Re: [PATCH 1/3] usb: gadget: f_tcm: take the TPG in tcm_bind() instead of tcm_alloc()

From: Greg KH

Date: Fri Nov 20 2026 - 05:58:24 EST


On Thu, Oct 22, 2026 at 03:49:24PM +0200, Slawomir Stepien wrote:
> tcm_alloc() runs when a TCM function is symlinked into a gadget
> configuration. It searches tpg_instances[] for the function instance
> and copies the TPG pointer into fu->tpg at that point, under
> tpg_instances_lock. This has two problems.
>
> The function and the TPG can be set up in either order. If the
> function is linked into a configuration before the TPG directory is
> created, tpg_instances[i].tpg is still NULL and fu->tpg stays NULL.
> Creating and enabling the TPG afterwards sets can_attach, so
> tcm_bind() succeeds, and the first USB request that reaches the
> target code dereferences the NULL fu->tpg.
>
> Removing a TPG also leaves can_attach set for configfs instances, so a
> later bind is not refused even though the TPG it would use is gone.
>
> Store the TPG in struct f_tcm_opts instead: set it in usbg_make_tpg()
> and clear it in usbg_drop_tpg(), under opts->dep_lock. In tcm_bind(),
> refuse to bind unless both can_attach and opts->tpg are set, and copy
> opts->tpg into fu->tpg there. Also clear can_attach in usbg_drop_tpg()
> for configfs instances. The legacy g_target driver sets can_attach once
> at module load, and its attach callback does not set it again, so
> leave it alone in that case or g_target could never bind again after a
> TPG is recreated.
>
> With fu->tpg no longer set in tcm_alloc(), the tpg_instances[] lookup
> there serves no purpose: f->fi already identifies the instance for the
> item's whole lifetime. Drop it, and with it the tpg_instances_lock
> acquisition. config_usb_cfg_link() calls tcm_alloc() with gi->lock
> held, so this also removes the gi->lock -> tpg_instances_lock edge
> from the following lockdep report:
>
> WARNING: possible circular locking dependency detected
> task is trying to acquire lock:
> (tpg_instances_lock){+.+.}-{4:4}, at: tcm_alloc
> but task is already holding lock:
> (&gi->lock){+.+.}-{4:4}, at: config_usb_cfg_link
>
> the existing dependency chain (in reverse order) is:
> -> #2 (&gi->lock):
> config_usb_cfg_link
> configfs_symlink
> -> #1 (&sb->s_type->i_mutex_key#24):
> configfs_depend_item_unlocked
> usbg_make_tpg
> target_fabric_make_tpg
> configfs_mkdir
> -> #0 (tpg_instances_lock):
> tcm_alloc
> usb_get_function
> config_usb_cfg_link
> configfs_symlink
>
> Fixes: 4bb8548df632 ("usb: gadget: f_tcm: add configfs support")
> Reported-by: syzbot+2799146a53c356f8017d@xxxxxxxxxxxxxxxxxxxxxxxxx
> Closes: https://syzkaller.appspot.com/bug?extid=2799146a53c356f8017d
> Link: https://syzkaller.appspot.com/ai_job?id=51e1f189-6d91-4321-a6ba-07d4eb3cd8cd
> Assisted-by: Gemini:gemini-4-argon Gemini:gemini-3.8-flash syzbot
> Signed-off-by: Slawomir Stepien <sst@xxxxxxxxx>
> ---
> drivers/usb/gadget/function/f_tcm.c | 29 +++++++++++------------------
> drivers/usb/gadget/function/u_tcm.h | 4 ++++
> 2 files changed, 15 insertions(+), 18 deletions(-)
>
> diff --git a/drivers/usb/gadget/function/f_tcm.c b/drivers/usb/gadget/function/f_tcm.c
> index 9e6d4f39900a..cc75d7e9792a 100644
> --- a/drivers/usb/gadget/function/f_tcm.c
> +++ b/drivers/usb/gadget/function/f_tcm.c
> @@ -1720,6 +1720,9 @@ static struct se_portal_group *usbg_make_tpg(struct se_wwn *wwn,
>
> tpg_instances[i].tpg = tpg;
> tpg->fi = tpg_instances[i].func_inst;
> + mutex_lock(&opts->dep_lock);
> + opts->tpg = tpg;
> + mutex_unlock(&opts->dep_lock);

Why is a lock needed to write one value?

> mutex_unlock(&tpg_instances_lock);
> return &tpg->se_tpg;
>
> @@ -1760,11 +1763,14 @@ static void usbg_drop_tpg(struct se_portal_group *se_tpg)
> opts = container_of(tpg_instances[i].func_inst,
> struct f_tcm_opts, func_inst);
> mutex_lock(&opts->dep_lock);
> - if (opts->has_dep)
> + opts->tpg = NULL;
> + if (opts->has_dep) {
> module_put(opts->dependent);

The whole module_put() logic should probably just be removed, right?


> - else
> + } else {
> + opts->can_attach = false;
> configfs_undepend_item_unlocked(
> &opts->func_inst.group.cg_item);
> + }
> mutex_unlock(&opts->dep_lock);
> }
> mutex_unlock(&tpg_instances_lock);
> @@ -2300,10 +2306,11 @@ static int tcm_bind(struct usb_configuration *c, struct usb_function *f)
> opts = container_of(f->fi, struct f_tcm_opts, func_inst);
>
> mutex_lock(&opts->dep_lock);
> - if (!opts->can_attach) {
> + if (!opts->can_attach || !opts->tpg) {
> mutex_unlock(&opts->dep_lock);
> return -ENODEV;
> }
> + fu->tpg = opts->tpg;
> mutex_unlock(&opts->dep_lock);

But opts->tpg could be changed now, but you saved it off? No reference
counting at all? That feels very wrong.

thanks,

greg k-h