From: Marek Olšák Date: Wed, 17 Aug 2016 21:49:27 +0000 (+0200) Subject: drm/amdgpu: throttle buffer migrations at CS using a fixed MBps limit (v2) X-Git-Url: https://git.stricted.de/?a=commitdiff_plain;h=95844d20ae024b5d553c9923a0d3145c3956bf69;p=GitHub%2Fmoto-9609%2Fandroid_kernel_motorola_exynos9610.git drm/amdgpu: throttle buffer migrations at CS using a fixed MBps limit (v2) The old mechanism used a per-submission limit that didn't take previous submissions within the same time frame into account. It also filled VRAM slowly when VRAM usage dropped due to a big eviction or buffer deallocation. This new method establishes a configurable MBps limit that is obeyed when VRAM usage is very high. When VRAM usage is not very high, it gives the driver the freedom to fill it quickly. The result is more consistent performance. It can't keep the BO move rate low if lots of evictions are happening due to VRAM fragmentation, or if a big buffer is being migrated. The amdgpu.moverate parameter can be used to set a non-default limit. Measurements can be done to find out which amdgpu.moverate setting gives the best results. Mainly APUs and cards with small VRAM will benefit from this. For F1 2015, anything with 2 GB VRAM or less will benefit. Some benchmark results - F1 2015 (Tonga 2GB): Limit MinFPS AvgFPS Old code: 14 32.6 128 MB/s: 28 41 64 MB/s: 15.5 43 32 MB/s: 28.7 43.4 8 MB/s: 27.8 44.4 8 MB/s: 21.9 42.8 (different run) Random drops in Min FPS can still occur (due to fragmented VRAM?), but the average FPS is much better. 8 MB/s is probably a good limit for this game & the current VRAM management. The random FPS drops are still to be tackled. v2: use a spinlock Signed-off-by: Marek Olšák Acked-by: Christian König Signed-off-by: Alex Deucher --- diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu.h b/drivers/gpu/drm/amd/amdgpu/amdgpu.h index 4cfcf9c37800..938ef1cb68cc 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu.h +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu.h @@ -64,6 +64,7 @@ extern int amdgpu_modeset; extern int amdgpu_vram_limit; extern int amdgpu_gart_size; +extern int amdgpu_moverate; extern int amdgpu_benchmarking; extern int amdgpu_testing; extern int amdgpu_audio; @@ -2034,6 +2035,14 @@ struct amdgpu_device { atomic64_t num_evictions; atomic_t gpu_reset_counter; + /* data for buffer migration throttling */ + struct { + spinlock_t lock; + s64 last_update_us; + s64 accum_us; /* accumulated microseconds */ + u32 log2_max_MBps; + } mm_stats; + /* display */ bool enable_virtual_display; struct amdgpu_mode_info mode_info; diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c index d80e5d3a4add..82927570333a 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c @@ -235,56 +235,115 @@ free_chunk: return ret; } -/* Returns how many bytes TTM can move per IB. +/* Convert microseconds to bytes. */ +static u64 us_to_bytes(struct amdgpu_device *adev, s64 us) +{ + if (us <= 0 || !adev->mm_stats.log2_max_MBps) + return 0; + + /* Since accum_us is incremented by a million per second, just + * multiply it by the number of MB/s to get the number of bytes. + */ + return us << adev->mm_stats.log2_max_MBps; +} + +static s64 bytes_to_us(struct amdgpu_device *adev, u64 bytes) +{ + if (!adev->mm_stats.log2_max_MBps) + return 0; + + return bytes >> adev->mm_stats.log2_max_MBps; +} + +/* Returns how many bytes TTM can move right now. If no bytes can be moved, + * it returns 0. If it returns non-zero, it's OK to move at least one buffer, + * which means it can go over the threshold once. If that happens, the driver + * will be in debt and no other