usb: mon: Fix a deadlock in usbmon between mmap and read
PATCH 1/1 usb: mon: Fix a deadlock in usbmon between mmap and read
The problem arises because our read() function grabs a lock of the circular buffer, finds something of interest, then invokes copy_to_user() straight from the buffer, which in turn takes mm->mmap_sem. In the same time, the callback mon_bin_vma_fault() is invoked under mm->mmap_sem. It attempts to take the fetch lock and deadlocks.
This patch does away with protecting of our page list with any semaphores, and instead relies on the kernel not close the device while mmap is active in a process.
In addition, we prohibit re-sizing of a buffer while mmap is active. This way, when (now unlocked) fault is processed, it works with the page that is intended to be mapped-in, and not some other random page. Note that this may have an ABI impact, but hopefully no legitimate program is this wrong.
--- drivers/usb/mon/mon_bin.c | 32 +++++++++++++++++++++----------- 1 file changed, 21 insertions(+), 11 deletions(-)
drivers/usb/mon/mon_bin.c +21 -11
@@ -1039,12 +1039,18 @@ static long mon_bin_ioctl(struct file *file, unsigned int cmd, unsigned long argmutex_lock (&rp->fetch_lock);spin_lock_irqsave (&rp->b_lock, flags); -mon_free_buff (rp->b_vec, rp->b_size/CHUNK_SIZE ); -kfree (rp->b_vec); - rp->b_vec = vec; - rp->b_size = size; - rp->b_read = rp->b_in = rp->b_out = rp->b_cnt =0 ; - rp->cnt_lost =0 ; +if (rp->mmap_active) { +mon_free_buff (vec, size/CHUNK_SIZE ); +kfree (vec); + ret = -EBUSY ; + }else { +mon_free_buff (rp->b_vec, rp->b_size/CHUNK_SIZE ); +kfree (rp->b_vec); + rp->b_vec = vec; + rp->b_size = size; + rp->b_read = rp->b_in = rp->b_out = rp->b_cnt =0 ; + rp->cnt_lost =0 ; + }spin_unlock_irqrestore (&rp->b_lock, flags);mutex_unlock (&rp->fetch_lock); }
@@ -1216,13 +1222,21 @@ mon_bin_poll(struct file *file, struct poll_table_struct *wait)static void mon_bin_vma_open (struct vm_area_struct *vma) {struct mon_reader_bin *rp = vma->vm_private_data; +unsigned long flags; + +spin_lock_irqsave (&rp->b_lock, flags); rp->mmap_active++; +spin_unlock_irqrestore (&rp->b_lock, flags); }static void mon_bin_vma_close (struct vm_area_struct *vma) { +unsigned long flags; +struct mon_reader_bin *rp = vma->vm_private_data; +spin_lock_irqsave (&rp->b_lock, flags); rp->mmap_active--; +spin_unlock_irqrestore (&rp->b_lock, flags); }/*
@@ -1234,16 +1248,12 @@ static vm_fault_t mon_bin_vma_fault(struct vm_fault *vmf)unsigned long offset, chunk_idx;struct page *pageptr; -mutex_lock (&rp->fetch_lock); offset = vmf->pgoff <<PAGE_SHIFT ; -if (offset >= rp->b_size) { -mutex_unlock (&rp->fetch_lock); +if (offset >= rp->b_size)return VM_FAULT_SIGBUS ; - } chunk_idx = offset /CHUNK_SIZE ; pageptr = rp->b_vec[chunk_idx].pg;get_page (pageptr); -mutex_unlock (&rp->fetch_lock); vmf->page = pageptr;return 0 ; }
Quoted 10 lines · Pete Zaitcev
The problem arises because our
read()function grabs a lock of the circular buffer, finds something of interest, then invokescopy_to_user()straight from the buffer, which in turn takesmm->mmap_sem. In the same time, the callbackmon_bin_vma_fault()is invoked undermm->mmap_sem. It attempts to take the fetch lock and deadlocks.This patch does away with protecting of our page list with any semaphores, and instead relies on the kernel not close the device while mmap is active in a process.
Quoted 8 lines
In addition, we prohibit re-sizing of a buffer while mmap is active. This way, when (now unlocked) fault is processed, it works with the page that is intended to be mapped-in, and not some other random page. Note that this may have an ABI impact, but hopefully no legitimate program is this wrong.
Signed-off-by Pete Zaitcev <[email protected]>Reported-by [email protected]
Also this should have:
Alan Stern
Quoted 4 lines · On Tue, 26 Nov 2019 10:20:14 -0500 (EST) Alan Stern…
Quoted 2 lines
Signed-off-by Pete Zaitcev <[email protected]>Reported-by [email protected]Reviewed-by Alan Stern <[email protected]>
Thanks.
Quoted 1 line
Fixes 46eb14a6e158 ("USB: fix usbmon BUG trigger")
Indeed... Either I didn't think that one through, or the copy_to_user used not to take the mmap_sem.
Quoted 1 line
CC <[email protected]>
Do we really need this? The problem was in the code for more than 10 years. It's not like anyone is exploiting systems because of it.
If we do need it, I should cc: the submission to the same place too, right?
-- Pete
Quoted 12 lines · Pete Zaitcev
Quoted 4 lines · On Tue, 26 Nov 2019 10:20:14 -0500 (EST) Alan Stern…
Quoted 2 lines
Signed-off-by Pete Zaitcev <[email protected]>Reported-by [email protected]Reviewed-by Alan Stern <[email protected]>Thanks.
Quoted 1 line
Fixes 46eb14a6e158 ("USB: fix usbmon BUG trigger")
Quoted 7 lines
Indeed... Either I didn't think that one through, or the
copy_to_userused not to take themmap_sem.Quoted 1 line
CC <[email protected]>Do we really need this? The problem was in the code for more than 10 years. It's not like anyone is exploiting systems because of it.
Well now we all have a simple reproducer for it, so yes, it should be backported. I'm doing that for all of the syzbot stuff.
Quoted 1 line
If we do need it, I should cc: the submission to the same place too, right?
Nope, the tag is just fine, that's all that is needed. I'll add the above to the patch when applying it to my trees.
thanks,
greg k-h
Quoted 6 lines · Pete Zaitcev
Quoted 3 lines · On Tue, 26 Nov 2019 10:20:14 -0500 (EST) Alan Stern…
Quoted 2 lines
Signed-off-by Pete Zaitcev <[email protected]>Reported-by [email protected]
Quoted 8 lines
Quoted 1 line
Reviewed-by Alan Stern <[email protected]>Thanks.
Quoted 1 line
Fixes 46eb14a6e158 ("USB: fix usbmon BUG trigger")Indeed... Either I didn't think that one through, or the
copy_to_userused not to take themmap_sem.
copy_to_user doesn't, but the fault handler does (the core handler, not the fault routine in mon_bin.c). After all, it doesn't want the memory map to change while a page is being read in to satisfy the fault.
Alan Stern