Skip to contentSkip to conversation

io_uring: fix possible use after free to sqd

PATCH In review Xiaoguang Wang 2 participants 5 messages Oct 15, 2020

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 io_uring: fix possible use after free to sqd

Reading codes finds a possible use after free issue to sqd:

          thread1              |       thread2
==> io_attach_sq_data()        |
===> sqd = ctx_attach->sq_data;|
                               | ==> io_put_sq_data()
                               | ===> refcount_dec_and_test(&sqd->refs)
                               |     If sqd->refs is zero, will free sqd.
                               |
===> refcount_inc(&sqd->refs); |
                               |
                               | ====> kfree(sqd);
===> now use after free to sqd |

Use refcount_inc_not_zero() to fix this issue.

---
 fs/io_uring.c | 6 +++++-
 1 file changed, 5 insertions(+), 1 deletion(-)
fs/io_uring.c +5 -1
index 33b5cf18bb51..48e230feb704 100644
@@ -6868,7 +6868,11 @@ static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
 		return ERR_PTR(-EINVAL);
 	}
 
-	refcount_inc(&sqd->refs);
+	if (!refcount_inc_not_zero(&sqd->refs)) {
+		fdput(f);
+		return ERR_PTR(-EINVAL);
+	}
+
 	fdput(f);
 	return sqd;
 }
Signature · 1 line
2.17.2
#1 lore raw
Quoted 13 lines · Xiaoguang Wang

Reading codes finds a possible use after free issue to sqd:

          thread1              |       thread2
==> io_attach_sq_data()        |
===> sqd = ctx_attach->sq_data;|
                               | ==> io_put_sq_data()
                               | ===> refcount_dec_and_test(&sqd->refs)
                               |     If sqd->refs is zero, will free sqd.
                               |
===> refcount_inc(&sqd->refs); |
                               |
                               | ====> kfree(sqd);
===> now use after free to sqd |

IIUC the io_attach_sq_data() is called only during the setup of an io_uring context, before that the fd is returned to the user space.

So, I'm not sure a second thread can call io_put_sq_data() while the first thread is in io_attach_sq_data().

Can you check if this situation can really happen?

Thanks,
Stefano

Quoted 27 lines

Use refcount_inc_not_zero() to fix this issue.

---
 fs/io_uring.c | 6 +++++-
 1 file changed, 5 insertions(+), 1 deletion(-)
fs/io_uring.c +5 -1
index 33b5cf18bb51..48e230feb704 100644
@@ -6868,7 +6868,11 @@ static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
 		return ERR_PTR(-EINVAL);
 	}
 
-	refcount_inc(&sqd->refs);
+	if (!refcount_inc_not_zero(&sqd->refs)) {
+		fdput(f);
+		return ERR_PTR(-EINVAL);
+	}
+
 	fdput(f);
 	return sqd;
 }
Signature · 1 line
2.17.2
#2 lore raw

hi,

Quoted 15 lines · Xiaoguang Wang
Quoted 4 lines

Reading codes finds a possible use after free issue to sqd:
thread1 | thread2
==> io_attach_sq_data() |
===> sqd = ctx_attach->sq_data;|

Quoted 9 lines
                                | ==> io_put_sq_data()
                                | ===> refcount_dec_and_test(&sqd->refs)
                                |     If sqd->refs is zero, will free sqd.
                                |
===> refcount_inc(&sqd->refs); |
                                |
                                | ====> kfree(sqd);

===> now use after free to sqd |

Quoted 2 lines

IIUC the io_attach_sq_data() is called only during the setup of an io_uring context, before that the fd is returned to the user space.

Sorry I didn't make it clear in commit message.
In io_attach_sq_data(), we'll try to attach to a previous io_uring instance indicated by p->wq_fd, this p->wq_fd could be closed independently.

Regards,
Xiaoguang Wang

Quoted 36 lines

So, I'm not sure a second thread can call io_put_sq_data() while the first thread is in io_attach_sq_data().

Can you check if this situation can really happen?

Thanks,
Stefano

Quoted 27 lines

Use refcount_inc_not_zero() to fix this issue.

