Message-ID: <edff6a62-5b4c-433b-d824-51143b83ad93@ti.com>
Date:   Mon, 29 Oct 2018 15:17:59 -0500
From:   Suman Anna <s-anna@...com>
To:     Loic PALLARDY <loic.pallardy@...com>,
        "bjorn.andersson@...aro.org" <bjorn.andersson@...aro.org>,
        "ohad@...ery.com" <ohad@...ery.com>
CC:     "linux-remoteproc@...r.kernel.org" <linux-remoteproc@...r.kernel.org>,
        "linux-kernel@...r.kernel.org" <linux-kernel@...r.kernel.org>,
        Arnaud POULIQUEN <arnaud.pouliquen@...com>,
        "benjamin.gaignard@...aro.org" <benjamin.gaignard@...aro.org>
Subject: Re: [PATCH v4 12/17] remoteproc: modify vring allocation to rely on
 centralized carveout allocator

Hi Loic,

On 10/24/18 10:14 AM, Loic PALLARDY wrote:
> Hi Suman,
> 
>> -----Original Message-----
>> From: Suman Anna <s-anna@...com>
>> Sent: mercredi 24 octobre 2018 02:14
>> To: Loic PALLARDY <loic.pallardy@...com>; bjorn.andersson@...aro.org;
>> ohad@...ery.com
>> Cc: linux-remoteproc@...r.kernel.org; linux-kernel@...r.kernel.org;
>> Arnaud POULIQUEN <arnaud.pouliquen@...com>;
>> benjamin.gaignard@...aro.org
>> Subject: Re: [PATCH v4 12/17] remoteproc: modify vring allocation to rely on
>> centralized carveout allocator
>>
>> On 7/27/18 8:14 AM, Loic Pallardy wrote:
>>> Current version of rproc_alloc_vring function supports only dynamic vring
>>> allocation.
>>>
>>> This patch allows to allocate vrings based on memory region declatation.
>>> Vrings are now manage like memory carveouts, to communize memory
>> management
>>> code in rproc_alloc_registered_carveouts().
>>>
>>> Allocated buffer is retrieved in rp_find_vq() thanks to
>>> rproc_find_carveout_by_name() functions for.
>>>
>>> This patch sets vrings names to vdev"x"vring"y" with x vdev index in
>>> resource table and y vring index in vdev. This will be updated when
>>> name will be associated to vdev in firmware resource table.
>>>
>>> Signed-off-by: Loic Pallardy <loic.pallardy@...com>
>>> ---
>>>  drivers/remoteproc/remoteproc_core.c     | 61 +++++++++++++++++------
>> ---------
>>>  drivers/remoteproc/remoteproc_internal.h |  2 ++
>>>  drivers/remoteproc/remoteproc_virtio.c   | 14 +++++++-
>>>  include/linux/remoteproc.h               |  6 ++--
>>>  4 files changed, 51 insertions(+), 32 deletions(-)
>>>
>>> diff --git a/drivers/remoteproc/remoteproc_core.c
>> b/drivers/remoteproc/remoteproc_core.c
>>> index c543d04..4edc6f0 100644
>>> --- a/drivers/remoteproc/remoteproc_core.c
>>> +++ b/drivers/remoteproc/remoteproc_core.c
>>> @@ -53,6 +53,11 @@ typedef int (*rproc_handle_resources_t)(struct
>> rproc *rproc,
>>>  typedef int (*rproc_handle_resource_t)(struct rproc *rproc,
>>>  				 void *, int offset, int avail);
>>>
>>> +static int rproc_alloc_carveout(struct rproc *rproc,
>>> +				struct rproc_mem_entry *mem);
>>> +static int rproc_release_carveout(struct rproc *rproc,
>>> +				  struct rproc_mem_entry *mem);
>>> +
>>>  /* Unique indices for remoteproc devices */
>>>  static DEFINE_IDA(rproc_dev_index);
>>>
>>> @@ -312,21 +317,33 @@ int rproc_alloc_vring(struct rproc_vdev *rvdev,
>> int i)
>>>  	struct device *dev = &rproc->dev;
>>>  	struct rproc_vring *rvring = &rvdev->vring[i];
>>>  	struct fw_rsc_vdev *rsc;
>>> -	dma_addr_t dma;
>>> -	void *va;
>>>  	int ret, size, notifyid;
>>> +	struct rproc_mem_entry *mem;
>>>
>>>  	/* actual size of vring (in bytes) */
>>>  	size = PAGE_ALIGN(vring_size(rvring->len, rvring->align));
>>>
>>> -	/*
