Merge branch 'linux-linaro-lsk-v4.4-android' of git://git.linaro.org/kernel/linux...
[firefly-linux-kernel-4.4.55.git] / drivers / platform / goldfish / goldfish_pipe_v2.c
diff --git a/drivers/platform/goldfish/goldfish_pipe_v2.c b/drivers/platform/goldfish/goldfish_pipe_v2.c
new file mode 100644 (file)
index 0000000..ad373ed
--- /dev/null
@@ -0,0 +1,889 @@
+/*
+ * Copyright (C) 2012 Intel, Inc.
+ * Copyright (C) 2013 Intel, Inc.
+ * Copyright (C) 2014 Linaro Limited
+ * Copyright (C) 2011-2016 Google, Inc.
+ *
+ * This software is licensed under the terms of the GNU General Public
+ * License version 2, as published by the Free Software Foundation, and
+ * may be copied, distributed, and modified under those terms.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ * GNU General Public License for more details.
+ *
+ */
+
+/* This source file contains the implementation of a special device driver
+ * that intends to provide a *very* fast communication channel between the
+ * guest system and the QEMU emulator.
+ *
+ * Usage from the guest is simply the following (error handling simplified):
+ *
+ *    int  fd = open("/dev/qemu_pipe",O_RDWR);
+ *    .... write() or read() through the pipe.
+ *
+ * This driver doesn't deal with the exact protocol used during the session.
+ * It is intended to be as simple as something like:
+ *
+ *    // do this _just_ after opening the fd to connect to a specific
+ *    // emulator service.
+ *    const char*  msg = "<pipename>";
+ *    if (write(fd, msg, strlen(msg)+1) < 0) {
+ *       ... could not connect to <pipename> service
+ *       close(fd);
+ *    }
+ *
+ *    // after this, simply read() and write() to communicate with the
+ *    // service. Exact protocol details left as an exercise to the reader.
+ *
+ * This driver is very fast because it doesn't copy any data through
+ * intermediate buffers, since the emulator is capable of translating
+ * guest user addresses into host ones.
+ *
+ * Note that we must however ensure that each user page involved in the
+ * exchange is properly mapped during a transfer.
+ */
+
+#include "goldfish_pipe.h"
+
+
+/*
+ * Update this when something changes in the driver's behavior so the host
+ * can benefit from knowing it
+ */
+enum {
+       PIPE_DRIVER_VERSION = 2,
+       PIPE_CURRENT_DEVICE_VERSION = 2
+};
+
+/*
+ * IMPORTANT: The following constants must match the ones used and defined
+ * in external/qemu/hw/goldfish_pipe.c in the Android source tree.
+ */
+
+/* List of bitflags returned in status of CMD_POLL command */
+enum PipePollFlags {
+       PIPE_POLL_IN    = 1 << 0,
+       PIPE_POLL_OUT   = 1 << 1,
+       PIPE_POLL_HUP   = 1 << 2
+};
+
+/* Possible status values used to signal errors - see goldfish_pipe_error_convert */
+enum PipeErrors {
+       PIPE_ERROR_INVAL  = -1,
+       PIPE_ERROR_AGAIN  = -2,
+       PIPE_ERROR_NOMEM  = -3,
+       PIPE_ERROR_IO     = -4
+};
+
+/* Bit-flags used to signal events from the emulator */
+enum PipeWakeFlags {
+       PIPE_WAKE_CLOSED = 1 << 0,  /* emulator closed pipe */
+       PIPE_WAKE_READ   = 1 << 1,  /* pipe can now be read from */
+       PIPE_WAKE_WRITE  = 1 << 2  /* pipe can now be written to */
+};
+
+/* Bit flags for the 'flags' field */
+enum PipeFlagsBits {
+       BIT_CLOSED_ON_HOST = 0,  /* pipe closed by host */
+       BIT_WAKE_ON_WRITE  = 1,  /* want to be woken on writes */
+       BIT_WAKE_ON_READ   = 2,  /* want to be woken on reads */
+};
+
+enum PipeRegs {
+       PIPE_REG_CMD = 0,
+
+       PIPE_REG_SIGNAL_BUFFER_HIGH = 4,
+       PIPE_REG_SIGNAL_BUFFER = 8,
+       PIPE_REG_SIGNAL_BUFFER_COUNT = 12,
+
+       PIPE_REG_OPEN_BUFFER_HIGH = 20,
+       PIPE_REG_OPEN_BUFFER = 24,
+
+       PIPE_REG_VERSION = 36,
+
+       PIPE_REG_GET_SIGNALLED = 48,
+};
+
+enum PipeCmdCode {
+       PIPE_CMD_OPEN = 1,      /* to be used by the pipe device itself */
+       PIPE_CMD_CLOSE,
+       PIPE_CMD_POLL,
+       PIPE_CMD_WRITE,
+       PIPE_CMD_WAKE_ON_WRITE,
+       PIPE_CMD_READ,
+       PIPE_CMD_WAKE_ON_READ,
+
+       /*
+        * TODO(zyy): implement a deferred read/write execution to allow parallel
+        *  processing of pipe operations on the host.