---
  fs/io_uring.c | 6 +++++-
  1 file changed, 5 insertions(+), 1 deletion(-)
fs/io_uring.c +5 -1
index 33b5cf18bb51..48e230feb704 100644
@@ -6868,7 +6868,11 @@ static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
  		return ERR_PTR(-EINVAL);
  	}
 
-	refcount_inc(&sqd->refs);
+	if (!refcount_inc_not_zero(&sqd->refs)) {
+		fdput(f);
+		return ERR_PTR(-EINVAL);
+	}
+
  	fdput(f);
  	return sqd;
  }
Signature · 1 line
2.17.2
#3 lore raw
Quoted 17 lines · Xiaoguang Wang

hi閿涳拷

Quoted 15 lines · Xiaoguang Wang
Quoted 13 lines

Reading codes finds a possible use after free issue to sqd:
thread1 | thread2
==> io_attach_sq_data() |
===> sqd = ctx_attach->sq_data;|

                                | ==> io_put_sq_data()
                                | ===> refcount_dec_and_test(&sqd->refs)
                                |     If sqd->refs is zero, will free sqd.
                                |
===> refcount_inc(&sqd->refs); |
                                |
                                | ====> kfree(sqd);

===> now use after free to sqd |

Quoted 5 lines
Quoted 2 lines

IIUC the io_attach_sq_data() is called only during the setup of an io_uring context, before that the fd is returned to the user space.

Sorry I didn't make it clear in commit message.
In io_attach_sq_data(), we'll try to attach to a previous io_uring instance indicated by p->wq_fd, this p->wq_fd could be closed independently.

Thanks to clarify! Got it.

Also in this case, IIUC io_put_sq_data() is called only when io_uring_release() is invoked on the previous io_uring instance.

Since we are taking a reference with fdget at the begin of io_attach_sq_data() and we release this reference only at the end of io_attach_sq_data(), can io_put_sq_data() runs in another thread while we have the reference?

Thanks,
Stefano

Quoted 39 lines

Regards,
Xiaoguang Wang

Quoted 35 lines

So, I'm not sure a second thread can call io_put_sq_data() while the first thread is in io_attach_sq_data().

Can you check if this situation can really happen?

Thanks,
Stefano

Quoted 26 lines

Use refcount_inc_not_zero() to fix this issue.

---
  fs/io_uring.c | 6 +++++-
  1 file changed, 5 insertions(+), 1 deletion(-)
fs/io_uring.c +5 -1
index 33b5cf18bb51..48e230feb704 100644
@@ -6868,7 +6868,11 @@ static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
  		return ERR_PTR(-EINVAL);
  	}
-	refcount_inc(&sqd->refs);
+	if (!refcount_inc_not_zero(&sqd->refs)) {
+		fdput(f);
+		return ERR_PTR(-EINVAL);
+	}
+
  	fdput(f);
  	return sqd;
  }
Signature · 1 line
2.17.2
#4 lore raw
Quoted 24 lines · Xiaoguang Wang
Quoted 17 lines

hi閿涳拷

Quoted 15 lines · Xiaoguang Wang
Quoted 13 lines

Reading codes finds a possible use after free issue to sqd:
thread1 | thread2
==> io_attach_sq_data() |
===> sqd = ctx_attach->sq_data;|

                                 | ==> io_put_sq_data()
                                 | ===> refcount_dec_and_test(&sqd->refs)
                                 |     If sqd->refs is zero, will free sqd.
                                 |
===> refcount_inc(&sqd->refs); |
                                 |
                                 | ====> kfree(sqd);

===> now use after free to sqd |

Quoted 5 lines
Quoted 2 lines

IIUC the io_attach_sq_data() is called only during the setup of an io_uring context, before that the fd is returned to the user space.

Sorry I didn't make it clear in commit message.
In io_attach_sq_data(), we'll try to attach to a previous io_uring instance indicated by p->wq_fd, this p->wq_fd could be closed independently.

Quoted 8 lines

Thanks to clarify! Got it.

Also in this case, IIUC io_put_sq_data() is called only when io_uring_release() is invoked on the previous io_uring instance.

Since we are taking a reference with fdget at the begin of io_attach_sq_data() and we release this reference only at the end of io_attach_sq_data(), can io_put_sq_data() runs in another thread while we have the reference?

Oh, right, thanks. Jens, you can ignore this patch, thanks.

Regards,
Xiaoguang Wang

Quoted 43 lines

Thanks,
Stefano

Quoted 39 lines

Regards,
Xiaoguang Wang

Quoted 35 lines

So, I'm not sure a second thread can call io_put_sq_data() while the first thread is in io_attach_sq_data().

Can you check if this situation can really happen?

Thanks,
Stefano

Quoted 26 lines

Use refcount_inc_not_zero() to fix this issue.

---
   fs/io_uring.c | 6 +++++-
   1 file changed, 5 insertions(+), 1 deletion(-)
fs/io_uring.c +5 -1
index 33b5cf18bb51..48e230feb704 100644
@@ -6868,7 +6868,11 @@ static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
   		return ERR_PTR(-EINVAL);
   	}
-	refcount_inc(&sqd->refs);
+	if (!refcount_inc_not_zero(&sqd->refs)) {
+		fdput(f);
+		return ERR_PTR(-EINVAL);
+	}
+
   	fdput(f);
   	return sqd;
   }
Signature · 1 line
2.17.2