Message ID | 20220329154948.10350-1-fmdefrancesco@gmail.com (mailing list archive) |
---|---|
State | Deferred |
Headers | show |
Series | scsi: sd: call device_del() if device_add_disk() fails | expand |
On Tue, Mar 29, 2022 at 05:49:48PM +0200, Fabio M. De Francesco wrote: > In sd_probe(), if device_add_disk() fails it simply calls put_device() > and jumps to the "out" label but the device is never deleted from system. > This leads to a memory leak as reported by Syzbot.[1] > > Fix this bug by calling device_del() soon before put_device() when > device_add_disk() fails. > > [1] [syzbot] memory leak in blk_mq_init_tags > https://lore.kernel.org/lkml/000000000000c341cc05db38c1b0@google.com/ > > Reported-by: syzbot+f08c77040fa163a75a46@syzkaller.appspotmail.com > Suggested-by: Dan Carpenter <dan.carpenter@oracle.com> > Fixes: 2a7a891f4c40 ("scsi: sd: Add error handling support for add_disk()") > Signed-off-by: Fabio M. De Francesco <fmdefrancesco@gmail.com> Thanks! regards, dan carpenter
On Thu, Mar 31, 2022 at 11:26:22AM -0400, 'Wenchao Hao' via syzkaller-bugs wrote: > I do not think it's necessary to call device_del() on this path. If the device > has been added, put_device() would delete it from sysfs. So the origin error > handle is ok with me. > No. The original is buggy and it was detected at runtime by syzbot. It's not static analysis, it is an actual bug found in testing. The device_put() unwinds device_initialize(). The device_del() unwinds device_add(). Take a look at the comments to device_add() or take a look at how device_register/unregister() work. The temptation was to call device_unregister() which is a combined device_del(); device_put(); but when the device_initialize() and device_add() are called separately, then I think it is more readable to call del and put separately as well. regards, dan carpenter
> The temptation was to call device_unregister() which is a combined > device_del(); device_put(); but when the device_initialize() and > device_add() are called separately, then I think it is more readable to > call del and put separately as well. I think we should also consolidate the initialization side. Using device_register and device_unregister would have prevented this bug and I should have switched to that before refactoring the code.
On giovedì? 31 marzo 2022 07:45:12 CEST Christoph Hellwig wrote: > > The temptation was to call device_unregister() which is a combined > > device_del(); device_put(); but when the device_initialize() and > > device_add() are called separately, then I think it is more readable to > > call del and put separately as well. > > I think we should also consolidate the initialization side. Using > device_register and device_unregister would have prevented this bug > and I should have switched to that before refactoring the code. > If I don't misunderstand what you wrote, I think you mean something like the following changes: diff --git a/drivers/scsi/sd.c b/drivers/scsi/sd.c index a390679cf458..7a000a9a9dbe 100644 --- a/drivers/scsi/sd.c +++ b/drivers/scsi/sd.c @@ -3431,7 +3431,7 @@ static int sd_probe(struct device *dev) sdkp->disk_dev.class = &sd_disk_class; dev_set_name(&sdkp->disk_dev, "%s", dev_name(dev)); - error = device_add(&sdkp->disk_dev); + error = device_register(&sdkp->disk_dev); if (error) { put_device(&sdkp->disk_dev); goto out; @@ -3474,7 +3474,7 @@ static int sd_probe(struct device *dev) error = device_add_disk(dev, gd, NULL); if (error) { - put_device(&sdkp->disk_dev); + device_unregister(&sdkp->disk_dev); goto out; } @Dan, @Christoph: what do you think of the changes that I've copy-pasted above? Thanks, Fabio M. De Francesco
On Thu, Mar 31, 2022 at 11:07:45AM +0200, Fabio M. De Francesco wrote: > If I don't misunderstand what you wrote, I think you mean something like > the following changes: > > diff --git a/drivers/scsi/sd.c b/drivers/scsi/sd.c > index a390679cf458..7a000a9a9dbe 100644 > --- a/drivers/scsi/sd.c > +++ b/drivers/scsi/sd.c > @@ -3431,7 +3431,7 @@ static int sd_probe(struct device *dev) > sdkp->disk_dev.class = &sd_disk_class; > dev_set_name(&sdkp->disk_dev, "%s", dev_name(dev)); > > - error = device_add(&sdkp->disk_dev); > + error = device_register(&sdkp->disk_dev); The device_initialize call about also need to go. > if (error) { > put_device(&sdkp->disk_dev); .. and this put_device > goto out; > @@ -3474,7 +3474,7 @@ static int sd_probe(struct device *dev) > > error = device_add_disk(dev, gd, NULL); > if (error) { > - put_device(&sdkp->disk_dev); > + device_unregister(&sdkp->disk_dev); > goto out; > } .. and then the cleanup patch would need the same logic. But thinking about it I don't think we actually can do that due to the split unregistration. So I take my suggestion back.
