[PATCH v2 0/2] misc: fastrpc: fix UAF and Oops around SSR teardown

[PATCH v2 0/2] misc: fastrpc: fix UAF and Oops around SSR teardown

[Jianping Li]

On Hamoa (X1E80100) IoT EVK the ADSP can restart repeatedly, and every
restart has a chance of taking the kernel down with it:

  Unable to handle kernel paging request at virtual address fffffdffc1ffffc0
  Internal error: Oops: 0000000096000006
  CPU: 5 UID: 0 PID: 2766 Comm: adsprpcd
  pc : ___free_pages+0x24/0xe0
  Call trace:
    ___free_pages
    __free_pages
    __dma_direct_free_pages
    dma_direct_free
    dma_free_attrs
    fastrpc_context_free [fastrpc]
    fastrpc_internal_invoke [fastrpc]
    fastrpc_device_ioctl [fastrpc]

The faulting address decodes to a struct page for a PFN that has no
vmemmap backing. It comes from dma_free_coherent() being called against
a qcom,fastrpc-compute-cb device that of_platform_depopulate() has
already unbound: with the IOMMU torn down, the call falls through to
dma_direct_free(), which takes the buffer's IOVA for a physical address
and hands the resulting page to the page allocator.

fastrpc_rpmsg_remove() wakes every pending invoke with -EPIPE and then
immediately depopulates the context banks, with no synchronisation in
between, so woken threads race the teardown on their way to
fastrpc_context_free().
The series is ordered so each patch stands on its own:

  1/2 makes fastrpc_notify_users() wake poll-mode waiters, which today
      keep spinning on a buffer the teardown is about to reclaim.

  2/2 counts in-flight invokes and drains them before touching any
      channel resource.

Jianping Li (2):
  misc: fastrpc: wake poll-mode waiters on SSR
  misc: fastrpc: drain in-flight invokes before tearing down context
    banks

 drivers/misc/fastrpc.c | 61 +++++++++++++++++++++++++++++++++++++++---
 1 file changed, 57 insertions(+), 4 deletions(-)

-- 
2.43.0

[PATCH v2 1/2] misc: fastrpc: wake poll-mode waiters on SSR

[Jianping Li]

fastrpc_notify_users() sets ctx->retval to -EPIPE and completes
ctx->work for every pending and interrupted context, which is enough to
release a thread blocked in fastrpc_wait_for_response().

A thread using polling mode does not wait on that completion. It spins
in poll_for_remote_response(), whose exit condition is:

  (val == FASTRPC_POLL_RESPONSE) || ctx->is_work_done

Since the DSP is already down, it will never write FASTRPC_POLL_RESPONSE
into the poll address, and is_work_done is left untouched by
fastrpc_notify_users(). The thread therefore keeps spinning until
FASTRPC_POLL_MAX_TIMEOUT_US expires instead of bailing out immediately.

Worse, the poll address lives inside ctx->buf, so a polling thread that
has not been told to stop is still reading a buffer that the SSR
teardown path is about to reclaim.

Set is_work_done along with retval so polling waiters observe the
termination on their next iteration, exactly like completion waiters do.

Fixes: 96e62cc0a84d ("misc: fastrpc: Add polling mode support for fastRPC driver")
Signed-off-by: Jianping Li <jianping.li@oss.qualcomm.com>
---
 drivers/misc/fastrpc.c | 1 +
 1 file changed, 1 insertion(+)

diff --git a/drivers/misc/fastrpc.c b/drivers/misc/fastrpc.c
index af18ff1992ee..d201f26d8e46 100644
--- a/drivers/misc/fastrpc.c
+++ b/drivers/misc/fastrpc.c
@@ -2663,6 +2663,7 @@ static void fastrpc_notify_users(struct fastrpc_user *user)
 	spin_lock(&user->lock);
 	list_for_each_entry(ctx, &user->pending, node) {
 		ctx->retval = -EPIPE;
+		ctx->is_work_done = true;
 		complete(&ctx->work);
 	}
 	spin_unlock(&user->lock);
-- 
2.43.0

[PATCH v2 2/2] misc: fastrpc: drain in-flight invokes before tearing down context banks

[Jianping Li]

fastrpc_rpmsg_remove() clears cctx->rpdev, wakes every pending invoke
with -EPIPE through fastrpc_notify_users(), and then goes straight on to
of_platform_depopulate() the context bank devices, with nothing
synchronising the two:

 - The only thing stopping new work is the unlocked rpdev check at the
   top of fastrpc_internal_invoke(). A thread can observe a valid rpdev,
   get preempted, and resume in the middle of the teardown.