buffer migrations can be done until that debt + * is repaid. + * + * This approach allows moving a buffer of any size (it's important to allow + * that). + * + * The currency is simply time in microseconds and it increases as the clock + * ticks. The accumulated microseconds (us) are converted to bytes and + * returned. */ static u64 amdgpu_cs_get_threshold_for_moves(struct amdgpu_device *adev) { - u64 real_vram_size = adev->mc.real_vram_size; - u64 vram_usage = atomic64_read(&adev->vram_usage); + s64 time_us, increment_us; + u64 max_bytes; + u64 free_vram, total_vram, used_vram; - /* This function is based on the current VRAM usage. + /* Allow a maximum of 200 accumulated ms. This is basically per-IB + * throttling. * - * - If all of VRAM is free, allow relocating the number of bytes that - * is equal to 1/4 of the size of VRAM for this IB. + * It means that in order to get full max MBps, at least 5 IBs per + * second must be submitted and not more than 200ms apart from each + * other. + */ + const s64 us_upper_bound = 200000; - * - If more than one half of VRAM is occupied, only allow relocating - * 1 MB of data for this IB. - * - * - From 0 to one half of used VRAM, the threshold decreases - * linearly. - * __________________ - * 1/4 of -|\ | - * VRAM | \ | - * | \ | - * | \ | - * | \ | - * | \ | - * | \ | - * | \________|1 MB - * |----------------| - * VRAM 0 % 100 % - * used used - * - * Note: It's a threshold, not a limit. The threshold must be crossed - * for buffer relocations to stop, so any buffer of an arbitrary size - * can be moved as long as the threshold isn't crossed before - * the relocation takes place. We don't want to disable buffer - * relocations completely. + if (!adev->mm_stats.log2_max_MBps) + return 0; + + total_vram = adev->mc.real_vram_size - adev->vram_pin_size; + used_vram = atomic64_read(&adev->vram_usage); + free_vram = used_vram >= total_vram ? 0 : total_vram - used_vram; + + spin_lock(&adev->mm_stats.lock); + + /* Increase the amount of accumulated us. */ + time_us = ktime_to_us(ktime_get()); + increment_us = time_us - adev->mm_stats.last_update_us; + adev->mm_stats.last_update_us = time_us; + adev->mm_stats.accum_us = min(adev->mm_stats.accum_us + increment_us, + us_upper_bound); + + /* This prevents the short period of low performance when the VRAM + * usage is low and the driver is in debt or doesn't have enough + * accumulated us to fill VRAM quickly. * - * The idea is that buffers should be placed in VRAM at creation time - * and TTM should only do a minimum number of relocations during - * command submission. In practice, you need to submit at least - * a dozen IBs to move all buffers to VRAM if they are in GTT. + * The situation can occur in these cases: + * - a lot of VRAM is freed by userspace + * - the presence of a big buffer causes a lot of evictions + * (solution: split buffers into smaller ones) * - * Also, things can get pretty crazy under memory pressure and actual - * VRAM usage can change a lot, so playing safe even at 50% does - * consistently increase performance. + * If 128 MB or 1/8th of VRAM is free, start filling it now by setting + * accum_us to a positive number. */ + if (free_vram >= 128 * 1024 * 1024 || free_vram >= total_vram / 8) { + s64 min_us; + + /* Be more aggresive on dGPUs. Try to fill a portion of free + * VRAM now. + */ + if (!(adev->flags & AMD_IS_APU)) + min_us = bytes_to_us(adev, free_vram / 4); + else + min_us = 0; /* Reset accum_us on APUs. */ + + adev->mm_stats.accum_us = max(min_us, adev->mm_stats.accum_us); + } - u64 half_vram = real_vram_size >> 1; - u64 half_free_vram = vram_usage >= half_vram ? 