On gioved? 31 marzo 2022 11:13:27 CEST Christoph Hellwig wrote: > > .. and then the cleanup patch would need the same logic. But thinking > about it I don't think we actually can do that due to the split > unregistration. So I take my suggestion back. > If I understand correctly, after thinking about it some more, you decided to withdraw your own suggestion. Dan had already approved this patch. Therefore, I'll leave the patch as it is now and wait for someone to place a "Reviewed-by" tag and Maintainers to merge (if, in the meantime, nobody else require further changes). Thanks, Fabio
On 2022/3/31 13:41, Dan Carpenter wrote: > On Thu, Mar 31, 2022 at 11:26:22AM -0400, 'Wenchao Hao' via syzkaller-bugs wrote: >> I do not think it's necessary to call device_del() on this path. If the device >> has been added, put_device() would delete it from sysfs. So the origin error >> handle is ok with me. >> > > No. The original is buggy and it was detected at runtime by syzbot. > It's not static analysis, it is an actual bug found in testing. > Yes, it's a bug, but the root reason is not we forget to call device_del(sdkp->disk_dev). It's because we did not cleanup gendisk. The leak memory is allocated in elevator_init_mq(), we should clean this memory via blk_cleanup_queue(). I summit a patch which would fix this memory leak: https://lore.kernel.org/linux-scsi/20220401011018.1026553-1-haowenchao@huawei.com/T/#u > The device_put() unwinds device_initialize(). The device_del() unwinds > device_add(). Take a look at the comments to device_add() or take a > look at how device_register/unregister() work. > You may read the implement of put_device(), it is based on kobj_xxx. If the kobj is still in sysfs, a cleanup would be performed. And device_del() seems would not decrease the reference count of kobj, the main aim is to make it invisibleto sysfs.
Wenchao Hao, what you're saying makes a lot of sense but it raises a lot of questions in turn. Fabio, did you test your patch? This is another reason why syzbot should display the whole dmesg because otherwise we can't ask people link to their test results if it just says "PASSED" with no additional information. If syzbot provided a dmesg at the end then I would require a link to it under the --- cut off line for patches that I review. In a way this gets back to the original testing that syzbot did do. If Wenchao's reading of the code is correct the Fabio's patch caused a series of use after free bugs but because the test results just said "PASSED" with no additional information. Either way, failing to call device_del() is still a bug. Also, I don't really understand why we don't have to call put_device(&sdkp->disk_dev) at the end of sd_remove(). regards, dan carpenter
On Thu, 2022-03-31 at 16:42 +0300, Dan Carpenter wrote: [...] > Also, I don't really understand why we don't have to call > put_device(&sdkp->disk_dev) at the end of sd_remove(). That's because the final put is done by the gendisk ->free_disk() function which is scsi_disk_free_disk(). Most of the gendisk functions we provide convert a gendisk to a scsi_disk (via the gendisk private_data), so the sdkp has to live as long as the gendisk. James
On Thu, Mar 31, 2022 at 10:19:57AM -0400, James Bottomley wrote: > On Thu, 2022-03-31 at 16:42 +0300, Dan Carpenter wrote: > [...] > > Also, I don't really understand why we don't have to call > > put_device(&sdkp->disk_dev) at the end of sd_remove(). > > That's because the final put is done by the gendisk ->free_disk() > function which is scsi_disk_free_disk(). Most of the gendisk functions > we provide convert a gendisk to a scsi_disk (via the gendisk > private_data), so the sdkp has to live as long as the gendisk. > > James Thanks James. Ah... And scsi_disk_free_disk() will not be called unless device_add_disk() succeeds. So Wenchao Hao's patch is correct. And after that there is just a missing device_del() and also I see another problem where if device_add() fails then we need to call put_disk(gd); on that error path. regards, dan carpenter
I do not think it's necessary to call device_del() on this path. If the device has been added, put_device() would delete it from sysfs. So the origin error handle is ok with me.