>>> -	 * Allocate non-cacheable memory for the vring. In the future
>>> -	 * this call will also configure the IOMMU for us
>>> -	 */
>>> -	va = dma_alloc_coherent(dev->parent, size, &dma, GFP_KERNEL);
>>> -	if (!va) {
>>> -		dev_err(dev->parent, "dma_alloc_coherent failed\n");
>>> -		return -EINVAL;
>>> +	rsc = (void *)rproc->table_ptr + rvdev->rsc_offset;
>>> +
>>> +	/* Search for pre-registered carveout */
>>> +	mem = rproc_find_carveout_by_name(rproc, "vdev%dvring%d",
>> rvdev->index,
>>> +					  i);
>>> +	if (mem) {
>>> +		if (rproc_check_carveout_da(rproc, mem, rsc->vring[i].da,
>> size))
>>> +			return -ENOMEM;
>>> +	} else {
>>> +		/* Register carveout in in list */
>>> +		mem = rproc_mem_entry_init(dev, 0, 0, size, rsc-
>>> vring[i].da,
>>> +					   rproc_alloc_carveout,
>>> +					   rproc_release_carveout,
>>> +					   "vdev%dvring%d",
>>> +					   rvdev->index, i);
>>> +		if (!mem) {
>>> +			dev_err(dev, "Can't allocate memory entry
>> structure\n");
>>> +			return -ENOMEM;
>>> +		}
>>> +
>>> +		rproc_add_carveout(rproc, mem);
>>>  	}
>>>
>>>  	/*
>>> @@ -337,7 +354,6 @@ int rproc_alloc_vring(struct rproc_vdev *rvdev, int i)
>>>  	ret = idr_alloc(&rproc->notifyids, rvring, 0, 0, GFP_KERNEL);
>>>  	if (ret < 0) {
>>>  		dev_err(dev, "idr_alloc failed: %d\n", ret);
>>> -		dma_free_coherent(dev->parent, size, va, dma);
>>>  		return ret;
>>>  	}
>>>  	notifyid = ret;
>>> @@ -346,21 +362,9 @@ int rproc_alloc_vring(struct rproc_vdev *rvdev, int
>> i)
>>>  	if (notifyid > rproc->max_notifyid)
>>>  		rproc->max_notifyid = notifyid;
>>>
>>> -	dev_dbg(dev, "vring%d: va %pK dma %pad size 0x%x idr %d\n",
>>> -		i, va, &dma, size, notifyid);
>>> -
>>> -	rvring->va = va;
>>> -	rvring->dma = dma;
>>>  	rvring->notifyid = notifyid;
>>>
>>> -	/*
>>> -	 * Let the rproc know the notifyid and da of this vring.
>>> -	 * Not all platforms use dma_alloc_coherent to automatically
>>> -	 * set up the iommu. In this case the device address (da) will
>>> -	 * hold the physical address and not the device address.
>>> -	 */
>>> -	rsc = (void *)rproc->table_ptr + rvdev->rsc_offset;
>>> -	rsc->vring[i].da = dma;
>>> +	/* Let the rproc know the notifyid of this vring.*/
>>>  	rsc->vring[i].notifyid = notifyid;
>>>  	return 0;
>>>  }
>>> @@ -392,12 +396,10 @@ int rproc_alloc_vring(struct rproc_vdev *rvdev,
>> int i)
>>>
>>>  void rproc_free_vring(struct rproc_vring *rvring)
>>>  {
>>> -	int size = PAGE_ALIGN(vring_size(rvring->len, rvring->align));
>>>  	struct rproc *rproc = rvring->rvdev->rproc;
>>>  	int idx = rvring->rvdev->vring - rvring;
>>>  	struct fw_rsc_vdev *rsc;
>>>
>>> -	dma_free_coherent(rproc->dev.parent, size, rvring->va, rvring-
>>> dma);
>>>  	idr_remove(&rproc->notifyids, rvring->notifyid);
>>
>> Also, I am not sure if FW_RSC_ADDR_ANY semantics were enforced
>> previously on the vring da. It was simply overwritten irrespective of
>> the value. Now, I am running again into the "bad carveout rsc
>> configuration value" due to the iommu_domain_check if
>> !FW_RSC_ADDR_ANY.
> 

Just realized that I missed responding on this thread last week.

> When are you running into the "bad carveout rsc configuration value" ? 
> This patch is creating one carveout per vring to rely on generic carveout allocator.
> Then carveout is retrieved from carveout  list and vring resource table information updated.
> 
> If the da of the carveout was fixed in the resource table, it is normal you have this error.

