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>] [day] [month] [year] [list]
Date:	Tue, 24 Nov 2015 17:10:07 -0800
From:	Jin Qian <jinqian@...roid.com>
To:	Greg Hackmann <ghackmann@...gle.com>,
	Greg Kroah-Hartman <gregkh@...uxfoundation.org>,
	Jason Hu <jia-cheng.hu@...el.com>,
	Joe Perches <joe@...ches.com>, Yu Ning <yu.ning@...el.com>,
	Peter Senna Tschudin <peter.senna@...il.com>,
	Christoffer Dall <christoffer.dall@...aro.org>,
	Alex Bennée <alex.bennee@...aro.org>,
	linux-kernel@...r.kernel.org
Cc:	Jin Qian <jinqian@...roid.com>
Subject: [PATCH 3/8] android_pipe: Pin pages to memory while copying and other cleanups

From: Christoffer Dall <christoffer.dall@...aro.org>

The existing code had a troubling TODO statement concerning the fact
that it just did a check if the page that the QEMU backend was going to
read from / write to was there before the call to the QEMU backend and
then relying on the fact that the page stayed around, even in a
preemptible SMP kernel.  Obviously the page could go away or be
reassigned, and strange things may happen.

Further, writes were not tracked, so any use of COW or KSM-like
features would break completely.  Probably that was never used by adbd
(the only current active user of the pipe), but could prove much more
dangerous for the GPU passthrough mechanism.

Instead, use get_user_pages() as the comment suggested and cleanup the
error path and add the set_page_dirt() call on a successful read
operation.

Also clarify the count used to return from successful read/write calls
and use Linux style commentary in various places of the file.

Note: The "just ignore error and return whatever we read so far" error
handling is really quite horrific.  I cannot change it without a more
careful study of all user space ABIs reliance on this 'feature'.

Signed-off-by: Christoffer Dall <christoffer.dall@...aro.org>
(cherry picked from commit ca8dafc623c6956fad77df80540a1e31f5f8fe2b)
Signed-off-by: Jin Qian <jinqian@...roid.com>
---
 drivers/platform/goldfish/goldfish_pipe.c | 129 +++++++++++++++++-------------
 1 file changed, 72 insertions(+), 57 deletions(-)

diff --git a/drivers/platform/goldfish/goldfish_pipe.c b/drivers/platform/goldfish/goldfish_pipe.c
index 0fb3a34..20a9337 100644
--- a/drivers/platform/goldfish/goldfish_pipe.c
+++ b/drivers/platform/goldfish/goldfish_pipe.c
@@ -2,6 +2,7 @@
  * Copyright (C) 2011 Google, Inc.
  * Copyright (C) 2012 Intel, Inc.
  * Copyright (C) 2013 Intel, Inc.
+ * Copyright (C) 2014 Linaro Limited
  *
  * This software is licensed under the terms of the GNU General Public
  * License version 2, as published by the Free Software Foundation, and
@@ -57,6 +58,7 @@
 #include <linux/slab.h>
 #include <linux/io.h>
 #include <linux/goldfish.h>
+#include <linux/mm.h>
 
 /*
  * IMPORTANT: The following constants must match the ones used and defined
@@ -257,17 +259,14 @@ static int access_with_param(struct goldfish_pipe_dev *dev, const int cmd,
 	return 0;
 }
 
-/* This function is used for both reading from and writing to a given
- * pipe.
- */
 static ssize_t goldfish_pipe_read_write(struct file *filp, char __user *buffer,
-				    size_t bufflen, int is_write)
+				       size_t bufflen, int is_write)
 {
 	unsigned long irq_flags;
 	struct goldfish_pipe *pipe = filp->private_data;
 	struct goldfish_pipe_dev *dev = pipe->dev;
 	unsigned long address, address_end;
-	int ret = 0;
+	int count = 0, ret = -EINVAL;
 
 	/* If the emulator already closed the pipe, no need to go further */
 	if (test_bit(BIT_CLOSED_ON_HOST, &pipe->flags))
@@ -290,30 +289,23 @@ static ssize_t goldfish_pipe_read_write(struct file *filp, char __user *buffer,
 	address_end = address + bufflen;
 
 	while (address < address_end) {
-		unsigned long  page_end = (address & PAGE_MASK) + PAGE_SIZE;
-		unsigned long  next     = page_end < address_end ? page_end
-								 : address_end;
-		unsigned long  avail    = next - address;
+		unsigned long page_end = (address & PAGE_MASK) + PAGE_SIZE;
+		unsigned long next     = page_end < address_end ? page_end
+								: address_end;
+		unsigned long avail    = next - address;
 		int status, wakeBit;
-
-		/* Ensure that the corresponding page is properly mapped */
-		/* FIXME: this isn't safe or sufficient - use get_user_pages */
-		if (is_write) {
-			char c;
-			/* Ensure that the page is mapped and readable */
-			if (__get_user(c, (char __user *)address)) {
-				if (!ret)
-					ret = -EFAULT;
-				break;
-			}
-		} else {
-			/* Ensure that the page is mapped and writable */
-			if (__put_user(0, (char __user *)address)) {
-				if (!ret)
-					ret = -EFAULT;
-				break;
-			}
-		}
+		struct page *page;
+
+		/*
+		 * We grab the pages on a page-by-page basis in case user
+		 * space gives us a potentially huge buffer but the read only
+		 * returns a small amount, then there's no need to pin that
+		 * much memory to the process.
+		 */
+		ret = get_user_pages(current, current->mm, address, 1,
+				     !is_write, 0, &page, NULL);
+		if (ret < 0)
+			return ret;
 