On gioved? 31 marzo 2022 15:42:10 CEST Dan Carpenter wrote: > Wenchao Hao, what you're saying makes a lot of sense but it raises a lot > of questions in turn. > > Fabio, did you test your patch? Yes, I did, Dan. I tested it the usual way with the "#syz test:" command. Obviously I have not the hardware to test code on it. Therefore, the messages that say "Syzbot tested the patch and it didn't trigger any issue" is all that I can know about the code being good or not. This is the criterion I've always used before sending patches for Syzbot's reports. However, my knowledge of these subsystems and the API that are related to this bug were very little, but now I can say that during the last couple of days it has improved to the point where I can affirm that Wenchao's patch seems to me to be the only correct solution. Thanks for all the help you and the the other developers provided. It was invaluable for a better understanding of this matter. Regards, Fabio M. De Francesco
On Thu, Mar 31, 2022 at 06:14:27PM +0200, Fabio M. De Francesco wrote: > On gioved? 31 marzo 2022 15:42:10 CEST Dan Carpenter wrote: > > Wenchao Hao, what you're saying makes a lot of sense but it raises a lot > > of questions in turn. > > > > Fabio, did you test your patch? > > Yes, I did, Dan. I tested it the usual way with the "#syz test:" command. > Obviously I have not the hardware to test code on it. > Yeah. What a nightmare. You posted a link to the first test. It said passed but definitely introduced some use after frees but how was anyone supposed to know? No way we would have figured this out. I'm working to make Smatch understand device_put() better but this one is way difficult. Sorry that you went through this. regards, dan carpenter
On gioved? 31 marzo 2022 18:24:16 CEST Dan Carpenter wrote: > On Thu, Mar 31, 2022 at 06:14:27PM +0200, Fabio M. De Francesco wrote: > > On gioved? 31 marzo 2022 15:42:10 CEST Dan Carpenter wrote: > > > Wenchao Hao, what you're saying makes a lot of sense but it raises a lot > > > of questions in turn. > > > > > > Fabio, did you test your patch? > > > > Yes, I did, Dan. I tested it the usual way with the "#syz test:" command. > > Obviously I have not the hardware to test code on it. > > > > Yeah. What a nightmare. You posted a link to the first test. It said > passed but definitely introduced some use after frees but how was anyone > supposed to know? Maybe that a "spare-time Linux developer" like me should leave these kinds of bug fixes to more experienced people. But we should also note that I tried two or three different patches and _all_ of them passed the tests. > > No way we would have figured this out. I think that something should change about the way Syzbot tests patches and about how it provides the results. The other four or five bugs that I have fixed were based mainly to the fact that they passed the Syzbot tests. Perhaps I've been lucky but my patches were good and they were merged. However, I began to trust Syzbot too much. This is not how I should approach and try to solve bugs. > I'm working to make Smatch > understand device_put() better but this one is way difficult. > > Sorry that you went through this. Please don't be sorry :) Believe me when I say that I cannot explain how many things I have learned during these days while working on this issue. I see no problems at all but only opportunities for learning. Thank you very much! Fabio M. De Francesco > > regards, > dan carpenter > >
diff --git a/drivers/scsi/sd.c b/drivers/scsi/sd.c index a390679cf458..13d96d0f9dde 100644 --- a/drivers/scsi/sd.c +++ b/drivers/scsi/sd.c @@ -3474,6 +3474,7 @@ static int sd_probe(struct device *dev) error = device_add_disk(dev, gd, NULL); if (error) { + device_del(&sdkp->disk_dev); put_device(&sdkp->disk_dev); goto out; }
In sd_probe(), if device_add_disk() fails it simply calls put_device() and jumps to the "out" label but the device is never deleted from system. This leads to a memory leak as reported by Syzbot.[1] Fix this bug by calling device_del() soon before put_device() when device_add_disk() fails. [1] [syzbot] memory leak in blk_mq_init_tags https://lore.kernel.org/lkml/000000000000c341cc05db38c1b0@google.com/ Reported-by: syzbot+f08c77040fa163a75a46@syzkaller.appspotmail.com Suggested-by: Dan Carpenter <dan.carpenter@oracle.com> Fixes: 2a7a891f4c40 ("scsi: sd: Add error handling support for add_disk()") Signed-off-by: Fabio M. De Francesco <fmdefrancesco@gmail.com> --- This patch replace the previous attempt to fix the bug reported by Syzbot. Therefore, the previous wrong patch at https://lore.kernel.org/lkml/20220328084452.11479-1-fmdefrancesco@gmail.com/ must be discarded. Many thanks to Dan Carpenter. drivers/scsi/sd.c | 1 + 1 file changed, 1 insertion(+)