+       */
+       PIPE_CMD_WAKE_ON_DONE_IO,
+};
+
+enum {
+       MAX_BUFFERS_PER_COMMAND = 336,
+       MAX_SIGNALLED_PIPES = 64,
+       INITIAL_PIPES_CAPACITY = 64
+};
+
+struct goldfish_pipe_dev;
+struct goldfish_pipe;
+struct goldfish_pipe_command;
+
+/* A per-pipe command structure, shared with the host */
+struct goldfish_pipe_command {
+       s32 cmd;                /* PipeCmdCode, guest -> host */
+       s32 id;                 /* pipe id, guest -> host */
+       s32 status;             /* command execution status, host -> guest */
+       s32 reserved;   /* to pad to 64-bit boundary */
+       union {
+               /* Parameters for PIPE_CMD_{READ,WRITE} */
+               struct {
+                       u32 buffers_count;                                      /* number of buffers, guest -> host */
+                       s32 consumed_size;                                      /* number of consumed bytes, host -> guest */
+                       u64 ptrs[MAX_BUFFERS_PER_COMMAND];      /* buffer pointers, guest -> host */
+                       u32 sizes[MAX_BUFFERS_PER_COMMAND];     /* buffer sizes, guest -> host */
+               } rw_params;
+       };
+};
+
+/* A single signalled pipe information */
+struct signalled_pipe_buffer {
+       u32 id;
+       u32 flags;
+};
+
+/* Parameters for the PIPE_CMD_OPEN command */
+struct open_command_param {
+       u64 command_buffer_ptr;
+       u32 rw_params_max_count;
+};
+
+/* Device-level set of buffers shared with the host */
+struct goldfish_pipe_dev_buffers {
+       struct open_command_param open_command_params;
+       struct signalled_pipe_buffer signalled_pipe_buffers[MAX_SIGNALLED_PIPES];
+};
+
+/* This data type models a given pipe instance */
+struct goldfish_pipe {
+       u32 id;                                                 /* pipe ID - index into goldfish_pipe_dev::pipes array */
+       unsigned long flags;                    /* The wake flags pipe is waiting for
+                                                                        * Note: not protected with any lock, uses atomic operations
+                                                                        *  and barriers to make it thread-safe.
+                                                                        */
+       unsigned long signalled_flags;  /* wake flags host have signalled,
+                                                                        *  - protected by goldfish_pipe_dev::lock */
+
+       struct goldfish_pipe_command *command_buffer;   /* A pointer to command buffer */
+
+       /* doubly linked list of signalled pipes, protected by goldfish_pipe_dev::lock */
+       struct goldfish_pipe *prev_signalled;
+       struct goldfish_pipe *next_signalled;
+
+       /*
+        * A pipe's own lock. Protects the following:
+        *  - *command_buffer - makes sure a command can safely write its parameters
+        *    to the host and read the results back.
+        */
+       struct mutex lock;
+
+       wait_queue_head_t wake_queue;   /* A wake queue for sleeping until host signals an event */
+       struct goldfish_pipe_dev *dev;  /* Pointer to the parent goldfish_pipe_dev instance */
+};
+
+struct goldfish_pipe_dev pipe_dev[1] = {};
+
+static int goldfish_cmd_locked(struct goldfish_pipe *pipe, enum PipeCmdCode cmd)
+{
+       pipe->command_buffer->cmd = cmd;
+       pipe->command_buffer->status = PIPE_ERROR_INVAL;        /* failure by default */
+       writel(pipe->id, pipe->dev->base + PIPE_REG_CMD);
+       return pipe->command_buffer->status;
+}
+
+static int goldfish_cmd(struct goldfish_pipe *pipe, enum PipeCmdCode cmd)
+{
+       int status;
+       if (mutex_lock_interruptible(&pipe->lock))
+               return PIPE_ERROR_IO;
+       status = goldfish_cmd_locked(pipe, cmd);
+       mutex_unlock(&pipe->lock);
+       return status;
+}
+
+/*
+ * This function converts an error code returned by the emulator through
+ * the PIPE_REG_STATUS i/o register into a valid negative errno value.