 		/* Now, try to transfer the bytes in the current page */
 		spin_lock_irqsave(&dev->lock, irq_flags);
@@ -332,33 +324,48 @@ static ssize_t goldfish_pipe_read_write(struct file *filp, char __user *buffer,
 		}
 		spin_unlock_irqrestore(&dev->lock, irq_flags);
 
+		if (status > 0 && !is_write)
+			set_page_dirty(page);
+		put_page(page);
+
 		if (status > 0) { /* Correct transfer */
-			ret += status;
+			count += status;
 			address += status;
 			continue;
-		}
-
-		if (status == 0)  /* EOF */
+		} else if (status == 0) { /* EOF */
+			ret = 0;
 			break;
-
-		/* An error occured. If we already transfered stuff, just
-		* return with its count. We expect the next call to return
-		* an error code */
-		if (ret > 0)
+		} else if (status < 0 && count > 0) {
+			/*
+			 * An error occurred and we already transferred
+			 * something on one of the previous pages.
+			 * Just return what we already copied and log this
+			 * err.
+			 *
+			 * Note: This seems like an incorrect approach but
+			 * cannot change it until we check if any user space
+			 * ABI relies on this behavior.
+			 */
+			pr_info_ratelimited("android_pipe: backend returned error %d on %s\n",
+					status, is_write ? "write" : "read");
+			ret = 0;
 			break;
+		}
 
-		/* If the error is not PIPE_ERROR_AGAIN, or if we are not in
-		* non-blocking mode, just return the error code.
-		*/
+		/*
+		 * If the error is not PIPE_ERROR_AGAIN, or if we are not in
+		 * non-blocking mode, just return the error code.
+		 */
 		if (status != PIPE_ERROR_AGAIN ||
 			(filp->f_flags & O_NONBLOCK) != 0) {
 			ret = goldfish_pipe_error_convert(status);
 			break;
 		}
 
-		/* We will have to wait until more data/space is available.
-		* First, mark the pipe as waiting for a specific wake signal.
-		*/
+		/*
+		 * The backend blocked the read/write, wait until the backend
+		 * tells us it's ready to process more data.
+		 */
 		wakeBit = is_write ? BIT_WAKE_ON_WRITE : BIT_WAKE_ON_READ;
 		set_bit(wakeBit, &pipe->flags);
 
@@ -372,22 +379,29 @@ static ssize_t goldfish_pipe_read_write(struct file *filp, char __user *buffer,
 		while (test_bit(wakeBit, &pipe->flags)) {
 			if (wait_event_interruptible(
 					pipe->wake_queue,
-					!test_bit(wakeBit, &pipe->flags)))
-				return -ERESTARTSYS;
+					!test_bit(wakeBit, &pipe->flags))) {
+				ret = -ERESTARTSYS;
+				break;
+			}
 
-			if (test_bit(BIT_CLOSED_ON_HOST, &pipe->flags))
-				return -EIO;
+			if (test_bit(BIT_CLOSED_ON_HOST, &pipe->flags)) {
+				ret = -EIO;
+				break;
+			}
 		}
 
 		/* Try to re-acquire the lock */
-		if (mutex_lock_interruptible(&pipe->lock))
-			return -ERESTARTSYS;
-
-		/* Try the transfer again */
-		continue;
+		if (mutex_lock_interruptible(&pipe->lock)) {
+			ret = -ERESTARTSYS;
+			break;
+		}
 	}
 	mutex_unlock(&pipe->lock);
-	return ret;
+
+	if (ret < 0)
+		return ret;
+	else
+		return count;
 }
 
 static ssize_t goldfish_pipe_read(struct file *filp, char __user *buffer,
@@ -440,10 +454,11 @@ static irqreturn_t goldfish_pipe_interrupt(int irq, void *dev_id)
 	unsigned long irq_flags;
 	int count = 0;
 
-	/* We're going to read from the emulator a list of (channel,flags)
-	* pairs corresponding to the wake events that occured on each
-	* blocked pipe (i.e. channel).
-	*/
+	/*
+	 * We're going to read from the emulator a list of (channel,flags)
+	 * pairs corresponding to the wake events that occurred on each
+	 * blocked pipe (i.e. channel).
+	 */
 	spin_lock_irqsave(&dev->lock, irq_flags);
 	for (;;) {
 		/* First read the channel, 0 means the end of the list */
-- 
2.6.0.rc2.230.g3dd15c0

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

Powered by blists - more mailing lists