Linaro-mm-sig

linaro-mm-sig@lists.linaro.org

24 participants
2728 discussions

Re: [Linaro-mm-sig] [PATCH] ASoC: fsl: imx-pcm-dma: Don't request dma channel in probe

by Lucas Stach

Am Dienstag, dem 20.04.2021 um 13:47 +0000 schrieb Robin Gong: > On 2021/04/19 17:46 Lucas Stach <l.stach(a)pengutronix.de> wrote: > > Am Montag, dem 19.04.2021 um 07:17 +0000 schrieb Robin Gong: > > > Hi Lucas, > > > > > > On 2021/04/14 Lucas Stach <l.stach(a)pengutronix.de> wrote: > > > > Hi Robin, > > > > > > > > Am Mittwoch, dem 14.04.2021 um 14:33 +0000 schrieb Robin Gong: > > > > > On 2020/05/20 17:43 Lucas Stach <l.stach(a)pengutronix.de> wrote: > > > > > > Am Mittwoch, den 20.05.2020, 16:20 +0800 schrieb Shengjiu Wang: > > > > > > > Hi > > > > > > > > > > > > > > On Tue, May 19, 2020 at 6:04 PM Lucas Stach > > > > > > > <l.stach(a)pengutronix.de> > > > > > > wrote: > > > > > > > > Am Dienstag, den 19.05.2020, 17:41 +0800 schrieb Shengjiu Wang: > > > > > > > > > There are two requirements that we need to move the > > > > > > > > > request of dma channel from probe to open. > > > > > > > > > > > > > > > > How do you handle -EPROBE_DEFER return code from the channel > > > > > > > > request if you don't do it in probe? > > > > > > > > > > > > > > I use the dma_request_slave_channel or dma_request_channel > > > > > > > instead of dmaengine_pcm_request_chan_of. so there should be > > > > > > > not -EPROBE_DEFER return code. > > > > > > > > > > > > This is a pretty weak argument. The dmaengine device might probe > > > > > > after you try to get the channel. Using a function to request > > > > > > the channel that doesn't allow you to handle probe deferral is > > > > > > IMHO a bug and should be fixed, instead of building even more > > > > > > assumptions on top > > > > of it. > > > > > > > > > > > > > > > - When dma device binds with power-domains, the power will > > > > > > > > > be enabled when we request dma channel. If the request of > > > > > > > > > dma channel happen on probe, then the power-domains will > > > > > > > > > be always enabled after kernel boot up, which is not good > > > > > > > > > for power saving, so we need to move the request of dma > > > > > > > > > channel to .open(); > > > > > > > > > > > > > > > > This is certainly something which could be fixed in the > > > > > > > > dmaengine driver. > > > > > > > > > > > > > > Dma driver always call the pm_runtime_get_sync in > > > > > > > device_alloc_chan_resources, the device_alloc_chan_resources > > > > > > > is called when channel is requested. so power is enabled on > > > > > > > channel > > > > request. > > > > > > > > > > > > So why can't you fix the dmaengine driver to do that RPM call at > > > > > > a later time when the channel is actually going to be used? This > > > > > > will allow further power savings with other slave devices than the audio > > PCM. > > > > > Hi Lucas, > > > > > Thanks for your suggestion. I have tried to implement runtime > > > > > autosuspend in fsl-edma driver on i.mx8qm/qxp with delay time (2 > > > > > sec) for this feature as below (or you can refer to > > > > > drivers/dma/qcom/hidma.c), and pm_runtime_get_sync/ > > > > > pm_runtime_put_autosuspend in all dmaengine driver interface like > > > > > device_alloc_chan_resources/device_prep_slave_sg/device_prep_dma_c > > > > > ycli > > > > > c/ > > > > > device_tx_status... > > > > > > > > > > > > > > > pm_runtime_use_autosuspend(fsl_chan->dev); > > > > > pm_runtime_set_autosuspend_delay(fsl_chan-> > > dev, > > > > 2000); > > > > > > > > > > That could resolve this audio case since the autosuspend could > > > > > suspend runtime after > > > > > 2 seconds if there is no further dma transfer but only channel > > > > request(device_alloc_chan_resources). > > > > > But unfortunately, it cause another issue. As you know, on our > > > > > i.mx8qm/qxp, power domain done by scfw > > > > > (drivers/firmware/imx/scu-pd.c) > > > > over mailbox: > > > > > imx_sc_pd_power()->imx_scu_call_rpc()-> > > > > > imx_scu_ipc_write()->mbox_send_message() > > > > > which means have to 'waits for completion', meanwhile, some driver > > > > > like tty will call dmaengine interfaces in non-atomic case as > > > > > below, > > > > > > > > > > static int uart_write(struct tty_struct *tty, const unsigned char > > > > > *buf, int count) { > > > > > ....... > > > > > port = uart_port_lock(state, flags); > > > > > ...... > > > > > __uart_start(tty); //call > > start_tx()->dmaengine_prep_slave_sg... > > > > > uart_port_unlock(port, flags); > > > > > return ret; > > > > > } > > > > > > > > > > Thus dma runtime resume may happen in that timing window and cause > > > > kernel alarm. > > > > > I'm not sure whether there are similar limitations on other driver > > > > > subsystem. But for me, It looks like the only way to resolve the > > > > > contradiction between tty and scu-pd (hardware limitation on > > > > > i.mx8qm/qxp) is to give up autosuspend and keep > > > > > pm_runtime_get_sync > > > > only in device_alloc_chan_resources because request channel is a > > > > safe non-atomic phase. > > > > > Do you have any idea? Thanks in advance. > > > > > > > > If you look closely at the driver you used as an example (hidma.c) > > > > it looks like there is already something in there, which looks very > > > > much like what you need > > > > here: > > > > > > > > In hidma_issue_pending() the driver tries to get the device to runtime > > resume. > > > > If this doesn't work, maybe due to the power domain code not being > > > > able to be called in atomic context, the actual work of waking up > > > > the dma hardware and issuing the descriptor is shunted to a tasklet. > > > > > > > > If I'm reading this right, this is exactly what you need here to be > > > > able to call the dmaengine code from atomic context: try the rpm get > > > > and issue immediately when possible, otherwise shunt the work to a > > > > non- atomic context where you can deal with the requirements of scu-pd. > > > Yes, I can schedule_work to worker to runtime resume edma channel by > > calling scu-pd. > > > But that means all dmaengine interfaces should be taken care, not only > > > issue_pending() but also > > > dmaengine_terminate_all()/dmaengine_pause()/dmaengine_resume()/ > > > dmaengine_tx_status(). Not sure why hidma only take care > > > issue_pending. Maybe their user case is just for memcpy/memset so that > > > no further complicate case as ALSA or TTY. > > > Besides, for autosuspend in cyclic, we have to add pm_runtime_get_sync > > > into interrupt handler as qcom/bam_dma.c. but how could resolve the > > > scu-pd's non-atmoic limitation in interrupt handler? > > > > Sure, this all needs some careful analysis on how those functions are called > > and what to do about atomic callers, but it should be doable. I don't see any > > fundamental issues here. > > > > I don't see why you would ever need to wake the hardware in an interrupt > > handler. Surely the hardware is already awake, as it wouldn't signal an > > interrupt otherwise. And for the issue with scu-pd you only care about the > > state transition of suspended->running. If the hardware is already > > running/awake, the runtime pm state handling is nothing more than bumping > > a refcount, which is atomic safe. Putting the HW in suspend is already handled > > asynchronously in a worker, so this is also atomic safe. > But with autosuspend used, in corner case, may runtime suspended before falling > Into edma interrupt handler if timeout happen with the delay value of > pm_runtime_set_autosuspend_delay(). Thus, can't touch any edma interrupt > status register unless runtime resume edma in interrupt handler while runtime > resume function based on scu-pd's power domain may block or sleep. > I have a simple workaround that disable runtime suspend in issue_pending worker > by calling pm_runtime_forbid() and then enable runtime auto suspend in > dmaengine_terminate_all so that we could easily regard that edma channel is always > in runtime resume between issue_pending and channel terminated and ignore the above > interrupt handler/scu-pd limitation. The IRQ handler is the point where you are informed by the hardware that a specific operation is complete. I don't see any use-case where it would be valid to drop the rpm refcount to 0 before the IRQ is handled. Surely the hardware needs to stay awake until the currently queued operations are complete and if the IRQ handler is the completion point the IRQ handler is the first point in time where your autosuspend timer should start to run. There should never be a situation where the timer expiry can get between IRQ signaling and the handler code running. Regards, Lucas