+ */
+static int goldfish_pipe_error_convert(int status)
+{
+       switch (status) {
+       case PIPE_ERROR_AGAIN:
+               return -EAGAIN;
+       case PIPE_ERROR_NOMEM:
+               return -ENOMEM;
+       case PIPE_ERROR_IO:
+               return -EIO;
+       default:
+               return -EINVAL;
+       }
+}
+
+static int pin_user_pages(unsigned long first_page, unsigned long last_page,
+       unsigned last_page_size, int is_write,
+       struct page *pages[MAX_BUFFERS_PER_COMMAND], unsigned *iter_last_page_size)
+{
+       int ret;
+       int requested_pages = ((last_page - first_page) >> PAGE_SHIFT) + 1;
+       if (requested_pages > MAX_BUFFERS_PER_COMMAND) {
+               requested_pages = MAX_BUFFERS_PER_COMMAND;
+               *iter_last_page_size = PAGE_SIZE;
+       } else {
+               *iter_last_page_size = last_page_size;
+       }
+
+       ret = get_user_pages_fast(
+                       first_page, requested_pages, !is_write, pages);
+       if (ret <= 0)
+               return -EFAULT;
+       if (ret < requested_pages)
+               *iter_last_page_size = PAGE_SIZE;
+       return ret;
+
+}
+
+static void release_user_pages(struct page **pages, int pages_count,
+       int is_write, s32 consumed_size)
+{
+       int i;
+       for (i = 0; i < pages_count; i++) {
+               if (!is_write && consumed_size > 0) {
+                       set_page_dirty(pages[i]);
+               }
+               put_page(pages[i]);
+       }
+}
+
+/* Populate the call parameters, merging adjacent pages together */
+static void populate_rw_params(
+       struct page **pages, int pages_count,
+       unsigned long address, unsigned long address_end,
+       unsigned long first_page, unsigned long last_page,
+       unsigned iter_last_page_size, int is_write,
+       struct goldfish_pipe_command *command)
+{
+       /*
+        * Process the first page separately - it's the only page that
+        * needs special handling for its start address.
+        */
+       unsigned long xaddr = page_to_phys(pages[0]);
+       unsigned long xaddr_prev = xaddr;
+       int buffer_idx = 0;
+       int i = 1;
+       int size_on_page = first_page == last_page
+                       ? (int)(address_end - address)
+                       : (PAGE_SIZE - (address & ~PAGE_MASK));
+       command->rw_params.ptrs[0] = (u64)(xaddr | (address & ~PAGE_MASK));
+       command->rw_params.sizes[0] = size_on_page;
+       for (; i < pages_count; ++i) {
+               xaddr = page_to_phys(pages[i]);
+               size_on_page = (i == pages_count - 1) ? iter_last_page_size : PAGE_SIZE;
+               if (xaddr == xaddr_prev + PAGE_SIZE) {
+                       command->rw_params.sizes[buffer_idx] += size_on_page;
+               } else {
+                       ++buffer_idx;
+                       command->rw_params.ptrs[buffer_idx] = (u64)xaddr;
+                       command->rw_params.sizes[buffer_idx] = size_on_page;
+               }
+               xaddr_prev = xaddr;
+       }
+       command->rw_params.buffers_count = buffer_idx + 1;
+}
+
+static int transfer_max_buffers(struct goldfish_pipe* pipe,
+       unsigned long address, unsigned long address_end, int is_write,
+       unsigned long last_page, unsigned int last_page_size,
+       s32* consumed_size, int* status)
+{
+       struct page *pages[MAX_BUFFERS_PER_COMMAND];
+       unsigned long first_page = address & PAGE_MASK;
+       unsigned int iter_last_page_size;
+       int pages_count = pin_user_pages(first_page, last_page,
+                       last_page_size, is_write,
+                       pages, &iter_last_page_size);
+       if (pages_count < 0)
+               return pages_count;
+
+       /* Serialize access to the pipe command buffers */
+       if (mutex_lock_interruptible(&pipe->lock))
+               return -ERESTARTSYS;
+
