| Message ID | 1449129208-63615-1-git-send-email-koro.chen@mediatek.com (mailing list archive) |
|---|---|
| State | Accepted |
| Commit | 8d6f88ce961cf62137696627448cfd6038f07f41 |
| Headers | show |
On Thu, 03 Dec 2015 08:53:28 +0100, Koro Chen wrote: > > Previously we recorded "last interrupt position" and used it in > pointer callback. This is not correct implementation, and it causes > underruns when user space monitors buffer level to decide when to > send next data chunk in low latency application. Using the last irq position isn't incorrect, per se. Many hardware can do only in this way. But it's suboptimal for your device, yes. > Remove position recording in IRQ handler and also hw_ptr in > struct mtk_afe_memif used to record that, and let pointer callback > reports current HW pointer instead. > > Signed-off-by: Koro Chen <koro.chen@mediatek.com> While reading this patch, I wondered how regmap can be used safely in an irq-disabled context. Mark, do we have any API for that? In this particular case, it would work since it specifies REGCACHE_NONE, so it drops all cache support. As its cost, the driver implements its own (kind of) cache in suspend/resume callbacks. It's not ideal, but still practically working. Takashi > --- > sound/soc/mediatek/mtk-afe-common.h | 1 - > sound/soc/mediatek/mtk-afe-pcm.c | 22 +++++++++++----------- > 2 files changed, 11 insertions(+), 12 deletions(-) > > diff --git a/sound/soc/mediatek/mtk-afe-common.h b/sound/soc/mediatek/mtk-afe-common.h > index cc4393c..9b1af1a 100644 > --- a/sound/soc/mediatek/mtk-afe-common.h > +++ b/sound/soc/mediatek/mtk-afe-common.h > @@ -92,7 +92,6 @@ struct mtk_afe_memif_data { > struct mtk_afe_memif { > unsigned int phys_buf_addr; > int buffer_size; > - unsigned int hw_ptr; /* Previous IRQ's HW ptr */ > struct snd_pcm_substream *substream; > const struct mtk_afe_memif_data *data; > const struct mtk_afe_irq_data *irqdata; > diff --git a/sound/soc/mediatek/mtk-afe-pcm.c b/sound/soc/mediatek/mtk-afe-pcm.c > index 7f71343..5399a0e 100644 > --- a/sound/soc/mediatek/mtk-afe-pcm.c > +++ b/sound/soc/mediatek/mtk-afe-pcm.c > @@ -175,8 +175,17 @@ static snd_pcm_uframes_t mtk_afe_pcm_pointer > struct snd_soc_pcm_runtime *rtd = substream->private_data; > struct mtk_afe *afe = snd_soc_platform_get_drvdata(rtd->platform); > struct mtk_afe_memif *memif = &afe->memif[rtd->cpu_dai->id]; > + unsigned int hw_ptr; > + int ret; > + > + ret = regmap_read(afe->regmap, memif->data->reg_ofs_cur, &hw_ptr); > + if (ret || hw_ptr == 0) { > + dev_err(afe->dev, "%s hw_ptr err\n", __func__); > + hw_ptr = memif->phys_buf_addr; > + } > > - return bytes_to_frames(substream->runtime, memif->hw_ptr); > + return bytes_to_frames(substream->runtime, > + hw_ptr - memif->phys_buf_addr); > } > > static const struct snd_pcm_ops mtk_afe_pcm_ops = { > @@ -602,7 +611,6 @@ static int mtk_afe_dais_hw_params(struct snd_pcm_substream *substream, > > memif->phys_buf_addr = substream->runtime->dma_addr; > memif->buffer_size = substream->runtime->dma_bytes; > - memif->hw_ptr = 0; > > /* start */ > regmap_write(afe->regmap, > @@ -737,7 +745,6 @@ static int mtk_afe_dais_trigger(struct snd_pcm_substream *substream, int cmd, > /* and clear pending IRQ */ > regmap_write(afe->regmap, AFE_IRQ_CLR, > 1 << memif->data->irq_clr_shift); > - memif->hw_ptr = 0; > return 0; > default: > return -EINVAL; > @@ -1081,7 +1088,7 @@ static const struct regmap_config mtk_afe_regmap_config = { > static irqreturn_t mtk_afe_irq_handler(int irq, void *dev_id) > { > struct mtk_afe *afe = dev_id; > - unsigned int reg_value, hw_ptr; > + unsigned int reg_value; > int i, ret; > > ret = regmap_read(afe->regmap, AFE_IRQ_STATUS, ®_value); > @@ -1097,13 +1104,6 @@ static irqreturn_t mtk_afe_irq_handler(int irq, void *dev_id) > if (!(reg_value & (1 << memif->data->irq_clr_shift))) > continue; > > - ret = regmap_read(afe->regmap, memif->data->reg_ofs_cur, > - &hw_ptr); > - if (ret || hw_ptr == 0) { > - dev_err(afe->dev, "%s hw_ptr err\n", __func__); > - hw_ptr = memif->phys_buf_addr; > - } > - memif->hw_ptr = hw_ptr - memif->phys_buf_addr; > snd_pcm_period_elapsed(memif->substream); > } > > -- > 1.7.9.5 > > _______________________________________________ > Alsa-devel mailing list > Alsa-devel@alsa-project.org > http://mailman.alsa-project.org/mailman/listinfo/alsa-devel >
On Thu, Dec 03, 2015 at 10:41:38AM +0100, Takashi Iwai wrote: > While reading this patch, I wondered how regmap can be used safely in > an irq-disabled context. Mark, do we have any API for that? We can use user supplied locks or spin_lock_irqsave().