4 years

Re: [Linaro-mm-sig] [PATCH v2] dma-buf: Add DmaBufTotal counter in meminfo

by Daniel Stone

Hi Peter, On Fri, 16 Apr 2021 at 13:34, Peter Enderborg <peter.enderborg(a)sony.com> wrote: > This adds a total used dma-buf memory. Details > can be found in debugfs, however it is not for everyone > and not always available. dma-buf are indirect allocated by > userspace. So with this value we can monitor and detect > userspace applications that have problems. > FWIW, this won't work super well for Android where gralloc is implemented as a system service, so all graphics usage will instantly be accounted to it. Cheers, Daniel

4 years

Re: [Linaro-mm-sig] [PATCH v5] dma-buf: Add DmaBufTotal counter in meminfo

by Daniel Vetter

On Tue, Apr 20, 2021 at 09:26:00AM +0000, Peter.Enderborg(a)sony.com wrote: > On 4/20/21 10:58 AM, Daniel Vetter wrote: > > On Sat, Apr 17, 2021 at 06:38:35PM +0200, Peter Enderborg wrote: > >> This adds a total used dma-buf memory. Details > >> can be found in debugfs, however it is not for everyone > >> and not always available. dma-buf are indirect allocated by > >> userspace. So with this value we can monitor and detect > >> userspace applications that have problems. > >> > >> Signed-off-by: Peter Enderborg <peter.enderborg(a)sony.com> > > So there have been tons of discussions around how to track dma-buf and > > why, and I really need to understand the use-cass here first I think. proc > > uapi is as much forever as anything else, and depending what you're doing > > this doesn't make any sense at all: > > > > - on most linux systems dma-buf are only instantiated for shared buffer. > > So there this gives you a fairly meaningless number and not anything > > reflecting gpu memory usage at all. > > > > - on Android all buffers are allocated through dma-buf afaik. But there > > we've recently had some discussions about how exactly we should track > > all this, and the conclusion was that most of this should be solved by > > cgroups long term. So if this is for Android, then I don't think adding > > random quick stop-gaps to upstream is a good idea (because it's a pretty > > long list of patches that have come up on this). > > > > So what is this for? > > For the overview. dma-buf today only have debugfs for info. Debugfs > is not allowed by google to use in andoid. So this aggregate the information > so we can get information on what going on on the system. > > And the LKML standard respond to that is "SHOW ME THE CODE". Yes. Except this extends to how exactly this is supposed to be used in userspace and acted upon. > When the top memgc has a aggregated information on dma-buf it is maybe > a better source to meminfo. But then it also imply that dma-buf requires memcg. > > And I dont see any problem to replace this with something better with it is ready. The thing is, this is uapi. Once it's merged we cannot, ever, replace it. It must be kept around forever, or a very close approximation thereof. So merging this with the justification that we can fix it later on or replace isn't going to happen. -Daniel > > > -Daniel > > > >> --- > >> drivers/dma-buf/dma-buf.c | 12 ++++++++++++ > >> fs/proc/meminfo.c | 5 ++++- > >> include/linux/dma-buf.h | 1 + > >> 3 files changed, 17 insertions(+), 1 deletion(-) > >> > >> diff --git a/drivers/dma-buf/dma-buf.c b/drivers/dma-buf/dma-buf.c > >> index f264b70c383e..4dc37cd4293b 100644 > >> --- a/drivers/dma-buf/dma-buf.c > >> +++ b/drivers/dma-buf/dma-buf.c > >> @@ -37,6 +37,7 @@ struct dma_buf_list { > >> }; > >> > >> static struct dma_buf_list db_list; > >> +static atomic_long_t dma_buf_global_allocated; > >> > >> static char *dmabuffs_dname(struct dentry *dentry, char *buffer, int buflen) > >> { > >> @@ -79,6 +80,7 @@ static void dma_buf_release(struct dentry *dentry) > >> if (dmabuf->resv == (struct dma_resv *)&dmabuf[1]) > >> dma_resv_fini(dmabuf->resv); > >> > >> + atomic_long_sub(dmabuf->size, &dma_buf_global_allocated); > >> module_put(dmabuf->owner); > >> kfree(dmabuf->name); > >> kfree(dmabuf); > >> @@ -586,6 +588,7 @@ struct dma_buf *dma_buf_export(const struct dma_buf_export_info *exp_info) > >> mutex_lock(&db_list.lock); > >> list_add(&dmabuf->list_node, &db_list.head); > >> mutex_unlock(&db_list.lock); > >> + atomic_long_add(dmabuf->size, &dma_buf_global_allocated); > >> > >> return dmabuf; > >> > >> @@ -1346,6 +1349,15 @@ void dma_buf_vunmap(struct dma_buf *dmabuf, struct dma_buf_map *map) > >> } > >> EXPORT_SYMBOL_GPL(dma_buf_vunmap); > >> > >> +/** > >> + * dma_buf_allocated_pages - Return the used nr of pages > >> + * allocated for dma-buf > >> + */ > >> +long dma_buf_allocated_pages(void) > >> +{ > >> + return atomic_long_read(&dma_buf_global_allocated) >> PAGE_SHIFT; > >> +} > >> + > >> #ifdef CONFIG_DEBUG_FS > >> static int dma_buf_debug_show(struct seq_file *s, void *unused) > >> { > >> diff --git a/fs/proc/meminfo.c b/fs/proc/meminfo.c > >> index 6fa761c9cc78..ccc7c40c8db7 100644 > >> --- a/fs/proc/meminfo.c > >> +++ b/fs/proc/meminfo.c > >> @@ -16,6 +16,7 @@ > >> #ifdef CONFIG_CMA > >> #include <linux/cma.h> > >> #endif > >> +#include <linux/dma-buf.h> > >> #include <asm/page.h> > >> #include "internal.h" > >> > >> @@ -145,7 +146,9 @@ static int meminfo_proc_show(struct seq_file *m, void *v) > >> show_val_kb(m, "CmaFree: ", > >> global_zone_page_state(NR_FREE_CMA_PAGES)); > >> #endif > >> - > >> +#ifdef CONFIG_DMA_SHARED_BUFFER > >> + show_val_kb(m, "DmaBufTotal: ", dma_buf_allocated_pages()); > >> +#endif > >> hugetlb_report_meminfo(m); > >> > >> arch_report_meminfo(m); > >> diff --git a/include/linux/dma-buf.h b/include/linux/dma-buf.h > >> index efdc56b9d95f..5b05816bd2cd 100644 > >> --- a/include/linux/dma-buf.h > >> +++ b/include/linux/dma-buf.h > >> @@ -507,4 +507,5 @@ int dma_buf_mmap(struct dma_buf *, struct vm_area_struct *, > >> unsigned long); > >> int dma_buf_vmap(struct dma_buf *dmabuf, struct dma_buf_map *map); > >> void dma_buf_vunmap(struct dma_buf *dmabuf, struct dma_buf_map *map); > >> +long dma_buf_allocated_pages(void); > >> #endif /* __DMA_BUF_H__ */ > >> -- > >> 2.17.1 > >> > >> _______________________________________________ > >> dri-devel mailing list > >> dri-devel(a)lists.freedesktop.org > >> https://urldefense.com/v3/__https://lists.freedesktop.org/mailman/listinfo/… > > _______________________________________________ > dri-devel mailing list > dri-devel(a)lists.freedesktop.org > https://lists.freedesktop.org/mailman/listinfo/dri-devel -- Daniel Vetter Software Engineer, Intel Corporation http://blog.ffwll.ch