+       populate_rw_params(pages, pages_count, address, address_end,
+               first_page, last_page, iter_last_page_size, is_write,
+               pipe->command_buffer);
+
+       /* Transfer the data */
+       *status = goldfish_cmd_locked(pipe,
+                                               is_write ? PIPE_CMD_WRITE : PIPE_CMD_READ);
+
+       *consumed_size = pipe->command_buffer->rw_params.consumed_size;
+
+       mutex_unlock(&pipe->lock);
+
+       release_user_pages(pages, pages_count, is_write, *consumed_size);
+
+       return 0;
+}
+
+static int wait_for_host_signal(struct goldfish_pipe *pipe, int is_write)
+{
+       u32 wakeBit = is_write ? BIT_WAKE_ON_WRITE : BIT_WAKE_ON_READ;
+       set_bit(wakeBit, &pipe->flags);
+
+       /* Tell the emulator we're going to wait for a wake event */
+       (void)goldfish_cmd(pipe,
+                       is_write ? PIPE_CMD_WAKE_ON_WRITE : PIPE_CMD_WAKE_ON_READ);
+
+       while (test_bit(wakeBit, &pipe->flags)) {
+               if (wait_event_interruptible(
+                               pipe->wake_queue,
+                               !test_bit(wakeBit, &pipe->flags)))
+                       return -ERESTARTSYS;
+
+               if (test_bit(BIT_CLOSED_ON_HOST, &pipe->flags))
+                       return -EIO;
+       }
+
+       return 0;
+}
+
+static ssize_t goldfish_pipe_read_write(struct file *filp,
+       char __user *buffer, size_t bufflen, int is_write)
+{
+       struct goldfish_pipe *pipe = filp->private_data;
+       int count = 0, ret = -EINVAL;
+       unsigned long address, address_end, last_page;
+       unsigned int last_page_size;
+
+       /* If the emulator already closed the pipe, no need to go further */
+       if (unlikely(test_bit(BIT_CLOSED_ON_HOST, &pipe->flags)))
+               return -EIO;
+       /* Null reads or writes succeeds */
+       if (unlikely(bufflen == 0))
+               return 0;
+       /* Check the buffer range for access */
+       if (unlikely(!access_ok(is_write ? VERIFY_WRITE : VERIFY_READ,
+                       buffer, bufflen)))
+               return -EFAULT;
+
+       address = (unsigned long)buffer;
+       address_end = address + bufflen;
+       last_page = (address_end - 1) & PAGE_MASK;
+       last_page_size = ((address_end - 1) & ~PAGE_MASK) + 1;
+
+       while (address < address_end) {
+               s32 consumed_size;
+               int status;
+               ret = transfer_max_buffers(pipe, address, address_end, is_write,
+                               last_page, last_page_size, &consumed_size, &status);
+               if (ret < 0)
+                       break;
+
+               if (consumed_size > 0) {
+                       /* No matter what's the status, we've transfered something */
+                       count += consumed_size;
+                       address += consumed_size;
+               }
+               if (status > 0)
+                       continue;
+               if (status == 0) {
+                       /* EOF */
+                       ret = 0;
+                       break;
+               }
+               if (count > 0) {
+                       /*
+                        * An error occured, but we already transfered
+                        * something on one of the previous iterations.
+                        * Just return what we already copied and log this
+                        * err.
+                        */
+                       if (status != PIPE_ERROR_AGAIN)
+                               pr_info_ratelimited("goldfish_pipe: backend error %d on %s\n",
+                                                                       status, is_write ? "write" : "read");
+                       break;
+               }
+
+               /*
+                * If the error is not PIPE_ERROR_AGAIN, or if we are in
+                * non-blocking mode, just return the error code.