On Thu, 03 Dec 2015 12:01:58 +0100, Mark Brown wrote: > > On Thu, Dec 03, 2015 at 10:41:38AM +0100, Takashi Iwai wrote: > > > While reading this patch, I wondered how regmap can be used safely in > > an irq-disabled context. Mark, do we have any API for that? > > We can use user supplied locks or spin_lock_irqsave(). I meant how to guarantee to make regmap_read() working in an already spin-locked context, typically in an irq handler? regmap_read() involves with the regcache and it may invoke kmalloc(). Takashi
On Thu, 2015-12-03 at 12:07 +0100, Takashi Iwai wrote: > On Thu, 03 Dec 2015 12:01:58 +0100, > Mark Brown wrote: > > > > On Thu, Dec 03, 2015 at 10:41:38AM +0100, Takashi Iwai wrote: > > > > > While reading this patch, I wondered how regmap can be used safely in > > > an irq-disabled context. Mark, do we have any API for that? > > > > We can use user supplied locks or spin_lock_irqsave(). > > I meant how to guarantee to make regmap_read() working in an already > spin-locked context, typically in an irq handler? regmap_read() > involves with the regcache and it may invoke kmalloc(). > Yes, we were hit by this before. When using devm_regmap_init_mmio() for regmap, it uses spin_lock_irqsave() before every read/write, and if cache type is RBTREE, then kmalloc will occur after spin_lock_irqsave() and it brings warning. That's why we changed RBTREE to NONE. Setting cache type to FLAT will also work, but we think our register accessing is fast enough without need of cache. > > Takashi > _______________________________________________ > Alsa-devel mailing list > Alsa-devel@alsa-project.org > http://mailman.alsa-project.org/mailman/listinfo/alsa-devel
On Thu, Dec 03, 2015 at 12:07:26PM +0100, Takashi Iwai wrote: > Mark Brown wrote: > > On Thu, Dec 03, 2015 at 10:41:38AM +0100, Takashi Iwai wrote: > > > While reading this patch, I wondered how regmap can be used safely in > > > an irq-disabled context. Mark, do we have any API for that? > > We can use user supplied locks or spin_lock_irqsave(). > I meant how to guarantee to make regmap_read() working in an already > spin-locked context, typically in an irq handler? regmap_read() > involves with the regcache and it may invoke kmalloc(). I know that's what you meant - that should be done by preallocating the cache (which can be done with defaults) and providing your own lock if there's a spinlock already held (since we use _irqsave() which IIRC isn't nestable). We can also use GFP_ATOMIC for some of the allocations in reasonable use cases but it's not in general supported.
