#include <linux/highmem.h>
 #include <linux/mmc/tmio.h>
+#include <linux/mutex.h>
 #include <linux/pagemap.h>
 #include <linux/spinlock.h>
 
 
        /* Track lost interrupts */
        struct delayed_work     delayed_reset_work;
-       spinlock_t              lock;
+       struct work_struct      done;
+
+       spinlock_t              lock;           /* protect host private data */
        unsigned long           last_req_ts;
+       struct mutex            ios_lock;       /* protect set_ios() context */
 };
 
 int tmio_mmc_host_probe(struct tmio_mmc_host **host,
 
 /* called with host->lock held, interrupts disabled */
 static void tmio_mmc_finish_request(struct tmio_mmc_host *host)
 {
-       struct mmc_request *mrq = host->mrq;
+       struct mmc_request *mrq;
+       unsigned long flags;
 
-       if (!mrq)
+       spin_lock_irqsave(&host->lock, flags);
+
+       mrq = host->mrq;
+       if (IS_ERR_OR_NULL(mrq)) {
+               spin_unlock_irqrestore(&host->lock, flags);
                return;
+       }
 
        host->cmd = NULL;
        host->data = NULL;
        cancel_delayed_work(&host->delayed_reset_work);
 
        host->mrq = NULL;
+       spin_unlock_irqrestore(&host->lock, flags);
 
-       /* FIXME: mmc_request_done() can schedule! */
        mmc_request_done(host->mmc, mrq);
 }
 
+static void tmio_mmc_done_work(struct work_struct *work)
+{
+       struct tmio_mmc_host *host = container_of(work, struct tmio_mmc_host,
+                                                 done);
+       tmio_mmc_finish_request(host);
+}
+
 /* These are the bitmasks the tmio chip requires to implement the MMC response
  * types. Note that R1 and R6 are the same in this scheme. */
 #define APP_CMD        0x0040
                        BUG();
        }
 
-       tmio_mmc_finish_request(host);
+       schedule_work(&host->done);
 }
 
 static void tmio_mmc_data_irq(struct tmio_mmc_host *host)
                                tasklet_schedule(&host->dma_issue);
                }
        } else {
-               tmio_mmc_finish_request(host);
+               schedule_work(&host->done);
        }
 
 out:
        if (ireg & (TMIO_STAT_CARD_INSERT | TMIO_STAT_CARD_REMOVE)) {
                tmio_mmc_ack_mmc_irqs(host, TMIO_STAT_CARD_INSERT |
                        TMIO_STAT_CARD_REMOVE);
-               mmc_detect_change(host->mmc, msecs_to_jiffies(100));
+               if (!work_pending(&host->mmc->detect.work))
+                       mmc_detect_change(host->mmc, msecs_to_jiffies(100));
                goto out;
        }
 
        struct tmio_mmc_data *pdata = host->pdata;
        unsigned long flags;
 
+       mutex_lock(&host->ios_lock);
+
        spin_lock_irqsave(&host->lock, flags);
        if (host->mrq) {
                if (IS_ERR(host->mrq)) {
                                host->mrq->cmd->opcode, host->last_req_ts, jiffies);
                }
                spin_unlock_irqrestore(&host->lock, flags);
+
+               mutex_unlock(&host->ios_lock);
                return;
        }
 
                        current->comm, task_pid_nr(current),
                        ios->clock, ios->power_mode);
        host->mrq = NULL;
+
+       mutex_unlock(&host->ios_lock);
 }
 
 static int tmio_mmc_get_ro(struct mmc_host *mmc)
                tmio_mmc_enable_sdio_irq(mmc, 0);
 
        spin_lock_init(&_host->lock);
+       mutex_init(&_host->ios_lock);
 
        /* Init delayed work for request timeouts */
        INIT_DELAYED_WORK(&_host->delayed_reset_work, tmio_mmc_reset_work);
+       INIT_WORK(&_host->done, tmio_mmc_done_work);
 
        /* See if we also get DMA */
        tmio_mmc_request_dma(_host, pdata);
                pm_runtime_get_sync(&pdev->dev);
 
        mmc_remove_host(host->mmc);
+       cancel_work_sync(&host->done);
        cancel_delayed_work_sync(&host->delayed_reset_work);
        tmio_mmc_release_dma(host);