+                */
+               if (status != PIPE_ERROR_AGAIN || (filp->f_flags & O_NONBLOCK) != 0) {
+                       ret = goldfish_pipe_error_convert(status);
+                       break;
+               }
+
+               status = wait_for_host_signal(pipe, is_write);
+               if (status < 0)
+                       return status;
+       }
+
+       if (count > 0)
+               return count;
+       return ret;
+}
+
+static ssize_t goldfish_pipe_read(struct file *filp, char __user *buffer,
+                               size_t bufflen, loff_t *ppos)
+{
+       return goldfish_pipe_read_write(filp, buffer, bufflen, /* is_write */ 0);
+}
+
+static ssize_t goldfish_pipe_write(struct file *filp,
+                               const char __user *buffer, size_t bufflen,
+                               loff_t *ppos)
+{
+       return goldfish_pipe_read_write(filp,
+                       /* cast away the const */(char __user *)buffer, bufflen,
+                       /* is_write */ 1);
+}
+
+static unsigned int goldfish_pipe_poll(struct file *filp, poll_table *wait)
+{
+       struct goldfish_pipe *pipe = filp->private_data;
+       unsigned int mask = 0;
+       int status;
+
+       poll_wait(filp, &pipe->wake_queue, wait);
+
+       status = goldfish_cmd(pipe, PIPE_CMD_POLL);
+       if (status < 0) {
+               return -ERESTARTSYS;
+       }
+
+       if (status & PIPE_POLL_IN)
+               mask |= POLLIN | POLLRDNORM;
+       if (status & PIPE_POLL_OUT)
+               mask |= POLLOUT | POLLWRNORM;
+       if (status & PIPE_POLL_HUP)
+               mask |= POLLHUP;
+       if (test_bit(BIT_CLOSED_ON_HOST, &pipe->flags))
+               mask |= POLLERR;
+
+       return mask;
+}
+
+static void signalled_pipes_add_locked(struct goldfish_pipe_dev *dev,
+       u32 id, u32 flags)
+{
+       struct goldfish_pipe *pipe;
+
+       BUG_ON(id >= dev->pipes_capacity);
+
+       pipe = dev->pipes[id];
+       if (!pipe)
+               return;
+       pipe->signalled_flags |= flags;
+
+       if (pipe->prev_signalled || pipe->next_signalled
+               || dev->first_signalled_pipe == pipe)
+               return; /* already in the list */
+       pipe->next_signalled = dev->first_signalled_pipe;
+       if (dev->first_signalled_pipe) {
+               dev->first_signalled_pipe->prev_signalled = pipe;
+       }
+       dev->first_signalled_pipe = pipe;
+}
+
+static void signalled_pipes_remove_locked(struct goldfish_pipe_dev *dev,
+       struct goldfish_pipe *pipe) {
+       if (pipe->prev_signalled)
+               pipe->prev_signalled->next_signalled = pipe->next_signalled;
+       if (pipe->next_signalled)
+               pipe->next_signalled->prev_signalled = pipe->prev_signalled;
+       if (pipe == dev->first_signalled_pipe)
+               dev->first_signalled_pipe = pipe->next_signalled;
+       pipe->prev_signalled = NULL;
+       pipe->next_signalled = NULL;
+}
+
+static struct goldfish_pipe *signalled_pipes_pop_front(struct goldfish_pipe_dev *dev,
+               int *wakes)
+{
+       struct goldfish_pipe *pipe;
+       unsigned long flags;
+       spin_lock_irqsave(&dev->lock, flags);
+
+       pipe = dev->first_signalled_pipe;
+       if (pipe) {
+               *wakes = pipe->signalled_flags;
+               pipe->signalled_flags = 0;
+               /*
+                * This is an optimized version of signalled_pipes_remove_locked() -
+                * we want to make it as fast as possible to wake the sleeping pipe
+                * operations faster
+                */
+               dev->first_signalled_pipe = pipe->next_signalled;
+               if (dev->first_signalled_pipe)
+                       dev->first_signalled_pipe->prev_signalled = NULL;
+               pipe->next_signalled = NULL;
+       }
+
+       spin_unlock_irqrestore(&dev->lock, flags);
+       return pipe;
+}
+
+static void goldfish_interrupt_task(unsigned long unused)
+{
+       struct goldfish_pipe_dev *dev = pipe_dev;
+       /* Iterate over the signalled pipes and wake them one by one */
+       struct goldfish_pipe *pipe;
+       int wakes;
+       while ((pipe = signalled_pipes_pop_front(dev, &wakes)) != NULL) {
+               if (wakes & PIPE_WAKE_CLOSED) {
+                       pipe->flags = 1 << BIT_CLOSED_ON_HOST;
+               } else {
+                       if (wakes & PIPE_WAKE_READ)
+                               clear_bit(BIT_WAKE_ON_READ, &pipe->flags);
+                       if (wakes & PIPE_WAKE_WRITE)
+                               clear_bit(BIT_WAKE_ON_WRITE, &pipe->flags);
+               }
+               /*
+                * wake_up_interruptible() implies a write barrier, so don't explicitly
+                * add another one here.