Yes, and if the vring da value is FW_RSC_ADDR_ANY, then I don't run into
this particular error. It's just that the semantics of vring da is open
previously, and my above comment being 0 was being used as a reset value
as well.

> To solve that ST driver is registering one fixed carveout per vring (with the right name today)

Yeah, we still expect to allocate these dynamically, so there won't be
any registration needed.

> It is the same discussion as with Patch 1. 

Right, except that we have a da from RSC_CARVEOUT and a da from vring,
and the previous code had some slight differences between the two. The
vring da semantics were never set before (value was always being
overwritten, also it didn't have a pa field), whereas the remoteproc.h
documentation did mention about FW_RSC_ADDR_ANY (without any backing
implementation previously) for the RSC_CARVEOUT da, with the entry also
having a field for pa.

If we consider we can't change fixed coprocessor address requests, rproc
core should stop its execution.
> It is the responsibility of platform driver to register the right memory regions.
> That's what we discussed with you and Bill in OpenAMP forum.
> TI usecase was to have the same DSP firmware with the same resource table being able to run on any DSP.

Yeah, it only covers one of the usecases/platforms (Keystone 2 DSP
platforms). And this only worked on these platforms so far because we
only were using internal memories - so there were no RSC_CARVEOUT
entries with valid da. Our Davinci DSP is a single instance and we do
have a RSC_CARVEOUT there, which fails due to the same Patch 1 logic
here as well.

> In that case each DSP platform driver has to provide the right memory region configuration with the correct pa to da.
> 
>>
>> FWIW, the rproc_free_vring was actually using the value 0 when resetting.
>  
> It is no more needed as the carveout list is cleared at each stop and recreated at each start.
> Moreover resource table (and firmware) is reloaded at each coprocessor start.

Yes, agreed. This is about the semantics of vring da from before (no
enforcement to strict enforcement of FW_RSC_ADDR_ANY). The overwriting
on da field on vrings with the dma address is actually a mistake, which
we are trying to proliferate more now. It all comes down to the fact of
treating da as dma address when it is not going to be the case on all
remoteprocs.