4 years

Re: [Linaro-mm-sig] [PATCH v5] dma-buf: Add DmaBufTotal counter in meminfo

by Daniel Vetter

On Sat, Apr 17, 2021 at 06:38:35PM +0200, Peter Enderborg wrote: > This adds a total used dma-buf memory. Details > can be found in debugfs, however it is not for everyone > and not always available. dma-buf are indirect allocated by > userspace. So with this value we can monitor and detect > userspace applications that have problems. > > Signed-off-by: Peter Enderborg <peter.enderborg(a)sony.com> So there have been tons of discussions around how to track dma-buf and why, and I really need to understand the use-cass here first I think. proc uapi is as much forever as anything else, and depending what you're doing this doesn't make any sense at all: - on most linux systems dma-buf are only instantiated for shared buffer. So there this gives you a fairly meaningless number and not anything reflecting gpu memory usage at all. - on Android all buffers are allocated through dma-buf afaik. But there we've recently had some discussions about how exactly we should track all this, and the conclusion was that most of this should be solved by cgroups long term. So if this is for Android, then I don't think adding random quick stop-gaps to upstream is a good idea (because it's a pretty long list of patches that have come up on this). So what is this for? -Daniel > --- > drivers/dma-buf/dma-buf.c | 12 ++++++++++++ > fs/proc/meminfo.c | 5 ++++- > include/linux/dma-buf.h | 1 + > 3 files changed, 17 insertions(+), 1 deletion(-) > > diff --git a/drivers/dma-buf/dma-buf.c b/drivers/dma-buf/dma-buf.c > index f264b70c383e..4dc37cd4293b 100644 > --- a/drivers/dma-buf/dma-buf.c > +++ b/drivers/dma-buf/dma-buf.c > @@ -37,6 +37,7 @@ struct dma_buf_list { > }; > > static struct dma_buf_list db_list; > +static atomic_long_t dma_buf_global_allocated; > > static char *dmabuffs_dname(struct dentry *dentry, char *buffer, int buflen) > { > @@ -79,6 +80,7 @@ static void dma_buf_release(struct dentry *dentry) > if (dmabuf->resv == (struct dma_resv *)&dmabuf[1]) > dma_resv_fini(dmabuf->resv); > > + atomic_long_sub(dmabuf->size, &dma_buf_global_allocated); > module_put(dmabuf->owner); > kfree(dmabuf->name); > kfree(dmabuf); > @@ -586,6 +588,7 @@ struct dma_buf *dma_buf_export(const struct dma_buf_export_info *exp_info) > mutex_lock(&db_list.lock); > list_add(&dmabuf->list_node, &db_list.head); > mutex_unlock(&db_list.lock); > + atomic_long_add(dmabuf->size, &dma_buf_global_allocated); > > return dmabuf; > > @@ -1346,6 +1349,15 @@ void dma_buf_vunmap(struct dma_buf *dmabuf, struct dma_buf_map *map) > } > EXPORT_SYMBOL_GPL(dma_buf_vunmap); > > +/** > + * dma_buf_allocated_pages - Return the used nr of pages > + * allocated for dma-buf > + */ > +long dma_buf_allocated_pages(void) > +{ > + return atomic_long_read(&dma_buf_global_allocated) >> PAGE_SHIFT; > +} > + > #ifdef CONFIG_DEBUG_FS > static int dma_buf_debug_show(struct seq_file *s, void *unused) > { > diff --git a/fs/proc/meminfo.c b/fs/proc/meminfo.c > index 6fa761c9cc78..ccc7c40c8db7 100644 > --- a/fs/proc/meminfo.c > +++ b/fs/proc/meminfo.c > @@ -16,6 +16,7 @@ > #ifdef CONFIG_CMA > #include <linux/cma.h> > #endif > +#include <linux/dma-buf.h> > #include <asm/page.h> > #include "internal.h" > > @@ -145,7 +146,9 @@ static int meminfo_proc_show(struct seq_file *m, void *v) > show_val_kb(m, "CmaFree: ", > global_zone_page_state(NR_FREE_CMA_PAGES)); > #endif > - > +#ifdef CONFIG_DMA_SHARED_BUFFER > + show_val_kb(m, "DmaBufTotal: ", dma_buf_allocated_pages()); > +#endif > hugetlb_report_meminfo(m); > > arch_report_meminfo(m); > diff --git a/include/linux/dma-buf.h b/include/linux/dma-buf.h > index efdc56b9d95f..5b05816bd2cd 100644 > --- a/include/linux/dma-buf.h > +++ b/include/linux/dma-buf.h > @@ -507,4 +507,5 @@ int dma_buf_mmap(struct dma_buf *, struct vm_area_struct *, > unsigned long); > int dma_buf_vmap(struct dma_buf *dmabuf, struct dma_buf_map *map); > void dma_buf_vunmap(struct dma_buf *dmabuf, struct dma_buf_map *map); > +long dma_buf_allocated_pages(void); > #endif /* __DMA_BUF_H__ */ > -- > 2.17.1 > > _______________________________________________ > dri-devel mailing list > dri-devel(a)lists.freedesktop.org > https://lists.freedesktop.org/mailman/listinfo/dri-devel -- Daniel Vetter Software Engineer, Intel Corporation http://blog.ffwll.ch

