Re: [PATCH] [v2] nvdimm-btt: fix several memleaks
> dinghao.liu@ wrote: > > > Ira Weiny wrote: > > > > Dinghao Liu wrote: [snip] > > > > > > > > This does not quite work. > > > > > > > > free_arenas() is used in the error paths of create_arenas() and > > > > discover_arenas(). In those cases devm_kfree() is probably a better way > > > > to clean up this. > > > > Here I'm a little confused about when devm_kfree() should be used. > > Over all it should be used whenever memory is allocated for the lifetime > of the device. > > > Code in btt_init() implies that resources allocated by devm_* could be > > auto freed in both error and success paths of probe/attach (e.g., btt > > allocated by devm_kzalloc is never freed by devm_kfree). > > Using devm_kfree() in free_arenas() is ok for me, but I want to know > > whether not using devm_kfree() constitutes a bug. > > Unfortunately I'm not familiar enough with this code to know for sure. > > After my quick checks before I thought it was. But each time I look at it > I get confused. This is why I was thinking maybe not using devm_*() and > using no_free_ptr() may be a better solution to make sure things don't > leak without changing the success path (which is likely working fine > because no bugs have been found.) We have the same confusion here... I find a discussion about this problem, which implies that not using devm_kfree() may delay the release, but the memory will be freed later and no memory is leaked: https://www.mail-archive.com/linux-kernel@vger.kernel.org/msg2009561.html > > > > > We might want to look at using no_free_ptr() in this code. See this > > > patch[1] for an example of how to inhibit the cleanup and pass the > > > pointer on when the function succeeds. > > > > > > [1] > > > https://lore.kernel.org/all/170261791914.1714654.6447680285357545638.st...@dwillia2-xfh.jf.intel.com/ > > > > > > Ira > > > > Thanks for this example. But it seems that no_free_ptr() is used to > > handle the scope based resource management. Changes in this patch does > > not introduce this feature. Do I understand this correctly? > > You do understand but I was thinking that perhaps using no_free_ptr() > rather than devm_*() might be an easier way to fix this bug without trying > to decode the lifetime of everything. > My concern is that no_free_ptr() may not be able to completely fix all memleaks because some of them are triggered in (part of) success paths (e.g., when btt_freelist_init succeeds but btt_rtt_init fails, discover_arenas still needs to clean up the memory allocated in btt_freelist_init). I checked the design of no_free_ptr(), and it seems that it will generate a new pointer on success and the memory still leaks in the above case. Therefore, I think using devm_*() is still the best solution for this bug. Regards, Dinghao
Re: [PATCH] [v2] nvdimm-btt: fix several memleaks
dinghao.liu@ wrote:
> > Ira Weiny wrote:
> > > Dinghao Liu wrote:
> >
> > [snip]
> >
> > -static int btt_freelist_init(struct arena_info *arena)
> > +static int btt_freelist_init(struct device *dev, struct arena_info *arena)
> >
> > Both struct arena_info and struct btt contain references to struct nd_btt
> > which is the device you are passing down this call stack.
> >
> > Just use the device in the arena/btt rather than passing a device
> > structure. That makes the code easier to read and ensures that the device
> > associated with this arena or btt is used.
>
> Thanks for this suggestion! I will fix this in the v3 patch.
>
> > [snip]
> > > >
> > > > -static int btt_maplocks_init(struct arena_info *arena)
> > > > +static int btt_maplocks_init(struct device *dev, struct arena_info
> > > > *arena)
> > > > {
> > > > u32 i;
> > > >
> > > > - arena->map_locks = kcalloc(arena->nfree, sizeof(struct
> > > > aligned_lock),
> > > > + arena->map_locks = devm_kcalloc(dev, arena->nfree,
> > > > sizeof(struct aligned_lock),
> > > > GFP_KERNEL);
> > > > if (!arena->map_locks)
> > > > return -ENOMEM;
> > > > @@ -805,9 +805,6 @@ static void free_arenas(struct btt *btt)
> > > >
> > > > list_for_each_entry_safe(arena, next, >arena_list, list) {
> > > > list_del(>list);
> > > > - kfree(arena->rtt);
> > > > - kfree(arena->map_locks);
> > > > - kfree(arena->freelist);
> > >
> > > This does not quite work.
