On Tue, Dec 14, 2021 at 04:25:11PM +0100, Greg KH wrote:

On Tue, Dec 14, 2021 at 03:59:57PM +0100, Jens Wiklander wrote:

quoted

On Tue, Dec 14, 2021 at 02:44:30PM +0100, Greg KH wrote:

quoted

On Tue, Dec 14, 2021 at 01:35:40PM +0100, Jens Wiklander wrote:

quoted

Since the tee subsystem does not keep a strong reference to its idle
shared memory buffers, it races with other threads that try to destroy a
shared memory through a close of its dma-buf fd or by unmapping the
memory.

In tee_shm_get_from_id() when a lookup in teedev->idr has been
successful, it is possible that the tee_shm is in the dma-buf teardown
path, but that path is blocked by the teedev mutex. Since we don't have
an API to tell if the tee_shm is in the dma-buf teardown path or not we
must find another way of detecting this condition.

Fix this by doing the reference counting directly on the tee_shm using a
new refcount_t refcount field. dma-buf is replaced by using
anon_inode_getfd() instead, this separates the life-cycle of the
underlying file from the tee_shm. tee_shm_put() is updated to hold the
mutex when decreasing the refcount to 0 and then remove the tee_shm from
teedev->idr before releasing the mutex. This means that the tee_shm can
never be found unless it has a refcount larger than 0.

So you are dropping dma-buf support entirely?  And anon_inode_getfd()
works instead?  Why do more people not do this as well?

I don't know, but it should be noted that we're not doing very much with
this file descriptor. We're only using it with mmap() and close().

quoted

quoted

Fixes: 967c9cca2cc5 ("tee: generic TEE subsystem")
Cc: stable@vger.kernel.org
Reviewed-by: Lars Persson <redacted>
Reviewed-by: Sumit Garg <redacted>
Reported-by: Patrik Lantz <redacted>
Signed-off-by: Jens Wiklander <jens.wiklander@linaro.org>
---
 drivers/tee/tee_shm.c   | 174 +++++++++++++++-------------------------
 include/linux/tee_drv.h |   2 +-
 2 files changed, 67 insertions(+), 109 deletions(-)

diff --git a/drivers/tee/tee_shm.c b/drivers/tee/tee_shm.c
index 8a8deb95e918..0c82cf981c46 100644
--- a/drivers/tee/tee_shm.c
+++ b/drivers/tee/tee_shm.c

@@ -1,20 +1,17 @@
 // SPDX-License-Identifier: GPL-2.0-only
 /*
- * Copyright (c) 2015-2016, Linaro Limited
+ * Copyright (c) 2015-2021, Linaro Limited

Nit, did Linaro really make a copyrightable change in 2017, 2018, 2019
and 2020 as well?  If not, please do not claim it.

Fair enough, I was a bit lazy 2018 shouldn't be there now that I've
checked the log. I'll fix.

quoted

quoted

  */
+#include <linux/anon_inodes.h>
 #include <linux/device.h>
-#include <linux/dma-buf.h>
-#include <linux/fdtable.h>
 #include <linux/idr.h>
+#include <linux/mm.h>
 #include <linux/sched.h>
 #include <linux/slab.h>
 #include <linux/tee_drv.h>
 #include <linux/uio.h>
-#include <linux/module.h>
 #include "tee_private.h"
 
-MODULE_IMPORT_NS(DMA_BUF);
-
 static void release_registered_pages(struct tee_shm *shm)
 {
 	if (shm->pages) {

@@ -31,16 +28,8 @@ static void release_registered_pages(struct tee_shm *shm)
 	}
 }
 
-static void tee_shm_release(struct tee_shm *shm)
+static void tee_shm_release(struct tee_device *teedev, struct tee_shm *shm)
 {
-	struct tee_device *teedev = shm->ctx->teedev;
-
-	if (shm->flags & TEE_SHM_DMA_BUF) {
-		mutex_lock(&teedev->mutex);
-		idr_remove(&teedev->idr, shm->id);
-		mutex_unlock(&teedev->mutex);
-	}
-
 	if (shm->flags & TEE_SHM_POOL) {
 		struct tee_shm_pool_mgr *poolm;
 