4 years

Re: [Linaro-mm-sig] [PATCH v4] dma-buf: Add DmaBufTotal counter in meminfo

by Christian König

Am 17.04.21 um 13:20 schrieb Peter.Enderborg(a)sony.com: > On 4/17/21 12:59 PM, Christian König wrote: >> Am 17.04.21 um 12:40 schrieb Peter Enderborg: >>> This adds a total used dma-buf memory. Details >>> can be found in debugfs, however it is not for everyone >>> and not always available. dma-buf are indirect allocated by >>> userspace. So with this value we can monitor and detect >>> userspace applications that have problems. >>> >>> Signed-off-by: Peter Enderborg <peter.enderborg(a)sony.com> >> Reviewed-by: Christian König <christian.koenig(a)amd.com> >> >> How do you want to upstream this? > I don't understand that question. The patch applies on Torvalds 5.12-rc7, > but I guess 5.13 is what we work on right now. Yeah, but how do you want to get it into Linus tree? I can push it together with other DMA-buf patches through drm-misc-next and then Dave will send it to Linus for inclusion in 5.13. But could be that you are pushing multiple changes towards Linus through some other branch. In this case I'm fine if you pick that way instead if you want to keep your patches together for some reason. Christian. > >>> --- >>> drivers/dma-buf/dma-buf.c | 13 +++++++++++++ >>> fs/proc/meminfo.c | 5 ++++- >>> include/linux/dma-buf.h | 1 + >>> 3 files changed, 18 insertions(+), 1 deletion(-) >>> >>> diff --git a/drivers/dma-buf/dma-buf.c b/drivers/dma-buf/dma-buf.c >>> index f264b70c383e..197e5c45dd26 100644 >>> --- a/drivers/dma-buf/dma-buf.c >>> +++ b/drivers/dma-buf/dma-buf.c >>> @@ -37,6 +37,7 @@ struct dma_buf_list { >>> }; >>> static struct dma_buf_list db_list; >>> +static atomic_long_t dma_buf_global_allocated; >>> static char *dmabuffs_dname(struct dentry *dentry, char *buffer, int buflen) >>> { >>> @@ -79,6 +80,7 @@ static void dma_buf_release(struct dentry *dentry) >>> if (dmabuf->resv == (struct dma_resv *)&dmabuf[1]) >>> dma_resv_fini(dmabuf->resv); >>> + atomic_long_sub(dmabuf->size, &dma_buf_global_allocated); >>> module_put(dmabuf->owner); >>> kfree(dmabuf->name); >>> kfree(dmabuf); >>> @@ -586,6 +588,7 @@ struct dma_buf *dma_buf_export(const struct dma_buf_export_info *exp_info) >>> mutex_lock(&db_list.lock); >>> list_add(&dmabuf->list_node, &db_list.head); >>> mutex_unlock(&db_list.lock); >>> + atomic_long_add(dmabuf->size, &dma_buf_global_allocated); >>> return dmabuf; >>> @@ -1346,6 +1349,16 @@ void dma_buf_vunmap(struct dma_buf *dmabuf, struct dma_buf_map *map) >>> } >>> EXPORT_SYMBOL_GPL(dma_buf_vunmap); >>> +/** >>> + * dma_buf_allocated_pages - Return the used nr of pages >>> + * allocated for dma-buf >>> + */ >>> +long dma_buf_allocated_pages(void) >>> +{ >>> + return atomic_long_read(&dma_buf_global_allocated) >> PAGE_SHIFT; >>> +} >>> +EXPORT_SYMBOL_GPL(dma_buf_allocated_pages); >>> + >>> #ifdef CONFIG_DEBUG_FS >>> static int dma_buf_debug_show(struct seq_file *s, void *unused) >>> { >>> diff --git a/fs/proc/meminfo.c b/fs/proc/meminfo.c >>> index 6fa761c9cc78..ccc7c40c8db7 100644 >>> --- a/fs/proc/meminfo.c >>> +++ b/fs/proc/meminfo.c >>> @@ -16,6 +16,7 @@ >>> #ifdef CONFIG_CMA >>> #include <linux/cma.h> >>> #endif >>> +#include <linux/dma-buf.h> >>> #include <asm/page.h> >>> #include "internal.h" >>> @@ -145,7 +146,9 @@ static int meminfo_proc_show(struct seq_file *m, void *v) >>> show_val_kb(m, "CmaFree: ", >>> global_zone_page_state(NR_FREE_CMA_PAGES)); >>> #endif >>> - >>> +#ifdef CONFIG_DMA_SHARED_BUFFER >>> + show_val_kb(m, "DmaBufTotal: ", dma_buf_allocated_pages()); >>> +#endif >>> hugetlb_report_meminfo(m); >>> arch_report_meminfo(m); >>> diff --git a/include/linux/dma-buf.h b/include/linux/dma-buf.h >>> index efdc56b9d95f..5b05816bd2cd 100644 >>> --- a/include/linux/dma-buf.h >>> +++ b/include/linux/dma-buf.h >>> @@ -507,4 +507,5 @@ int dma_buf_mmap(struct dma_buf *, struct vm_area_struct *, >>> unsigned long); >>> int dma_buf_vmap(struct dma_buf *dmabuf, struct dma_buf_map *map); >>> void dma_buf_vunmap(struct dma_buf *dmabuf, struct dma_buf_map *map); >>> +long dma_buf_allocated_pages(void); >>> #endif /* __DMA_BUF_H__ */