+                */
+               wake_up_interruptible(&pipe->wake_queue);
+       }
+}
+DECLARE_TASKLET(goldfish_interrupt_tasklet, goldfish_interrupt_task, 0);
+
+/*
+ * The general idea of the interrupt handling:
+ *
+ *  1. device raises an interrupt if there's at least one signalled pipe
+ *  2. IRQ handler reads the signalled pipes and their count from the device
+ *  3. device writes them into a shared buffer and returns the count
+ *      it only resets the IRQ if it has returned all signalled pipes,
+ *      otherwise it leaves it raised, so IRQ handler will be called
+ *      again for the next chunk
+ *  4. IRQ handler adds all returned pipes to the device's signalled pipes list
+ *  5. IRQ handler launches a tasklet to process the signalled pipes from the
+ *      list in a separate context
+ */
+static irqreturn_t goldfish_pipe_interrupt(int irq, void *dev_id)
+{
+       u32 count;
+       u32 i;
+       unsigned long flags;
+       struct goldfish_pipe_dev *dev = dev_id;
+       if (dev != pipe_dev)
+               return IRQ_NONE;
+
+       /* Request the signalled pipes from the device */
+       spin_lock_irqsave(&dev->lock, flags);
+
+       count = readl(dev->base + PIPE_REG_GET_SIGNALLED);
+       if (count == 0) {
+               spin_unlock_irqrestore(&dev->lock, flags);
+               return IRQ_NONE;
+       }
+       if (count > MAX_SIGNALLED_PIPES)
+               count = MAX_SIGNALLED_PIPES;
+
+       for (i = 0; i < count; ++i)
+               signalled_pipes_add_locked(dev,
+                       dev->buffers->signalled_pipe_buffers[i].id,
+                       dev->buffers->signalled_pipe_buffers[i].flags);
+
+       spin_unlock_irqrestore(&dev->lock, flags);
+
+       tasklet_schedule(&goldfish_interrupt_tasklet);
+       return IRQ_HANDLED;
+}
+
+static int get_free_pipe_id_locked(struct goldfish_pipe_dev *dev)
+{
+       int id;
+       for (id = 0; id < dev->pipes_capacity; ++id)
+               if (!dev->pipes[id])
+                       return id;
+
+       {
+               /* Reallocate the array */
+               u32 new_capacity = 2 * dev->pipes_capacity;
+               struct goldfish_pipe **pipes =
+                               kcalloc(new_capacity, sizeof(*pipes),
+                                       GFP_ATOMIC);
+               if (!pipes)
+                       return -ENOMEM;
+               memcpy(pipes, dev->pipes, sizeof(*pipes) * dev->pipes_capacity);
+               kfree(dev->pipes);
+               dev->pipes = pipes;
+               id = dev->pipes_capacity;
+               dev->pipes_capacity = new_capacity;
+       }
+       return id;
+}
+
+/**
+ *     goldfish_pipe_open - open a channel to the AVD
+ *     @inode: inode of device
+ *     @file: file struct of opener
+ *
+ *     Create a new pipe link between the emulator and the use application.
+ *     Each new request produces a new pipe.
+ *
+ *     Note: we use the pipe ID as a mux. All goldfish emulations are 32bit
+ *     right now so this is fine. A move to 64bit will need this addressing
+ */
+static int goldfish_pipe_open(struct inode *inode, struct file *file)
+{
+       struct goldfish_pipe_dev *dev = pipe_dev;
+       unsigned long flags;
+       int id;
+       int status;
+
+       /* Allocate new pipe kernel object */
+       struct goldfish_pipe *pipe = kzalloc(sizeof(*pipe), GFP_KERNEL);
+       if (pipe == NULL)
+               return -ENOMEM;
+
+       pipe->dev = dev;
+       mutex_init(&pipe->lock);
+       init_waitqueue_head(&pipe->wake_queue);
+
+       /*
+        * Command buffer needs to be allocated on its own page to make sure it is
+        * physically contiguous in host's address space.
