg(gpu_dbg_info, "g=%p", g); + + d = debugfs_create_file( + "arb_stats", + S_IRUGO, + gpu_root, + g, + &nvgpu_clk_arb_stats_fops); + if (!d) + return -ENOMEM; + + return 0; +} +#endif diff --git a/drivers/gpu/nvgpu/clk/clk_arb.h b/drivers/gpu/nvgpu/clk/clk_arb.h index 95749369..717cca9b 100644 --- a/drivers/gpu/nvgpu/clk/clk_arb.h +++ b/drivers/gpu/nvgpu/clk/clk_arb.h @@ -22,13 +22,13 @@ struct nvgpu_clk_session; int nvgpu_clk_arb_init_arbiter(struct gk20a *g); int nvgpu_clk_arb_get_arbiter_clk_range(struct gk20a *g, u32 api_domain, - u64 *min_hz, u64 *max_hz); + u16 *min_mhz, u16 *max_mhz); -int nvgpu_clk_arb_get_arbiter_actual_hz(struct gk20a *g, - u32 api_domain, u64 *actual_hz); +int nvgpu_clk_arb_get_arbiter_actual_mhz(struct gk20a *g, + u32 api_domain, u16 *actual_mhz); -int nvgpu_clk_arb_get_arbiter_effective_hz(struct gk20a *g, - u32 api_domain, u64 *actual_hz); +int nvgpu_clk_arb_get_arbiter_effective_mhz(struct gk20a *g, + u32 api_domain, u16 *effective_mhz); int nvgpu_clk_arb_get_arbiter_clk_f_points(struct gk20a *g, u32 api_domain, u32 *max_points, u16 *fpoints); @@ -46,19 +46,21 @@ int nvgpu_clk_arb_init_session(struct gk20a *g, void nvgpu_clk_arb_release_session(struct gk20a *g, struct nvgpu_clk_session *session); -int nvgpu_clk_arb_apply_session_constraints(struct gk20a *g, - struct nvgpu_clk_session *session, int *completion_fd); +int nvgpu_clk_arb_commit_request_fd(struct gk20a *g, + struct nvgpu_clk_session *session, int request_fd); -int nvgpu_clk_arb_set_session_target_hz(struct nvgpu_clk_session *session, - u32 api_domain, u64 target_hz); +int nvgpu_clk_arb_set_session_target_mhz(struct nvgpu_clk_session *session, + int fd, u32 api_domain, u16 target_mhz); -int nvgpu_clk_arb_get_session_target_hz(struct nvgpu_clk_session *session, - u32 api_domain, u64 *target_hz); +int nvgpu_clk_arb_get_session_target_mhz(struct nvgpu_clk_session *session, + u32 api_domain, u16 *target_mhz); int nvgpu_clk_arb_install_event_fd(struct gk20a *g, struct nvgpu_clk_session *session, int *event_fd); +int nvgpu_clk_arb_install_request_fd(struct gk20a *g, + struct nvgpu_clk_session *session, int *event_fd); - +void nvgpu_clk_arb_schedule_vftable_update(struct gk20a *g); #endif /* _CLK_ARB_H_ */ diff --git a/drivers/gpu/nvgpu/clk/clk_mclk.c b/drivers/gpu/nvgpu/clk/clk_mclk.c index 86f4ff6d..6ad6c054 100644 --- a/drivers/gpu/nvgpu/clk/clk_mclk.c +++ b/drivers/gpu/nvgpu/clk/clk_mclk.c @@ -2222,7 +2222,7 @@ int clk_mclkseq_init_mclk_gddr5(struct gk20a *g) return 0; } -int clk_mclkseq_change_mclk_gddr5(struct gk20a *g, enum gk20a_mclk_speed speed) +int clk_mclkseq_change_mclk_gddr5(struct gk20a *g, u16 val) { struct clk_mclk_state *mclk; struct pmu_payload payload = { {0} }; @@ -2236,6 +2236,7 @@ int clk_mclkseq_change_mclk_gddr5(struct gk20a *g, enum gk20a_mclk_speed speed) #ifdef CONFIG_DEBUG_FS u64 t0, t1; #endif + enum gk20a_mclk_speed speed; gk20a_dbg_info(""); @@ -2246,6 +2247,13 @@ int clk_mclkseq_change_mclk_gddr5(struct gk20a *g, enum gk20a_mclk_speed speed) if (!mclk->init) goto exit_status; + /* TODO thia should be done according to VBIOS tables */ + + speed = (val <= MCLK_LOW_SPEED_LIMIT) ? gk20a_mclk_low_speed : + (val <= MCLK_MID_SPEED_LIMIT) ? gk20a_mclk_mid_speed : + gk20a_mclk_high_speed; + + if (speed == mclk->speed) goto exit_status; @@ -2374,20 +2382,13 @@ exit_status: #ifdef CONFIG_DEBUG_FS static int mclk_debug_speed_set(void *data, u64 val) { - enum gk20a_mclk_speed speed; struct gk20a *g = (struct gk20a *) data; struct clk_mclk_state *mclk; mclk = &g->clk_pmu.clk_mclk; - /* TODO thia should be done according to VBIOS tables */ - - speed = (val <= MCLK_LOW_SPEED_LIMIT) ? gk20a_mclk_low_speed : - (val <= MCLK_MID_SPEED_LIMIT) ? gk20a_mclk_mid_speed : - gk20a_mclk_high_speed; - if (mclk->change) - return mclk->change(g, speed); + return mclk->change(g, (u16) val); return 0; } diff --git a/drivers/gpu/nvgpu/clk/clk_mclk.h b/drivers/gpu/nvgpu/clk/clk_mclk.h index 9d193c96..e3e6c1ee 100644 --- a/drivers/gpu/nvgpu/clk/clk_mclk.h +++ b/drivers/gpu/nvgpu/clk/clk_mclk.h @@ -22,9 +22,12 @@ enum gk20a_mclk_speed { gk20a_mclk_high_speed, }; -#define DEFAULT_BOOT_MCLK_SPEED gk20a_mclk_high_speed #define MCLK_LOW_SPEED_LIMIT 405 #define MCLK_MID_SPEED_LIMIT 810 +#define MCLK_HIGH_SPEED_LIMIT 3003 + +#define DEFAULT_BOOT_MCLK_SPEED MCLK_HIGH_SPEED_LIMIT + struct clk_mclk_state { enum gk20a_mclk_speed speed; struct mutex mclk_mutex; @@ -32,7 +35,7 @@ struct clk_mclk_state { bool init; /* function pointers */ - int (*change)(struct gk20a *g, enum gk20a_mclk_speed speed); + int (*change)(struct gk20a *g, u16 val); #ifdef CONFIG_DEBUG_FS s64 switch_max; @@ -45,7 +48,6 @@ struct clk_mclk_state { }; int clk_mclkseq_init_mclk_gddr5(struct gk20a *g); -int clk_mclkseq_change_mclk_gddr5(struct gk20a *g, - enum gk20a_mclk_speed speed); +int clk_mclkseq_change_mclk_gddr5(struct gk20a *g, u16 val); #endif diff --git a/drivers/gpu/nvgpu/gp106/clk_arb_gp106.c b/drivers/gpu/nvgpu/gp106/clk_arb_gp106.c index 112cb588..d1cbb32b 100644 --- a/drivers/gpu/nvgpu/gp106/clk_arb_gp106.c +++ b/drivers/gpu/nvgpu/gp106/clk_arb_gp106.c @@ -23,7 +23,7 @@ static u32 gp106_get_arbiter_clk_domains(struct gk20a *g) } static int gp106_get_arbiter_clk_range(struct gk20a *g, u32 api_domain, - u64 *min_hz, u64 *max_hz) + u16 *min_mhz, u16 *max_mhz) { enum nv_pmu_clk_clkwhich clkwhich; struct clk_set_info *p0_info; @@ -52,14 +52,14 @@ static int gp106_get_arbiter_clk_range(struct gk20a *g, u32 api_domain, if (!p0_info) return -EINVAL; - *min_hz = (u64)(p5_info->min_mhz) * (u64)MHZ; - *max_hz = (u64)(p0_info->max_mhz) * (u64)MHZ; + *min_mhz = p5_info->min_mhz; + *max_mhz = p0_info->max_mhz; return 0; } static int gp106_get_arbiter_clk_default(struct gk20a *g, u32 api_domain, - u64 *default_hz) + u16 *default_mhz) { enum nv_pmu_clk_clkwhich clkwhich; struct clk_set_info *p0_info; @@ -82,7 +82,7 @@ static int gp106_get_arbiter_clk_default(struct gk20a *g, u32 api_domain, if (!p0_info) return -EINVAL; - *default_hz = (u64)p0_info->max_mhz * (u64)MHZ; + *default_mhz = p0_info->max_mhz; return 0; } diff --git a/drivers/gpu/nvgpu/pstate/pstate.c b/drivers/gpu/nvgpu/pstate/pstate.c index 0dc15201..f01b52ad 100644 --- a/drivers/gpu/nvgpu/pstate/pstate.c +++ b/drivers/gpu/nvgpu/pstate/pstate.c @@ -153,10 +153,6 @@ int gk20a_init_pstate_pmu_support(struct gk20a *g) if (err) return err; - err = clk_set_boot_fll_clk(g); - if (err) - return err; - err = pmgr_domain_pmu_setup(g); return err; } -- cgit v1.2.2