4 years

Re: [Linaro-mm-sig] [PATCH v4] dma-buf: Add DmaBufTotal counter in meminfo

by Christian König

Am 20.04.21 um 09:46 schrieb Michal Hocko: > On Tue 20-04-21 09:32:14, Christian König wrote: >> Am 20.04.21 um 09:04 schrieb Michal Hocko: >>> On Mon 19-04-21 18:37:13, Christian König wrote: >>>> Am 19.04.21 um 18:11 schrieb Michal Hocko: > [...] >>> What I am trying to bring up with NUMA side is that the same problem can >>> happen on per-node basis. Let's say that some user consumes unexpectedly >>> large amount of dma-buf on a certain node. This can lead to observable >>> performance impact on anybody on allocating from that node and even >>> worse cause an OOM for node bound consumers. How do I find out that it >>> was dma-buf that has caused the problem? >> Yes, that is the direction my thinking goes as well, but also even further. >> >> See DMA-buf is also used to share device local memory between processes as >> well. In other words VRAM on graphics hardware. >> >> On my test system here I have 32GB of system memory and 16GB of VRAM. I can >> use DMA-buf to allocate that 16GB of VRAM quite easily which then shows up >> under /proc/meminfo as used memory. > This is something that would be really interesting in the changelog. I > mean the expected and extreme memory consumption of this memory. Ideally > with some hints on what to do when the number is really high (e.g. mount > debugfs and have a look here and there to check whether this is just too > many users or an unexpected pattern to be reported). > >> But that isn't really system memory at all, it's just allocated device >> memory. > OK, that was not really clear to me. So this is not really accounted to > MemTotal? It depends. In a lot of embedded systems you only have system memory and in this case that value here is indeed really useful. > If that is really the case then reporting it into the oom > report is completely pointless and I am not even sure /proc/meminfo is > the right interface either. It would just add more confusion I am > afraid. I kind of agree. As I said a DMA-buf could be backed by system memory or device memory. In the case when it is backed by system memory it does make sense to report this in an OOM dump. But only the exporting driver knows what the DMA-buf handle represents, the framework just provides the common ground for inter driver communication. >>> See where I am heading? >> Yeah, totally. Thanks for pointing this out. >> >> Suggestions how to handle that? > As I've pointed out in previous reply we do have an API to account per > node memory but now that you have brought up that this is not something > we account as a regular memory then this doesn't really fit into that > model. But maybe I am just confused. Well does that API also has a counter for memory used by device drivers? If yes then the device driver who exported the DMA-buf should probably use that API. If no we might want to create one. I mean the author of this patch seems to have an use case where this is needed and I also see that we have some hole in how we account memory. Christian.

4 years

Re: [Linaro-mm-sig] [PATCH v4] dma-buf: Add DmaBufTotal counter in meminfo

by Christian König

Am 20.04.21 um 09:04 schrieb Michal Hocko: > On Mon 19-04-21 18:37:13, Christian König wrote: >> Am 19.04.21 um 18:11 schrieb Michal Hocko: > [...] >>> The question is not whether it is NUMA aware but whether it is useful to >>> know per-numa data for the purpose the counter is supposed to serve. >> No, not at all. The pages of a single DMA-buf could even be from different >> NUMA nodes if the exporting driver decides that this is somehow useful. > As the use of the counter hasn't been explained yet I can only > speculate. One thing that I can imagine to be useful is to fill gaps in > our accounting. It is quite often that the memroy accounted in > /proc/meminfo (or oom report) doesn't add up to the overall memory > usage. In some workloads the workload can be huge! In many cases there > are other means to find out additional memory by a subsystem specific > interfaces (e.g. networking buffers). I do assume that dma-buf is just > one of those and the counter can fill the said gap at least partially > for some workloads. That is definitely useful. Yes, completely agree. I'm just not 100% sure if the DMA-buf framework should account for that or the individual drivers exporting DMA-bufs. See below for a further explanation. > What I am trying to bring up with NUMA side is that the same problem can > happen on per-node basis. Let's say that some user consumes unexpectedly > large amount of dma-buf on a certain node. This can lead to observable > performance impact on anybody on allocating from that node and even > worse cause an OOM for node bound consumers. How do I find out that it > was dma-buf that has caused the problem? Yes, that is the direction my thinking goes as well, but also even further. See DMA-buf is also used to share device local memory between processes as well. In other words VRAM on graphics hardware. On my test system here I have 32GB of system memory and 16GB of VRAM. I can use DMA-buf to allocate that 16GB of VRAM quite easily which then shows up under /proc/meminfo as used memory. But that isn't really system memory at all, it's just allocated device memory. > See where I am heading? Yeah, totally. Thanks for pointing this out. Suggestions how to handle that? Regards, Christian.