 - Invokes that were already admitted are woken and head for the bail:
   label, where fastrpc_context_put() releases ctx->buf through
   dma_free_coherent(). If the depopulate wins, the call runs against an
   unbound context bank and falls through to dma_direct_free(), which
   treats the IOVA as a physical address.

Close both windows under cctx->lock:

 - Add cctx->teardown, set at the start of fastrpc_rpmsg_remove().
   fastrpc_internal_invoke() checks it and, in the same critical
   section, increments cctx->invoke_cnt, so an invoke is either rejected
   or counted. The count is dropped on every return path.

 - fastrpc_rpmsg_remove() waits for invoke_cnt to reach zero before
   clearing rpdev and before touching any other channel resource.

The gate lives in fastrpc_internal_invoke() rather than in the ioctl
dispatcher so that it also covers fastrpc_device_release() ->
fastrpc_release_current_dsp_process(), which sends INIT_RELEASE outside
of any ioctl.

Teardown is kept separate from rpdev because the two have different
lifetimes: the gate must close immediately, whereas rpdev has to stay
valid until the admitted invokes have finished rpmsg_send(). Clearing
rpdev up front, as the code does today, lets an invoke that was sleeping
in an allocation dereference a NULL rpdev:

  Unable to handle kernel NULL pointer dereference at virtual address
0000000000000358
  CPU: 2 UID: 0 PID: 2003 Comm: adsprpcd
  pc : fastrpc_internal_invoke+0xb40/0x1398 [fastrpc]
  Call trace:
    fastrpc_internal_invoke [fastrpc]
    fastrpc_device_release [fastrpc]
    __fput
    __arm64_sys_close

The faulting address is the offset of ept within struct rpmsg_device,
i.e. the cctx->rpdev->ept dereference in rpmsg_send().

Signed-off-by: Jianping Li <jianping.li@oss.qualcomm.com>
---
 drivers/misc/fastrpc.c | 60 +++++++++++++++++++++++++++++++++++++++---
 1 file changed, 56 insertions(+), 4 deletions(-)

diff --git a/drivers/misc/fastrpc.c b/drivers/misc/fastrpc.c
index d201f26d8e46..cffe96a181a2 100644
--- a/drivers/misc/fastrpc.c
+++ b/drivers/misc/fastrpc.c
@@ -328,6 +328,15 @@ struct fastrpc_channel_ctx {
 	atomic_t ctx_seq;
 	u64 dma_mask;
 	const struct fastrpc_soc_data *soc_data;
+	/*
+	 * Set once fastrpc_rpmsg_remove() starts tearing the channel down.
+	 * Checked under @lock so that no new invoke can begin afterwards.
+	 */
+	atomic_t teardown;
+	/* Number of in-flight invokes; guarded by @lock */
+	int invoke_cnt;
+	/* Woken whenever @invoke_cnt drops to zero */
+	wait_queue_head_t ssr_wait_queue;
 };
 
 struct fastrpc_device {
@@ -1350,12 +1359,23 @@ static int fastrpc_wait_for_completion(struct fastrpc_invoke_ctx *ctx,
 	return fastrpc_wait_for_response(ctx, kernel);
 }
 