> > >
> > > free_arenas() is used in the error paths of create_arenas() and
> > > discover_arenas(). In those cases devm_kfree() is probably a better way
> > > to clean up this.
>
> Here I'm a little confused about when devm_kfree() should be used.
Over all it should be used whenever memory is allocated for the lifetime
of the device.
> Code in btt_init() implies that resources allocated by devm_* could be
> auto freed in both error and success paths of probe/attach (e.g., btt
> allocated by devm_kzalloc is never freed by devm_kfree).
> Using devm_kfree() in free_arenas() is ok for me, but I want to know
> whether not using devm_kfree() constitutes a bug.
Unfortunately I'm not familiar enough with this code to know for sure.
After my quick checks before I thought it was. But each time I look at it
I get confused. This is why I was thinking maybe not using devm_*() and
using no_free_ptr() may be a better solution to make sure things don't
leak without changing the success path (which is likely working fine
because no bugs have been found.)
>
> > >
> > > However...
> > >
> > > > debugfs_remove_recursive(arena->debugfs_dir);
> > > > kfree(arena);
> > >
> > > Why can't arena be allocated with devm_*?
> > >
> > > We need to change this up a bit more to handle the error path vs regular
> > > device shutdown free (automatic devm frees).
> >
>
> At first, I think the use of arena is correct. Therefore, allocating arena
> with devm_* should be a code style optimization. However, I rechecked
> discover_arenas and found arena might also be leaked (e.g., if
> alloc_arena() fails in the second loop, the previously allocated
> resources in the loop is leaked). The correct code could be found in
> create_arenas(), where free_arenas is called on failure of
> alloc_arena().
Yea I've not reached that level of detail in my analysis.
>
> To fix this issue, I think it's better to change the 'goto out_super'
> tag to 'goto out'. We could also use devm_* for arena to simplify the
> error path in discover_arenas().
I think it is your call at this point as I don't have time to dig in more
than I have. Sorry.
>
> > We might want to look at using no_free_ptr() in this code. See this
> > patch[1] for an example of how to inhibit the cleanup and pass the
> > pointer on when the function succeeds.
> >
> > [1]
> > https://lore.kernel.org/all/170261791914.1714654.6447680285357545638.st...@dwillia2-xfh.jf.intel.com/
> >
> > Ira
>
> Thanks for this example. But it seems that no_free_ptr() is used to
> handle the scope based resource management. Changes in this patch does
> not introduce this feature. Do I understand this correctly?
You do understand but I was thinking that perhaps using no_free_ptr()
rather than devm_*() might be an easier way to fix this bug without trying
to decode the lifetime of everything.
Ira
>
> Regards,
> Dinghao
Re: [PATCH] [v2] nvdimm-btt: fix several memleaks
> Ira Weiny wrote:
> > Dinghao Liu wrote:
>
> [snip]
>
> -static int btt_freelist_init(struct arena_info *arena)
> +static int btt_freelist_init(struct device *dev, struct arena_info *arena)
>
> Both struct arena_info and struct btt contain references to struct nd_btt
> which is the device you are passing down this call stack.
>
> Just use the device in the arena/btt rather than passing a device
> structure. That makes the code easier to read and ensures that the device
> associated with this arena or btt is used.
Thanks for this suggestion! I will fix this in the v3 patch.
> [snip]
> > >
> > > -static int btt_maplocks_init(struct arena_info *arena)
> > > +static int btt_maplocks_init(struct device *dev, struct arena_info
> > > *arena)
> > > {
> > > u32 i;
> > >
> > > - arena->map_locks = kcalloc(arena->nfree, sizeof(struct aligned_lock),
> > > + arena->map_locks = devm_kcalloc(dev, arena->nfree, sizeof(struct
> > > aligned_lock),
> > > GFP_KERNEL);
> > > if (!arena->map_locks)
> > > return -ENOMEM;
> > > @@ -805,9 +805,6 @@ static void free_arenas(struct btt *btt)
> > >
> > > list_for_each_entry_safe(arena, next, >arena_list, list) {
> > > list_del(>list);
> > > - kfree(arena->rtt);
> > > - kfree(arena->map_locks);
> > > - kfree(arena->freelist);
> >
> > This does not quite work.
