mmc: sdhci: fix low memory corruption

When dma mapping (dma_map_sg) fails in sdhci_pre_dma_transfer, -EINVAL
is returned. There are 3 callers of sdhci_pre_dma_transfer:
* sdhci_pre_req and sdhci_adma_table_pre: handle negative return
* sdhci_prepare_data: handles 0 (error) and "else" (good) only

sdhci_prepare_data is therefore broken. When it receives -EINVAL from
sdhci_pre_dma_transfer, it assumes 1 sg mapping was mapped. Later,
this non-existent mapping with address 0 is kmap'ped and written to:
Corrupted low memory at ffff880000001000 (1000 phys) = 22b7d67df2f6d1cf
Corrupted low memory at ffff880000001008 (1008 phys) = 63848a5216b7dd95
Corrupted low memory at ffff880000001010 (1010 phys) = 330eb7ddef39e427
Corrupted low memory at ffff880000001018 (1018 phys) = 8017ac7295039bda
Corrupted low memory at ffff880000001020 (1020 phys) = 8ce039eac119074f
...

So teach sdhci_prepare_data to understand negative return values from
sdhci_pre_dma_transfer and disable DMA in that case, as well as for
zero.

It was introduced in 348487c (mmc:
sdhci: use pipeline mmc requests to improve performance). The commit
seems to be suspicious also by assigning host->sg_count both in
sdhci_pre_dma_transfer and sdhci_adma_table_pre.

Signed-off-by: Jiri Slaby <jslaby@suse.cz>
Cc: stable@vger.kernel.org # 4.0+
Fixes: 348487c
Cc: Ulf Hansson <ulf.hansson@linaro.org>
Cc: Haibo Chen <haibo.chen@freescale.com>
Signed-off-by: Ulf Hansson <ulf.hansson@linaro.org>

drivers/mmc/host/sdhci.c

@@ -834,7 +834,7 @@ static void sdhci_prepare_data(struct sdhci_host *host, struct mmc_command *cmd)
 			int sg_cnt;
 
 			sg_cnt = sdhci_pre_dma_transfer(host, data, NULL);
-			if (sg_cnt == 0) {
+			if (sg_cnt <= 0) {
 				/*
 				 * This only happens when someone fed
 				 * us an invalid request.