brcmsmac: allocate ucode with GFP_KERNEL

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 <tiwai@suse.de>
Acked-by: Arend van Spriel <arend.vanspriel@broadcom.com>
Signed-off-by: Kalle Valo <kvalo@codeaurora.org>

diff --git a/drivers/net/wireless/broadcom/brcm80211/brcmsmac/mac80211_if.c b/drivers/net/wireless/broadcom/brcm80211/brcmsmac/mac80211_if.c
index 8e58f68..ecc89e7 100644 (file)
--- 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;