> >
> > free_arenas() is used in the error paths of create_arenas() and
> > discover_arenas(). In those cases devm_kfree() is probably a better way
> > to clean up this.
Here I'm a little confused about when devm_kfree() should be used.
Code in btt_init() implies that resources allocated by devm_* could be
auto freed in both error and success paths of probe/attach (e.g., btt
allocated by devm_kzalloc is never freed by devm_kfree).
Using devm_kfree() in free_arenas() is ok for me, but I want to know
whether not using devm_kfree() constitutes a bug.
> >
> > However...
> >
> > > debugfs_remove_recursive(arena->debugfs_dir);
> > > kfree(arena);
> >
> > Why can't arena be allocated with devm_*?
> >
> > We need to change this up a bit more to handle the error path vs regular
> > device shutdown free (automatic devm frees).
>
At first, I think the use of arena is correct. Therefore, allocating arena
with devm_* should be a code style optimization. However, I rechecked
discover_arenas and found arena might also be leaked (e.g., if alloc_arena()
fails in the second loop, the previously allocated resources in the loop is
leaked). The correct code could be found in create_arenas(), where free_arenas
is called on failure of alloc_arena().
To fix this issue, I think it's better to change the 'goto out_super' tag
to 'goto out'. We could also use devm_* for arena to simplify the error path
in discover_arenas().
> We might want to look at using no_free_ptr() in this code. See this
> patch[1] for an example of how to inhibit the cleanup and pass the pointer
> on when the function succeeds.
>
> [1]
> https://lore.kernel.org/all/170261791914.1714654.6447680285357545638.st...@dwillia2-xfh.jf.intel.com/
>
> Ira
Thanks for this example. But it seems that no_free_ptr() is used to handle
the scope based resource management. Changes in this patch does not introduce
this feature. Do I understand this correctly?
Regards,
Dinghao
Re: [PATCH] [v2] nvdimm-btt: fix several memleaks
Ira Weiny wrote:
> Dinghao Liu wrote:
[snip]
> >
> > -static int btt_maplocks_init(struct arena_info *arena)
> > +static int btt_maplocks_init(struct device *dev, struct arena_info *arena)
> > {
> > u32 i;
> >
> > - arena->map_locks = kcalloc(arena->nfree, sizeof(struct aligned_lock),
> > + arena->map_locks = devm_kcalloc(dev, arena->nfree, sizeof(struct
> > aligned_lock),
> > GFP_KERNEL);
> > if (!arena->map_locks)
> > return -ENOMEM;
> > @@ -805,9 +805,6 @@ static void free_arenas(struct btt *btt)
> >
> > list_for_each_entry_safe(arena, next, >arena_list, list) {
> > list_del(>list);
> > - kfree(arena->rtt);
> > - kfree(arena->map_locks);
> > - kfree(arena->freelist);
>
> This does not quite work.
>
> free_arenas() is used in the error paths of create_arenas() and
> discover_arenas(). In those cases devm_kfree() is probably a better way
> to clean up this.
>
> However...
>
> > debugfs_remove_recursive(arena->debugfs_dir);
> > kfree(arena);
>
> Why can't arena be allocated with devm_*?
>
> We need to change this up a bit more to handle the error path vs regular
> device shutdown free (automatic devm frees).
We might want to look at using no_free_ptr() in this code. See this
patch[1] for an example of how to inhibit the cleanup and pass the pointer
on when the function succeeds.
