The brcms_ucode_init_buf() duplicates the ucode chunks via kmemdup() with GFP_ATOMIC as a precondition of wl->lock acquired. This caused allocation failures sometimes as reported in the bugzilla below.
When looking at the the real usage, one can find that it's called solely from brcms_request_fw(), and it's obviously outside the lock. Hence we can use GFP_KERNEL there safely for avoiding such allocation errors. Bugzilla: http://bugzilla.suse.com/show_bug.cgi?id=1085174 Signed-off-by: Takashi Iwai <[email protected]> --- drivers/net/wireless/broadcom/brcm80211/brcmsmac/mac80211_if.c | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/drivers/net/wireless/broadcom/brcm80211/brcmsmac/mac80211_if.c b/drivers/net/wireless/broadcom/brcm80211/brcmsmac/mac80211_if.c index ddfdfe177e24..a14fccf8a902 100644 --- a/drivers/net/wireless/broadcom/brcm80211/brcmsmac/mac80211_if.c +++ b/drivers/net/wireless/broadcom/brcm80211/brcmsmac/mac80211_if.c @@ -1563,7 +1563,7 @@ void brcms_free_timer(struct brcms_timer *t) } /* - * precondition: perimeter lock has been acquired + * precondition: no locking required */ int brcms_ucode_init_buf(struct brcms_info *wl, void **pbuf, u32 idx) { @@ -1578,7 +1578,7 @@ int brcms_ucode_init_buf(struct brcms_info *wl, void **pbuf, u32 idx) if (le32_to_cpu(hdr->idx) == idx) { pdata = wl->fw.fw_bin[i]->data + le32_to_cpu(hdr->offset); - *pbuf = kmemdup(pdata, len, GFP_ATOMIC); + *pbuf = kmemdup(pdata, len, GFP_KERNEL); if (*pbuf == NULL) goto fail; -- 2.16.2