+/* Caller must hold cctx->lock */
+static void fastrpc_channel_update_invoke_cnt(struct fastrpc_channel_ctx *cctx,
+					      bool enter)
+{
+	if (enter)
+		cctx->invoke_cnt++;
+	else if (--cctx->invoke_cnt == 0)
+		wake_up(&cctx->ssr_wait_queue);
+}
+
 static int fastrpc_internal_invoke(struct fastrpc_user *fl,  u32 kernel,
 				   u32 handle, u32 sc,
 				   struct fastrpc_invoke_args *args)
 {
 	struct fastrpc_invoke_ctx *ctx = NULL;
 	struct fastrpc_buf *buf, *b;
+	unsigned long flags;
 
 	int err = 0;
 
@@ -1365,14 +1385,25 @@ static int fastrpc_internal_invoke(struct fastrpc_user *fl,  u32 kernel,
 	if (!fl->cctx->rpdev)
 		return -EPIPE;
 
+	spin_lock_irqsave(&fl->cctx->lock, flags);
+	if (atomic_read(&fl->cctx->teardown)) {
+		spin_unlock_irqrestore(&fl->cctx->lock, flags);
+		return -EPIPE;
+	}
+	fastrpc_channel_update_invoke_cnt(fl->cctx, true);
+	spin_unlock_irqrestore(&fl->cctx->lock, flags);
+
 	if (handle == FASTRPC_INIT_HANDLE && !kernel) {
 		dev_warn_ratelimited(fl->sctx->dev, "user app trying to send a kernel RPC message (%d)\n",  handle);
-		return -EPERM;
+		err = -EPERM;
+		goto out;
 	}
 
 	ctx = fastrpc_context_alloc(fl, kernel, sc, args);
-	if (IS_ERR(ctx))
-		return PTR_ERR(ctx);
+	if (IS_ERR(ctx)) {
+		err = PTR_ERR(ctx);
+		goto out;
+	}
 
 	err = fastrpc_get_args(kernel, ctx);
 	if (err)
@@ -1428,6 +1459,10 @@ static int fastrpc_internal_invoke(struct fastrpc_user *fl,  u32 kernel,
 
 	if (err)
 		dev_dbg(fl->sctx->dev, "Error: Invoke Failed %d\n", err);
+out:
+	spin_lock_irqsave(&fl->cctx->lock, flags);
+	fastrpc_channel_update_invoke_cnt(fl->cctx, false);
+	spin_unlock_irqrestore(&fl->cctx->lock, flags);
 
 	return err;
 }
@@ -2635,6 +2670,9 @@ static int fastrpc_rpmsg_probe(struct rpmsg_device *rpdev)
 	INIT_LIST_HEAD(&data->invoke_interrupted_mmaps);
 	spin_lock_init(&data->lock);
 	idr_init(&data->ctx_idr);
+	atomic_set(&data->teardown, 0);
+	data->invoke_cnt = 0;
+	init_waitqueue_head(&data->ssr_wait_queue);
 	data->domain_id = domain_id;
 	data->rpdev = rpdev;
 	dev_set_drvdata(&rpdev->dev, data);
@@ -2678,11 +2716,25 @@ static void fastrpc_rpmsg_remove(struct rpmsg_device *rpdev)
 
 	/* No invocations past this point */
 	spin_lock_irqsave(&cctx->lock, flags);
-	cctx->rpdev = NULL;
+	atomic_set(&cctx->teardown, 1);
 	list_for_each_entry(user, &cctx->users, user)
 		fastrpc_notify_users(user);
 	spin_unlock_irqrestore(&cctx->lock, flags);
 
+	/*
+	 * Wait for every invoke that was already past the gate to finish.
+	 * They have all just been woken with -EPIPE, and no new one can be
+	 * counted, so this is guaranteed to make progress.
+	 */
+	spin_lock_irqsave(&cctx->lock, flags);
+	while (cctx->invoke_cnt > 0) {
+		spin_unlock_irqrestore(&cctx->lock, flags);
+		wait_event(cctx->ssr_wait_queue, cctx->invoke_cnt == 0);
+		spin_lock_irqsave(&cctx->lock, flags);
+	}
+	cctx->rpdev = NULL;
+	spin_unlock_irqrestore(&cctx->lock, flags);
+
 	if (cctx->fdevice)
 		misc_deregister(&cctx->fdevice->miscdev);
 
-- 
2.43.0