#include <linux/seq_file.h> // For seq_* functions and types
#include <linux/version.h>  // Macros to detect kernel version

#include "nvdebug.h"

#define RUNLIST_PROCFS_NAME "runlist"
#define DETAILED_CHANNEL_INFO

/* Print channel details using PCCSR (Programmable Channel Control System RAM?)
 * @param s      Pointer to state from seq_file subsystem to pass to seq_printf
 * @param g      Pointer to our internal GPU state
 * @param chid   ID of channel to print details on, range [0, 512)
 * @param prefix Text string to prefix each line with, or empty string
 */
#ifdef DETAILED_CHANNEL_INFO
static int runlist_detail_seq_show_chan(struct seq_file *s, struct nvdebug_state *g, uint32_t chid, char *prefix) {
	channel_ctrl_t chan;
	char *loc_txt;
	u64 instance_ptr;
	chan.raw = nvdebug_readq(g, NV_PCCSR_CHANNEL_INST(chid));
	loc_txt = target_to_text(chan.inst_target);
	if (!loc_txt)
		return -EIO;
	instance_ptr = chan.inst_ptr;
	instance_ptr <<= 12;
	seq_printf(s, "%s+- Channel Info %-4d -+\n", prefix, chid);
	seq_printf(s, "%s| Enabled:           %d|\n", prefix, chan.enable);
	seq_printf(s, "%s| Next:              %d|\n", prefix, chan.next);
	seq_printf(s, "%s| Force CTX Reload:  %d|\n", prefix, chan.force_ctx_reload);
	seq_printf(s, "%s| Enable set:        %d|\n", prefix, chan.enable_set);
	seq_printf(s, "%s| Enable clear:      %d|\n", prefix, chan.enable_clear);
	seq_printf(s, "%s| PBDMA Faulted:     %d|\n", prefix, chan.pbdma_faulted);
	seq_printf(s, "%s| ENG Faulted:       %d|\n", prefix, chan.eng_faulted);
	seq_printf(s, "%s| Status:           %2d|\n", prefix, chan.status);
	seq_printf(s, "%s| Busy:              %d|\n", prefix, chan.busy);
	seq_printf(s, "%s| Instance PTR:       |\n", prefix);
	seq_printf(s, "%s| %#018llx  |\n", prefix, instance_ptr);
	seq_printf(s, "%s| %-20s|\n", prefix, loc_txt);
	seq_printf(s, "%s| Instance bound:    %d|\n", prefix, chan.inst_bind);
	// START TEMP
	// "runlist_id -1 is synonym for the ENGINE_GR_GK20A runlist id"
	// GR, GRCE, and ASYNC_CE
	// Note that this appears to be broken??
	// Peek into the channel instance RAM
	if (chan.inst_target == TARGET_SYS_MEM_COHERENT) {
		seq_printf(s, "%s| Target Engine:    %2d|\n", prefix, *(uint32_t*)phys_to_virt(instance_ptr + 4/*bytes for 32bits*/*43/*NV_RAMFC_TARGET*/) & 0x1f);
		seq_printf(s, "%s| PDB LO:   %#08x|\n", prefix, *(uint32_t*)phys_to_virt(instance_ptr + 4/*bytes for 32bits*/*128/*NV_RAMIN_PAGE_DIR_BASE_LO*/) & 0xfffff000);
		seq_printf(s, "%s| Num subcontexts:  %2ld|\n", prefix, hweight64(*(uint64_t*)phys_to_virt(instance_ptr + 4/*bytes for 32bits*/*166/*NV_RAMIN_SC_PDB_VALID*/)));
		// This appears to be unset on Xavier
		//seq_printf(s, "%s| PAS ID:     %8ld|\n", prefix, *(uint32_t*)phys_to_virt(instance_ptr + 4/*bytes for 32bits*/*135/*NV_RAMIN_PASID*/) & 0xfffff);
	}
	// END TEMP
	seq_printf(s, "%s+---------------------+\n", prefix);
	return 0;
}
#endif

#if LINUX_VERSION_CODE < KERNEL_VERSION(4,19,0)
// Bug workaround. See comment in runlist_file_seq_start()
static loff_t pos_fixup;
#endif

static void *runlist_file_seq_start(struct seq_file *s, loff_t *pos) {
	static struct runlist_iter rl_iter;
	struct nvdebug_state *g = &g_nvdebug_state[file2parentgpuidx(s->file)];
	// *pos == 0 for first call after read of file
	if (*pos == 0) {
		int err = get_runlist_iter(g, seq2gpuidx(s), &rl_iter);
		if (err)
			return ERR_PTR(err);
		// Don't try to print an empty runlist
		if (rl_iter.rl_info.len <= 0)
			return NULL;
		return &rl_iter;
	}
	// If we're resuming an earlier print
	if (*pos < rl_iter.rl_info.len) {
#if LINUX_VERSION_CODE < KERNEL_VERSION(4,19,0)
		// There's a nasty bug prior to 4.19-rc1 that if the buffer overflows, the
		// last update to `pos` is not saved. Work around that here by reloading a
		// saved copy of `pos`.
		if (!pos_fixup)
			return NULL;
		*pos = pos_fixup;
#endif
		return &rl_iter;
	}
	// When called with *pos != 0, we already traversed the runlist
	return NULL;
}