4 years

Re: [Linaro-mm-sig] [PATCH v4] dma-buf: Add DmaBufTotal counter in meminfo

by Christian König

Am 19.04.21 um 18:11 schrieb Michal Hocko: > On Mon 19-04-21 17:44:13, Christian König wrote: >> Am 19.04.21 um 17:19 schrieb Peter.Enderborg(a)sony.com: >>> On 4/19/21 5:00 PM, Michal Hocko wrote: >>>> On Mon 19-04-21 12:41:58, Peter.Enderborg(a)sony.com wrote: >>>>> On 4/19/21 2:16 PM, Michal Hocko wrote: >>>>>> On Sat 17-04-21 12:40:32, Peter Enderborg wrote: >>>>>>> This adds a total used dma-buf memory. Details >>>>>>> can be found in debugfs, however it is not for everyone >>>>>>> and not always available. dma-buf are indirect allocated by >>>>>>> userspace. So with this value we can monitor and detect >>>>>>> userspace applications that have problems. >>>>>> The changelog would benefit from more background on why this is needed, >>>>>> and who is the primary consumer of that value. >>>>>> >>>>>> I cannot really comment on the dma-buf internals but I have two remarks. >>>>>> Documentation/filesystems/proc.rst needs an update with the counter >>>>>> explanation and secondly is this information useful for OOM situations >>>>>> analysis? If yes then show_mem should dump the value as well. >>>>>> >>>>>> From the implementation point of view, is there any reason why this >>>>>> hasn't used the existing global_node_page_state infrastructure? >>>>> I fix doc in next version. Im not sure what you expect the commit message to include. >>>> As I've said. Usual justification covers answers to following questions >>>> - Why do we need it? >>>> - Why the existing data is insuficient? >>>> - Who is supposed to use the data and for what? >>>> >>>> I can see an answer for the first two questions (because this can be a >>>> lot of memory and the existing infrastructure is not production suitable >>>> - debugfs). But the changelog doesn't really explain who is going to use >>>> the new data. Is this a monitoring to raise an early alarm when the >>>> value grows? Is this for debugging misbehaving drivers? How is it >>>> valuable for those? >>>> >>>>> The function of the meminfo is: (From Documentation/filesystems/proc.rst) >>>>> >>>>> "Provides information about distribution and utilization of memory." >>>> True. Yet we do not export any random counters, do we? >>>> >>>>> Im not the designed of dma-buf, I think global_node_page_state as a kernel >>>>> internal. >>>> It provides a node specific and optimized counters. Is this a good fit >>>> with your new counter? Or the NUMA locality is of no importance? >>> Sounds good to me, if Christian Koenig think it is good, I will use that. >>> It is only virtio in drivers that use the global_node_page_state if >>> that matters. >> DMA-buf are not NUMA aware at all. On which node the pages are allocated >> (and if we use pages at all and not internal device memory) is up to the >> exporter and importer. > The question is not whether it is NUMA aware but whether it is useful to > know per-numa data for the purpose the counter is supposed to serve. No, not at all. The pages of a single DMA-buf could even be from different NUMA nodes if the exporting driver decides that this is somehow useful. Christian.

4 years

Re: [Linaro-mm-sig] [PATCH v4] dma-buf: Add DmaBufTotal counter in meminfo

by Christian König

Am 19.04.21 um 17:19 schrieb Peter.Enderborg(a)sony.com: > On 4/19/21 5:00 PM, Michal Hocko wrote: >> On Mon 19-04-21 12:41:58, Peter.Enderborg(a)sony.com wrote: >>> On 4/19/21 2:16 PM, Michal Hocko wrote: >>>> On Sat 17-04-21 12:40:32, Peter Enderborg wrote: >>>>> This adds a total used dma-buf memory. Details >>>>> can be found in debugfs, however it is not for everyone >>>>> and not always available. dma-buf are indirect allocated by >>>>> userspace. So with this value we can monitor and detect >>>>> userspace applications that have problems. >>>> The changelog would benefit from more background on why this is needed, >>>> and who is the primary consumer of that value. >>>> >>>> I cannot really comment on the dma-buf internals but I have two remarks. >>>> Documentation/filesystems/proc.rst needs an update with the counter >>>> explanation and secondly is this information useful for OOM situations >>>> analysis? If yes then show_mem should dump the value as well. >>>> >>>> From the implementation point of view, is there any reason why this >>>> hasn't used the existing global_node_page_state infrastructure? >>> I fix doc in next version. Im not sure what you expect the commit message to include. >> As I've said. Usual justification covers answers to following questions >> - Why do we need it? >> - Why the existing data is insuficient? >> - Who is supposed to use the data and for what? >> >> I can see an answer for the first two questions (because this can be a >> lot of memory and the existing infrastructure is not production suitable >> - debugfs). But the changelog doesn't really explain who is going to use >> the new data. Is this a monitoring to raise an early alarm when the >> value grows? Is this for debugging misbehaving drivers? How is it >> valuable for those? >> >>> The function of the meminfo is: (From Documentation/filesystems/proc.rst) >>> >>> "Provides information about distribution and utilization of memory." >> True. Yet we do not export any random counters, do we? >> >>> Im not the designed of dma-buf, I think global_node_page_state as a kernel >>> internal. >> It provides a node specific and optimized counters. Is this a good fit >> with your new counter? Or the NUMA locality is of no importance? > Sounds good to me, if Christian Koenig think it is good, I will use that. > It is only virtio in drivers that use the global_node_page_state if > that matters. DMA-buf are not NUMA aware at all. On which node the pages are allocated (and if we use pages at all and not internal device memory) is up to the exporter and importer. Christian. > > >>> dma-buf is a device driver that provides a function so I might be >>> on the outside. However I also see that it might be relevant for a OOM. >>> It is memory that can be freed by killing userspace processes. >>> >>> The show_mem thing. Should it be a separate patch? >> This is up to you but if you want to expose the counter then send it in >> one series. >>

4 years

Re: [Linaro-mm-sig] [PATCH] ASoC: fsl: imx-pcm-dma: Don't request dma channel in probe