+        */
+       pipe->command_buffer =
+                       (struct goldfish_pipe_command*)__get_free_page(GFP_KERNEL);
+       if (!pipe->command_buffer) {
+               status = -ENOMEM;
+               goto err_pipe;
+       }
+
+       spin_lock_irqsave(&dev->lock, flags);
+
+       id = get_free_pipe_id_locked(dev);
+       if (id < 0) {
+               status = id;
+               goto err_id_locked;
+       }
+
+       dev->pipes[id] = pipe;
+       pipe->id = id;
+       pipe->command_buffer->id = id;
+
+       /* Now tell the emulator we're opening a new pipe. */
+       dev->buffers->open_command_params.rw_params_max_count =
+                       MAX_BUFFERS_PER_COMMAND;
+       dev->buffers->open_command_params.command_buffer_ptr =
+                       (u64)(unsigned long)__pa(pipe->command_buffer);
+       status = goldfish_cmd_locked(pipe, PIPE_CMD_OPEN);
+       spin_unlock_irqrestore(&dev->lock, flags);
+       if (status < 0)
+               goto err_cmd;
+       /* All is done, save the pipe into the file's private data field */
+       file->private_data = pipe;
+       return 0;
+
+err_cmd:
+       spin_lock_irqsave(&dev->lock, flags);
+       dev->pipes[id] = NULL;
+err_id_locked:
+       spin_unlock_irqrestore(&dev->lock, flags);
+       free_page((unsigned long)pipe->command_buffer);
+err_pipe:
+       kfree(pipe);
+       return status;
+}
+
+static int goldfish_pipe_release(struct inode *inode, struct file *filp)
+{
+       unsigned long flags;
+       struct goldfish_pipe *pipe = filp->private_data;
+       struct goldfish_pipe_dev *dev = pipe->dev;
+
+       /* The guest is closing the channel, so tell the emulator right now */
+       (void)goldfish_cmd(pipe, PIPE_CMD_CLOSE);
+
+       spin_lock_irqsave(&dev->lock, flags);
+       dev->pipes[pipe->id] = NULL;
+       signalled_pipes_remove_locked(dev, pipe);
+       spin_unlock_irqrestore(&dev->lock, flags);
+
+       filp->private_data = NULL;
+       free_page((unsigned long)pipe->command_buffer);
+       kfree(pipe);
+       return 0;
+}
+
+static const struct file_operations goldfish_pipe_fops = {
+       .owner = THIS_MODULE,
+       .read = goldfish_pipe_read,
+       .write = goldfish_pipe_write,
+       .poll = goldfish_pipe_poll,
+       .open = goldfish_pipe_open,
+       .release = goldfish_pipe_release,
+};
+
+static struct miscdevice goldfish_pipe_dev = {
+       .minor = MISC_DYNAMIC_MINOR,
+       .name = "goldfish_pipe",
+       .fops = &goldfish_pipe_fops,
+};
+
+static int goldfish_pipe_device_init_v2(struct platform_device *pdev)
+{
+       char *page;
+       struct goldfish_pipe_dev *dev = pipe_dev;
+       int err = devm_request_irq(&pdev->dev, dev->irq, goldfish_pipe_interrupt,
+                               IRQF_SHARED, "goldfish_pipe", dev);
+       if (err) {
+               dev_err(&pdev->dev, "unable to allocate IRQ for v2\n");
+               return err;
+       }
+
+       err = misc_register(&goldfish_pipe_dev);
+       if (err) {
+               dev_err(&pdev->dev, "unable to register v2 device\n");
+               return err;
+       }
+
+       dev->first_signalled_pipe = NULL;
+       dev->pipes_capacity = INITIAL_PIPES_CAPACITY;
+       dev->pipes = kcalloc(dev->pipes_capacity, sizeof(*dev->pipes), GFP_KERNEL);
+       if (!dev->pipes)
+               return -ENOMEM;
+
+       /*
+        * We're going to pass two buffers, open_command_params and
+        * signalled_pipe_buffers, to the host. This means each of those buffers
+        * needs to be contained in a single physical page. The easiest choice is
+        * to just allocate a page and place the buffers in it.