[1]
https://lore.kernel.org/all/170261791914.1714654.6447680285357545638.st...@dwillia2-xfh.jf.intel.com/
Ira
Re: [PATCH] [v2] nvdimm-btt: fix several memleaks
Dinghao Liu wrote:
> Resources allocated by kcalloc() in btt_freelist_init(),
> btt_rtt_init(), and btt_maplocks_init() are not correctly
> released in their callers when an error happens. For
> example, when an error happens in btt_freelist_init(), its
> caller discover_arenas() will directly free arena, which makes
> arena->freelist a leaked memory. Fix these memleaks by using
> devm_kcalloc() to make the memory auto-freed on driver detach.
Thanks for this work!
>
> Fixes: 5212e11fde4d ("nd_btt: atomic sector updates")
> Signed-off-by: Dinghao Liu
> ---
>
> Changelog:
>
> v2: -Use devm_kcalloc() to fix the memleaks.
> -Fix the potential leaked memory in btt_rtt_init()
> and btt_maplocks_init().
> ---
> drivers/nvdimm/btt.c | 35 ---
> 1 file changed, 16 insertions(+), 19 deletions(-)
>
> diff --git a/drivers/nvdimm/btt.c b/drivers/nvdimm/btt.c
> index d5593b0dc700..c55231f42617 100644
> --- a/drivers/nvdimm/btt.c
> +++ b/drivers/nvdimm/btt.c
> @@ -531,13 +531,13 @@ static int arena_clear_freelist_error(struct arena_info
> *arena, u32 lane)
> return ret;
> }
>
> -static int btt_freelist_init(struct arena_info *arena)
> +static int btt_freelist_init(struct device *dev, struct arena_info *arena)
Both struct arena_info and struct btt contain references to struct nd_btt
which is the device you are passing down this call stack.
Just use the device in the arena/btt rather than passing a device
structure. That makes the code easier to read and ensures that the device
associated with this arena or btt is used.
> {
> int new, ret;
> struct log_entry log_new;
> u32 i, map_entry, log_oldmap, log_newmap;
>
> - arena->freelist = kcalloc(arena->nfree, sizeof(struct free_entry),
> + arena->freelist = devm_kcalloc(dev, arena->nfree, sizeof(struct
> free_entry),
... devm_kcalloc(>nd_btt.dev, ...)
> GFP_KERNEL);
> if (!arena->freelist)
> return -ENOMEM;
> @@ -718,20 +718,20 @@ static int log_set_indices(struct arena_info *arena)
> return 0;
> }
>
> -static int btt_rtt_init(struct arena_info *arena)
> +static int btt_rtt_init(struct device *dev, struct arena_info *arena)
> {
> - arena->rtt = kcalloc(arena->nfree, sizeof(u32), GFP_KERNEL);
> + arena->rtt = devm_kcalloc(dev, arena->nfree, sizeof(u32), GFP_KERNEL);
> if (arena->rtt == NULL)
> return -ENOMEM;
>
> return 0;
> }
>
> -static int btt_maplocks_init(struct arena_info *arena)
> +static int btt_maplocks_init(struct device *dev, struct arena_info *arena)
> {
> u32 i;
>
> - arena->map_locks = kcalloc(arena->nfree, sizeof(struct aligned_lock),
> + arena->map_locks = devm_kcalloc(dev, arena->nfree, sizeof(struct
> aligned_lock),
> GFP_KERNEL);
> if (!arena->map_locks)
> return -ENOMEM;
> @@ -805,9 +805,6 @@ static void free_arenas(struct btt *btt)
>
> list_for_each_entry_safe(arena, next, >arena_list, list) {
> list_del(>list);
> - kfree(arena->rtt);
> - kfree(arena->map_locks);
> - kfree(arena->freelist);
This does not quite work.
free_arenas() is used in the error paths of create_arenas() and
discover_arenas(). In those cases devm_kfree() is probably a better way
to clean up this.
However...
> debugfs_remove_recursive(arena->debugfs_dir);
> kfree(arena);
Why can't arena be allocated with devm_*?
We need to change this up a bit more to handle the error path vs regular
device shutdown free (automatic devm frees).
Ira
