linux-kernel - Re: [PATCH v3 3/3] misc: fastrpc: Support mapping userspace-allocated buffers

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives
Hash Suite: Windows password security audit tool. GUI, reports in PDF.
[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <7e52f1d4-c05c-45ea-878c-cea2e6354c3f@oss.qualcomm.com>
Date: Fri, 30 Jan 2026 17:22:05 +0530
From: Ekansh Gupta <ekansh.gupta@....qualcomm.com>
To: rob.clark@....qualcomm.com
Cc: Dmitry Baryshkov <dmitry.baryshkov@....qualcomm.com>, srini@...nel.org,
        linux-arm-msm@...r.kernel.org, gregkh@...uxfoundation.org,
        quic_bkumar@...cinc.com, linux-kernel@...r.kernel.org,
        quic_chennak@...cinc.com, dri-devel@...ts.freedesktop.org,
        arnd@...db.de
Subject: Re: [PATCH v3 3/3] misc: fastrpc: Support mapping userspace-allocated
 buffers



On 1/30/2026 3:41 AM, Rob Clark wrote:
> On Thu, Jan 29, 2026 at 2:39 AM Ekansh Gupta
> <ekansh.gupta@....qualcomm.com> wrote:
>>
>>
>> On 1/6/2026 8:21 AM, Dmitry Baryshkov wrote:
>>> On Tue, Dec 30, 2025 at 04:32:25PM +0530, Ekansh Gupta wrote:
>>>> Currently, FastRPC only supports mapping buffers allocated by the
>>>> kernel. This limits flexibility for applications that allocate memory
>>>> in userspace using rpcmem or DMABUF and need to share it with the DSP.
>>> Hmm, for DMABUF we need _import_ support rather than support for mapping
>>> of userspace-allocated buffers.
>>>
>>>> Add support for mapping and unmapping userspace-allocated buffers to
>>>> the DSP through SMMU. This includes handling map requests for rpcmem
>>>> and DMABUF-backed memory and providing corresponding unmap
>>>> functionality.
>>> For me this definitely looks like a step back. For drm/accel we are
>>> going to have GEM-managed buffers only. Why do we need to handle
>>> userspace-allocated buffers here?
>> That's correct, GEM-PRIME will handle it properly. Here, the reason to add this
>> change is to enable routing of DSP logs to HLOS which is done by using a shared
>> buffer between userspace process and DSP PD. The buffer can be allocated from
>> both fastrpc driver's DMA-BUF or DMABUF heap(eg. system heap).
>>
>> So this shared buffer is getting mapped to both process's IOMMU device and DSP PD
>> with this change.
> So, a mmap'd dma-buf is not necessarily pinned.  Or even backed with
> pages.  So you wouldn't want to try to map a userspace vaddr from a
> dma-buf to the device.
>
> But looking at the patch, this looks more like mapping an imported
> dmabuf?  Presumably going thru dma_buf_map_attachment() somewhere in
> the existing fastrpc code?
yes, when the fd is passed to this call, first fastrpc_map_create is called which is
calling dma_buf_map_attachment[1]. After this the buffer is mapped onto DSP.

[1] https://git.kernel.org/pub/scm/linux/kernel/git/next/linux-next.git/tree/drivers/misc/fastrpc.c#n781

>
> BR,
> -R
>
>>>> Signed-off-by: Ekansh Gupta <ekansh.gupta@....qualcomm.com>
>>>> ---
>>>>  drivers/misc/fastrpc.c | 97 +++++++++++++++++++++++++++++++++++++-----
>>>>  1 file changed, 86 insertions(+), 11 deletions(-)
>>>>
>>>> @@ -1989,25 +2020,69 @@ static int fastrpc_req_buf_alloc(struct fastrpc_user *fl,
>>>>      return err;
>>>>  }
>>>>
>>>> -static int fastrpc_req_mmap(struct fastrpc_user *fl, char __user *argp)
>>>> +static int fastrpc_req_map_create(struct fastrpc_user *fl,
>>>> +                              struct fastrpc_req_mmap req,
>>>> +                              char __user *argp)
>>>>  {
>>>> -    struct fastrpc_req_mmap req;
>>>> +    struct fastrpc_map *map = NULL;
>>>> +    struct device *dev = fl->sctx->dev;
>>>> +    u64 raddr = 0;
>>>>      int err;
>>>>
>>>> -    if (copy_from_user(&req, argp, sizeof(req)))
>>>> -            return -EFAULT;
>>>> +    err = fastrpc_map_create(fl, req.fd, req.size, 0, &map);
>>>> +    if (err) {
>>>> +            dev_err(dev, "failed to map buffer, fd = %d\n", req.fd);
>>>> +            return err;
>>>> +    }
>>>> +
>>>> +    err = fastrpc_req_map_dsp(fl, map->phys, map->size, req.flags,
>>>> +                              req.vaddrin, &raddr);
>>>> +    if (err)
>>>> +            goto err_invoke;
>>>>
>>>> -    if (req.flags != ADSP_MMAP_ADD_PAGES && req.flags != ADSP_MMAP_REMOTE_HEAP_ADDR) {
>>>> -            dev_err(fl->sctx->dev, "flag not supported 0x%x\n", req.flags);
>>>> +    /* update the buffer to be able to deallocate the memory on the DSP */
>>>> +    map->raddr = (u64)raddr;
>>> Which type are you converting? And why?
>> I'll drop this.
>>>> -            return -EINVAL;
>>>> +    /* let the client know the address to use */
>>>> +    req.vaddrout = raddr;
>>>> +    dev_dbg(dev, "mmap OK: raddr=%p [len=0x%08llx]\n",
>>>> +            (void *)(unsigned long)map->raddr, map->size);
>>>> +
>>>> +    if (copy_to_user(argp, &req, sizeof(req))) {
>>>> +            err = -EFAULT;
>>>> +            goto err_copy;
>>>>      }
>>>>
>>>> -    err = fastrpc_req_buf_alloc(fl, req, argp);
>>>> +    return 0;
>>>> +err_copy:
>>>> +    fastrpc_req_munmap_dsp(fl, map->raddr, map->size);
>>>> +err_invoke:
>>>> +    fastrpc_map_put(map);
>>>>
>>>>      return err;
>>>>  }
>>>>
>>>> +static int fastrpc_req_mmap(struct fastrpc_user *fl, char __user *argp)
>>>> +{
>>>> +    struct fastrpc_req_mmap req;
>>>> +    int err;
>>>> +
>>>> +    if (copy_from_user(&req, argp, sizeof(req)))
>>>> +            return -EFAULT;
>>>> +
>>>> +    if ((req.flags == ADSP_MMAP_ADD_PAGES ||
>>>> +         req.flags == ADSP_MMAP_REMOTE_HEAP_ADDR)) {
>>> Side note: why are these flags not defined in the uABI header?
>> Ack. These should be part of uABI. I'll create a separate patch for this.
>>>> +            err = fastrpc_req_buf_alloc(fl, req, argp);
>>>> +            if (err)
>>>> +                    return err;