+        */
+       BUG_ON(sizeof(*dev->buffers) > PAGE_SIZE);
+       page = (char*)__get_free_page(GFP_KERNEL);
+       if (!page) {
+               kfree(dev->pipes);
+               return -ENOMEM;
+       }
+       dev->buffers = (struct goldfish_pipe_dev_buffers*)page;
+
+       /* Send the buffer addresses to the host */
+       {
+               u64 paddr = __pa(&dev->buffers->signalled_pipe_buffers);
+               writel((u32)(unsigned long)(paddr >> 32), dev->base + PIPE_REG_SIGNAL_BUFFER_HIGH);
+               writel((u32)(unsigned long)paddr, dev->base + PIPE_REG_SIGNAL_BUFFER);
+               writel((u32)MAX_SIGNALLED_PIPES, dev->base + PIPE_REG_SIGNAL_BUFFER_COUNT);
+
+               paddr = __pa(&dev->buffers->open_command_params);
+               writel((u32)(unsigned long)(paddr >> 32), dev->base + PIPE_REG_OPEN_BUFFER_HIGH);
+               writel((u32)(unsigned long)paddr, dev->base + PIPE_REG_OPEN_BUFFER);
+       }
+       return 0;
+}
+
+static void goldfish_pipe_device_deinit_v2(struct platform_device *pdev) {
+       struct goldfish_pipe_dev *dev = pipe_dev;
+       misc_deregister(&goldfish_pipe_dev);
+       kfree(dev->pipes);
+       free_page((unsigned long)dev->buffers);
+}
+
+static int goldfish_pipe_probe(struct platform_device *pdev)
+{
+       int err;
+       struct resource *r;
+       struct goldfish_pipe_dev *dev = pipe_dev;
+
+       BUG_ON(sizeof(struct goldfish_pipe_command) > PAGE_SIZE);
+
+       /* not thread safe, but this should not happen */
+       WARN_ON(dev->base != NULL);
+
+       spin_lock_init(&dev->lock);
+
+       r = platform_get_resource(pdev, IORESOURCE_MEM, 0);
+       if (r == NULL || resource_size(r) < PAGE_SIZE) {
+               dev_err(&pdev->dev, "can't allocate i/o page\n");
+               return -EINVAL;
+       }
+       dev->base = devm_ioremap(&pdev->dev, r->start, PAGE_SIZE);
+       if (dev->base == NULL) {
+               dev_err(&pdev->dev, "ioremap failed\n");
+               return -EINVAL;
+       }
+
+       r = platform_get_resource(pdev, IORESOURCE_IRQ, 0);
+       if (r == NULL) {
+               err = -EINVAL;
+               goto error;
+       }
+       dev->irq = r->start;
+
+       /*
+        * Exchange the versions with the host device
+        *
+        * Note: v1 driver used to not report its version, so we write it before
+        *  reading device version back: this allows the host implementation to
+        *  detect the old driver (if there was no version write before read).
+        */
+       writel((u32)PIPE_DRIVER_VERSION, dev->base + PIPE_REG_VERSION);
+       dev->version = readl(dev->base + PIPE_REG_VERSION);
+       if (dev->version < PIPE_CURRENT_DEVICE_VERSION) {
+               /* initialize the old device version */
+               err = goldfish_pipe_device_init_v1(pdev);
+       } else {
+               /* Host device supports the new interface */
+               err = goldfish_pipe_device_init_v2(pdev);
+       }
+       if (!err)
+               return 0;
+
+error:
+       dev->base = NULL;
+       return err;
+}
+
+static int goldfish_pipe_remove(struct platform_device *pdev)
+{
+       struct goldfish_pipe_dev *dev = pipe_dev;
+       if (dev->version < PIPE_CURRENT_DEVICE_VERSION)
+               goldfish_pipe_device_deinit_v1(pdev);
+       else
+               goldfish_pipe_device_deinit_v2(pdev);
+       dev->base = NULL;
+       return 0;
+}
+
+static const struct acpi_device_id goldfish_pipe_acpi_match[] = {
+       { "GFSH0003", 0 },
+       { },
+};
+MODULE_DEVICE_TABLE(acpi, goldfish_pipe_acpi_match);
+
+static const struct of_device_id goldfish_pipe_of_match[] = {
+       { .compatible = "google,android-pipe", },
+       {},
+};
+MODULE_DEVICE_TABLE(of, goldfish_pipe_of_match);
+
+static struct platform_driver goldfish_pipe_driver = {
+       .probe = goldfish_pipe_probe,
+       .remove = goldfish_pipe_remove,
+       .driver = {
+               .name = "goldfish_pipe",
+               .of_match_table = goldfish_pipe_of_match,
+               .acpi_match_table = ACPI_PTR(goldfish_pipe_acpi_match),
+       }
+};
+
+module_platform_driver(goldfish_pipe_driver);
+MODULE_AUTHOR("David Turner <digit@google.com>");
+MODULE_LICENSE("GPL");