0 : half_vram - vram_usage; - u64 bytes_moved_threshold = half_free_vram >> 1; - return max(bytes_moved_threshold, 1024*1024ull); + /* This returns 0 if the driver is in debt to disallow (optional) + * buffer moves. + */ + max_bytes = us_to_bytes(adev, adev->mm_stats.accum_us); + + spin_unlock(&adev->mm_stats.lock); + return max_bytes; +} + +/* Report how many bytes have really been moved for the last command + * submission. This can result in a debt that can stop buffer migrations + * temporarily. + */ +static void amdgpu_cs_report_moved_bytes(struct amdgpu_device *adev, + u64 num_bytes) +{ + spin_lock(&adev->mm_stats.lock); + adev->mm_stats.accum_us -= bytes_to_us(adev, num_bytes); + spin_unlock(&adev->mm_stats.lock); } static int amdgpu_cs_bo_validate(struct amdgpu_cs_parser *p, @@ -297,15 +356,10 @@ static int amdgpu_cs_bo_validate(struct amdgpu_cs_parser *p, if (bo->pin_count) return 0; - /* Avoid moving this one if we have moved too many buffers - * for this IB already. - * - * Note that this allows moving at least one buffer of - * any size, because it doesn't take the current "bo" - * into account. We don't want to disallow buffer moves - * completely. + /* Don't move this buffer if we have depleted our allowance + * to move it. Don't move anything if the threshold is zero. */ - if (p->bytes_moved <= p->bytes_moved_threshold) + if (p->bytes_moved < p->bytes_moved_threshold) domain = bo->prefered_domains; else domain = bo->allowed_domains; @@ -494,6 +548,8 @@ static int amdgpu_cs_parser_bos(struct amdgpu_cs_parser *p, goto error_validate; } + amdgpu_cs_report_moved_bytes(p->adev, p->bytes_moved); + fpriv->vm.last_eviction_counter = atomic64_read(&p->adev->num_evictions); diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c index 1ef4034b3be5..847583d8a3b3 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c @@ -1490,6 +1490,7 @@ int amdgpu_device_init(struct amdgpu_device *adev, { int r, i; bool runtime = false; + u32 max_MBps; adev->shutdown = false; adev->dev = &pdev->dev; @@ -1549,6 +1550,7 @@ int amdgpu_device_init(struct amdgpu_device *adev, spin_lock_init(&adev->didt_idx_lock); spin_lock_init(&adev->gc_cac_idx_lock); spin_lock_init(&adev->audio_endpt_idx_lock); + spin_lock_init(&adev->mm_stats.lock); INIT_LIST_HEAD(&adev->shadow_list); mutex_init(&adev->shadow_list_lock); @@ -1660,6 +1662,14 @@ int amdgpu_device_init(struct amdgpu_device *adev, adev->accel_working = true; + /* Initialize the buffer migration limit. */ + if (amdgpu_moverate >= 0) + max_MBps = amdgpu_moverate; + else + max_MBps = 8; /* Allow 8 MB/s. */ + /* Get a log2 for easy divisions. */ + adev->mm_stats.log2_max_MBps = ilog2(max(1u, max_MBps)); + amdgpu_fbdev_init(adev); r = amdgpu_ib_pool_init(adev); diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c index 1b787d974515..6fed75454800 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c @@ -62,6 +62,7 @@ int amdgpu_vram_limit = 0; int amdgpu_gart_size = -1; /* auto */ +int amdgpu_moverate = -1; /* auto */ int amdgpu_benchmarking = 0; int amdgpu_testing = 0; int amdgpu_audio = -1; @@ -100,6 +101,9 @@ module_param_named(vramlimit, amdgpu_vram_limit, int, 0600); MODULE_PARM_DESC(gartsize, "Size of PCIE/IGP gart to setup in megabytes (32, 64, etc., -1 = auto)"); module_param_named(gartsize, amdgpu_gart_size, int, 0600); +MODULE_PARM_DESC(moverate, "Maximum buffer migration rate in MB/s. (32, 64, etc., -1=auto, 0=1=disabled)"); +module_param_named(moverate, amdgpu_moverate, int, 0600); + MODULE_PARM_DESC(benchmark, "Run benchmark"); module_param_named(benchmark, amdgpu_benchmarking, int, 0444);