@@ -67,45 +56,6 @@ static void tee_shm_release(struct tee_shm *shm)
 	tee_device_put(teedev);
 }
 
-static struct sg_table *tee_shm_op_map_dma_buf(struct dma_buf_attachment
-			*attach, enum dma_data_direction dir)
-{
-	return NULL;
-}
-
-static void tee_shm_op_unmap_dma_buf(struct dma_buf_attachment *attach,
-				     struct sg_table *table,
-				     enum dma_data_direction dir)
-{
-}
-
-static void tee_shm_op_release(struct dma_buf *dmabuf)
-{
-	struct tee_shm *shm = dmabuf->priv;
-
-	tee_shm_release(shm);
-}
-
-static int tee_shm_op_mmap(struct dma_buf *dmabuf, struct vm_area_struct *vma)
-{
-	struct tee_shm *shm = dmabuf->priv;
-	size_t size = vma->vm_end - vma->vm_start;
-
-	/* Refuse sharing shared memory provided by application */
-	if (shm->flags & TEE_SHM_USER_MAPPED)
-		return -EINVAL;
-
-	return remap_pfn_range(vma, vma->vm_start, shm->paddr >> PAGE_SHIFT,
-			       size, vma->vm_page_prot);
-}
-
-static const struct dma_buf_ops tee_shm_dma_buf_ops = {
-	.map_dma_buf = tee_shm_op_map_dma_buf,
-	.unmap_dma_buf = tee_shm_op_unmap_dma_buf,
-	.release = tee_shm_op_release,
-	.mmap = tee_shm_op_mmap,
-};
-
 struct tee_shm *tee_shm_alloc(struct tee_context *ctx, size_t size, u32 flags)
 {
 	struct tee_device *teedev = ctx->teedev;

@@ -140,6 +90,7 @@ struct tee_shm *tee_shm_alloc(struct tee_context *ctx, size_t size, u32 flags)
 		goto err_dev_put;
 	}
 
+	refcount_set(&shm->refcount, 1);
 	shm->flags = flags | TEE_SHM_POOL;
 	shm->ctx = ctx;
 	if (flags & TEE_SHM_DMA_BUF)

@@ -153,10 +104,7 @@ struct tee_shm *tee_shm_alloc(struct tee_context *ctx, size_t size, u32 flags)
 		goto err_kfree;
 	}
 
-
 	if (flags & TEE_SHM_DMA_BUF) {
-		DEFINE_DMA_BUF_EXPORT_INFO(exp_info);
-
 		mutex_lock(&teedev->mutex);
 		shm->id = idr_alloc(&teedev->idr, shm, 1, 0, GFP_KERNEL);
 		mutex_unlock(&teedev->mutex);

@@ -164,28 +112,11 @@ struct tee_shm *tee_shm_alloc(struct tee_context *ctx, size_t size, u32 flags)
 			ret = ERR_PTR(shm->id);
 			goto err_pool_free;
 		}
-
-		exp_info.ops = &tee_shm_dma_buf_ops;
-		exp_info.size = shm->size;
-		exp_info.flags = O_RDWR;
-		exp_info.priv = shm;
-
-		shm->dmabuf = dma_buf_export(&exp_info);
-		if (IS_ERR(shm->dmabuf)) {
-			ret = ERR_CAST(shm->dmabuf);
-			goto err_rem;
-		}
 	}
 
 	teedev_ctx_get(ctx);
 
 	return shm;
-err_rem:
-	if (flags & TEE_SHM_DMA_BUF) {
-		mutex_lock(&teedev->mutex);
-		idr_remove(&teedev->idr, shm->id);
-		mutex_unlock(&teedev->mutex);
-	}
 err_pool_free:
 	poolm->ops->free(poolm, shm);
 err_kfree:

@@ -246,6 +177,7 @@ struct tee_shm *tee_shm_register(struct tee_context *ctx, unsigned long addr,
 		goto err;
 	}
 
+	refcount_set(&shm->refcount, 1);
 	shm->flags = flags | TEE_SHM_REGISTER;
 	shm->ctx = ctx;
 	shm->id = -1;

@@ -306,22 +238,6 @@ struct tee_shm *tee_shm_register(struct tee_context *ctx, unsigned long addr,
 		goto err;
 	}
 
-	if (flags & TEE_SHM_DMA_BUF) {
-		DEFINE_DMA_BUF_EXPORT_INFO(exp_info);
-
-		exp_info.ops = &tee_shm_dma_buf_ops;
-		exp_info.size = shm->size;
-		exp_info.flags = O_RDWR;
-		exp_info.priv = shm;
-
-		shm->dmabuf = dma_buf_export(&exp_info);
-		if (IS_ERR(shm->dmabuf)) {
-			ret = ERR_CAST(shm->dmabuf);
-			teedev->desc->ops->shm_unregister(ctx, shm);
-			goto err;
-		}
-	}
-
 	return shm;
 err:
 	if (shm) {

@@ -339,6 +255,35 @@ struct tee_shm *tee_shm_register(struct tee_context *ctx, unsigned long addr,
 }
 EXPORT_SYMBOL_GPL(tee_shm_register);
 
+static int tee_shm_fop_release(struct inode *inode, struct file *filp)
+{
+	tee_shm_put(filp->private_data);
+	return 0;
+}
+
+static int tee_shm_fop_mmap(struct file *filp, struct vm_area_struct *vma)
+{
+	struct tee_shm *shm = filp->private_data;
+	size_t size = vma->vm_end - vma->vm_start;
+
+	/* Refuse sharing shared memory provided by application */
+	if (shm->flags & TEE_SHM_USER_MAPPED)
+		return -EINVAL;
+
+	/* check for overflowing the buffer's size */
+	if (vma->vm_pgoff + vma_pages(vma) > shm->size >> PAGE_SHIFT)
+		return -EINVAL;
+
+	return remap_pfn_range(vma, vma->vm_start, shm->paddr >> PAGE_SHIFT,
+			       size, vma->vm_page_prot);
+}
+
+static const struct file_operations tee_shm_fops = {
+	.owner = THIS_MODULE,
+	.release = tee_shm_fop_release,
+	.mmap = tee_shm_fop_mmap,
+};
+
 /**
  * tee_shm_get_fd() - Increase reference count and return file descriptor
  * @shm:	Shared memory handle

@@ -351,10 +296,11 @@ int tee_shm_get_fd(struct tee_shm *shm)
 	if (!(shm->flags & TEE_SHM_DMA_BUF))
 		return -EINVAL;
 
-	get_dma_buf(shm->dmabuf);
-	fd = dma_buf_fd(shm->dmabuf, O_CLOEXEC);
+	/* matched by tee_shm_put() in tee_shm_op_release() */
+	refcount_inc(&shm->refcount);
+	fd = anon_inode_getfd("tee_shm", &tee_shm_fops, shm, O_RDWR);
 	if (fd < 0)
-		dma_buf_put(shm->dmabuf);
+		tee_shm_put(shm);
 	return fd;
 }
 

@@ -364,17 +310,7 @@ int tee_shm_get_fd(struct tee_shm *shm)
  */
 void tee_shm_free(struct tee_shm *shm)
 {
-	/*
-	 * dma_buf_put() decreases the dmabuf reference counter and will
-	 * call tee_shm_release() when the last reference is gone.
-	 *
-	 * In the case of driver private memory we call tee_shm_release
-	 * directly instead as it doesn't have a reference counter.
-	 */
-	if (shm->flags & TEE_SHM_DMA_BUF)
-		dma_buf_put(shm->dmabuf);
-	else
-		tee_shm_release(shm);
+	tee_shm_put(shm);
 }
 EXPORT_SYMBOL_GPL(tee_shm_free);
 