static void* runlist_file_seq_next(struct seq_file *s, void *raw_rl_iter,
				   loff_t *pos) {
	struct runlist_iter* rl_iter = raw_rl_iter;
	void *ret = NULL;
	struct nvdebug_state *g = &g_nvdebug_state[file2parentgpuidx(s->file)];
	// Advance by one TSG or channel
	(*pos)++;
	rl_iter->curr_entry += NV_RL_ENTRY_SIZE(g);
	// Verify we haven't reached the end of the runlist
	// rl_info.len is the num of tsg entries + total num of channel entries
	if (*pos < rl_iter->rl_info.len) {
		ret = rl_iter;
	}
#if LINUX_VERSION_CODE < KERNEL_VERSION(4,19,0)
	// Bug workaround. See comment in runlist_file_seq_start()
	pos_fixup = ret ? *pos : 0;
#endif
	return ret;
}

static void runlist_file_seq_stop(struct seq_file *s, void *raw_rl_iter) {
	// No cleanup needed
}

static int runlist_file_seq_show(struct seq_file *s, void *raw_rl_iter) {
	struct runlist_iter *rl_iter = raw_rl_iter;
	void *entry = rl_iter->curr_entry;
	struct nvdebug_state *g = &g_nvdebug_state[file2parentgpuidx(s->file)];
	if (entry_type(g, entry) == ENTRY_TYPE_TSG) {
		if (rl_iter->channels_left_in_tsg) {
			printk(KERN_WARNING "[nvdebug] Found TSG ID%d @ %px when %d channels were still expected under the previous TSG in the runlist!\n", tsgid(g, entry), entry, rl_iter->channels_left_in_tsg);
			while (rl_iter->channels_left_in_tsg--)
				seq_printf(s, "[missing channel]\n");
		}
		rl_iter->channels_left_in_tsg = tsg_length(g, entry);
		seq_printf(s, "+---- TSG Entry %-3d---+\n", tsgid(g, entry));
		seq_printf(s, "| Scale: %-13d|\n", timeslice_scale(g, entry));
		seq_printf(s, "| Timeout: %-11d|\n", timeslice_timeout(g, entry));
		seq_printf(s, "| Length: %-12d|\n", tsg_length(g, entry));
		seq_printf(s, "+---------------------+\n");
	} else {
		char *indt = "";
#ifndef DETAILED_CHANNEL_INFO
		u64 instance_ptr = 0;
#endif
		if (rl_iter->channels_left_in_tsg) {
			indt = "  ";
			rl_iter->channels_left_in_tsg--;
		}
#ifdef DETAILED_CHANNEL_INFO
		runlist_detail_seq_show_chan(s, g, chid(g, entry), indt);
#else
		// Reconstruct pointer to channel instance block
		if (g->chip_id >= NV_CHIP_ID_VOLTA) {
			instance_ptr = ((struct gv100_runlist_chan*)entry)->inst_ptr_hi;
			instance_ptr <<= 32;
		}
		instance_ptr |= inst_ptr_lo(g, entry) << 12;

		seq_printf(s, "%s+- Channel Entry %-4d-+\n", indt, chid(g, entry));
		if (g->chip_id >= NV_CHIP_ID_VOLTA)
			seq_printf(s, "%s| Runqueue Selector: %d|\n", indt,
				   ((struct gv100_runlist_chan*)entry)->runqueue_selector);
		seq_printf(s, "%s| Instance PTR:       |\n", indt);
		seq_printf(s, "%s| %#018llx  |\n", indt, instance_ptr);
		seq_printf(s, "%s| %-20s|\n", indt, target_to_text(inst_target(g, entry)));
		seq_printf(s, "%s+---------------------+\n", indt);
#endif
	}
	return 0;
}

static const struct seq_operations runlist_file_seq_ops = {
	.start = runlist_file_seq_start,
	.next = runlist_file_seq_next,
	.stop = runlist_file_seq_stop,
	.show = runlist_file_seq_show,
};

static int runlist_file_open(struct inode *inode, struct file *f) {
	return seq_open(f, &runlist_file_seq_ops);
}

struct file_operations runlist_file_ops = {
	.open = runlist_file_open,
	.read = seq_read,
	.llseek = seq_lseek,
	.release = seq_release,
};

ssize_t preempt_tsg_file_write(struct file *f, const char __user *buffer,
			       size_t count, loff_t *off) {
	uint32_t target_tsgid;
	// Passing 0 as the base to kstrtou32 indicates autodetect hex/octal/dec
	int err = kstrtou32_from_user(buffer, count, 0, &target_tsgid);
	struct nvdebug_state *g = &g_nvdebug_state[file2gpuidx(f)];
	if (err)
		return err;

	// TSG IDs are a 12-bit field, so make sure the request is in-range
	if (target_tsgid > MAX_TSGID)
		return -ERANGE;