On Thu, 03 Dec 2015 15:56:12 +0100, Mark Brown wrote: > > On Thu, Dec 03, 2015 at 12:07:26PM +0100, Takashi Iwai wrote: > > Mark Brown wrote: > > > On Thu, Dec 03, 2015 at 10:41:38AM +0100, Takashi Iwai wrote: > > > > > While reading this patch, I wondered how regmap can be used safely in > > > > an irq-disabled context. Mark, do we have any API for that? > > > > We can use user supplied locks or spin_lock_irqsave(). > > > I meant how to guarantee to make regmap_read() working in an already > > spin-locked context, typically in an irq handler? regmap_read() > > involves with the regcache and it may invoke kmalloc(). > > I know that's what you meant - that should be done by preallocating the > cache (which can be done with defaults) and providing your own lock > if there's a spinlock already held (since we use _irqsave() which IIRC > isn't nestable). The extra lock should be fine unless any mutex is involved silently in regmap code. But preallocation sounds not so intuitive. > We can also use GFP_ATOMIC for some of the allocations > in reasonable use cases but it's not in general supported. Yeah, but we have already map->alloc_flags, so we can use it appropriately as a fallback? Takashi
diff --git a/sound/soc/mediatek/mtk-afe-common.h b/sound/soc/mediatek/mtk-afe-common.h index cc4393c..9b1af1a 100644 --- a/sound/soc/mediatek/mtk-afe-common.h +++ b/sound/soc/mediatek/mtk-afe-common.h @@ -92,7 +92,6 @@ struct mtk_afe_memif_data { struct mtk_afe_memif { unsigned int phys_buf_addr; int buffer_size; - unsigned int hw_ptr; /* Previous IRQ's HW ptr */ struct snd_pcm_substream *substream; const struct mtk_afe_memif_data *data; const struct mtk_afe_irq_data *irqdata; diff --git a/sound/soc/mediatek/mtk-afe-pcm.c b/sound/soc/mediatek/mtk-afe-pcm.c index 7f71343..5399a0e 100644 --- a/sound/soc/mediatek/mtk-afe-pcm.c +++ b/sound/soc/mediatek/mtk-afe-pcm.c @@ -175,8 +175,17 @@ static snd_pcm_uframes_t mtk_afe_pcm_pointer struct snd_soc_pcm_runtime *rtd = substream->private_data; struct mtk_afe *afe = snd_soc_platform_get_drvdata(rtd->platform); struct mtk_afe_memif *memif = &afe->memif[rtd->cpu_dai->id]; + unsigned int hw_ptr; + int ret; + + ret = regmap_read(afe->regmap, memif->data->reg_ofs_cur, &hw_ptr); + if (ret || hw_ptr == 0) { + dev_err(afe->dev, "%s hw_ptr err\n", __func__); + hw_ptr = memif->phys_buf_addr; + } - return bytes_to_frames(substream->runtime, memif->hw_ptr); + return bytes_to_frames(substream->runtime, + hw_ptr - memif->phys_buf_addr); } static const struct snd_pcm_ops mtk_afe_pcm_ops = { @@ -602,7 +611,6 @@ static int mtk_afe_dais_hw_params(struct snd_pcm_substream *substream, memif->phys_buf_addr = substream->runtime->dma_addr; memif->buffer_size = substream->runtime->dma_bytes; - memif->hw_ptr = 0; /* start */ regmap_write(afe->regmap, @@ -737,7 +745,6 @@ static int mtk_afe_dais_trigger(struct snd_pcm_substream *substream, int cmd, /* and clear pending IRQ */ regmap_write(afe->regmap, AFE_IRQ_CLR, 1 << memif->data->irq_clr_shift); - memif->hw_ptr = 0; return 0; default: return -EINVAL; @@ -1081,7 +1088,7 @@ static const struct regmap_config mtk_afe_regmap_config = { static irqreturn_t mtk_afe_irq_handler(int irq, void *dev_id) { struct mtk_afe *afe = dev_id; - unsigned int reg_value, hw_ptr; + unsigned int reg_value; int i, ret; ret = regmap_read(afe->regmap, AFE_IRQ_STATUS, ®_value); @@ -1097,13 +1104,6 @@ static irqreturn_t mtk_afe_irq_handler(int irq, void *dev_id) if (!(reg_value & (1 << memif->data->irq_clr_shift))) continue; - ret = regmap_read(afe->regmap, memif->data->reg_ofs_cur, - &hw_ptr); - if (ret || hw_ptr == 0) { - dev_err(afe->dev, "%s hw_ptr err\n", __func__); - hw_ptr = memif->phys_buf_addr; - } - memif->hw_ptr = hw_ptr - memif->phys_buf_addr; snd_pcm_period_elapsed(memif->substream); }
Previously we recorded "last interrupt position" and used it in pointer callback. This is not correct implementation, and it causes underruns when user space monitors buffer level to decide when to send next data chunk in low latency application. Remove position recording in IRQ handler and also hw_ptr in struct mtk_afe_memif used to record that, and let pointer callback reports current HW pointer instead. Signed-off-by: Koro Chen <koro.chen@mediatek.com> --- sound/soc/mediatek/mtk-afe-common.h | 1 - sound/soc/mediatek/mtk-afe-pcm.c | 22 +++++++++++----------- 2 files changed, 11 insertions(+), 12 deletions(-)