@@ -481,10 +417,15 @@ struct tee_shm *tee_shm_get_from_id(struct tee_context *ctx, int id)
 	teedev = ctx->teedev;
 	mutex_lock(&teedev->mutex);
 	shm = idr_find(&teedev->idr, id);
+	/*
+	 * If the tee_shm was found in the IDR it must have a refcount
+	 * larger than 0 due to the guarantee in tee_shm_put() below. So
+	 * it's safe to use refcount_inc().
+	 */
 	if (!shm || shm->ctx != ctx)
 		shm = ERR_PTR(-EINVAL);
-	else if (shm->flags & TEE_SHM_DMA_BUF)
-		get_dma_buf(shm->dmabuf);
+	else
+		refcount_inc(&shm->refcount);
 	mutex_unlock(&teedev->mutex);
 	return shm;
 }

@@ -496,7 +437,24 @@ EXPORT_SYMBOL_GPL(tee_shm_get_from_id);
  */
 void tee_shm_put(struct tee_shm *shm)
 {
-	if (shm->flags & TEE_SHM_DMA_BUF)
-		dma_buf_put(shm->dmabuf);
+	struct tee_device *teedev = shm->ctx->teedev;
+	bool do_release = false;
+
+	mutex_lock(&teedev->mutex);
+	if (refcount_dec_and_test(&shm->refcount)) {
+		/*
+		 * refcount has reached 0, we must now remove it from the
+		 * IDR before releasing the mutex. This will guarantee that
+		 * the refcount_inc() in tee_shm_get_from_id() never starts
+		 * from 0.
+		 */
+		if (shm->flags & TEE_SHM_DMA_BUF)
+			idr_remove(&teedev->idr, shm->id);
+		do_release = true;

As you are using a refcount in the "traditional" way, why not just use a
kref instead?  That solves your "do_release" mess here.

Yes, but it adds another problem. I don't want to hold the mutex when
calling tee_shm_release() so that would mean moving idr_remove() to
tee_shm_release() again and then use kref_get_unless_zero() in
tee_shm_get_from_id() instead.

Why does the idr have anything to do with it here?  Once you are "done"
with the object, remove the entry from the idr and that's it.  You
should never have to mess with kref_get_unless_zero(), that's what locks
are for.

I hope it becomes more clear below.

quoted

With this approach the tee_shm is removed from the IDR so it cannot be
seen any longer when the refcount is 0. I tried implementing it in both
ways before and in my opinion this turned out better.

Really?  Something feels wrong here.  tee_release_device() should drop
the reference from the idr structure as then you know that userspace can
not grab any new references to it and should not need to look it up
again from anything.  Then in your final kref_put() call, in the release
callback for that, free the memory.

I assume you mean tee_shm_release().

As I understand it you're describing more or less how it worked before
this patch, the only significant difference I see is that it was the
release callback from the DMA-buf that called tee_shm_release() instead
of kref_put() as you suggest. With that we have a window where the
reference counter is 0, but the tee_shm is still in the IDR. If another
thread, perhaps malicious, calls tee_shm_get_from_id() it may be able to
grab the mutex before tee_shm_release(). If we just increase the
reference counter without checking if it's 0 to start with or not then
we're in trouble.

I see two ways around this, one is doing like in this patch and the
other is to check that the reference counter isn't 0 before increasing
it in tee_shm_get_from_id().

To me it seems preferable to avoid having invalid tee_shm's reachable
from the IDR.

When is the idr being used by any other code path to look up the object
that will be alive after your char dev's release call is made?

Once the tee_device is released then all the tee_shm's should have been
removed from the IDR. The tee_device will live as long as it has any
tee_shm's.

Thanks,
Jens