by Lucas Stach

Am Montag, dem 19.04.2021 um 07:17 +0000 schrieb Robin Gong: > Hi Lucas, > > On 2021/04/14 Lucas Stach <l.stach(a)pengutronix.de> wrote: > > Hi Robin, > > > > Am Mittwoch, dem 14.04.2021 um 14:33 +0000 schrieb Robin Gong: > > > On 2020/05/20 17:43 Lucas Stach <l.stach(a)pengutronix.de> wrote: > > > > Am Mittwoch, den 20.05.2020, 16:20 +0800 schrieb Shengjiu Wang: > > > > > Hi > > > > > > > > > > On Tue, May 19, 2020 at 6:04 PM Lucas Stach > > > > > <l.stach(a)pengutronix.de> > > > > wrote: > > > > > > Am Dienstag, den 19.05.2020, 17:41 +0800 schrieb Shengjiu Wang: > > > > > > > There are two requirements that we need to move the request of > > > > > > > dma channel from probe to open. > > > > > > > > > > > > How do you handle -EPROBE_DEFER return code from the channel > > > > > > request if you don't do it in probe? > > > > > > > > > > I use the dma_request_slave_channel or dma_request_channel instead > > > > > of dmaengine_pcm_request_chan_of. so there should be not > > > > > -EPROBE_DEFER return code. > > > > > > > > This is a pretty weak argument. The dmaengine device might probe > > > > after you try to get the channel. Using a function to request the > > > > channel that doesn't allow you to handle probe deferral is IMHO a > > > > bug and should be fixed, instead of building even more assumptions on top > > of it. > > > > > > > > > > > - When dma device binds with power-domains, the power will be > > > > > > > enabled when we request dma channel. If the request of dma > > > > > > > channel happen on probe, then the power-domains will be always > > > > > > > enabled after kernel boot up, which is not good for power > > > > > > > saving, so we need to move the request of dma channel to > > > > > > > .open(); > > > > > > > > > > > > This is certainly something which could be fixed in the > > > > > > dmaengine driver. > > > > > > > > > > Dma driver always call the pm_runtime_get_sync in > > > > > device_alloc_chan_resources, the device_alloc_chan_resources is > > > > > called when channel is requested. so power is enabled on channel > > request. > > > > > > > > So why can't you fix the dmaengine driver to do that RPM call at a > > > > later time when the channel is actually going to be used? This will > > > > allow further power savings with other slave devices than the audio PCM. > > > Hi Lucas, > > > Thanks for your suggestion. I have tried to implement runtime > > > autosuspend in fsl-edma driver on i.mx8qm/qxp with delay time (2 sec) > > > for this feature as below (or you can refer to > > > drivers/dma/qcom/hidma.c), and pm_runtime_get_sync/ > > > pm_runtime_put_autosuspend in all dmaengine driver interface like > > > device_alloc_chan_resources/device_prep_slave_sg/device_prep_dma_cycli > > > c/ > > > device_tx_status... > > > > > > > > > pm_runtime_use_autosuspend(fsl_chan->dev); > > > pm_runtime_set_autosuspend_delay(fsl_chan->dev, > > 2000); > > > > > > That could resolve this audio case since the autosuspend could suspend > > > runtime after > > > 2 seconds if there is no further dma transfer but only channel > > request(device_alloc_chan_resources). > > > But unfortunately, it cause another issue. As you know, on our > > > i.mx8qm/qxp, power domain done by scfw (drivers/firmware/imx/scu-pd.c) > > over mailbox: > > > imx_sc_pd_power()->imx_scu_call_rpc()-> > > > imx_scu_ipc_write()->mbox_send_message() > > > which means have to 'waits for completion', meanwhile, some driver > > > like tty will call dmaengine interfaces in non-atomic case as below, > > > > > > static int uart_write(struct tty_struct *tty, const unsigned char > > > *buf, int count) { > > > ....... > > > port = uart_port_lock(state, flags); > > > ...... > > > __uart_start(tty); //call start_tx()->dmaengine_prep_slave_sg... > > > uart_port_unlock(port, flags); > > > return ret; > > > } > > > > > > Thus dma runtime resume may happen in that timing window and cause > > kernel alarm. > > > I'm not sure whether there are similar limitations on other driver > > > subsystem. But for me, It looks like the only way to resolve the > > > contradiction between tty and scu-pd (hardware limitation on > > > i.mx8qm/qxp) is to give up autosuspend and keep pm_runtime_get_sync > > only in device_alloc_chan_resources because request channel is a safe > > non-atomic phase. > > > Do you have any idea? Thanks in advance. > > > > If you look closely at the driver you used as an example (hidma.c) it looks like > > there is already something in there, which looks very much like what you need > > here: > > > > In hidma_issue_pending() the driver tries to get the device to runtime resume. > > If this doesn't work, maybe due to the power domain code not being able to > > be called in atomic context, the actual work of waking up the dma hardware > > and issuing the descriptor is shunted to a tasklet. > > > > If I'm reading this right, this is exactly what you need here to be able to call the > > dmaengine code from atomic context: try the rpm get and issue immediately > > when possible, otherwise shunt the work to a non- atomic context where you > > can deal with the requirements of scu-pd. > Yes, I can schedule_work to worker to runtime resume edma channel by calling scu-pd. > But that means all dmaengine interfaces should be taken care, not only issue_pending() > but also dmaengine_terminate_all()/dmaengine_pause()/dmaengine_resume()/ > dmaengine_tx_status(). Not sure why hidma only take care issue_pending. Maybe > their user case is just for memcpy/memset so that no further complicate case as ALSA > or TTY. > Besides, for autosuspend in cyclic, we have to add pm_runtime_get_sync into interrupt > handler as qcom/bam_dma.c. but how could resolve the scu-pd's non-atmoic limitation > in interrupt handler? Sure, this all needs some careful analysis on how those functions are called and what to do about atomic callers, but it should be doable. I don't see any fundamental issues here. I don't see why you would ever need to wake the hardware in an interrupt handler. Surely the hardware is already awake, as it wouldn't signal an interrupt otherwise. And for the issue with scu-pd you only care about the state transition of suspended->running. If the hardware is already running/awake, the runtime pm state handling is nothing more than bumping a refcount, which is atomic safe. Putting the HW in suspend is already handled asynchronously in a worker, so this is also atomic safe. Regards, Lucas

4 years

Jump to page:

2025

2024

2023

2022

2021

2020

2019

2018

2017

2016

2015

2014

2013

2012

2011

Linaro-mm-sig