	// Execute preemption
	err = preempt_tsg(g, target_tsgid);
	if (err)
		return err;

	return count;
}

struct file_operations preempt_tsg_file_ops = {
	.write = preempt_tsg_file_write,
	.llseek = default_llseek,
};

ssize_t disable_channel_file_write(struct file *f, const char __user *buffer,
				   size_t count, loff_t *off) {
	uint32_t target_channel;
	channel_ctrl_t chan;
	int err;
	runlist_info_t rl_info;
	runlist_disable_t rl_disable;
	struct nvdebug_state *g = &g_nvdebug_state[file2gpuidx(f)];
	// Passing 0 as the base to kstrtou32 indicates autodetect hex/octal/dec
	err = kstrtou32_from_user(buffer, count, 0, &target_channel);
	if (err)
		return err;

	if (target_channel > MAX_CHID)
		return -ERANGE;

	// Disable channel
	chan.raw = nvdebug_readq(g, NV_PCCSR_CHANNEL_INST(target_channel));
	chan.enable_clear = true;
	// disable sched
	rl_info.raw = nvdebug_readl(g, NV_PFIFO_RUNLIST);
	rl_disable.raw = nvdebug_readl(g, NV_PFIFO_SCHED_DISABLE);
	rl_disable.raw |= BIT(rl_info.id);
	nvdebug_writel(g, NV_PFIFO_SCHED_DISABLE, rl_disable.raw);
	// disable chan
	nvdebug_writeq(g, NV_PCCSR_CHANNEL_INST(target_channel), chan.raw);
	// enable sched
	rl_disable.raw &= ~BIT(rl_info.id);
	nvdebug_writel(g, NV_PFIFO_SCHED_DISABLE, rl_disable.raw);

	return count;
}

struct file_operations disable_channel_file_ops = {
	.write = disable_channel_file_write,
	.llseek = default_llseek,
};

ssize_t enable_channel_file_write(struct file *f, const char __user *buffer,
				   size_t count, loff_t *off) {
	uint32_t target_channel;
	channel_ctrl_t chan;
	int err;
	struct nvdebug_state *g = &g_nvdebug_state[file2gpuidx(f)];
	// Passing 0 as the base to kstrtou32 indicates autodetect hex/octal/dec
	err = kstrtou32_from_user(buffer, count, 0, &target_channel);
	if (err)
		return err;

	if (target_channel > MAX_CHID)
		return -ERANGE;

	// Disable channel
	chan.raw = nvdebug_readq(g, NV_PCCSR_CHANNEL_INST(target_channel));
	chan.enable_set = true;
	nvdebug_writeq(g, NV_PCCSR_CHANNEL_INST(target_channel), chan.raw);

	return count;
}

struct file_operations enable_channel_file_ops = {
	.write = enable_channel_file_write,
	.llseek = default_llseek,
};

ssize_t switch_to_tsg_file_write(struct file *f, const char __user *buffer,
				   size_t count, loff_t *off) {
	uint32_t target_tsgid;
	struct gv100_runlist_chan* chan;
	channel_ctrl_t chan_ctl;
	struct runlist_iter rl_iter;
	int err;
	loff_t pos = 0;
	struct nvdebug_state *g = &g_nvdebug_state[file2gpuidx(f)];
	// Passing 0 as the base to kstrtou32 indicates autodetect hex/octal/dec
	err = kstrtou32_from_user(buffer, count, 0, &target_tsgid);
	if (err)
		return err;

	if (target_tsgid > MAX_TSGID)
		return -ERANGE;

	err = get_runlist_iter(g, 0, &rl_iter);
	if (err)
		return err;

	// Iterate through all TSGs
	while (pos < rl_iter.rl_info.len) {
		if (tsgid(g, rl_iter.curr_entry) == target_tsgid) {
			// Enable channels of target TSG
			for_chan_in_tsg(g, chan, rl_iter.curr_entry) {
				chan_ctl.raw = nvdebug_readq(g, NV_PCCSR_CHANNEL_INST(chan->chid));
				chan_ctl.enable_set = true;
				nvdebug_writeq(g, NV_PCCSR_CHANNEL_INST(chan->chid), chan_ctl.raw);
			}
		} else {
			// XXX: Fix for bare channels. Maybe a "for_chan_until_tsg" macro?
			// Disable all other channels
			// (This is how the Jetson nvgpu driver disables TSGs)
			for_chan_in_tsg(g, chan, rl_iter.curr_entry) {
				chan_ctl.raw = nvdebug_readq(g, NV_PCCSR_CHANNEL_INST(chan->chid));
				chan_ctl.enable_clear = true;
				nvdebug_writeq(g, NV_PCCSR_CHANNEL_INST(chan->chid), chan_ctl.raw);
			}
		}
		pos += 1 + tsg_length(g, rl_iter.curr_entry);
		rl_iter.curr_entry = next_tsg(g, rl_iter.curr_entry);
	}
	// Switch to next TSG with active channels (should be our TSG)
	err = preempt_tsg(g, target_tsgid);
	if (err)
		return err;

	return count;
}

struct file_operations switch_to_tsg_file_ops = {
	.write = switch_to_tsg_file_write,
	.llseek = default_llseek,
};