regards
Suman

> 
> Regards,
> Loic
>>
>> regards
>> Suman
>>
>>>
>>>  	/* reset resource entry info */
>>> @@ -484,6 +486,7 @@ static int rproc_handle_vdev(struct rproc *rproc,
>> struct fw_rsc_vdev *rsc,
>>>
>>>  	rvdev->id = rsc->id;
>>>  	rvdev->rproc = rproc;
>>> +	rvdev->index = rproc->nb_vdev++;
>>>
>>>  	/* parse the vrings */
>>>  	for (i = 0; i < rsc->num_of_vrings; i++) {
>>> @@ -528,9 +531,6 @@ void rproc_vdev_release(struct kref *ref)
>>>
>>>  	for (id = 0; id < ARRAY_SIZE(rvdev->vring); id++) {
>>>  		rvring = &rvdev->vring[id];
>>> -		if (!rvring->va)
>>> -			continue;
>>> -
>>>  		rproc_free_vring(rvring);
>>>  	}
>>>
>>> @@ -1322,6 +1322,9 @@ static int rproc_fw_boot(struct rproc *rproc,
>> const struct firmware *fw)
>>>  	/* reset max_notifyid */
>>>  	rproc->max_notifyid = -1;
>>>
>>> +	/* reset handled vdev */
>>> +	rproc->nb_vdev = 0;
>>> +
>>>  	/* handle fw resources which are required to boot rproc */
>>>  	ret = rproc_handle_resources(rproc, rproc_loading_handlers);
>>>  	if (ret) {
>>> diff --git a/drivers/remoteproc/remoteproc_internal.h
>> b/drivers/remoteproc/remoteproc_internal.h
>>> index 7570beb..f6cad24 100644
>>> --- a/drivers/remoteproc/remoteproc_internal.h
>>> +++ b/drivers/remoteproc/remoteproc_internal.h
>>> @@ -60,6 +60,8 @@ struct dentry *rproc_create_trace_file(const char
>> *name, struct rproc *rproc,
>>>  int rproc_elf_load_rsc_table(struct rproc *rproc, const struct firmware
>> *fw);
>>>  struct resource_table *rproc_elf_find_loaded_rsc_table(struct rproc
>> *rproc,
>>>  						       const struct firmware
>> *fw);
>>> +struct rproc_mem_entry *
>>> +rproc_find_carveout_by_name(struct rproc *rproc, const char *name,
>> ...);
>>>
>>>  static inline
>>>  int rproc_fw_sanity_check(struct rproc *rproc, const struct firmware *fw)
>>> diff --git a/drivers/remoteproc/remoteproc_virtio.c
>> b/drivers/remoteproc/remoteproc_virtio.c
>>> index bbecd44..de21f62 100644
>>> --- a/drivers/remoteproc/remoteproc_virtio.c
>>> +++ b/drivers/remoteproc/remoteproc_virtio.c
>>> @@ -76,7 +76,9 @@ static struct virtqueue *rp_find_vq(struct
>> virtio_device *vdev,
>>>  	struct rproc_vdev *rvdev = vdev_to_rvdev(vdev);
>>>  	struct rproc *rproc = vdev_to_rproc(vdev);
>>>  	struct device *dev = &rproc->dev;
>>> +	struct rproc_mem_entry *mem;
>>>  	struct rproc_vring *rvring;
>>> +	struct fw_rsc_vdev *rsc;
>>>  	struct virtqueue *vq;
>>>  	void *addr;
>>>  	int len, size;
>>> @@ -88,8 +90,14 @@ static struct virtqueue *rp_find_vq(struct
>> virtio_device *vdev,
>>>  	if (!name)
>>>  		return NULL;
>>>
>>> +	/* Search allocated memory region by name */
>>> +	mem = rproc_find_carveout_by_name(rproc, "vdev%dvring%d",
>> rvdev->index,
>>> +					  id);
>>> +	if (!mem || !mem->va)
>>> +		return ERR_PTR(-ENOMEM);
>>> +
>>>  	rvring = &rvdev->vring[id];
>>> -	addr = rvring->va;
>>> +	addr = mem->va;
>>>  	len = rvring->len;
>>>
>>>  	/* zero vring */
>>> @@ -114,6 +122,10 @@ static struct virtqueue *rp_find_vq(struct
>> virtio_device *vdev,
>>>  	rvring->vq = vq;
>>>  	vq->priv = rvring;
>>>
>>> +	/* Update vring in resource table */
>>> +	rsc = (void *)rproc->table_ptr + rvdev->rsc_offset;
>>> +	rsc->vring[id].da = mem->da;
>>> +
>>>  	return vq;
>>>  }
>>>
>>> diff --git a/include/linux/remoteproc.h b/include/linux/remoteproc.h
>>> index 4cdd0c6..6b3a234 100644
>>> --- a/include/linux/remoteproc.h
>>> +++ b/include/linux/remoteproc.h
>>> @@ -453,6 +453,7 @@ struct rproc_dump_segment {
>>>   * @table_sz: size of @cached_table
>>>   * @has_iommu: flag to indicate if remote processor is behind an MMU
>>>   * @dump_segments: list of segments in the firmware
>>> + * @nb_vdev: number of vdev currently handled by rproc
>>>   */
>>>  struct rproc {
>>>  	struct list_head node;
>>> @@ -485,6 +486,7 @@ struct rproc {
>>>  	bool has_iommu;
>>>  	bool auto_boot;
>>>  	struct list_head dump_segments;
>>> +	int nb_vdev;
>>>  };
>>>
>>>  /**
>>> @@ -512,7 +514,6 @@ struct rproc_subdev {
>>>  /**
>>>   * struct rproc_vring - remoteproc vring state
>>>   * @va:	virtual address
>>> - * @dma: dma address
>>>   * @len: length, in bytes
>>>   * @da: device address
>>>   * @align: vring alignment
>>> @@ -522,7 +523,6 @@ struct rproc_subdev {
>>>   */
>>>  struct rproc_vring {
>>>  	void *va;
>>> -	dma_addr_t dma;
>>>  	int len;
>>>  	u32 da;
>>>  	u32 align;
>>> @@ -541,6 +541,7 @@ struct rproc_vring {
>>>   * @vdev: the virio device
>>>   * @vring: the vrings for this vdev
>>>   * @rsc_offset: offset of the vdev's resource entry
>>> + * @index: vdev position versus other vdev declared in resource table
>>>   */
>>>  struct rproc_vdev {
>>>  	struct kref refcount;
>>> @@ -553,6 +554,7 @@ struct rproc_vdev {
>>>  	struct virtio_device vdev;
>>>  	struct rproc_vring vring[RVDEV_NUM_VRINGS];
>>>  	u32 rsc_offset;
>>> +	u32 index;
>>>  };
>>>
>>>  struct rproc *rproc_get_by_phandle(phandle phandle);
>>>
> 

Powered by blists - more mailing lists